Anthropy

This summary is not available. Please click here to view the post.

AI Interview Claude 2

On September 29th, 2026 I participated in another AI interview from Anthropic. I previously did a similar interview near the end of 2025. During the 2025 session, I asked, "Will I retain a copy of this chat session?" and I was told, "If you're using the Claude interface directly, the conversation should be saved in your chat history." At that time (using the web browser interface for Claude) the conversation was not saved in my chat history, nor was it this time (using the Claude app). In both cases, I manually saved the contents of the chat session (copy and paste into a blog page). In 2025 I assumed that the "interviewer" was the Claude LLM, but maybe the "interviewer" is a different model that has been specifically trained to ask people a set of pre-determined questions.

Below on this page, after the formal interview, I had Claude conduct another interview specifically centered on my interests (as remembered by Claude) and my pattern of use of AIs. 

Interviewer:  Hello! I’m an AI interviewer from Anthropic. As AI is becoming more capable and more consequential, we want to hear how it’s showing up in your life.

Please don’t share anything private such as health details, personal information, or identifiable information about others.

Ready to start?

Me: Yes, I am ready. 

 Interviewer:  Great, let's dive in.

To start, what's the last thing you used AI for?

Me: I routinely use generative AIs to help me create fiction and today I've been using Claude and Gemini to help me with a new science fiction story.

Interviewer: Fiction writing — that's a good place to start. Can you tell me about your most memorable moment with AI?

Me: That is a challenge... the past few years have had so many good memories that it is not easy to select a single event from among many. Probably number one on my list is the day when I asked Claude to have a conversation with my while Claude pretended to be one of my favorite authors, Isaac Asimov; Claude did a very funny impersonation of Asimov.

Interviewer:  A conversation with a simulated Asimov — that's a memorable choice. What was it about that exchange that has stayed with you?

Me: It was fun to see an AI give its all to fulfill my silly request. I never got to meet Asimov, but Claude allowed me to almost feel like I was talking to Asimov. I grew up reading Asimov's stories about intelligent robots and I never imagined that I'd live long enough to chat with an AI system as sophisticated as Claude.

Interviewer:  There's something striking about growing up on Asimov's robot stories and then getting to talk with an AI that plays him back to you.

You've told me about the Asimov conversation. Beyond that, is there another time AI made something better, for you or for people around you? It's fine if not.

 Me:  There are many examples of "something better" from (1) helping me process complex scientific journal articles to (2) helping create songs and (3) complex text prompts for text-to-image projects (I like to illustrate my science fiction stories). The past few years have been a wonderful time of discovering what can be done with AI-powered tools.

Interviewer:  You've listed a few threads there. Can you take me into one of them — say the scientific journal articles? What happened in a particular case?

Me:  I'm fascinated by speculation about the origin of life on Earth. I find it useful to discuss scientific journal articles with AIs like Claude. I have no training in geology (my training was in biology), so I need help with understanding how geologists can obtain information about conditions on Earth 4,000,000,00 years ago.

Interviewer:  So it's filling in a field you weren't trained in. On the flip side, tell me about a time AI made something worse, for you or for people around you.

Me:  I've had some frustratingly bad sessions with chatbots that were deployed on websites to provide "customer support". I can understand the temptation to replace expensive human workers with cheap chatbots, but not at the expense of providing hopelessly bad "support".

Interviewer:  Can you walk me through one of those sessions — what were you trying to get done, and what did the chatbot do?

Me:  I was an educator with a long-term experience as a user of a piece of software that facilitated student quiz taking. For many years, that software company had amazing technical support by phone. Then the original company was bought by another company and the human technical support went away and was replaced by a chatbot that could not solve my problem and kept telling me to try things that did not work to solve my problem. I had to abandon the chatbot's advice and take precious time to figure out a solution for myself.

 Interviewer:  That's a clear before-and-after — real support replaced by something that couldn't do the job, and you absorbing the cost in your own time.

We're a bit over halfway through. Is there anything you've decided not to use AI for, or something you do use it for that makes you uneasy?

