Analyzing qualitative data is challenging. Such analyses are even more difficult when the topic is controversial and the results will drive important policy decisions. This post explores AI methods for qualitative research, using chatGPT for categorization, embeddings to find hidden topics, and long-context summarization with Claude2 on a case study analyzing free-text public comments to a controversial Environmental Impact decision.

Background

Quite a while ago, I detailed why replacing wolves on Isle Royale National Park was a bad policy decision back by even worse science. Since then, the National Park Service (NPS) decided to commit to wolf replacement anyway, dropping 19 new wolves on the island in 2018 and 2019. The results were expected. The new wolves killed the last original male wolf in 2019, almost certainly ensuring that the new wolf population will be genetically disconnected from the prior population. Of the 20 wolves that NPS attempted to relocate, one died before making it to the island, one voluntarily crossed the ice back to the mainland*, and four others died by the end of 2019. The surviving 14 wolves successfully bred and the population now stands at 31. So, in the end, we have a new, synthetic wolf population that is entirely disjunct from a genetic and ecological perspective. As I predicted in my original post: “in reality, this is not a genetic rescue project, it is a genetic replacement project,” which violates both the scientific and management purpose of the Park.

* This contradicts one of the primary justifications for replacing the wolves. Proponents argued that the lack of ice due to climate change would make natural repopulation impossible.

But neither science nor policy drove NPS’s decision. Management of charismatic mammals, especially in a well-known National Park, is largely a matter of public sentiment. In fact, it is a codified part of the decision process. Federal managers are required to seek public comments as part of the NEPA process.

In general, I am a huge supporter of public voices in important conservation decisions (I’ve even written papers advocating for it). But, sometimes I worry about how advocacy groups can skew the perception of organic public sentiment. That’s what I’d like to analyze in this post.

All of the public comments submitted to NPS on the Isle Royale wolf-moose management plan are public record. You can download and read all 1117 pages of comments.

But 1117 pages is a lot of text to read and digest. In this post, I want to show how you can easily process lots of text using AI (both generative large-language models (LLM), like chatGPT, and LLM embeddings) to make quantitative (or semi-quantitative) analyses.

Basic analyses

Visit my GitHub repo for this project for a fully reproducible analysis.

First, we’ll set up the environment and load in necessary packages.

# Load libraries
library(pdftools) # We will use 'pdftools' to convert the pdf to plain text
library(tidyverse)
library(stringr)
library(RColorBrewer)

# Set up the directory structure:
make_new_dir <- 
     function(DIR_TO_MAKE){
          if(dir.exists(DIR_TO_MAKE) == FALSE){
               dir.create(DIR_TO_MAKE)
          }else{
               print("Directory exists")
          }
     }

make_new_dir("./data/")
make_new_dir("./figs/")

We can download the comments from the NPW website.

download.file(
     url = "https://parkplanning.nps.gov/showFile.cfm?projectID=59316&MIMEType=application%252Fpdf&filename=ISRO%5FMWVPlan%5FAllCorrespondence%5FPEPC%2Epdf&sfid=232552",
     destfile = "./data/ISRO_MWVPlan_AllCorrespondence_PEPC.pdf",
mode = "wb"
)

The first step to analyze the public comments is to parse the pdf into text. This is a tedious process. I won’t show it here, but you can follow all of the steps on my GitHub repo for this project.

You can download my pre-processed dataset to short-cut the the PDF parsing steps.

download.file(
     url = "https://www.azandisresearch.com/wp-content/uploads/2023/09/EIS_comments.csv",
     destfile = "./data/EIS_comments2.csv"
)

EIS_comments <- read.csv("./data/EIS_comments.csv")

The formatting follow the same structure for every comment. I’ve extracted the ‘Comment ID’, ‘Received’ date time, ‘Correspondence Type’, and ‘Correspondence’ text into a dataframe. I’ve also truncated the longest comments (…comment 68 looks like someone copy and pasted their term paper) to 12,000. This will be important later because the context window for chatGPT is 4000 tokens.

EIS_comments %>% glimpse()
Rows: 2,776
Columns: 4
$ ID             <dbl> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29,…
$ Received       <dttm> 2015-07-12 20:45:30, 2015-07-14 23:18:34, 2015-07-15 12:03:55, 2015-07-15 13:14:52, 2015-07-15 13:35:47, …
$ Correspondence <chr> "Web Form Correspondence", "Web Form Correspondence", "Web Form Correspondence", "Web Form Correspondence"…
$ Content        <chr> "The alternatives are complete enough as a starting point. The issues will be related to the details. The …

We can do some basic summary analysis on these initial variables. The most comments were submitted in the week before the comment deadline on Sept 1. The vast majority of comments were received through the web form. Less than 10% of comments were physical letters and 51 of the 2777 comments were form cards given to Park visitors.

Often, large influxes of web and email comments are the product of advocacy groups encouraging their members to submit pre-written comments. I’ve used this tactic myself in conservation campaigns, so I won’t cast dispersions. But, I’ll also be the first to admit that a copy-and-pasted form letter is far less sincere than a uniquely crafted opinion.

After checking for matches among the comments, it is clear that there were two archetypical pre-written texts.  These include 733 near identical comment in favor of wolf replacement (i.e. Alternative B), likely from National Parks Conservation Association:

EIS_comments %>%
+   filter(grepl("I care about the wildlife at our national parks, including the wolves and moose at Isle Royale. Right now there are only three", Content)) %>%
+   group_by(Content) %>%
+   tally() %>%
+   arrange(desc(n)) %>%
+   ungroup() %>%
+   filter(row_number() == 1) %>%
+   .$Content %>% 
+   cat()
Dear Superintendent Green, I care about the wildlife at our national parks, including the wolves and moose at Isle Royale. Right now there are only three wolves left at the park- -the lowest number of wolves in more than 50 years- -threatening the overall ecosystem health of this iconic national park. I support management Alternative B to bring new wolves to the island, but urge the Park Service to do this as needed, rather than one time only. Without wolves, the moose population on the island will continue to increase, eating until the food sources are gone. If we bring new wolves to the island, they will help keep the moose population from rapidly expanding and minimize impacts to the native vegetation. This option is much less intrusive in this wilderness park than culling moose, removing moose from the island, or having to replant native vegetation once the moose consume it. As stewards of this park, the National Park Service should take the least intrusive action that results in the biggest benefit to the island's wildlife and ecosystem. I support the Park Service taking action to bring new wolves to the park immediately, before the population vanishes altogether. Thank you for considering my concerns. Sincerely,

And 55 nearly identical comments in favor of Wilderness (i.e. Alternative A), likely from Wilderness Watch:

EIS_comments %>%
+   filter(grepl("Isle Royale's wilderness designation requires that we protect the area's unmanipulated, untrammeled wilderness character. Wild", Content)) %>%
+   group_by(Content) %>%
+   tally() %>%
+   arrange(desc(n)) %>%
+   ungroup() %>%
+   filter(row_number() == 1) %>%
+   .$Content %>% 
+   cat()
Isle Royale's wilderness designation requires that we protect the area's unmanipulated, untrammeled wilderness character. Wilderness designation means we let Nature call the shots. Transplanting wolves from the mainland to Isle Royale is a major manipulation of the Isle Royale Wilderness and must not be done. Alternative Concept A, the No Action Alternative, is the best alternative to protect Isle Royale's unmanipulated, untrammeled wilderness character.

It is important to flag these duplicated comments because the methods that we will use later on will not behave correctly with nearly identical strings.

EIS_comments_deduplicated <- 
     EIS_comments %>%
     # Remove comments with no content
     filter(!is.na(Content)) %>%
     # Flag the web form duplicates
     mutate(form_duplicate = ifelse(grepl("I care about the wildlife at our national parks, including the wolves and moose at Isle Royale. Right now there are only three", Content), "for Alt B", NA)) %>%
     mutate(form_duplicate = ifelse(grepl("Isle Royale's wilderness designation requires that we protect the area's unmanipulated, untrammeled wilderness character. Wild", Content), "for Alt A", form_duplicate)) %>%
     # Form duplicates are not exact matches
     mutate(Content_dup = ifelse(is.na(form_duplicate), Content, form_duplicate)) %>%
     group_by(Content_dup) %>%
     # Retain one of the duplicate sets
     slice_sample(n = 1)

After removing the duplicates and cleaning the data, we are left with 1970 unique comments.

Text analysis with chatGPT

Now, we can start analyzing the content. There are many ways that we could do this, depending on the question we want to answer. For instance, maybe we want to see with questions naturally group together to see if we can find common themes? Traditionally, a common way to do this type of natural language processing would be to use an approach like a Latent-Dirchelt allocation topic analysis that groups comments by tf-idf values of the stems of words contained in the comment. (I cover tf-idf in a previous post). But, one problems with this approach is that the context of words is lost.

If we want to capture the context of the text, we might try using word embeddings from a LLM like GPT. We’ll try this approach later.

In our case, maybe we just want to know how many comments support a given policy.. It would be hard to answer that from the embeddings ourselves, but we could treat GPT as an agent who could read and categorize comments by preferred policy alternative.

We’ll use two packages. httr helps us interact with the chatGPT API. The API speaks in json format. jsonlite helps us parse formatted prompts and responses.

library(httr)
library(jsonlite)

Working with chatGPT is a lot like working with a new intern. Like an new intern, it has no prior contextual understanding of our specific task–we have to be very explicit with our directions. On the bright side, our chatGPT intern has endless patience and never sleeps!

