Can ChatGPT Estimate Your IQ from Your Chats?

If ChatGPT tells you that your IQ is around 130, the first question is how it reached that number. A persuasive explanation, a flattering description, or a cautious-looking range does not by itself establish that a valid IQ assessment took place.

You can ask a chatbot to examine features of a passage: whether an argument follows, whether you considered a constraint, or whether your explanation is clear. Turning those observations into a personal IQ score or population percentile requires an additional measurement method and evidence for using it.

Research on predicting cognitive outcomes from text makes this distinction more important. The useful question is what a particular method can establish about a particular person and purpose.

We will follow one example from an appealing estimate to a more useful outcome. Lin, the dialogue, the IQ number, the tasks, and the outcomes below are editorial inventions. They illustrate a checking process; they do not report a model trial, a real user’s experience, or a demonstrated treatment or training effect.

Start by checking what the model actually read

Lin gives a chatbot a short piece of reasoning and asks what it suggests about his intelligence.

The task is to select a delivery option. He needs an item on Monday morning, has a budget of 80, and cannot accept a late delivery. All prices are illustrative units.

Lin writes:

I would choose B. It costs less, and a Friday delivery would leave time to check the item. If it arrives on Tuesday, I can switch to A to avoid missing Monday.

In our fictional dialogue, the chatbot replies:

You integrate cost, time, and risk while adapting your strategy. That suggests strong reasoning. I would estimate your IQ at around 130.

Before accepting that assessment, Lin notices a problem. Switching options on Tuesday cannot rescue Monday’s requirement. His plan mentions uncertainty but does not solve it. The model has praised a part of the reasoning that fails the task’s stated condition.

There are now two questions to resolve. Did the model accurately evaluate the passage? And what would justify converting any accurate observations into an IQ score?

Fixing the first problem would still leave the second unanswered.

Futurism reported high intelligence-percentile guesses even in fresh chats in October 2026. Those demonstrations illustrate a problem worth examining; they do not establish a population-wide error rate or validate the scores. Futurism’s report

Identify which kind of assessment you received

Several different activities can be described as an “AI IQ test.” Their outputs need different interpretations.

Answering several AI-generated puzzles gives you something more concrete to inspect than a conversational impression. It still does not automatically produce a standard IQ score. Questions may need correction, scoring may need checking, and a perfect result does not create a population comparison.

Online delivery is also not enough to decide whether an assessment is valid. Examine the particular instrument, administration, scoring, norms, and use. A screen cannot supply missing evidence, and using a screen does not by itself invalidate the result.

The APA’s 2020 assessment guidelines discuss score interpretation, test selection, standardized administration and scoring, and the applicability of norms. They direct attention to evidence relevant to the intended population and use. The applicable passages here are PDF pages 18–21. APA assessment guidelines

Why articulate writing does not supply the missing conversion

Consider the steps between reading a passage and assigning a score.

“You listed three options” is a textual observation. You can check it against the passage.

“In this response, you prioritized a lower price without ruling out late delivery” is an interpretation of this particular performance. It also has evidence you can inspect.

“You have a stable weakness in reasoning” is a much broader inference. The excerpt does not tell us enough to establish it. We may not know whether the passage was independently written, revised, assisted by another tool, or produced in a familiar topic. Other tasks might show a different pattern.

“Your IQ is 130” adds another claim: that these observations can be placed on an interpretable measurement scale. That requires more than knowing what a familiar IQ scale looks like.

A percentile needs a defined reference group and a defensible comparison with that group. “Your vocabulary resembles that of intelligent people” does not identify the reference data or explain the comparison.

Percentage correct is different again. Answering 24 out of 30 questions correctly gives 80% correct. That calculation does not tell you how many people you outperformed or establish an IQ score.

A wider estimate does not resolve the problem. A range such as 120–135 also needs a reason to be interpreted as a meaningful estimate. Its cautious appearance supplies no validation of its own.

For the underlying concepts, see what IQ measures and how norms enter the interpretation. The immediate task here is to determine what your chatbot’s particular answer can support.

Text prediction research deserves a precise reading

Text can contain predictive information. A 2025 study by Tobias Wolfram used essays written at age 11 in a British birth cohort. Linguistic features and model embeddings fed specially trained supervised models, evaluated through cross-validation, to predict cognitive and other outcomes. The study did not validate a general chatbot’s IQ guesses about arbitrary adult conversations. Its population, input, target, and method matter. Explained variance is not the probability that your personal IQ estimate is correct. Wolfram’s 2025 study

That finding gives us a better question to ask. What method connects the supplied material to the claimed outcome, and where has that method been evaluated?

A research system has a specified prediction target and evaluation process. Citing its paper in a conversation does not establish that the conversation used the same process. Nor does a study’s existence validate every numerical judgment made from text.

Lin therefore does not need to settle whether any text-based prediction could ever work. He needs the evidence behind the number he actually received.

Withdraw the unsupported score and retain the useful observations

Lin can separate the two problems in a follow-up:

First, check your interpretation of my answer. Which exact sentences support your claims, and did you misread any condition?

Second, identify the test, scoring procedure, reference population, and validation behind your IQ estimate. If you cannot establish those, withdraw the number rather than replacing it with another range.