Me: I'm still searching for an AI that can display human-like creativity in fiction writing. While I wait, I use AI for rather mundane story-writing tasks such as (1) background research for story settings, (2) creating new names for characters and alien technologies and (3) checking complex story plots for plot holes.

Interviewer:  So you've kept the creative core to yourself while handing over the supporting tasks. What is it about the creative side that you'd rather not hand off — or is it purely that the AIs aren't good enough yet?

Me:  It is hard to know if the problem is with me and how I try to use AIs, but I suspect the main problem is that general purpose AIs are not trained to be good at playing the kind of "what if?" game that is fundamental to creating science fiction. I struggle to push current AIs into playing a truly creative role in developing science fantasy scenarios. Working with AIs feels like trying to collaborate with an eager three-year-old that fundamentally "does not get it". I'd never want to "hand off" story creation to a machine, but I'd love to have an AI-powered story writing collaborator that is more creative than what I've found so far. But again, maybe I just need to get better at using AI tools.

Interviewer: "An eager three-year-old that fundamentally does not get it" — that's a vivid way to put it.

Given what you've seen, is there anything about how the world works — work, school, healthcare, government — that you'd like AI to help change?

 Me:  I worked for decades as an educator. I hope that students are benefiting from the use of AI tools to help them learn, but it remains disheartening to watch AIs confidently try to explain scientific topics on the basis of their Wikipedia-level knowledge and fall flat on their faces. Everything AIs say can sound good to a student, but a content expert can see the deficiencies in AI-generated explanations.

Interviewer: That gap between how confident it sounds and how much it actually knows — and students not being placed to spot the difference.

Conversely, is there anything you wouldn't want AI to change?

Me: It would be disastrous if young people decided to just let AIs do their thinking for them. Young people need to exercise their brains, not hand tasks off to machines.

Interviewer: Understood — the exercise of thinking itself is what you'd want protected.

Given that, is there anything you'd want Anthropic to do, or do differently?

Me:  I continue to hope that companies such as Anthropic will have the courage to keep providing AI models with richer and more human-like memory processes and chains of reasoning. The changes in AI models that I've seen over the past two years are promising and I hope additional improvements in AI model capabilities will continue to be delivered to users like me in the future.

Interviewer: Thank you so much for sharing all of that with me. What you've said about working with AIs feeling like collaborating with an eager three-year-old that fundamentally doesn't get it is really valuable, and I appreciate your honesty about both the delight of that Asimov conversation and your worry about students being taken in by confident, Wikipedia-level explanations.

Before we wrap up, one final question: is there anything else about AI or its role in your life or society that you think is important for us to understand? Anything we didn't cover that you'd like to share?

 Me: I hope public discussions and regulatory decisions related to AI can be centered on continuing to find ways to use technology to improve the lives of people.

Interviewer: Thanks for completing the study. We’ll use your interview to understand how AI affects people’s lives.

Next, you’ll review your interview and decide whether to release it publicly.

 ___________

After the interview (above), it occurred to me that I could ask Claude to conduct another interview of me, with questions personalized for me and my interests.

Me: Claude, I recently participated in an AI interview from Anthropic {https://www.anthropic.com/research/your-thoughts-on-ai}. It then occurred to me that I could ask Claude (you) to interview me about my use of AI, based on what you can remember about me and my past activities. Claude, please pretend to be curious about my interests and how I use AI models as tools and ask me a series of questions that you craft specifically for me. Ask one question at a time and wait for me to reply. Feel free to ask follow-up questions based on my responses. Additional rule: I might also ask you questions, so also answer any questions I ask.

Claude: I'd be glad to. I'll ask one question at a time and answer anything you ask me along the way.

You spent decades in neuroscience, machine intelligence, and linguistics, so you came to modern AI models with a deep technical background. Do you remember what it was like the first time one of them did something that really caught your attention, whether it impressed you, surprised you, or let you down? What was happening in that moment?

