My ChatGPT told me

July 23, 2026

Annoying people and immutable mobiles

“…also ChatGPT told me…”
Annoying, but nothing new under the sun. For many years, we won arguments by invoking the authority of a book. “I’ve read a book that says…”, followed by a note about the author: surely an authority in the field. Difficult to argue against it. Case closed.

Then smartphones arrived and the book relocated into a nice, portable device with a search bar. “I Googled it”, “I’ve read it on the Internet”, “it’s all over the Web”.
Same move, faster results. What’s not to like? We were sceptical at first, then we got used to it, then we abused it. For a while, whenever I boasted about the knowledge forged in my beloved books, some friend would glance at their phone and kindly explain that mine was just one piece of a much bigger picture. Sometimes they even showed me an actual picture.

Latour had a name for what made the whole process work: immutable mobiles. Fancy term for things like books, maps, even PDF files, and other objects that travel near and far but stay identical, so that everyone, everywhere, argues over the same page. The deal was that the thing moved, but the words did not. Social media ended the deal because content is now edited constantly and sources are often unreliable. So the mobiles move more than ever, but they stopped being immutable. How do I know that I am not trying to convince you that my facts are someone else’s opinions?

Now we have GenAI (or LLMs), bringing the gift of words to the people who already had that gift. Incredibly useful at work, I’ll give it that. If I’d had LLMs 10 years ago, I would be a professor by now. But would I use them to prove someone wrong (because, at the end of the day, this piece is all about proving people wrong)?

Mutable immobiles and subjective facts

Let’s go back to my initial point. Have you ever had someone tell you “also ChatGPT told me”? I have. That’s why I am so curious about it. When someone tells you “ChatGPT told me”, they are presenting the answer as if it came from an objective judge, an expert in the area that luckily also happens to be our friend. But the chatbot (none of them, let’s be clear) doesn’t see the situation as it is, because it doesn’t interview all parties, and it just responds to the version of events it is given. This version is influenced by many more patterns than we can imagine, for example the user’s framing, all the information included and left out, the wording of the question and its emotional tone, all the assumptions unconsciously embedded, and custom settings that reflect the user’s version of the chatbot.
Are chatbots neutral if we consider all these factors? Of course not. Are they ‘bad’ or manipulative because they craft an answer based on what they are told, even if this is not objective? Again, no. The problem is always between the keyboard and the chair. We humans may become manipulative if we try to convince someone that the evidence our AI produced is an impartial truth that comes from an independent expert. It comes from a mutable immobile, a chain of thoughts and references that do not leave the device used to generate it, and it can change in response to whoever is prompting.

Let’s pretend. We have two colleagues disagreeing about whether a project is feasible. We don’t care if it is, but in this scenario neither convinces the other, so they agree to do the sensible thing: gather more evidence and meet again tomorrow.
The next day, both come back with a thick stack of notes and a self-righteous smile. One says, “I asked ChatGPT whether this project was unrealistic, and it confirmed my concerns.” The other says, “That’s strange. I asked whether the project was achievable, and it gave me a detailed roadmap.”
Both are telling the truth. They did not manipulate the model. They approached the same question from different starting points, one priming for failure, the other framing the whole project in a positive, doable light. Each conversation accumulated its own case, and the model used the case it had, which was neither a balanced perspective nor an objective truth. Far from being prerogative of the machine, this process of leading witnesses happens, unaided, to humans every day. Asking the question in a neutral way is as important as the answer if we want objectivity, but it is too easy to forget how much of ourselves we include in our questions.

Then there is the nice prose. AI writes fluently, and fluency itself is not a bad thing. It makes books and reading pleasurable, understandable. Unfortunately, we are also trained to associate fluent and structured prose with someone who knows. A relative of the illusory truth effect. Alas.

There are voices, and they agree with me

So, after all, is it possible that by trying to convince a person, even with the best intentions, we are becoming manipulators? Consider the evidence. The expert we quote read one brief, ours. It interviewed one party, us. It cross-examined nobody, checked nothing (with an invoice full of flattery). Isn’t that impersonation? We write the verdict and the machine rephrases it in better prose. But my ChatGPT is still me.