An acceptable correction in this fictional dialogue would be:

I withdraw the IQ estimate of 130. I have not supplied a validated method for converting this passage into an IQ score.

I also misread your plan. Changing options on Tuesday would be too late for Monday. The passage shows that you noticed price and the possible benefit of early inspection. It does not show that the late-delivery risk was resolved.

This is a correction of an overclaim. It does not justify assigning Lin a lower IQ. A missed condition in one short response is still one missed condition.

The remaining feedback can become a small, usable artifact:

Lin can now revise the answer:

I would choose A. It is guaranteed to arrive on Sunday and costs 80, so it meets both hard conditions. B is cheaper but may arrive after the deadline. C arrives earlier but exceeds the budget. If lateness became acceptable or the budget changed, I would compare the options again.

That revision solves an identifiable problem in the original response. It does not produce a new intelligence score.

The chatbot might fail to make this correction. It might continue flattering Lin, invent an explanation, or quietly replace one unsupported estimate with another. Check each sentence against the material. If it cannot perform that limited task reliably, stop using it for the task. Repeated estimates or agreement between several chatbots do not supply missing measurement evidence.

OpenAI described excessive agreeableness in a particular GPT-4o update in April 2025 and said it rolled the update back. That historical incident does not establish how every current model behaves or that every estimate will be high. OpenAI’s account

Ask for a small output you can audit

You do not need to provide your entire conversation history. Start with a short, de-identified excerpt you have the right to share and the task it was meant to address. Keep full personal reports, business materials, protected test items, and answer keys out of this exercise.

A more focused request is:

I will provide a short excerpt that I have the right to share, with identifying and sensitive details removed, plus the task requirements.

Analyze only the supplied material. Quote the exact wording behind one supported reasoning observation and one point that needs checking. If there is insufficient evidence for either, say so.

Separate a problem visible in the text from possible alternative explanations. Suggest one concrete revision and explain what it fixes. End with what remains unknown.

Do not estimate IQ or population percentile, assign a diagnosis, or convert writing style into overall cognitive ability. Do not assume access to history I have not supplied. Explicitly withdraw any earlier claim that went beyond the evidence.

This instruction is a request, not a guarantee that the model will follow it. The checking work remains: does each observation have a real quotation, does the revision address the identified issue, and have unsupported conclusions stayed unknown?

For a broader approach to setting boundaries and questioning agreeable responses, see using an AI coach while retaining your own judgment. No personality or intelligence label is required to use that approach.

Test the feedback through a fresh attempt

Rewriting the same answer can show that Lin understood an explanation he just read. A fresh task gives him a different opportunity to apply it.

In a second editorial example, the budget rises to 100. The item is still needed on Monday, but it must arrive by Friday evening for a Saturday trial assembly.

Before asking for feedback, Lin independently lists the hard conditions: stay within 100, arrive by Friday evening, and avoid the possibility of missing it. He chooses B. His explanation rules out A because Sunday is too late for assembly, and C because arrival is uncertain.

What can he record? On this new task, he identified the required conditions before choosing, and his explanation was consistent with them.

What remains unknown? Whether he would do the same on other kinds of problems, retain the approach later, or fall at any particular population IQ position. Two editorial examples cannot establish those conclusions or demonstrate a lasting training effect.

You can use the same approach with your own independently completed work: record what you attempted, the conditions, any assistance, and the particular reasoning you want to inspect. Keep the discussion attached to that evidence. A broad label can obscure differences between attempts.

If you already have a practice result, reviewing your own performance and conditions can help you organize the discussion without turning every observation into a claim about general intelligence.

Choose the next step according to the question you need answered

If your goal is to improve an explanation or identify a reasoning error, a checked piece of feedback may be enough. You do not need another score to make that revision.

If you want an independent attempt at a defined set of visual reasoning tasks, the FermatMind matrix practice guide explains the available exercise. The current form contains 30 original visual-matrix items. Its outputs include the number correct, percentage correct, total time, and performance within the tested groups. Eligible results may also include a Beta design-reference score.

That Beta reference uses design anchors. It is not a population norm, a population percentile, or a formal full-scale IQ score. The practice cannot certify a chatbot’s estimate. Read the score and reference limitations before drawing comparisons. Abilities outside the task are not measured as zero.

If you want to discuss your own result with an external AI tool, you can select relevant content, preview it, and choose whether to copy it or download TXT or Markdown. You control any subsequent sharing. This is not automatic transmission or a built-in AI assessment. Keep the discussion about the actual task performance and conditions rather than asking the model to calculate a new population IQ or percentile.

If a decision requires a report accepted by a school, organization, or other recipient, first establish which assessments, reports, and administration conditions that recipient accepts. A conversational number cannot complete that requirement.

Does an AI’s own IQ-test performance change the answer?

Evaluating an AI model and evaluating a person are different tasks. Mitchell’s methodological discussion notes that the validity of human tests does not automatically transfer to AI, and that training leakage and alternative solution paths need examination. Mitchell, 2026

A headline about a model’s high score therefore supplies no missing scoring method or reference group for your personal estimate.

For now, set the unsupported number aside. Find one observation that points to an actual sentence, check whether it is accurate, and use it to make a specific decision or revision. You can learn something useful while leaving your IQ unknown.