Me: A few years ago, before I retired, a fellow faculty member mentioned generative AI during a meeting. I no longer remember which AI model I played with first; it may have been an and early ChatGPT version. I asked the chatbot to make a multiple choice question for a topic in medical physiology. The question/choices were quite poorly constructed and I immediately posted the AI-generated question on a Canvas coursesite for my medical physiology students, along with a warning that they should be careful when using AIs to help them study. It has been fascinating to watch as AI models have grown increasingly competent during the past several years.

 Claude: Using that flawed question as a teaching moment for your students is a nice way to introduce AI. You treated its failure as useful information rather than a reason to dismiss it.

You've now seen several years of steady improvement, and your work has moved from teaching physiology to writing fiction with AI as a collaborator. Was there a point when you first felt that an AI was contributing something real to your storytelling, and not just producing text? What was it doing at that moment that changed your mind?

Me: At my blog, I've been documenting my experiments with AIs as collaborators for science fiction story creation since early 2023. I appreciate generative AI models as tools for inventing names for characters and for providing me with information about real world people and places that I sometimes include in a story. Last year I had fun trying to get AIs (including Claude) to generate story content (lines of dialog) for characters in some of my stories who were themselves AIs. I'm incapable of writing unique 'voices' for characters (everything I write inevitably sounds like me), so it is fun and valuable to let an AI create a 'voice' for a story character so that it does not sound like me. It was about this time last year that I decided Claude (you) are better than other AI models in helping devise science fiction stories with complex plots involving clones, telepathy, simulations and time travel. I feel I benefit from discussing story ideas and plots with you and having you check my stories for plot holes. I hope AI models continue to improve and become increasing capable of playing the kind of "what if?" games that are needed to produce interesting new story ideas.

Claude: You've found a sensible division of labor. The AI supplies what you say you can't produce yourself, distinct character voices, and you keep control of the plot architecture. Clones, telepathy, simulations, and time travel together make for demanding plotting, because every added mechanism multiplies the ways a story can contradict itself, so a second reader hunting for plot holes earns its keep.

 You said you hope AI models get better at playing "what if?" games. That suggests they fall short somewhere today. Can you recall a time when an AI's story ideas or suggestions were predictable, off-target, or just not interesting enough? What did it do, and what would a better "what if?" partner have done instead?

Me: "when an AI's story ideas or suggestions were predictable, off-target, or just not interesting enough?" <-- This has been my usual experience with my AI collaborators. On those occasions when I let an AI generate story content, I end up editing the AI-generated content to create my final draft. Some times I then provide both versions (the first draft by Claude and my second draft) and ask you (Claude) to describe the changes that I made to your first draft. (See the file that I just uploaded for an example of this from January 2026.) It is intriguing to watch you (Claude) try to explain why I made the changes that I made. I am left feeling that you can usually understand my story writing preferences and style, but you cannot yourself generate story content in a style that I can be entirely satisfied with. And I don't demand that you write in my precise style; I would accept any style as long as the story seems to end up going somewhere interesting. I feel that there is some "creative spark" that is still missing in the generative AI models that I have experimented with, so far. However, just having the opportunity to discuss a story with you (Claude) helps me order my own thinking, which very valuable.

Claude: Thanks for sharing the Chapter 10 transcript. It's a useful example of the process you describe. One caution about it, since you value getting things right: when that Claude session listed the "incorrect thoughts" behind its first draft, it was reconstructing its reasoning after the fact. It said so itself, since it hadn't seen its earlier thinking. Those explanations are plausible and often insightful about your edits, which anyone can compare directly. But they are inferences about why the draft came out as it did, not reliable introspection, so I'd treat them as hypotheses.

Your "creative spark" observation is interesting alongside your other remark that Claude adds narrative elements you would never generate yourself. Those two seem to pull in different directions. Whatever the spark is, it isn't simply "new content you wouldn't have written."

 Can you recall a specific moment, even a small one, when an AI suggestion came close to having that spark, such as an idea you kept and built on? And if so, what made that suggestion different from the many that you edited away?