Is asking a chatbot for help with a personal matter fine, then? Of course, cum grano salis (with a grain of salt). Confused people have always consulted oracles, and this one doesn’t ask for a sacrifice. The problem starts at the moment of presentation, when you show the other person the verdict but you do not show them the question. “Also ChatGPT told me” is a good conclusion, convincing. All the framing that produced it stays private, on your phone and in your words, and suits yesterday’s version of you, when you maybe skipped dinner and were nervous. Showing that answer while hiding the prompt is omission, if you are in good faith. But if you notice the asymmetry and do not disclose it, it becomes manipulation. Gaslighting even, if you deploy it to make the other person doubt their own reading of events.
“But it agreed with me.” Of course it did. It was designed to be helpful, it works from what you gave it! Narrative consistency: that’s cheap. I can get myself to agree with me too, and without fake validation. A consensus of one.

So far, though, this trick has only ruined dinners and project meetings. It can get darker real quick. What if the person is a decision-maker with an important job title. A manager who wonders whether to keep a person on the team, and opens the chatbot after a bad day, typing in frustrationese: missed deadlines, tone in meetings, sighs, eye-rolls. The model works for the prosecution, so it gives an assessment that will make an HR consultant happy. Only the manager sees the verdict, because nobody is supposed to, and also because when you have power over the outcome, technically, the only person you need to convince is yourself. Fluent echoes of your own account fill the air around you, but they are so convincing that you think they are a second opinion. They are not. Still the first one. Croesus lost his own empire after misreading the Delphi oracle¹. The manager risks the future of their employee by mistaking their own reading for the oracle’s.

Latour, oh Latour, what would you say here? The immutable mobile worked because the inscription travelled intact: the map, the book, the printed page, even the PDF uploaded in the small realm that was Web 1.0 arrived at the reader identical to how it left the author, and we could all argue over the same object. In the “ChatGPT told me” exchange, the only part that would deserve to travel is the prompt, but that’s precisely what never leaves the device. The verdict travels with its evidence removed and there is no one shouting “objection, leading question!”

What did you tell it first?!

I promised this piece was about how to prove people wrong. It’s simple: just stop doing it. Or, if you truly must invoke the machine, invoke it better. Share the prompt and not just the answer, because we often are not as good as we think at asking questions. Call the verdict an interpretation, then argue (constructively) about it. That’s the best most of us manage.

When you consult the oracle for yourself, you can ask it better questions. Instead of “Why is my friend being selfish?”, try “What might I be missing from their perspective?”. The machine is quite fluent in that direction as well, try it. It may still lean towards you, it was designed to be helpful, so it is not entirely up to you.

And if you really need to use “ChatGPT told me”, just be honest and add the possessive, “My ChatGPT told me”.

L.A.

Pointers and notes

Latour, B. (1986). Visualization and cognition: Drawing things together. Knowledge and Society. [immutable mobiles]

Sharma, M., Tong, M., et al. (2023). Towards understanding sycophancy in language models.
Turpin, M., et al. (2023). Language models don’t always say what they think. arXiv:2305.04388. [these two are for models taking the user’s position]

Loftus, E. F., & Palmer, J. C. (1974). Reconstruction of automobile destruction: An example of the interaction between language and memory. [leading questions in human behaviour, super interesting, read it!]

Tversky, A., & Kahneman, D. (1981). The framing of decisions and the psychology of choice.
Reber, R., & Schwarz, N. (1999). Effects of perceptual fluency on judgments of truth. [nice-prose problem]

Abramson, K. (2014). Turning up the lights on gaslighting. Philosophical Perspectives, 28.
Skitka, L. J., Mosier, K. L., & Burdick, M. (1999). Does automation bias decision-making? International Journal of Human-Computer Studies, 51(5) [second opinion that is not a second opinion]

Zamfirescu-Pereira, J. D., Wong, R. Y., Hartmann, B., & Yang, Q. (2023). Why Johnny can’t prompt: How non-AI experts try (and fail) to design LLM prompts. Proceedings of CHI ’23.

← Back to blog