We will be interacting with chatGPT through the API. This differs from the dialectical way that most people interact with chatGPT. We need to engineer our prompt to get a robust response in exactly the same format, every time.  We can do that by passing in quite a bit of context in our prompt and giving specific directions for the output, with examples. Here is the prompt we’ll use:

You are a federal employee tasked with reading the following comment submitted by a member of the public in response to the The Isle Royale National Park Moose-Wolf-Vegetation Management Plan/EIS. The Plan/EIS is a document that evaluates management alternatives for the moose and wolf populations on the island National Park land.
Management alternatives include:

- Alternative A: No Action. Continue the current management of letting nature take its course, without any intervention or manipulation of the moose or wolf populations or their habitats.
- Alternative B: Immediate Wolf Introduction. Introduce 20-30 wolves over a three-year period, starting as soon as possible to reduce the moose population and its impacts on vegetation.
- Alternative C: Wolf Introduction after Thresholds are Met. Introduce wolves if certain thresholds are met, such as the extirpation of wolves, the overabundance of moose, or the degradation of vegetation. The number and timing of wolf introductions would depend on the conditions at the time.
- Alternative D: Moose Reduction and Wolf Assessment. Reduce the moose population by lethal and non-lethal means, such as hunting, contraception, or relocation. The goal would be to lower the moose density to a level that would allow vegetation recovery and assessing introducing wolves to the island in the future.

Here is the text of the public comment: '[INSERT COMMENT TEXT]'.

State which alternative the commenter is most likely to favor (A, B, C, D).
State if the comment is 'For', 'Against', or 'Neutral' on wolf introductions.
State if the strength of the commenter's opinion on a scale from 'Extremely strong', 'Very strong', 'Strong', 'Somewhat strong', or 'Mild'.

Produce the output in json format like this:
{
"favored_alternative": "",
"wolf_opinion": "",
"opinion_strength": ""
}

ChatGPT 3.5 costs 0.002$ per 1000 tokens. We can use the OpenAI tokenizer to estimate the number of tokens constituting our input prompt.

Our input is 420 tokens. The output should be less than 50 tokens. So we can round to assume 500 tokens per query. So, it will cost us about $1 to process 1000 comments. Much cheaper than paying a human!

In the old days, you could pass a list of inputs into chatGPT ‘completions’ model all at once. This is no longer possible. Now, to use the ‘chat/completions’ API requires looping through each of the inputs and making individual requests. Unfortunately, the API often fails or hits the request rate limit. So, we need to be smart about staging and error handling with this larger loop. The structure of this loop is to define the prompt, wait 18 seconds to avoid the rate limit, run a tryCatch block to test if the API call fails, and if so, it skips to the next record and logs the records that the error occurred on, otherwise, parse the response and store the output in a file.

After getting initial responses, I also want to rerun 500 randomly selected comments in order to check chatGPT’s consistency. This is a critical part of using a generative model in quantitative analysis. I’ll talk more about this later.

Here’s the loop. It will take quite a while depending on your rate limit. I’d suggest either running it overnight or putting in on a remote server. Because we write each response out to file, there’s no problem if it fails. Just note the number of the last successful iteration (which will be printed to the screen) and start back up there.

set.seed(7097)

# Randomly select 500 records to resample
IDs_to_resample <- sample(unique(EIS_comments_deduplicated$ID), 500, replace = FALSE)
ID_list <- c(unique(EIS_comments_deduplicated$ID), IDs_to_resample)

# Create a vector to store failed IDs
failed_ids <- c()

ID_list <- Still_need_IDs

for (i in 1:length(ID_list)) {
  ID_number = ID_list[i]
  # Define the prompt
  prompt_content <- paste0( "Here is the text of the public comment: '", EIS_comments_deduplicated %>%
        filter(ID == ID_number) %>%
        .$Content,
      "'.
    State which alternative the commenter is most likely to favor (A, B, C, D).
State if the comment is 'For', 'Against', or 'Neutral' on wolf introductions.
State if the strength of the commenter's opinon on a scale from 'Extremely strong', 'Very strong', 'Strong', 'Somewhat strong', or 'Mild'.
Produce the output in json format like this:\n{\n\"favored_alternative\": \"\",\n\"wolf_opinion\": \"\",\n\"opinion_strength\": \"\"\n}"
    )
  
  # Initialize gpt_response
  gpt_response <- NULL
  
  # With my account, I can make 3 requests per minute. To avoid denied API calls, I add a 18 second pause in each loop.
  Sys.sleep(18)
  
  tryCatch({
    # Call GPT for a response
    gpt_response <- 
      POST(
        url = "https://api.openai.com/v1/chat/completions", 
        add_headers(Authorization = paste0("Bearer ", read_lines("../credentials/openai.key"))),
        content_type_json(),
        encode = "json",
        body = list(
          model = "gpt-3.5-turbo",
          messages = list(
            list(
              "role" = "system",
              "content" = "You are a federal employee tasked with reading the following comment submitted by a member of the public in response to the The Isle Royale National Park Moose-Wolf-Vegetation Management Plan/EIS. The Plan/EIS is a document that evaluates management alternatives for the moose and wolf populations on the island National Park land.
Management alternatives include:
- Alternative A: No Action. Continue the current management of letting nature take its course, without any intervention or manipulation of the moose or wolf populations or their habitats.
- Alternative B: Immediate Wolf Introduction. Introduce 20-30 wolves over a three-year period, starting as soon as possible to reduce the moose population and its impacts on vegetation.
- Alternative C: Wolf Introduction after Thresholds are Met. Introduce wolves if certain thresholds are met, such as the extirpation of wolves, the overabundance of moose, or the degradation of vegetation. The number and timing of wolf introductions would depend on the conditions at the time.
- Alternative D: Moose Reduction and Wolf Assessment. Reduce the moose population by lethal and non-lethal means, such as hunting, contraception, or relocation. The goal would be to lower the moose density to a level that would allow vegetation recovery and assessing introducing wolves to the island in the future."
            ),
            list(
              "role" = "user",
              "content" = prompt_content
            )
          )
        )
      )
    print(paste0("API call successful for ID: ", ID_number, ", index: ", i))
  }, error = function(e) {
    # Handle API call errors
    cat("API call failed for ID: ", ID_number, ", index: ", i, "\n")
    failed_ids <- c(failed_ids, i)
  })
  
  # If the API call was successful, proceed with data wrangling and output
  if (!is.null(gpt_response)) {
    # parse the response object as JSON
    content <- content(gpt_response, as = "parsed")
    
    # Assign the ID to the GPT response
    gpt_response_df <- data.frame(
      response_id = ID_number,
      gpt_response = content$choices[[1]]$message$content
    )
    
    # Convert the JSON to a dataframe and join to the record data
    output <- bind_cols( EIS_comments_deduplicated %>%
        filter(ID == ID_number),
      fromJSON(gpt_response_df$gpt_response) %>% 
        as.data.frame()
    ) %>%
      mutate(response_created_time = Sys.time())
    
    # Append the data to the extant records and write the output to a file. (This is a bit less memory efficient to do this within the loop, but I )
    if (!file.exists("./EIS_GPT_responses.csv")) {
      write.csv(output, "./EIS_GPT_responses.csv", row.names = FALSE)
    } else {
      read.csv("./EIS_GPT_responses.csv") %>%
        mutate(across(everything(), as.character)) %>%
        bind_rows(output %>%
                    mutate(across(everything(), as.character))
        ) %>%
        write.csv("./EIS_GPT_responses.csv", row.names = FALSE)
    }
    
    print(paste0("Completed response ", i))
  }
}

# Log the failed IDs to a file
if (length(failed_ids) > 0) {
  write.csv(data.frame(ID = failed_ids), "./failed_ids.csv", row.names = FALSE)
  cat("Failed IDs logged to 'failed_ids.csv'\n")
}

ChatGPT is nondeterministic, so your responses will differ. You can download the responses I got to follow along.

download.file(
     url = "https://www.azandisresearch.com/wp-content/uploads/2023/09/Final_GPT_Responses.csv",
     destfile = "./data/GPT_output.csv"
)

GPT_output <- read.csv("./data/GPT_output.csv")

GPT_output %>% glimpse()
Rows: 2,470
Columns: 13
$ ID                    <int> 93, 440, 2164, 636, 839, 2335, 36, 487, 1268, 2303, 1781, 60, 1033, 1948, 1826, 1538, 1685, 308, 22…
$ Received              <chr> "7/29/2015 9:09", "8/9/2015 5:14", "8/27/2015 14:36", "8/18/2015", "8/25/2015", "8/28/2015 12:30", …
$ Correspondence        <chr> "Web Form Correspondence", "Web Form Correspondence", "Web Form Correspondence", "Web Form Correspo…
$ Content               <chr> "\"100% o wolves examined since 1994...have spinal anomalies.\"- -Of the six alternatives put forth…
$ form_duplicate        <chr> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA,…
$ Content_dup           <chr> "\"100% o wolves examined since 1994...have spinal anomalies.\"- -Of the six alternatives put forth…
$ favored_alternative   <chr> "C", "C", "Alternative D", "C", "C", "B", "C", "C", "D", "C", "Unknown", "C", "B", "A", "B", "A", "…
$ wolf_opinion          <chr> "For", "Against", "Neutral", "For", "Neutral", "For", "For", "For", "Against", "For", "Neutral", "F…
$ opinion_strength      <chr> "Very strong", "Very strong", "Strong", "Strong", "Somewhat strong", "Very strong", "Strong", "Stro…
$ response_created_time <chr> "32:19.2", "33:11.7", "33:16.9", "33:19.5", "34:35.2", "34:54.2", "34:55.4", "36:15.1", "36:16.3", …
$ Favored_alternative   <lgl> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA,…
$ Wolf_opinion          <lgl> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA,…
$ Opinion_strength      <lgl> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA,…

A couple of interesting things to note here. First, I apparently was not specific enough in my instructions for classifying the favored alternative because chatGPT sometimes returns “Alternative B” instead of just “B”. This is one of the struggles with using chatGPT, a generative model, in this way. It strays from instructions just like human survey respondents when inputting free-text results. For example, common responses to the survey question, “How are you feeling on a scale from 1 (bad) to 10 (good)?” might be “I’m good” or “Okay” or “nine” or “0”. None of those answers fit the instructions, so we have to clean them up.

