The prompt was pretty basic:

“Could you please create an annotated infographic, using the attached screenshot of a transcript from a news interview? The infographic should present details from a conversation-analytic study of the interview based on research by Steve Clayman and John Heritage.”

What we’re not drowning in is money. It’d be great to hire students and post-docs for projects, but in the absence of that, I’ll gladly use tools that allow me to do more with less. I don’t see this as unethical.

There are interactional phenomena that are relatively rare or elusive yet analytically significant. Gene Lerner’s approach is to build collections of these over years or decades as one comes across specimens. An AI that can watch and listen might find hundreds of these in days or weeks.

I’d say we’re well on our way to artificial general intelligence (AGI), systems that are as intelligent as humans across all domains, including CA. The real breakthrough will be when AIs can analyze video data directly and help us build and analyze collections.

AI models are good at literature reviews. What I like about this test is that coming up with a good idea for a project, or even a good place to start, is a harder task.

As a lecturer, I tend to size up AIs by academic level, from undergrad to PhD candidate. Where would you place these AIs on that scale?

These observations are not entirely dissimilar from some early-stage CA projects: spotting a "media disclaimer" in the data could very well serve as a point of departure for further analysis. It's not an uninteresting observation.

The phenomena the models found (e.g. “weather talk”) lack the fine-grained specificity CA generally demands, but they did identify multiple instances of "a something". Questions about and assessments of the weather are indeed recurrent in the NB calls, for example.

Projected co-presence work by Claude Opus 4

Opus identified “projected co-presence work” as a phenomenon in which callers negotiate future face-to-face meetings and manage relational and practical dimensions of reunion.

docs.google.com/document/d/1...

Media distancing disclaimers by ChatGPT o3

o3 identified “media-distancing disclaimers” as a recurrent move: speakers emphatically refuse to watch distressing TV coverage to project moral self-care, recruit affiliation, and pivot out of sequences of shared lament.

docs.google.com/document/d/1...

Weather talk by ChatGPT o3

o3 identified “weather talk” as a systematic practice: immediately after greetings, callers offer stance-laden weather assessments (e.g. “beautiful,” “lousy”) to secure affiliation and signal the shift into substantive conversation.

docs.google.com/document/d/1...

Yes, I'd steer my AI students away from those terms as well, but I can see where they've picked them up (e.g., onlinelibrary.wiley.com/doi/full/10...., www.taylorfrancis.com/books/mono/1...). Collecting data will be tricky, but once they can use computers, they'll be able to access online corpora.

This reminds me of what happens when two Claudes talk to each other for an extended period of time.

www.astralcodexten.com/p/the-claude...

What do you think of these AI-generated PhD proposals? Would you supervise an AI if it wanted to do one of these projects? What about a human guided by an AI?

The quality of the proposals varies by AI: some show signs of insight and creativity, while others miss the mark. Keep in mind that this is as bad as AI will ever be. Future models will only become smarter and more capable.