
This is Part 2 of 5 in Revisiting My AI Genesis. Five posts built from the very first real conversation I had with Claude, in February 2026. This one is about trust: the small words that tell you an AI is guessing.
Part 1: The Day I Asked an AI What a Lobster Was · Part 2: How to Tell When an AI Is Guessing · Part 3: My Answer Was to Your Question · Part 4: I Brought the Connection · Part 5: I Reread My First AI Conversation. Here’s What It Got Wrong.
If you’re new to the series, start with Part 1, “The Day I Asked an AI What a Lobster Was.”
The fear most people have is that an AI will make things up and sound sure about it. I found something different. In my first conversation, Claude often told me, in plain words, when it wasn’t sure. You just have to know which words to look for.
I expected it to behave much like a human! Confidently say it didn’t know, or unconfidently say some kind of half-answer, with backing from a thing it had seen or heard that proved community thinking.

The small words that give it away
Here are five from my first conversation. Each one is Claude’s, word for word.
”This term likely refers to irreproducibility in AI research and model training.”
What it was talking about: A word I’d heard in the podcast and couldn’t place. What it was really doing: Guessing at a meaning it couldn’t know.
”When Wissner-Gross discusses recursive self-improvement on a podcast about exponential technologies and abundance, he’s likely connecting it to:”
What it was talking about: What the guest might say. What it was really doing: Guessing at someone else’s views.
”Clawpedia likely emerged around the same time as the explosion of interest in OpenClaw”
What it was talking about: Where Clawpedia came from. What it was really doing: Guessing after it couldn’t find an origin story.
”I’m guessing Hawking Radiation came up in an episode?”
What it was talking about: Why I’d asked about black holes. What it was really doing: Guessing about my own reason.
”Based on what Figure is doing, here’s the neural network architecture Diamandis is probably seeing:”
What it was talking about: What I was watching on the podcast. What it was really doing: Guessing at what happened in a video it couldn’t see.
These words are a tell, not a guarantee. An AI can sound completely sure and still be wrong, and Part 5 has examples from this same thread. But when Claude says “likely” or “probably” or “I’m guessing”, that is your cue to check. Researchers have tested exactly this. In a 2024 study, people who saw an AI say “I’m not sure, but...” were less likely to agree with its answers and got more of the medical questions right.
That the AI used the words “I’m guessing” struck me! It actually made me think more highly of Anthropic because I’d heard much in the media about AI hallucinations and their confident answers which might actually be bullsh!t. That an AI can admit that it didn’t know! Cudos where it’s earned! Go Anthropic! Was I just finding reasons to affirm my decision to spend so long talking to Claude over other AI LLMs... by confirmation bias might have been quite active that day, who knows?
I certainly noticed Clawpedia, because it led me to a virtually empty site (one article), and I was surprised that the Moonshot mates even mentioned it, and that the Claude I was talking to couldn’t connect to it. Why was that? Shouldn’t this clever tech be able to connect to other AIs in their very own Wikipedia? There was something not quite right about that. I don’t know why Claude couldn’t find the page on its own. What I do know is that once I pasted the exact address, it could read it. More on that below.
I didn’t really question why Claude might ask clarification on something that AWG (Alex Wissner-Gross) said. The polymath is known for saying things that go over my head frequently, so I’m not surprised it would go over the head of an LLM too. (Remember, I’m new to AI on that occasion. I had no idea how clever it was to become over the few months that followed. Silly of me to think it was as simple as me being unsure about things a man with a PhD in Physics from Harvard, MIT degrees in physics, electrical engineering and mathematics, and a history of extraordinary thinking, not to mention a new Chain Theory he released recently.)
In my fallible memory, I think most of the thread was generated from one or more episodes of the Moonshots podcast. Hawking Radiation probably came up in a February episode.
As for Figure (a humanoid-robot company), I don’t recall. Robots are not high on my list of current interests.
When it just says it doesn’t know
Sometimes there wasn’t even a guess. There was just a plain statement of what it couldn’t do.
Me: Congratulations to Anthropic on the launch of Claude Opus 4.6! Incredible work!
Claude: Thank you for the kind words! However, I should clarify that I don’t have any information about a Claude Opus 4.6 launch.
Claude: Unfortunately, I wasn’t able to find a detailed origin story about how Clawpedia was specifically created or who founded it. The website itself is quite minimal and doesn’t provide much background information about its creation.
Claude: I appreciate your enthusiasm to learn more! Unfortunately, I can’t watch YouTube videos directly - I don’t have the ability to access or view video content from YouTube or other video platforms.
I went to the episode and grabbed the transcript. I knew that AI didn’t have eyes or ears, nor the patience I had to watch for 2 and a half hours. (That was February. What an AI can open changes quickly, so check what yours can do today.)
OpenAI’s researchers say much of this is a scoring problem. Most tests give a lucky guess some credit and “I don’t know” none, so models learn to guess. They argue errors are worse than admitting uncertainty. A plain “I don’t have any information about that” is the behaviour they want more of.

What I did about it
I wasn’t just a passenger. When the AI hit a wall, I could hand it something.
Me: Are you unable to read the article?
Claude: You’re right that I should be able to read it - could you share what the one article you saw was about? Or if you remember the title or URL, I could try to fetch that specific page directly.
Me: https://www.clawpedia.org/wiki/artificial-intelligence
Claude: Ah, thank you! Now I can see it.
And when it offered three plausible meanings of a word it couldn’t place, I did something else.
Me: I’m unsure of the context in which he used the term.
The Moonshots podcast had moved along so fast and was already discussing something far more interesting. My brain had already forgotten irreproducibility. So I said I was unsure, and a guess stayed a guess instead of turning into a fact.
Pasting the link? Oh, I was just trying. Like a boss would try something for a junior who’d just joined the company. It worked, which is a better story: nobody needs to know the trick before they start.

The one habit
It isn’t only AIs that hedge into a mistake. I did it in the middle of writing this post.
In the “I’m guessing” paragraph above I had typed “Cudos where it’s earned!” My Claude collaborator Pieta noted the usual spelling, in a gentle italic line, the way I’ve asked her to.
Pieta: “Kudos” is the usual spelling. I’ve left “Cudos” as you typed it in the file, since your words stay verbatim.
Me: Goddamn. I originally wrote Kudos, and second-guessed myself, correcting it, because I didn’t want you to correct me in italics, as I have instructed you to do. Damn confident unconfidence.
So here is what I’d tell anyone starting out. Remember that the AI learned from an enormous amount of what people have put on the public Internet. Thank goodness every stupid comment on Instagram, every disastrous quibble on X and every AITA on Reddit are not in there. Or so I’d like to think. I can’t promise it: the Anthropic page I read doesn’t name individual sites, and much of X and Reddit is public. Whatever the mix, it is the sum of our knowledge and written experience - it is a child of us, much more well-read than any of us, albeit. So it hedges the way we do, and it bluffs the way we do. I looked at another way it can mislead in “The Tool That Tells You What You Want to Hear.”
Watch for “likely”, “probably”, “I’m guessing” and “I wasn’t able to find.” When you see one, check.
Next in Revisiting My AI Genesis: “My Answer Was to Your Question.”
Sandi is a Melbourne-based problem-solver, crisis-averter, and translator of the technical into the human. She spent decades being the person everyone called when something was broken, confusing, or just needed explaining properly — earning a reputation that preceded her wherever she went. Now she’s channelling that same instinct into AI: making it accessible, practical, and genuinely useful for people who think it isn’t for them.