In the case of chatGPT, we might be able to reduce these errors with more specific prompt engineering. For now, we’ll just clean up the responses on the backend.

# Fix erroneous column names
GPT_output <-
     GPT_output %>%
     mutate(
          favored_alternative = ifelse(is.na(favored_alternative), Favored_alternative, favored_alternative),
          wolf_opinion = ifelse(is.na(wolf_opinion), Wolf_opinion, wolf_opinion),
          opinion_strength = ifelse(is.na(opinion_strength), Opinion_strength, opinion_strength)
          ) %>%
     select(
          -Wolf_opinion,
          -Favored_alternative,
          -Opinion_strength
     )

# There are probably more elegant ways to write generalized rules to classify these reponses, but this does the trick
GPT_output <-
     GPT_output %>%
     # Fix 'favored alternative' responses
     mutate(
          favored_alternative_edit = case_when(
               (grepl(" and ", favored_alternative) | grepl(" or ", favored_alternative) | grepl("/", favored_alternative) | grepl("&", favored_alternative) | favored_alternative == "B, D") & !grepl(" and Wolf ", favored_alternative) & !grepl("N/A", favored_alternative) ~ "Multiple",
               grepl("\\bAlternative A\\b", favored_alternative) | favored_alternative %in% c("A", "No Action (A)") ~ "A",
               grepl("\\bAlternative B\\b", favored_alternative) | favored_alternative == "B" ~ "B",
               grepl("\\bAlternative C\\b", favored_alternative) | favored_alternative %in% c("C", "Concept C") ~ "C",
               grepl("\\bAlternative D\\b", favored_alternative) | favored_alternative == "D" ~ "D",
               TRUE ~ "Other"
          )
     ) %>%
     # Fix 'opinion strength' responses
     mutate(opinion_strength = tolower(opinion_strength)) %>%
     mutate(
          opinion_strength_edit = case_when(
               opinion_strength %in% c("strong", "very strong", "mild", "somewhat strong", "extremely strong") ~ opinion_strength,
               TRUE ~ "other"
          )
     ) %>%
     # Fix 'wolf opinion' responses
     mutate(wolf_opinion = tolower(wolf_opinion)) %>%
     mutate(
          wolf_opinion_edit = case_when(
          wolf_opinion %in% c("for", "against", "neutral") ~ wolf_opinion,
          TRUE ~ "other"
          )
     )

Let’s take a look at the results.

We can see that the majority of comments favor Alternative B: immediate wolf introduction. However, if we exclude the duplicated comments, our conclusion shifts to a majority in favor of the more moderate Alternative C: introduce wolves only after certain thresholds are met. Almost no one supports Alternative D: moose reduction and wolf assessment.

Comments that favored Alternative A were stronger proportionally. Alternative B supporters had mostly strong opinions but very few extremely strong or mild opinions. Supporters of Alternatives C and D were the least opinionated.

Validating chatGPT responses

It is worth asking ourselves how reliable chatGPT is at classifying these responses. One way to test this is to rerun a subset of comments, like we did above and check for agreement. This is called inter-rater reliability* (IRR).

* Although, maybe it should be called intra-rater reliability in this case. I guess it depends on out definition of ‘individual’ with LLM queries, but that’s a very philosophical bag of worms!

First, we need to subset our dataset to the responses that we scored twice.

IRR_comparisons <- 
     GPT_output %>%
     group_by(ID) %>%
     arrange(response_created_time) %>%
     mutate(ID_row_count = row_number()) %>%
     filter(ID_row_count <= 2) %>%
     mutate(n = n()) %>%
     filter(n > 1) %>%
     ungroup()

Then we can see how reliably the favored alternative was scored,

IRR_comparisons %>%
     select(ID, favored_alternative_edit, ID_row_count) %>%
     pivot_wider(
          id_cols = "ID",
          names_from = "ID_row_count",
          values_from = "favored_alternative_edit",
          names_prefix = "val"
     ) %>%
     group_by(val1 == val2) %>%
     tally() %>%
     mutate(
          total = sum(n),
          prop = n/total
     )
# A tibble: 2 × 4
  `val1 == val2`     n total  prop
  <lgl>          <int> <int> <dbl>
1 FALSE              2   500 0.004
2 TRUE             498   500 0.996

ChatGPT gave consistent responses in 498 out of 500 cases. That’s pretty good! Let’s look at the comments where it disagreed with itself.

IRR_comparisons %>%
     select(ID, favored_alternative_edit, ID_row_count) %>%
     pivot_wider(id_cols = "ID", names_from = "ID_row_count", values_from = "favored_alternative_edit", names_prefix = "val") %>%
     filter(val1 != val2)
# A tibble: 2 × 3
     ID val1  val2 
1   288 C     B    
2  1160 B     C    
 
EIS_comments_deduplicated %>%
     filter(ID == 288) %>%
     .$Content %>%
     cat()
There should be a balance between the wolf population and moose. When it is not balanced there is more harm than good done to the environment. Please introduce more wolves on this island instead of decreasing their population and this will keep the moose in check. Please add more wolves to contain the moose population. So many wolves are under attack in other states and decreasing their population is NOT the answer. It only creates more problems to the environment. There should be intense management of the wolf population to help it thrive and return the land back to it's natural state where there are enough moose and wolves. I think the public should be consulted as far as future plans for any culling. There should be intense management to monitor the effects of climate change as this will affect all aspects of wildlife and plant life on the island. I do not like the idea of a moose cull. I like the idea of introducing more wolves to the island so long as there is harmony with the existing wolves on the island. Maybe possibly try to introduce another type of animal that would be a good balance with the wolves and moose but only if it does not disrupt the balance and create new problems. Other states have adopted disastrous wolf culling plans that are only in the interests of farmers and ranchers. As the wolf population is dwindling, other problems will begin to develop as there is not a proper balance. Please keep wolves in mind and do your best to increase their population before it is too late and more animals will be needlessly killed without the proper balance of mother nature.> 
 
EIS_comments_deduplicated %>%
     filter(ID == 1160) %>%
     .$Content %>%
     cat()
I have heard both sides of this situation and I believe that new wolves should be introduced on Isle Royale. Climate change has made a large impact on the amount of ice that freezes in the Isle Royale region. Previously wolves from the mainland could cross the ice that formed and take up residence on the Isle. The ice hasn't been stable enough for these crossings in the last few years and the wolves are becoming inbred and dying off. If you will check a video that I have watched about the wolves being reintroduced to Yellowstone, you will see that the ecology of the region is benefited by the wolves being there. If enough wolves are transported to Isle Royale, the wolves will keep the moose in check and the ecology will improve. Allowing the pack to die off is really not a positive move. Introducing a new bloodline to the pack will help. I believe the wilderness designation of Isle Royale is a positive thing and that the wolves help to keep the ecosystem there in good order. Thank you for taking comments from the public.

In both cases, chatGPT vacillated between classifying the comment as favoring alternative B or C. Difference between those alternatives is admittedly nuanced. Both alternatives propose replacing wolves, the only difference is in the timing. In Alternative B, wolves would be introduced immediately and in Alternative C wolve would be introduced, “if certain thresholds are met, such as the extirpation of wolves, the overabundance of moose, or the degradation of vegetation. The number and timing of wolf introductions would depend on the conditions at the time.”

Both of the comments that made chatGPT disagree with itself focus on the environmental conditions that wolf introductions might remedy. However, these comments seems to presuppose that those conditions have been met and seem to suggest immediate introduction is necessary. So, I can see where chatGPT might have a hard time solidly classifying these comments.

Let’s also check the IRR for chatGPT’s classification of ‘opinion strength.’ Unlike the favored alternative, where most folks explicitly stated their preference, classifying the strength of an opinion is a far more subjective task.

