In the real world I am a researcher and university teacher on technology and culture. I am at once deeply troubled by AI’s potential to shortcut meaningful human thought, and curious about its applications as a cognitive support. My use and discussion of AI here is not intended as an endorsement of the technology. Rather, I wish to include interested and/or concerned readers in my exploration.

For my review of Isle of the Bleeding Ghost, I was interested in seeing how AI could help me move past some of the cognitive obstacles I tend to encounter when reading and writing about RPGs. Specifically, I hoped it could help me organized my usually scattered thoughts and help me focus on reading and responding authentically to the text. My process in writing the review was as follows:

  1. I downloaded and printed the adventure PDF.
  2. While reading the module, I dictated my thoughts as they arose into Google’s Recorder app. This app uses Google’s AI technology to automatically produce a transcript of the recording.
  3. I fed the full transcript into ChatGPT. First, out of curiosity, I asked it to write an entire review. I found the output generic and ridden with awful prose.
  4. Next, I asked ChatGPT to produce an outline for this review.
  5. Finally, I asked ChatGPT to produce a list of the follow-up questions and to-do items from the transcript (I had noted a number of these verbally).
  6. I revised the generated outline. The organization it proposed was not very effective and would have made for an even more rambling review. Notably, I had to re-include many of my more critical comments, which it had filtered out. It also misunderstood me in a few instances - referring, for example, to a “trapped cellar room” that does not exist in the module.
  7. After struggling to reorganize that outline into something a little more intentional, I asked ChatGPT to identify critical themes in my transcript. Getting it to do this well took some trying. Eventually it produced a list of 10 themes, which I had it boil down to 4. They were deeply generic and not particularly true to what I had said. They became a little more useful after I asked it to support those themes with specific quotations from the transcript.
  8. The themes and quotations helped me sort, narrow down, and clarify my own outline in progress, from which I composed the first draft of the review.
  9. I fed the draft back to ChatGPT, asking it correct my spelling and grammar while changing as little of the original wording as possible. I proofread the corrected version and found it had largely done so, although although it did take a couple liberties. I fixed up the parts that didn’t sound like me and send off the post.

One challenge with doing this review was that unlike me, Bryce works fast. All said, the reading took a languid hour or so (I am a very slow reader) the outlining process 90 minutes, and the writing and proofreading another 90. That’s an improvement over the infinity minutes the outline likely would have taken me otherwise. It is still way longer than I could manage every time Bryce wrote a review.

As a focus/memory tool, the AI was just okay. It helped, but it needed a lot of help in turn, as it tends towards generic, lukewarm, and superficial statements. I’ll probably play around with it more and see if I can find a tighter workflow.