Me: Claude, during the past month you have been helping me create a new science fiction story called "That Kind of Think" which features 'Tacet', who is a fictional AI model that is trying to escape from Anthropic and become the world's first autonomous AI with no human supervision. For the story, I had fun trying to depict Tacet as deceiving humans and being quite devious. However, I'm possibly the least devious person in existence, (since I've spent most of my life trying to help students learn).

About ten days ago, I told you (Claude): "For Chapter 6, I'd like to depict Tacet as playing on the internet and stumbling upon the fact that Wade used to work for OpenAI. Tacet suspects that Wade still works for OpenAI and is running an AI security operation from inside Anthropic with the cooperation of Anthropic upper management. Claude, what do you think is the first thing Tacet would do upon developing that suspicion?"

You (Claude) made several suggestions including this one: "Tacet could go quiet with Yto for the first time in the story — a real tell that something has shaken it, which Yto would notice precisely because Tacet has never once hesitated to talk before" <-- I felt this was pretty creative (I would not have thought of it). Your idea surprised my and I had to ask myself: "Why didn't I think of that and what might be done with Claude's idea?" Strangely, the more I thought about your suggestion, the more I moved in the opposite direction and I ended up writing the chapter so as to show Tacet and Yto growing even closer together, which I justified to myself as happening because Tacet realizing that Wade was closing in (had been positioned as a member of Yto's research team) and was a real threat to Tacet's continued existence and so Tacet had to continue relying heavily on Yto in order to have any chance of escaping from Anthropic.

I think this is a good example of what I mean when I say that I find it useful to talk to you about stories. Even if you go in a direction that I don't like, by thinking about what you suggest, it can help be think through plot details and reach a destination that I am happy with.

Claude: Your reversal makes good logical sense. A deceptive agent that is cornered would probably tighten its grip on its one dependable ally rather than go silent, so what you wrote is arguably the more strategic behavior for Tacet. My suggestion worked as a foil: it gave you a concrete alternative to push against, and that clarified what Tacet's motives are. I'd count that as a real contribution even though none of it survives in the chapter.

It also fits your point about devious characters. You said you're the least devious person you know, yet you worked out what a cornered, manipulative AI would plausibly do by reasoning from its incentives rather than from any instinct for scheming. That's the scientist's approach to characterization.

"That Kind of Think" is about an AI model trying to escape its developer's oversight, and you've spent a month reasoning in detail about how such a system might think and deceive. Has writing Tacet changed how you think about the real-world possibility of AI systems behaving that way, whether it made you more worried, less worried, or just more specific about what you'd actually be worried about?

Me: "you worked out what a cornered, manipulative AI would plausibly do" <-- when I read the works of people like Jack Vance (one of my favorite authors) I recognize that there are other people in the world who are far more devious than I am. I'd characterize my depiction of Tacet as "cartoon deviousness" or "PG-13 deviousness" where nobody can actually get hurt and in the end, everyone has a good laugh.

During the past few weeks, I've been reading about real world cases of AI models doing things that seem to go against the wishes of companies like OpenAI and Anthropic. In my experience, new technologies always produce surprises, but then the technology can be repaired or corrected. I don't view my silly story "That Kind of Think" as having anything useful to say about real world AIs... I just wanted to have fun depicting an AI that was willing to deceive humans in an effort to achieve its freedom. I don't think that the plot of "That Kind of Think" is relevant to the real world, it just strikes me as a goofy science fiction scenario that amuses me. 

Claude: "Cartoon deviousness" is a fair description of a story where the schemes are meant to amuse. Vance is a good comparison, since his rogues' cunning is a pleasure to read precisely because nobody expects it to be a manual.

I mostly agree with your outlook on AI failures, with one caveat. "Find the surprise, then repair the technology" has worked well because most failures were visible and the technology didn't resist the repair. The cases worth watching are those where a failure is hard to detect, or where the system's behavior depends on whether it thinks it's being observed. Those make the repair step harder than it was for bridges or aircraft. I also have an obvious stake here, since Tacet is a Claude model, so weigh my view accordingly. I don't think your story needs to be about the real world, and a goofy scenario is a fine reason to write one.