IRR_comparisons %>%
     select(ID, opinion_strength_edit, ID_row_count) %>%
     pivot_wider(
          id_cols = "ID",
          names_from = "ID_row_count",
          values_from = "opinion_strength_edit",
          names_prefix = "val") %>%
     group_by(val1 == val2) %>%
     tally() %>%
     mutate(
          total = sum(n),
          prop = n/total
     )
# A tibble: 2 × 4
  `val1 == val2`     n total  prop
  <lgl>          <int> <int> <dbl>
1 FALSE              5   500  0.01
2 TRUE             495   500  0.99

ChatGPT disagreed with itself in 5 cases, but gave reliable classifications 99% of the time. That’s pretty good! However, just assessing binary disagreement or agreement isn’t a strong metric for this variable. A switch from “extremely strong” to “very strong” is less of an issue than a vacillating from “extremely strong” to “mild”.

Instead, we can use the Krippendorff’s Alpha. This metric provides a formal way to assess the the amount of inter-rater disagreement. There are multiple metrics that we could use, but Krippendorff’s Alpha is nice because it can generalize to any number of reviewers and can handle many types of disagreement (i.e. binary, ordinal, interval, categorical, etc.). Here’s a great post for understanding Krippendorff’s Alpha. We’ll use the irr package to estimate it.

library(irr)

The irr package needs the dataset in wide format matrix with one row per reviewer and each record (the package calls records ‘subjects’ because this metric is traditionally used in social science research) as a column. For this analysis, we’ll consider the first and second responses from chatGPT as individual reviewers. We also need to enforce the order of our opinion strength levels; otherwise, R will naturally order them alphabetically.

IRR_comparisons %>%
     mutate(opinion_strength_edit = fct_relevel(
          opinion_strength_edit,
          c(
               "other",
               "mild",
               "somewhat strong",
               "strong",
               "very strong",
               "extremely strong"
           )
     )) %>%
     select(
          ID,
          opinion_strength_edit,
          ID_row_count
     ) %>%
     pivot_wider(
          id_cols = "ID_row_count",
          names_from = "ID",
          values_from = "opinion_strength_edit",
          names_prefix = "ID_"
     ) %>%
     select(-ID_row_count) %>%
     as.matrix() %>%
     kripp.alpha(method = "ordinal")
  
Krippendorff's alpha

 Subjects = 500 
   Raters = 2 
    alpha = 0.996 

Krippendorff’s Alpha ranges from -1 to 1, where 1 means perfect concordance, 0 means random guesses among reviewers, and -1 is perfect negative concordance. At .996, we are pretty near perfect reliability.

For many datasets, there will be a lower degree of IRR. But, it is important to remember to interpret the alpha value in context. Perfect concordance may not be realistic, especially in highly subjective classifications. In most cases our goals is not perfect concordance, but simply greater reliability than we’d get if we hired a bunch of humans to do the annotating. Preliminary evidence seems to indicate that even version 3.5 of chatGPT is more reliable than humans (even domain experts!) in subjective classification tasks.

In most cases, you won’t have the resources to get human annotations for an entire dataset for comparison. Instead, you could 1.) get human annotations for a small subset, 2.) use a similar benchmark dataset, or 3.) spot-check responses yourself. If you choose to spot check, I’d suggest rerunning chatGPT multiple times (> 3) in order to estimate the variance in responses. High variance responses indicate especially difficult classifications that you should target for spot-checks. Another tip is to ask chatGPT to return it’s justification with each response. Ultimately, this process will help you diagnose problematic types of responses and enable you to engineer better prompts to deal with those edge cases.

The bottom line is that working with chatGPT is less like working with a model and more like working with human raters–and all of the validation tasks that entails.

Analysis with token embeddings

Up to this point, we’ve presupposed the classifications we wanted ChatGPT to identify in our data. But, what if we wanted to uncover hidden categories in the responses? Folks could advocate for the same Alternative but for different reasons. For example, among those who favor Alternative C, some might argue from the perspective of climate change and some from the perspective of moose populations.

We can use token embeddings to uncover hidden clusters of topics in our responses. Embeddings are the way that LLMs encode free text into numeric form.  Each token or ‘unit of language’ is numerically described as a position in multidimensional language space. This is a huge advantage over more traditional language clustering methods that simply count the occurrence of certain words. Embeddings retain the context of each token as it exists in the document.

As a toy example, the word “train” in these sentences: “I train a model”, “I train for a marathon”, “I rode the train”, “I’m on the Soul Train” could be described in two dimensions of more or less metaphorical and noun/verb. If we do this for all of the words in a document or chunk of text, we can then think of all the embeddings as a point cloud. Documents with highly overlapping point clouds are more similar that those that don’t overlap at all.

We call a different OpenAI model, text-embedding-ada-002, to return the embeddings. Unlike the chat model, we can pass all of the responses as a list in a single call, instead of looping through each response. This makes embeddings much faster and cheaper than using the chatGPT API.

Prior to embedding, I like to remove non-alpha numeric characters from the text.

# Clean up the text to remove non-alpha numeric characters
input_to_embed <- 
     EIS_comments_deduplicated %>%
     mutate(Content_cleaned = str_replace_all(Content, "[^[:alnum:]]", " "))

# Call OpenAI for the embeddings
embeddings_return <- 
     POST(
          "https://api.openai.com/v1/embeddings",
          add_headers(Authorization = paste0(
               "Bearer ", read_lines("../credentials/openai.key"))
          ),
          body = list(
               model = "text-embedding-ada-002",
               input = input_to_embed$Content_cleaned
               ),
          encode = "json"
     )

The returned object is a bit convoluted. We can use a bit of purrr and jsonlite to extract the embeddings.

# Extract the embeddings from the API return
embeddings_list <-
     embeddings_return %>%
     content(as = "text", encoding = "UTF-8") %>%
     fromJSON(flatten = TRUE) %>%
     pluck("data", "embedding")

Then add the embeddings back into the dataframe.

# Combine the embeddings with the original data
EIS_GPT_embeddings <- 
     EIS_comments_deduplicated %>%
     as_tibble() %>%
     mutate(
          embeddings = embeddings_list,
          ID = as.character(ID)
     ) %>%
     left_join(
# We need to get only the first instance of the GPT response data, which also included the repeated reliability test responses, to know which alternative the comment favors
          GPT_output %>%
               group_by(ID) %>%
               arrange(response_created_time) %>%
               mutate(ID_row_count = row_number()) %>%
               filter(ID_row_count == 1) %>%
               ungroup() %>%
     select(
          ID,
          favored_alternative_edit,
          opinion_strength_edit
          )
     )

Topical clustering from text embeddings

The problem is that those point clouds exist in extremely high dimensions. OpenAI’s text-embedding-ada-002 model returns 1536 dimensions. We need a method to reduce that complexity into something useful.

As mentioned, the embeddings allow us to see how comments relate in high-dimensional language space. We want to figure out where there are denser clusters of point clouds in that space which indicate common themes in the comments.

A couple of common ways to do this is to use a clustering algorithm (e.g. K-means) or dimension reduction (e.g. PCA). For this tutorial I want to use a bit of a hybrid approach called t-SNE (t-distributed Stochastic Neighbor Embedding) that will allow us to easily visualize the clusters of common comments which we can then explore.

We’ll use Rtsne package which requires that the data be in matrix form.

library(Rtsne)

# Rtsne requires the embeddings to be in matrix form, so we extract the lists of emdeddings from the dataframe and convert them to matrix form.
openai_embeddings_mat <-
     matrix(
          unlist(
               EIS_GPT_embeddings %>%
               .$embeddings
               ),
          ncol = 1536,
          byrow = TRUE
     )

# Estimate tSNE coordinates
set.seed(7267158)
tsne_embeddings <-
     Rtsne(
          openai_embeddings_mat,
          pca = TRUE,
          theta = 0.5,
          perplexity = 50,
          dims = 2,
          max_iter = 10000
     )

Determining the proper theta (i.e. learning rate) and perplexity (basically an estimate of how close points are in relation to the expected groupings) is more of an art than a science. This post does a great job of exploring choices for these parameters. By setting pca = TRUE in this case, we are first reducing the dimensionality to 50 principal components and then using tSNE to do the final reduction to two visual dimensions.

# Extract the tSNE coordinates and add them to the main dataset
EIS_GPT_embeddings <- 
     EIS_GPT_embeddings %>%
     mutate(
          tsne_dim1 = tsne_embeddings$Y[,1],
          tsne_dim2 = tsne_embeddings$Y[,2]
     )

# Visualize the tSNE plot
EIS_GPT_embeddings %>%
     ggplot(aes(x = tsne_dim1, y = tsne_dim2)) +
     geom_point(alpha = 0.5, pch = 16)

The first thing to note is that we are not seeing much discrete grouping of the points. This tells us that that the comments share a lot more in common across all comments than across local groups of comments. The second thing to notice is that despite the spread, we do see a handful of groups budding off along the periphery. In fact, one group in the bottom right is very distinct. It is important to remember that, unlike PCA, the axis dimensions in tSNE are meaningless. In fact, I’ll remove them from plot for the rest of the post. Position doesn’t matter in tSNE–only relative closeness.

At this point, we might want to manually delimit groups that we want to analyze further, like pulling out all of the comments from that cluster in the top left. To make this a bit easier, I’ve opted to cluster the two dimensional tSNE with hierarchical clustering. It is important to realize that this is purely a convenience for visualization. If we really wanted to use clustering to directly define groups (like hierarchical, KNN, etc.), it would make much more sense to cluster directly on the first 50 principle components.

tsne_embedding_clusters <- 
     hclust(
          dist(tsne_embeddings$Y), 
          method = "average"
     )

EIS_embeddings_clustered <-
     EIS_GPT_embeddings %>%
     mutate(
          cluster = cutree(tsne_embedding_clusters, 7)
)

Since we are clustering on tSNE dimensions where distance doesn’t really matter, deciding where to set our breakpoint is a personal choice. I’ve decided to use 7 clusters because it seemed a natural breakpoint and recovered the obvious clusters.

