Verbal IQ of ChatGPT: What the 155 Estimate Means

What does verbal IQ measure in a human test?

Verbal scores are not simply a vocabulary counter. In a Wechsler battery, language-based tasks sample crystallized knowledge, verbal concept formation, comprehension of social situations, and the ability to explain relationships. They are administered with standardized instructions, time limits, scoring rules, and a controlled interaction with an examiner.

ChatGPT has an unusual advantage on written verbal items: it was trained on enormous amounts of text and can generate a polished response immediately. That advantage is relevant to language performance, but it is not the same as a human’s accumulated knowledge, episodic experience, or ability to understand a situation. A test item can also be leaked into training data or reconstructed from familiar patterns.

Related

Cognitive Abilities - Memory, Speed, Verbal & Performance IQ

A modern IQ score isn't one number but four cognitive abilities: verbal comprehension, fluid reasoning, working memory, and processing speed.

What did later research find about ChatGPT and verbal intelligence?

The 155 estimate is not the only evidence. A peer-reviewed PLOS ONE study administered a computerized adaptive vocabulary test three times to ChatGPT-3.5 and a Bing system based on GPT-4. The authors emphasized that IQ is a relative indicator tied to a norming sample. Their design is useful because it repeats a defined task, but vocabulary ability is still only one slice of verbal intelligence.

Other research has tested large language models on verbal and visual IQ-style tasks. These studies can compare systems and show where a model succeeds or fails, but they do not turn a benchmark into a clinical diagnosis. Different prompts, decoding settings, model updates, and access to tools can change the answer. A result for GPT-3.5 should never be silently presented as a result for GPT-4, GPT-4o, or a later model.

Why an AI “IQ” is not the same as a person’s IQ

Human IQ is designed to compare a person’s performance with a norm group under a standard administration. The test taker has a fixed history, cannot query a database, and is evaluated under rules that are meant to be repeatable. A language model is a trained statistical system whose outputs depend on its training corpus, system prompt, sampling, and interface.

That difference creates several validity problems:

1. Training contamination. The model may have encountered an item, an answer key, or a close paraphrase before.
2. No stable identity across versions. “ChatGPT” names a product interface that can route to different models over time.
3. Written-only administration. Removing spoken interaction, timing, and nonverbal subtests changes the construct being sampled.
4. No human life history. An AI answer cannot be interpreted as a stable trait in the same way as a person’s score.
5. Prompt sensitivity. Small wording changes can produce different reasoning paths and scores.

These limitations do not make the experiment useless. They make its conclusion narrower. It is a benchmark of language-model performance on selected human-style items, not evidence that the model possesses a human mind or a permanent place on the IQ distribution.

Can you compare your score with ChatGPT’s 155?

Not directly. A human’s verbal IQ comes from a standardized assessment with age-based norms and an examiner who follows the manual. ChatGPT’s 155 came from a researcher adapting written items and estimating a result. The numbers share a scale label but not an identical measurement process.

If you want to know your own verbal reasoning level, use a properly normed assessment rather than asking a chatbot to score you. If you use ChatGPT for practice, treat its explanations as feedback and its “IQ estimate” as entertainment. As of 2026, the strongest claim the evidence supports is that some language models can perform exceptionally well on selected verbal tasks; it does not establish a human-equivalent IQ.

Frequently asked questions

Q: What is ChatGPT’s verbal IQ?

A: An early experiment estimated 155 on five written WAIS-III verbal subtests, but ChatGPT has no official or stable verbal IQ score. The estimate depends on the model version and administration method.

Q: Did ChatGPT really score higher than 99.9% of people?

A: The 155 estimate was above 99.9% of the WAIS-III standardization sample used for comparison. That result applies to selected written subtests in one experiment; it is not a claim about every human ability or every ChatGPT release.

Q: Can ChatGPT take a real IQ test?

A: It can answer some text-based test items, but it cannot take a human clinical IQ test in the ordinary sense. Full assessments rely on standardized administration, timing, norms, and tasks that include more than written language.

Q: Is ChatGPT smarter than a person with an IQ of 155?

A: That comparison is not valid. The AI estimate and a person’s clinical score come from different conditions and measure different things. High language performance does not establish superiority in judgment, experience, visual reasoning, or general intelligence.

Q: Does a high verbal score mean ChatGPT understands language like a human?

A: No. It shows that the model can produce highly rated language responses under a task’s scoring rules. Understanding, consciousness, lived experience, and reliable real-world judgment are separate questions.

References

• Wai, J. — I Gave ChatGPT an IQ Test. Here’s What I Discovered (Scientific American, 2023)
• Klein, B., & Kovacs, K. — The performance of ChatGPT and Bing on a computerized adaptive test of verbal intelligence (PLOS ONE, 2024)
• The feasibility of large language models in verbal comprehension assessment (JMIR, 2025)
• American Psychological Association — Intelligence

Last updated: July 19, 2026