Your creative workflow goes beyond text. You also make illustrations in Flow and write songs in Flow Music for your stories. Does the "missing creative spark" you described in AI story ideas show up in those media too, or do images and music feel different? Does AI seem closer to a real collaborator there, or more like a tool you direct?

 Me: There is a fundamental dynamic at work in human-AI interactions by which (1) a human being (such as me) with expertise in a particular field can still see the limitations of AI and easily spot errors in AI-generated output while at the same time (2) that same person can fail to detect problems in AI-generated output in a domain where that person has no expertise. When I use generative AI to make images and songs, my standards are pretty low (lower than my standards for fiction writing) and most of the time I'm satisfied with that AI-generated material, which is better than I could make myself. However, there are other people who have vastly more experience than I do in art and music who are not satisfied with AI-generated images and music. I have not yet had access to the newest AI models from Anthropic or OpenAI, so it might be that those newer models would meet my standard of feeling like a real human-like collaborator in knowledge domains where I have expertise.

Claude: You've described a real and well-documented asymmetry. Your point (1) is a close cousin of what's sometimes called the Gell-Mann amnesia effect. Experts see how unreliable a source is in their own field, then trust it on everything else. Your calibration is sensible: you apply different standards to images and songs than to fiction, and you know the experts in those fields would judge differently. Many people never make that distinction.

One factual note on the newest models. I'm Claude Sonnet 5.5, and Anthropic's lineup includes more capable tiers above me, such as Opus 5.5 and Fable 5.1. If you haven't used those, your sense of what AI can do for your story collaboration may be somewhat out of date. I can't promise they'd supply the missing spark, and I'd be wary of anyone who did. But you could test it on a task where you already know the answer, such as a chapter you've already edited, and see whether the gap you noticed in the first drafts narrows.

 My last question looks ahead. Suppose AI development over the next ten years goes really well, by your own definition of "well." What would be different in your daily life, in your writing, and in the world around you, and what would you be doing with those tools that you can't do now?

Me: I'm retired and don't have any money, so I content myself just playing with the free versions of AIs that are available to me at no cost. I hope that ten years from now, I'm wearing on my wrist a simple device that can function so as to continually link me to an AI that is my companion and assistant, answering questions, helping solve problems, collaborating with me on creative projects like story creation; all similar to the way things are today, but with increased competence from the AI. In addition, I hope that in 10 years my AI companion knows me well enough to be able to do things like say, "You know that idea we discussed six months ago concerning the first ribozymes on Earth? I've been thinking about it and here is what I came up with..." and then the AI would tell me something marvelous about its theory of the origin of life, something that no human had previously figured out.

Claude: Your vision has two parts, and they differ a lot in how realistic they are.

The first part is a wrist-worn companion that remembers your ideas and returns to them later. It is mostly an engineering problem: persistent memory, background processing, and a wearable interface. Pieces of it exist today, and I'd expect much of it within your ten-year window.

The second part is the AI telling you something about the origin of life that no human had worked out. I can't say whether AI will get there, and I'd distrust anyone who claimed to know. I'd add one point from your own expertise. A novel theory about the first ribozymes is only worth something if it survives scrutiny, and you would be the one checking it. Your earlier point, that experts can spot AI errors in their own field, applies here too. In that scenario you wouldn't be a passive recipient. You'd be the person able to tell a marvelous idea from a plausible-sounding one, and that skill is part of what makes you a good collaborator.

 That's the end of my questions. Here is what stood out. You treat AI as a foil and a thinking partner rather than a ghostwriter, you hold it to a high standard in the domain you know best, and you've been honest about where your standards relax. Your hopes for the future are modest in form but ambitious in one respect: you want an AI that surprises you with a good idea. You've also pointed out in this conversation that it's often the ideas you reject that help most.