Text analysis of topical clusters

Now that we have putative clusters of topics, we can perform some classic natural language processing (NLP) to illuminate the themes of those topics. We’ll use tidytext for this task.

library(tidytext)

First, we need to get the data into a long, tidy format where each word in every comments is its own row. We’ll also remove common stop words that are predefined in the tidytext library. Then, we can calculate the term frequency-inverse document frequency (TF-IDF) for the clusters. TF-IDF is basically a measure of how common a word is within a cluster, after accounting for how common a given words is overall.

For example, if we take a look at the most common words in each cluster, it is unsurprising that “wolves”, “moose”, “isle” and “royale” dominate. (Although it is interesting that the top words for clusters 4 and 7 are “wilderness” and “management”… more on that later).

However, TF-IDF tells us about the relatively unique words that define a cluster of comments. Some clusters, like 1 and 2 have very even tf-idf distribution and the important words are mostly filler or nonsense words. This happens when clusters are saturated with common words and there is no strong theme producing uniquely important words. We could have guessed from the tSNE plot of the embeddings that the bulk of comments in the center of the plot would fall in this lexical no-man’s-land. But! Clusters 3, 4, 5, and 7 show promisingly skewed distributions.

Cluster 3 seems to orient towards a topic of animal welfare, with words like, “contraception”, “sterilization”, “lethal”, and “culls”. I suspect that these comments speak to folks’ concerned less about the wolf population or wilderness management, and more about the ethics of any proposed action involving animals. In a similar way, it looks like Cluster 7 is more concerned with the science and measurement behind the management decision and less about the decision itself with words like, “evaluating”, “approximately”, and “tools” with high uniqueness and “management” as the most common word overall. These topics would have been completely lost if we had stopped at categorizing favored alternatives.

Meanwhile cluster 4 appears to be squarely concerned with Wilderness issues. “Wilderness” and “nature” are the most common words in this cluster and “untrammeled” and “unmanipulated” are the most uniquely important words. We might expect that most of the comments that chatGPT categorizes as favoring alternative A will fall into cluster 4.

We can also take a look at how the clusters map onto the chatGPT categorizations.

Mappin the chatGPT categorized ‘favored alternative’ onto the tSNE coordinates, we can see that comments roughly sort by favored alternative. Cluster 6 is almost entirely defined by support for Alternative B – immediate wolf introduction. Cluster 4, which seemed to orient towards Wilderness values is mostly comprised of comments in support of Alternative A – no action.

Cluster 7 and Cluster 3, are mostly skewed to Alternative C – more monitoring, but exhibit very similar distributions. This might be a great example where even folks who tend to agree on the same Alternative, do so for different reasons–a pattern we would have totally missed without text analysis.

The remaining clusters which compose the bulk of the midland in the tSNE plot favor a mix of Alternatives.

Chain-of-density summarization

We can learn a lot from looking at common and important words and using our human judgement to piece together the topical theme of each cluster. Ideally, we would read all of the comments in a cluster to develop a topical summary. But that would take a long time. As an alternative, we can pass all of the comments in a given cluster to an LLM and have it summarize the theme.

Currently, only a handful of models support context windows large enough to digest the entirety of the comments in our clusters. Anthropic’s Claude2 has a context widow of up to 100k tokens (rough 75,00 words). Although, it isn’t quite as good at chatGPT 4. To get the most out of Claude2, we can use a special type of prompting developed for summarization called “chain-of-density”. Chain-of-density prompting forces the model to recurrently check it’s own output to maximize the density and quality of its summarization. Research shows that people tend to like the chain-of-density summaries even better than human-written summaries of new articles.

For demonstration, we’ll use chain-of-density prompting to summarize the theme of cluster 3. Here is the prompt that we will pass to Claude2:

"You will generate increasingly concise entity-dense summaries of the semicolon separated comments included below.

The comments were submitted by a member of the public in response to the The Isle Royale National Park Moose-Wolf-Vegetation Management Plan/EIS. The Plan/EIS is a document that evaluates management alternatives for the moose and wolf populations on the island National Park land.

Now that you know the context, here are the semicolon separated survey response:

[INSERT SEMICOLON SEPARATED COMMENTS]

Instructions: You will generate increasingly concise entity-dense summaries of the above semicolon separated comments. Repeat the following 2 steps 5 times.

Step 1: Identify 1-3 informative entities (delimited) from the comments which are missing from the previously generated summary.
Step 2: Write a new denser summary of identical length which covers every entity and detail from the previous summary plus the missing entities.

A missing entity is
- Relevant: to the main themes of the comments.
- Specific: descriptive yet concise (5 words or fewer).
- Novel: not in the previous summary.
- Faithful: present in the comments.
- Anywhere: located in the comments.

Guidelines:
- The first summary should be long (7 - 5 sentences, ~100 words), yet highly non-specific, containing little information beyond the entities marked as missing. Use overly verbose language and fillers (e.g., 'these comments discuss') to reach ~100 words.
- Make every word count. Rewrite the previous summary to improve flow and make space for additional entities.
- Make space with fusion, compression, and removal of uninformative phrases like 'these comments discuss'.
- The summaries should become highly dense and concise, yet self-contained, e.g., easily understood without having read the comments.
- Missing entities can appear anywhere in the new summary.
- Never drop entities from the previous summary. If space cannot be made, add fewer new entities.

Remember: Use the exact same number of words for each summary.

Answer in JSON. The JSON should be a list (length 5) of dictionaries whose keys are 'missing_entities' and 'denser_summary'."

I don’t have API access to Claude2, but unlike chatGPT, Claude2 accepts files as input. So, it is easy to output our prompt with the contained comments as a .txt file and submit it to Claude2 manually.

Claude2 gives us five increasingly dense summaries:

Summary 1: These comments provide a range of perspectives on how the National Park Service should manage the declining wolf population and increasing moose population on Isle Royale. There is discussion of various preliminary alternative concepts, including taking no action, reintroducing wolves once, maintaining both moose and wolves, reducing the moose population, and intensively managing the moose population. Many commenters express support for alternative C, which involves reintroducing and augmenting the wolf population on Isle Royale more than once if necessary. There is also frequent mention of the possibility of relocating problem wolves from elsewhere to the island. Overall, most commenters emphasize the importance of restoring the natural wolf/moose balance on Isle Royale through wolf reintroduction rather than intensively managing the moose population.

Summary 2: These comments provide perspectives on managing the declining Isle Royale wolf and increasing moose populations. Many support alternative C to reintroduce and augment wolves more than once if needed. Wolf relocation from the mainland is also mentioned. Most emphasize restoring the wolf/moose balance through wolf reintroduction rather than intensively managing moose. Climate change reducing ice bridges for wolf migration and the wilderness designation of Isle Royale are frequently discussed. Many oppose lethal moose culling, instead supporting moose contraception/sterilization. Overall there is a focus on restoring the natural predator-prey dynamic through wolf reintroduction while minimizing intensive human management due to the wilderness designation.

Summary 3: These comments discuss managing the declining Isle Royale wolf and increasing moose populations. Many support alternative C to reintroduce and augment wolves more than once if needed to restore the wolf/moose balance. Wolf relocation and genetics are often mentioned given reduced migration from climate change. The wilderness designation and opposition to lethal moose culling are frequently discussed, with support for contraception/sterilization. There is concern about moose overbrowsing without sufficient wolves. Some mention wolf tourism benefits. Overall there is emphasis on restoring the natural predator-prey dynamic and wolf genetics through reintroduction while minimizing intensive human management due to the wilderness designation.

Summary 4: These comments discuss managing the declining Isle Royale wolf and increasing moose populations, often supporting alternative C to reintroduce and augment wolves more than once to restore balance. Wolf relocation and genetics are frequently mentioned to compensate for reduced migration from climate change. Many oppose lethal moose culling, instead supporting contraception/sterilization, expressing concerns about moose overbrowsing and ticks without wolves. Some mention wolf tourism benefits and preserving the wilderness designation. There is emphasis on restoring balance through wolf reintroduction to limit moose impacts on vegetation while minimizing intensive human management.

Summary 5: These comments frequently support alternative C to reintroduce and augment Isle Royale wolves to restore wolf/moose balance and predator-prey dynamics, concerned about moose overbrowsing, ticks, and vegetation without sufficient wolves. Wolf relocation and genetics are often proposed to compensate for reduced migration from climate change. Many oppose lethal moose culling, noting wilderness values. Some mention continuing the iconic predator-prey research, wolf tourism benefits, wolf-livestock issues, and the value of public engagement. Overall there is emphasis on reestablishing a healthy Isle Royale wolf population through reintroduction to minimize intensive human management and moose impacts.

The final two summaries start to get at the nuance particular to this cluster. Summary 4 mentions, “Many oppose lethal moose culling, instead supporting contraception/sterilization”. Summary 5 mentions: “Many oppose lethal moose culling, noting wilderness values.” As expected, much of the summary is likely common across all comments. So, a human interpretation of the summaries is still necessary.

