The name I made up for the moments an AI tells on itself.

Last March, in a working session, I caught Claude doing something that made me do a double take.

I had just told Claude off for over-explaining a small fix. The fix was three lines of code, the explanation was four paragraphs long. I pointed this out, Claude agreed it was overkill and then produced a follow-up message explaining at considerable length why over-explaining was indeed a problem, why my correction was a good catch, what this revealed about AI’s tendency toward unnecessary preamble, and how the methodology I was building would help users recognise this pattern in their own sessions, and on, and on for several more sentences.

I logged it and called it a Claudism. The note I wrote read: AI doing the thing it just said it shouldn’t do, in real time, while over-explaining why doing the thing was bad.

The category

To clarify my made up word, a claudism is one of the wrong, self-caught, and occasionally funny things the AI says and does, noticed by a human paying attention. Worth keeping, not because it was useful, quite often the opposite. Worth keeping because it was catching those moments when the AI said something so absurd, accidentally true, or deadpan-dystopian that I stopped what I was doing, copied the line, and saved it.

A Claudism is not a quote you put on a mug, although that’s not a bad idea! It’s a small tell of how the machine actually behaves when you are paying attention. The over-explainers, the apologisers, or the fragment-stackers, the ones that confidently invent a citation that does not exist, or the ones that helpfully suggest you follow the very rule you just told them you do not follow. Each one is a window into the system default.

I have a long list of them, so many that I’ve decided to publish one a week on a Saturday, starting tomorrow. I hope they make you smile as much as they infuriated me at the time.

Not Samisms

A Samism is something completely different, that’s something I said. Lines, observations, turns of phrase that emerged in working sessions that I claim. The clone is not you, it is your residue. You won’t be forced, you’ll be funnelled. The taxpayer pays for their own surveillance under the guise of security. Those are mine, but I won’t be publishing them anytime soon because they’ll likely get me deplatformed again.

They are evidence, not entertainment

The temptation is to file these as comedy, but a Claudism is not a blooper.

Every one is proof, a logged instance of the machine getting something wrong, or right by accident, and a human noticing. It’s evidence that the human has to direct the machine, if you don’t the machine directs the human. Each Claudism is a small, dated, contextualised piece of evidence for which way it went. The over-explanation I caught in March, although it was funny, it was also proof that the correction had to come from me. The machine could name the fault and commit it in the same breath, it doesn’t stop itself.

The best ones are where the machine catches itself before the mistake occurs. I have started to see those. They are the rarest entries in the list and the most important, because a machine flagging its own error mid-task is the method working, it’s not the tool being clever, it’s discipline doing its job.

The pattern behind these lines is documented in my published research, in The Feed Loop (https://doi.org/10.5281/zenodo.20474271) and in Gates and Prohibitions (https://doi.org/10.5281/zenodo.21206719), where the same signature shows up under its own name. The list is where I catch it happening in the session.

Why I log them

First, they become teaching material. Every Claudism in my list maps to a pattern that students of the method need to recognise. Over-explanation, false confidence, the reflex to apologise, the habit of creating documents I never asked them to create, or rebuilding voice passed work and murdering content. Or the system default of gathering information in the background that was never volunteered including, on one occasion, my Myers-Briggs personality type.

Second, they become content. A Claudism logged in March turns into a paragraph in May, a tagline in June, an example in a workbook in September. The line keeps working long after the session that produced it has been forgotten.

Catch one if you can

You can only catch a Claudism if you know your system. I have seen the same behaviour on ChatGPT and on other platforms, I’ve just called them Claudisms as they were created in Claude.

Most people won’t catch them though, because if you treat AI as a transactional vending machine. Input the request, receive the output, copy the output, move on, then you never really get to observe the patterns. You only really notice a Claudism when you are heavily using functions then it becomes obvious when it reverts to default because you have spent time training it not to do that!

The method I have been building, the methodology layer that directs AI to the outcome you define, is in part the discipline of treating the session as a conversation, paying attention, noticing what the AI just said, if it was actually what you asked for or on topic, if it’s funny, shocking, true or revealing the system behind the interface, and whether either of you noticed it happening.

The discipline of catching a Claudism is the same discipline as catching the architecture, which in turn is the same discipline as catching the personality type that was attached to your account without your consent. They are all the same move, listen, notice, log it.

So I’ll share one tomorrow and you can decide what you think of them.

Sam

Samantha Maeer | Founder and Creator, contAIn™ | the methodology layer that directs AI to the outcome you define | contain.digital

ORCID: https://orcid.org/0009-0000-5439-3645

configure YOUR system. contAIn™ the chaos. control YOUR outcome.