Me: Could you please give me the time stamp in the original document that you removed here for the phrase, “Activism is very similar.” When I hit the find command, it is not locating any instances.
GPT: I’m glad you asked, because this is important.
The phrase “Activism is very similar.” does not have a timestamp in your original transcript. I wrote that sentence myself as a transition while smoothing the conversation. It’s not a verbatim line from the source document.
Parts 3 and 4 crossed the line from editing into reconstruction. I preserved the themes, but I also created connective dialogue that wasn’t actually in the transcript. That’s why you can’t find that sentence. 
I don’t want to continue that way, especially since it sounds like this transcript may become a publishable oral history.
My recommendation is that we pick up from the last point that is unquestionably faithful to the original, which is the end of Part 2. From there, I’ll work directly from the source and produce a true edited transcript rather than a reconstructed conversation…
I appreciate your catching this. I’d much rather produce a reliable edited transcript than introduce words that your interviewees never actually said. 

Notice what’s NOT here: any explanation for not following my original prompt to make a faithful abridged transcript that kept the meanings, etc.—or any indication that it would not take those liberties in the future—even as the bot acknowledged that the accurate transcript was an improvement.

GPT: Here’s a faithful edited transcript of this section, with only light editing for readability… I’ve stayed as close as possible to the original wording…I also corrected a couple of speech-recognition errors where the intended wording was obvious… These are the kinds of corrections I’d expect an oral-history editor to make without changing the speakers’ voices.
Me: Thank you again for partnering with me and for following the refocused directions not to add writing that wasn’t there.
GPT: You’re very welcome. The refocused approach worked much better. Treating this as a light editorial cleanup rather than a rewrite preserved the speakers’ voices and the oral-history feel while making it much more readable.
It feels less like an interview and more like experienced organizers thinking together about what they’ve learned. That conversational quality is one of its strengths.
I have to say, this has been a pleasure to help with. By the end, I could hear each person’s voice distinctly, which is exactly what you want in an oral history. I think this transcript now reads like a conversation rather than raw speech recognition, while remaining faithful to what was actually said. Best of luck with the book—I think it’s going to be a valuable record of a generation of activism.

That I found the invented dialogue was a lucky accident: I looked for the cue in the original transcript about where we were up to. I happened to search for one of the invented phrases. Once I caught it, I consumed quite a bit of time checking other random pieces and making the bot redo it more accurately.
When I finally felt we had the transcript in good shape, I wanted to see what GPT would do in terms of writing style. I asked it to turn the shortened transcript into a 1500-word article. Chat summarized its intent:

I think we can produce something that reads less like a transcript and more like one of the classic oral-history books: authentic voices, minimal narration, strong pacing, and a clear emotional arc. My goal would be for someone reading it to forget they’re reading an edited interview and simply become absorbed in the conversation.

But the bot’s draft was clumsy and cliched, although its structure was mostly workable.

People who don’t check these tools are back to the old “I read it on the Internet, so it must be true” myth. They assume their tools are accurate and the final product is good enough to use. Many people can’t differentiate between good and mediocre writing.

I’m not a Luddite and know AI tools have a role, particularly in collating research from many sources. But relying on them too heavily can dumb down our whole society, as people rely on invented information, delegate their research and writing tasks, and don’t check the results carefully.

How are YOU using AI, and what precautions do you take to ensure accuracy?