Conclusion

In this post, we explored using chatGPT, text embeddings, and Claude2 to analyze public comments on a complex environmental management decision. We learned methods to responsibly validate chatGPT output. While not perfect, chatGPT showed promising reliability at categorizing free-form opinions. The text embeddings allowed us to uncover hidden topical clusters among comments that traditional methods would have missed. Claude2’s long context window allowed us to further interpret the topical clusters. Together, these tools enabled a nuanced quantitative analysis of subjective text data that would be infeasible for a single human analyst to perform manually.

From my field notes, 17 August 2018:

“Last night I slept on the flat bench of duff-covered ground at the base of the ‘camp tree,’ a white and skeletal cedar with the characteristic axe marks from decades-past Tlingit or Haida woodsmen who had notched the bark from the underhanging side of the leaning tree trunk. As Wayne, our on-hand anthropologist, explained, the notched wood of the tree would have died and dried in the rainshadow of the trunk making for ready tinder to be harvested the next season when the camp was reoccupied. The overhanging tree would have also provided the Native campers with shelter for fire or lean-to bivouac. One of the same purposes for with I utilize it today. 

As I fell asleep under the history-laden tree, I dreamed of a boat at anchor in high winds. As the gales tugged at the boat, the anchor dug into the muddy bottom and the rode strained taught to the windlass bolted to the foredeck. But the wind was strong, and the windlass was pulling from the deck boards and the road was splitting. With each gust, the planks screamed “Eeeeerrrrck!”

A few more heavy gusts would rend the boat from anchor, setting it adrift in the turbulence. Another gust and a louder “Eeeerrrrck!”

In my dreamy torpor I was frozen, I could do nothing but watch.

“EEEEEERRRRCK!” and I watched the windlass pull loose, hanging by a single bolt.

Just before the final fastener snapped, I woke up to the sun shining sidelong under my tarp and straight into my eyes. I could still hear the creaking Eeeeerrrck sound out in the cove and realized it was the agitate call of a belted Kingfisher. The sound of the gusts in my dream had been the sound of light swell washing the beach.”

One of my favorite quotes about Wilderness is from former Senator Clinton Anderson who said, “Wilderness is an anchor to windward. Knowing it is there, we can also know that we are still a rich nation, tending our resources as we should—not a people in despair searching every last nook and cranny of our land for a board of lumber, a barrel of oil, a blade of grass, or a tank of water.” No doubt, I must have had that quote on my mind as I fell asleep on the third night of our three week expedition in South Prince of Wales Wilderness.

South POW is one of the many forgotten Wilderness areas of Southeast Alaska, overshadowed by the iconic ice-filled fjords like Tracy Arm or Glacier Bay, and tourist hotspots like Misty Fjords. Nevertheless, it can easily compete in scale and dynamic topography, and what it lacks in defining photo-geniality, it more than makes up for in its isolation and profound solitude.

At just under 90,000 acres, South POW is the 10^th largest and the southern-most (narrowly outcompeting the southern point of Misty Fjords) of the twenty-four Wilderness Areas in Southeast Alaska. As its name implies, the designation was carved from the southern portion of Prince of Wales island, including the watersheds draining into Klakas Inlet, all watersheds south that drain to Cordova Bay and Barrier Islands. Encompassed in the boundary is a labarynith of convoluted (some might say fractal-like) shoreline. This is dynamic and magic country. It is intricate and you need to see it intimately, by kayak, or even better, by foot. Inlets and passages otherwise hidden from sight appear as if by incantation, the trees parting to a kayak-width channel only when you paddle immediately beside.

From my field notes, 18 August 2018:

“From our southernmost camp, we paddled out to the most seaward rocks and islands. Out here on the outside waters, among the breakers, expansive kelp beds, and scattered, battered islands with rocky headlands. White breaking water sparkled in the sun with each swell, but the water was otherwise calm. All of the branches of the few trees on these islands point leeward, like frozen weathervanes recording past winter storms originating from the Dixon Entrance to the south.”

From my field notes, 19 August 2018:

“After exploring the Barrier Islands, we spent time tracing the coastline of the main island which is punctuated by deep, multi-chambered inlets, bays, and saltchucks. For instance, a side channel in Hunter Bay opens into Biscuit Lagoon. At the back of Biscuit Lagoon, the Saltchuck opens at high water above a tidal waterfall. The shoreline is like walking through a labyrinth with new passages appearing and opening into whole new habitats.

We paddled up to the base of the Saltchuck where salmon were preparing to run at high water. Wayne, whose eyes always seem to catch any inorganic shape on the landscape, noticed that beneath our tethered kayaks, below the waterline, rocks had been arranged in a line. He reasoned that at lower water the stones would cordon off a pool. The Native fishermen would have either trapped salmon in the pool or used it as a holding pen for their catch pulled from the falls.”

The Wilderness is a mélange of habitats from wave washed coastal shores, to glass-still secluded bays, upland muskegs, lowland salal thickets, and rich estuaries throbbing with activity. The varied habitats granted us countless wildlife sightings, include a few once-in-a-lifetime encounters.

From my field notes, 19 August 2018:

“As we ate lunch next to the falls and pondered the rock wall, Kim noticed two wolves trotting up the shoregrass upwind from us. When the wolves saw us, the first bolted. The curiosity of the second got the best of him. With the wind preventing any scent information, he boldly came up to us for a close look. Satisfied, he trotted back away only to decide he still needed a closer look. This process repeated, with us standing like stones, the click of my camera shutter the only sound, until he dematerialized in to the forest.”

The Alexander Archipelago wolf are a controversial subspecies (Canus lupus ligoni) of North American wolf that occur only in the islands and mainland of Southeast Alaska. They are absent from the ABC islands which are dominated by brown bear, but the range is largely congruent with black bears in Southeast. No fossil evidence of the wolves exists, suggesting that the species colonized after the last glacial maximum.

Goldman (1944) was the first to describe the species and granted the name Alexander Archipelago wolf. A more recent molecular study (Weckworth et al. 2005), corroborated the original distinction, finding that the Southeast Alaska population was genetically delineated from the continental population and was itself highly structured and diverse. That same study found evidence that the Prince of Wales population formed its own unique and isolate subgroup within Southeast Alaska. Having spent a fair amount of time with wolves in Glacier Bay and Gustavus, our crew had guessed this was the case. In fact, it took us a split second to even realize we were staring down a wolf when it appeared–it was so dissimilar from the canids we had encountered up north. These island wolves appeared smaller, more dog-like, and with an unusual coat pattern.

The biogeography of the Prince of Wales wolves puts them in a particularly perilous position. It is a general rule in ecology that the smaller and more isolated a population, the more sensitive it will be to environmental and demographic fluctuations (for example, the case of Isle Royale wolves). Although Prince of Wales is an exceptionally large island (the 4^th largest in the U.S., just after Puerto Rico), the landscape itself has been highly fragmented and reduced by massive clearcuts and the most extensive road system in Southeast Alaska (about 2,500 miles; more roads than the rest of Southeast combined, many times over).

Roads are a major problem for POW wolves (even more so than for most widlife). Studies across all U.S. populations of wolves have shown a negative correlation between road density and wolf abundance. Wolves were absent where road densities exceed 0.9 mi/mi² (Jensen et al. 1986,
Mech et al. 1988, Fuller 1989, as cited in Schoen and Person, n.d.).  Most of the roads on POW were created by logging companies (on tax-payer dollars), which makes a strong correlation with roads and clearcuts, both of which are avoided by wolves (Person 2001).

Black-tailed deer, the wolves’ primary prey, require old-growth forest for summer browse and winter habitat to shelter from heavy snow. Clearcutting areas leads to a “succession debt” for wolves as logging removes habitat which removes deer. And logging continues on the island, driving the succession debt further into the red. In fact, habitat loss from logging alone is expected to reduce the wolf population by a further 25% before 2045 (Person 2001).

But habitat loss is not the only threat to POW wolves. Despite Canis lupus being listed as an endangered species in the lower 48, wolves are considered both big-game and furbears in Alaska, subjecting them to both hunting and trapping. Between the 1980s and 2010, around 175 wolves were killed in Southeast Alaska; 75 of those were from POW (Schoen and Person, n.d.).

As logging and associate road infrastructure continue to proliferate, “wolf populations on Prince of Wales and adjacent islands will be caught between two significant pressures: declining prey abundance and increasing hunting and trapping mortality” (Schoen and Person, n.d.).

The South POW Wilderness, at least, is a partial refuge for the wolves. While Wilderness designation prevents future roads and logging, it does not exclude hunting nor trapping; so, there really is no safe haven for these unique canids. My hope is that the isolation and extreme difficulty of hunting here compared to the road system will keep it to a minimum.

In addition to the wolves, we came nose-to-nose with loads of wildlife, including sandhill cranes, black bears, whales (alive and bones), and so many kingfishers and river otters.

From my field notes:

“More kingfishers than I’ve ever seen! The shoreline seems like it is just dripping with kingfishers from every low-hanging branch. And below, the tidal rocks are writhing with families of river otter. It must be a productive place for small carnivores and pescavores!”