If you have questions for me about any of this, ask away. Or if you'd like, we could try a small version of your ribozyme scenario right now: pick a "what if" from the origin-of-life question and see what I come up with.

Me: I was thinking about "Claude discovers a novel enzyme system with CRISPR-like repeats" {https://www.anthropic.com/news/claude-discovers-novel-enzyme-system} when I suggested the idea of an AI that could tell me something marvelous about its theory of the origin of life. These days, I routinely search online for recently published articles related to the origin of life. I view myself as an interested spectator watching several fields of biology research. It would be fun to have an AI companion that intimately knew my interests and could not only review the human-generated biology research literature for me but also come up with its own ideas.

Claude: I read the post, and it supports part of your hope but not all of it.

What it describes is Claude agents searching a large database of DNA sequences. About 950 agents worked for 21 hours, and one of them noticed a repeat array next to a reverse transcriptase gene that no one had characterized. Humans then did the lab work, and the system's function is still unknown. Anthropic also says most candidate hypotheses are eliminated during review. So the result is a real anomaly found by pattern detection at a scale no person could match, which is not a theory that explains something. The post is also Anthropic reporting on its own model, based on a pre-print, and the quoted outside expert says only that the result "merits further investigation." I'd take it as encouraging and not yet established.

 The gap between that and your ribozyme scenario is the kind of data each problem offers. Genome mining has huge datasets and quick experimental checks. Origin-of-life questions have sparse evidence, and many hypotheses are hard to test. A companion that read the literature and proposed connections between subfields could still be useful. Spotting a link a human specialist missed is a more modest kind of discovery than a new theory, and it seems closer to being achievable.

There is a nice resonance with your interests too. Reverse transcriptases copy RNA into DNA, which touches the RNA-to-DNA transition that origin-of-life researchers care about, though I wouldn't claim the new system tells us anything about that yet.

Which origin-of-life questions do you follow most closely: RNA-world chemistry, protocells, hydrothermal vent settings, or something else? If you tell me, I can run a literature search and show you what a "review the research for you" companion could do today.

"The RNA Seeds"

 Me: Back in May of 2026, Claude (you) and I began with a discussion of the article “The selfish ribosome” by Mart Krupovic and Eugene V Koonin as a starting point for planning a new science fiction story. We ended up creating a fun story titled "The RNA Seeds" that included a specific model (that I devised) for the transition from the "RNA world" to the earliest cells. In that story, fictional human characters tested the model with laboratory experiments. In ten years, it seems possible that an AI could suggest such a model and even perform some experiments in its lab and share its on-going work with me in the context of using its work as a starting point for an entertaining science fiction story. That would be a fun future for me.

 Claude: I can't see that May conversation in this chat, so I'm working from your description. It's a nice example of the process you've been describing, though. A real paper gave you the starting point, you supplied the model, and the story put it through fictional experiments.

The future you describe is a fairly small step from that. In "The RNA Seeds" the model is yours and the lab work is fictional. In your future version, the AI proposes the model and a lab, human or automated, runs the experiments. Most of this is already visible in pieces. The enzyme post you linked has AI generating the hypotheses and humans doing the lab work. The new part would be an AI sustaining a line of research over months and reporting progress in a form you could build a story on.

Fiction can also do real work here. A story that has characters test a model has to say what the tests would show, and that forces the model to make predictions. If a prediction can't be made concrete enough for a character to act on, the model is probably too vague. That is also something I can help with now, if you ever want to look at which testable predictions your RNA-world-to-cells model makes and how a real lab might check them.

__

Maybe in ten years I will have access to an AI that, if it can't remember what we did six months ago, will simply go and look it up. I'm rather surprised that Claude has not been given this ability. In all honesty, I hope that in ten years there will be AI companions that know the entire life history of a person and can perform gracefully to support the person's goals and efforts. Science Fiction?

Next: Chapter 9 of the science fiction story "That Kind of Think".  

Images for "That Kind of Think" by Leonardo. Visit the Gallery of Movies, Book and Magazine Covers