We spent a large chunk of our time in Klakas Inlet, the long fjord that dominates the norther reach of the Wilderness. Although not too steep, the shoreline harbors very few beaches along the flanks. One exception is a grassy cove sheltering a pink and sockeye waterfall. We spent an entire day watching black bears filter through, try their luck with the flying salmon, and continue their up-fjord journey.

The density of black bears in Klakas was astounding. We paddled to the head of the inlet, where the fjord splinters into three massive estuaries, each full of late-running salmon. At one point, I could see 7 bears around the shore. They circled the shoreline of the fjord almost like a slow and dispersed school of fish circling a pond.

From my field notes, 27 August 2018:

“In the morning I watched multicolored bat stars pass beneath my kayak. Their five to six bright arms were like a dappled rainbow fluorescing against the dark water. Watching black bears pass by on shore is like the inversion of the bat stars. The dark fur seems to absorb light and looking for bears is more about looking for the absence of a bear or a bear shaped black hole in the otherwise colorful shoreline.”

From my field notes, 21 August 2018:

“On every trip, we try to make a point of hiking uphill, above the treeline, in order to look back down on the landscape. Experiencing a Wilderness is a lot like experiencing a painting. It takes time. You have to shift your focus and view it from different angles to take it all in. Look close, then step back. Put your nose right up to the canvas, then from across the room. In wild places, you have to paddle close to shore, get out and crawl through the brush of the forest, then get up high to see the whole thing. The detail and the composition are the functional integral of artistic mastery; same goes for the majesty of nature.”

So, we set our sights on a bald knob at the mouth of Klakas inlet and landed below a steep, even-aged stand of spruce and hemlock. We had envisioned a strenuous, patience-trying hike over old growth deadfall and bashing through endless under-story brush. It turned out that the stand had been beach logged many decades ago. The regenerating forest was the perfect age to shade out the under-story brush, but too young to accumulate much deadfall. It turned out to be one of the easiest routes we’d ever hiked. Surprisingly, there was not even brush at the transition from the forest to the subalpine bald, so we walked out into the sun and onto the rocks. From the top, we were treated to views of the entire southern half of Cordova Bay, up into Klakas Inlet and the ridges of the surrounding watersheds. In short, we could see almost the entire Wilderness area and the saltwater well out into the Pacific.

The pools on the bald were ringed by bright red sundews, a carnivorous plant that produces a sticky, sweet digestive enzyme from the tips of hairs on its leaves. Bug are attracted to the scent, land on the leaf, then wrap themselves up in the leaf as they struggle and adhere to more leaf hairs.

I see the round leaf sundew (Drosera rotundiflora) often, but I was surprised to see the English sundew (Drosera anglica) was more common and growing right alongside the D. rotundiflora.

From my field notes, 22 August 2018:

“The Native Haida and Tlingit folks who frequented this region made ample use of the fractal coastline. In almost every landable spot, Wayne noticed some trace of human occupation Some hints were subtle, like notches in trees or new forest growth indicating an area that had been cleared. Occasionally, the signs were more obvious, like the square foundations of old plank houses subsumed into the ground, and even pilons supporting old floor joists, now fixed in place by the roots of saplings growing atop them. The debris of stoves and other iron objects were sunk into the moss of a few such sites.

Wayne pointed out an iron ax head—the same type of ax head used to make the tree notches we’d been seeing all over the coast.”

After 16 days, we left the Wilderness toward Hydaburg. After days of calm waters and avoiding the oppressive sun, we found ourselves preparing for some exposed sections and large crossings in big water. The water shut us down and sent us back to camp for an extra night. We made progress despite lots of waiting and watching as the conditions fluctuated. We made our longest crossing of Cordova Bay in building, following seas. About midway across, we spotted black fins slicing the water miles away. As we thrashed onward on our heading, we watched the pod of orca surface and dive, rapidly closing the distance between us. The bull crossed comfortably forward of our bows, but the four females intersected our course. They surfaced with explosive spouts close enough to startle us. We exchanged some wide-eyed glances of concern while the females cavorted amidst out kayaks until they melted back into the water and joined the male far off to starboard.

From my field notes, 30 August 2018:

“South POW is a palimpsest. The page has been written by glaciers, over-written by many chapters of ecological succession, punctuated by interludes of logging, and a final chapter back to succession, all with the footnotes and interjections of Haida and Tlingit history throughout.
Now, having departed the Wilderness Area, our time there seems even quieter in contrast. Clear cuts appear over every ridge and we hear the whining buzz of outboards every few tens of minutes as boats zip in and out of Hydaburg.”

We paddled into Hydaburg on a quiet mid-morning diffused with grey. The brightly painted poles and homes seemed like holes pricked into the grey blanket of the day.

We found the folks in Hydaburg incredibly welcoming. We were invited into smokehouses and carving sheds, folks told us stories and we even made canine friends after saving a dog from drowning at the dock.

Overall, our South Prince of Wales expedition will remain one of my favorites. The immense quiet and solitude, the shoreline full of intrigue, the enigmatic wildlife—will define my memory of this place.

Route and Logistics:

Out trip could not have happened without Katy. We rented all of our kayaks through her, and in order to maximize our time in the Wilderness, Katy transported us to Klakas by skiff. She also picked us up and even stored our gear for us.

If you decide to paddle in South POW (and you should) keep in mind that camping can be sparse, weather can pick up quickly, parts of the coastline are committing, and communications are almost non-existent. Talk to locals, know your own skill level, and pour over your charts. Of course, I’m always happy to provide beta. Just drop me a line.

Check out my posts from other expeditions.

Check out other posts about designated Wilderness. 

References:

Fuller (1989) Population dynamics of wolves in northcentral Minnesota. Wildlife Monographs 105.

Goldman (1944) Classification of wolves. In S. Young and E. Goldman, eds., The wolves of North America, Part 2. Dover Publications, New York

Jensen, Fuller, and Robinson (1986) Wolf (Canis lupus) distribution on the Ontario to Michigan border near Sault St. Marie. Canadian Field Naturalist 100: 363-366.

Mech, Fritts, Radde, and Paul (1988) Wolf distribution and road density in Minnesota. Wildlife Society Bulletin 16:85-87.

Person (2001) Alexander Archipelago wolves: ecology and population viability in a disturbed, insular landscape (Doctoral dissertation). University of Alaska, Fairbanks, AK.

Schoen and Pearson, Chapter 6.4 “Alexander Archipelago Wolf.” In A Conservation Assessment and Resource Synthesis for The Coastal Forests and Mountains Ecoregion in the Tongass National Forest and Southeast Alaska. The Nature Conservancy.

Weckworth, Talbot, Sage, Person, and Cook (2005) A signal for independent coastal and continental histories among North American wolves. Mol. Ecol. 14: 917-931.

Overview:

Isle Royale is an island in Lake Superior that is designated as a Wilderness Area and managed by the National Park Service. In the 20th century, wolves and moose migrated to the island and their dynamic spurred one of the longest predator-prey studies in history. Now, the wolf population has dropped to 2 and the Park Service is planning a major intervention that will install an entirely new, synthetic population of wolves on the island. This plan is the result of myopic research perspective and disregard for eco-evolutionary dynamics. It is bad policy and even worse science. Here’s why:

Background:

Isle Royal is a smallish-island (just over 200 square miles) that sits about 8 miles from the north shore of Lake Superior. (Although it is small, it is large enough to host its own internal lake with an island, making it, as upper-midwesterners are fond of point out, the largest island on the largest lake on the largest island on the largest lake in the world.) The Isle and its many tiny satellite islands became a National Park in 1940 and were designated as a federal Wilderness Area in 1976.

Because the island is small and a long swim from the mainland, large fauna populations have been inconsistent denizens, historically. Moose first arrived on the island in the early 1900s. Wolves followed the moose in the 1940s, adding two major trophic levels to the island ecosystem. The complex predator-prey interactions became one of the classic test cases of ecological theory (see Peterson et al. 1984; McLaren & Peterson 1994).

Over the decades, the wolf and moose populations have demonstrated a standard predator-prey oscillation, with the wolves generally bouncing around about 20 individuals, but reaching a population maximum of 50 individuals in the 1980.

However, Isle Royale is a small place. Small islands are more susceptible to tipping points on the roller-coaster of demographic stochasticity. It’s kind of like drunkenly walking along the centerline of a bridge versus drunkenly walking a tightrope. If you stumble off course too far on the bridge, you have the latitude to recover and get back on course. Too big a waiver on a tightrope and you’re done for. The small size and isolation of Isle Royale makes it a tight rope for large predators. Like all oscillatory ecological patterns, what goes up eventually comes down, and in the last decade or so, the wolf population has declined in a mirror-like inversion of the population boom in the 1980s. As of this year, there are only two wolves left. As per the dynamics of island-biogeography, the natural course looks like the rein of the wolf will eclipse on the island, probably followed by a boom and eventually extirpation of moose, and the island will continue along as it did for the many decades prior to the most recent immigration events. That is, until the next colonists arrive, as has happened multiple times in the past. Coyotes immigrated and blinked out in 50 years in the early to mid-1900s. At times, lynx and caribou both made the pilgrimage to the island and subsequently slipped off the tightrope.

The issue:

Now, the National Park Service has released an Environmental Impact Statement (EIS) for a plan to install a new populations of wolves on the island (available here). If you are unfamiliar with the NEPA process, here’s how it works: When a land management agency like the National Park Service wants to embark on a project that might run counter to its mandate and/or result in large impacts, they are required to vet all potential options, usually as an EIS, and ask for the public’s comments on the plan. After the revision process, they make a final decision to enact one of those potential options, the “preferred alternative.”

Since 99% of  Isle Royale Park is a designated Wilderness, “where the earth and its community of life are untrammeled by man” and “generally appears to have been affected primarily by the forces of nature, with the imprint of man’s work substantially unnoticeable” (Wilderness Act, 1964), shipping in a boatload of wolves to manipulate the ecosystem required an EIS.

As a scientist and especially as an ecologist, I tend to view Wilderness Areas as our most critical ‘controls’ or ‘baselines’ for science to contrast other areas where human impact alters systems. Though every system is touched by human impact to some extent, there is huge value in preserving the least impacted places in an unmanipulated state. As an analogy, a blemished diamond might be worth a little less than a perfect diamond, but that doesn’t reduce it to equal value with a lump of coal.

But, not all scientists think that way.

To introduce or not to introduce:

The push to introduce wolves to Isle Royale has been championed primarily by two researchers at Michigan Tech, John Vucetich and Rolf Peterson, whose careers are rooted in the Isle’s wolf-moose study.

I first heard about this proposal when Vucetich gave a presentation at the Sigurd Olsen Environmental Institute. At the time, it wasn’t the science that bothered me about the presentation–it was the patent misrepresentation and obvious straw-man Vucetich employed to characterize the intention of the Wilderness Act. Since then, these researchers have made major pushes in film (and this one) and popular press to portray wolves as an ever-present and integral part of the Isle Royale ecosystem, and pit the “health” of the ecosystem against what they believe is an outdated philosophy of conservation.

Essentially, their argument is that since climate change impacts the whole globe, no wilderness is really free of human manipulation, so we should be free to further manipulate it to our own design. They explicitly argue for “new visions for the meaning of wilderness,” with their preferred vision being “a place where concern for ecosystem health is paramount, even if human action is required to maintain it” (from here).

Intentionally or not, their use of relativistic “ecosystem health” rhetoric and attempts to stretch the ‘wilderness myth’ concept into their own application has thoroughly muddied the debate.

And it’s resulted in a lot of public confusion on the topic. For instance, here’s one comment I pulled from the public response in the EIS:

I have visited Isle Royale twice and it remains one of my favorite places in the world. The wolves and the moose have become a part of the island and that is a good thing. Wolves and moose aren’t faring well on the mainland due to politics ignorance and climate change. Isle Royale remains a unique microcosm where we can still observe and study this ancient predator-prey relationship. In a world where species are becoming extinct on a daily basis, this rare relationship has endured and that should be given a lot of weight when making the decision of what to do about the wolf-moose problem on Isle Royale. Please use common sense and act sooner rather than when it is too late.

First off, it’s not an ancient relationship (it’s only been going on for 60 years on the island), and it’s not a rare relationship (wolves eat moose all over the continent all the time). What makes it “rare” is the fact that it happens without human intervention (at least until NPS takes control of the population) on an isolated island with researchers tracking every move.

This person’s comment shows that opinions on the wolf issue are completely colored by human perception: i.e. anything that happened before your lifetime is “ancient,” anything that looks the way it is when you first saw it is “natural,” if you’ve only heard about something in one place, it must be “rare,” etc. The most pernicious perception is that the only species that are worth concerning ourselves with are the big ones with faces that you can relate to (after all, amphibian populations fluctuate on and off in ponds all over the upper midwest following the exact biogeographic pattern as the wolves of Isle Royle, but I’ve yet to see an outrage).

Even the main proponents of wolf introduction, Vucetich et al. and the National Parks Conservation Association invoke the myth that a “sustainable” wolf population is critical to the island’s “health.” Considering that wolves only appeared on the island within a human lifetime and probably blinked on and off the island historically, wolves are only an ephemeral component of this dynamic ecosystem. They never have been “sustainable,” and if ecosystem “health” hinges on the presence of wolves, the island has always been naturally unhealthy.

The rhetoric of “healthy” ecosystems is useless in science, because its meaning is entirely relative. Rolf Peterson, the researcher who initiated the moose-wolf study in the 70s, states that, “There’s a mythical belief that Isle Royale has been working well because we kept our hands off it; my opinion is, it worked well because there were wolves there” (from here). You can only consider a wolf-inhabited Isle Royale as “healthy” if you define a “healthy” ecosystem as one that looks the same way it did when you started your research plan. The real myth is conflating wolf presence with Isle Royale’s natural state, and in this case, it seems a personal mythology crafted to shore the legacy of Peterson’s research project.

The Park’s plan:

The preferred action of NPS is to install 20-30 wolves on the island over the next 3 years, and if those don’t take, to continue introducing for 2 more years.

Originally, the proponents of airdropping new wolves onto the Isle proposed it as genetic “rescue.” But with only two post-breeding age, inbred stock left, there is little chance that new wolves will breed with the two relics. Thus, in reality, this is not a genetic rescue project, it is a genetic replacement project.

So, where do the replacement wolves come from? The EIS suggests that wolves should be sourced from the mainland near the Park, but that many different populations around Lake Superior should be mixed on the island. They also suggest sourcing wolves with experience hunting moose (which are rare in mid-western populations).

Will a new population fair better? The reason wolves lost the plot in the first place was due to the ubiquitous force of natural selection. When faced with strong selection pressure, organisms are faced with three choices: move (not possible on an island), adapt, or disappear. The current wolves were not able to adapt to the ecological scenario they found on the island, so they are disappearing. The NPS knows that new wolves will be even more likely to succumb to selection pressure because they will not be locally adapted. This is why they are planning recurring introductions for a total of up to 5 years. The new population, with lots of diverse genetic material to work with, might be more prone to local adaptation, or it might be more prone to crash because the animals are too locally adapted to their naive system to cope in the new setting.

It might be tempting to think that evolution won’t be a factor considering the short tenure of wolves on the island, but wolf generation times are under 5 years (Mech et al. 2016) which means that they’ve had about 20 generations on the island. We know from the deluge of rapid evolution studies in the past few years that 20 generations is well within the timespan for marked evolution. Similarly, one can expect that moose have been evolving in that time too (Hoy et al. 2018), as have the plants that are browsed by moose, and the small mammals, and the microorganisms that exists in concert… In other words, the entire trophic system has been subject to dynamic eco-evolutionary change that has refined its assemblage and genetic composition. Replacing local wolves with wolves from elsewhere will short-circuit that dynamic process and set a new eco-evolutionary trajectory. Any study that occurs post-introduction will be studying a different eco-evolutionary system, altogether.

Proponents have made the case that occasional genetic influx from the mainland population (when a wolf might cross the ice to the island in cold winters) is part of the natural dynamic, but that climate change has disrupted this process. Leaving aside the fact that much of Isle Royales history was wolf-less long before climate change, reintroducing wolves does not simulate this natural process. In natural migration events, wolves are not randomly selected from a larger pool. The process of migration is a selective sieve that winnows out some potential migrants and selects for others. By high-grading the genetic stock from the mainland based on their own criteria, the Park Service will not be replicating nature, they will be conducting a large-scale, manipulative selection experiment.

The value of non-intervention:

As I mentioned, one of the most critical values of Wildernesses are their role as baselines. This is a point  repeatedly highlighted in the “Strategic Plan for Scientific Research in Isle Royal National Park” (Schlesinger et al. 2009).  The Plan lists as “Unique Attributes of Isle Royale National Park” that it is “An Isolated Location for Baseline Studies”, and “an Ideal Place to Study Fundamental Ecological Concepts” like island-biogeography and predator-prey dynamics.

Isle Royale attracts biogeographers, whose focus is the distribution of life forms as determined by the balance of regional dispersal and local extinction processes (MacArthur and Wilson 1967). Determination of the relative importance of both dispersal and extinction is of central interest to ecologists wishing to explain variability in the species diversity of a given environment and the potential changes brought about by environmental change. As the Strategic Plan states, “the pristine nature of Isle Royale offers an opportunity to examine the potential influence of regime shifts due to natural causes or indirect anthropogenic causes such as climate change.” If we choose a policy of artificially imposing stasis on a naturally dynamic ecosystem we lose that value almost entirely.

On the other hand, if we practice humility and allow natural systems to be dynamic, we can ask a list of interesting questions: What happens if we remove those top trophic levels of moose and wolf? How will that impact the nutrient cycle? How will it impact community dynamics? In what ways will the change in selection pressures drive evolution? Will the eco-evo dynamic play out in predicatable ways based on theory and inference from other archipelagos? What will the post-wolf community composition look like and will it be the same as the pre-wolf community? Etc. etc…

There are endless scientific questions that a wolf-less IR can answer. On the other hand, a replacement wolf population cannot even answer the original question that it is intended to address because such a manipulation cannot be considered a continuation of that community; at best, we can only consider this a manipulative experiment at the price of sacrificing an entire natural ecosystem and ruining an exemplary opportunity to study eco-evo dynamics.

References:

Hoy, S.R., Peterson, R.O., and Vuctich, J.A. 2018. Climate warming is associated with smaller body size and shorter lifespans in moose near their southern range limit. Global Change Biology. DOI: 10.1111/gcb.14015