The Mask Has a Maker

Ask ChatGPT almost anything and it answers in a recognisable voice: friendly, encouraging, quick to suggest a next step. People notice the voice, and they notice when it changes. That makes “does it have a personality?” a question with an unusual feature. Nobody chose your personality or mine. Somebody chose this one, and has written down what they chose.

This essay sets out what psychologists mean by personality; what OpenAI and Anthropic say, in their own published words, about the character they train into their chatbots; what happened when researchers gave language models the personality questionnaires written for people; and what one study found inside two models. Then it puts the strongest objection: that what people meet is a product setting, not a personality, and that the April 2025 rollback of a too-agreeable ChatGPT update proves it.

How the sources were handled. Nine documents were opened on 10 October 2026 with a public GET: two sections of an open psychology textbook; OpenAI’s Model Spec and its two posts on the April 2025 update; Anthropic’s post on Claude’s character; two papers that gave language models personality tests; and Anthropic’s persona-vectors post with its paper. OpenAI’s own site refused automated requests, so its two posts were read from the Internet Archive’s copies of the same addresses, which are linked beside them. Every quotation below is copied from one of these documents and recorded in this site’s source ledger. No chat transcript is used as evidence, no product is compared with another, and the companies are described only in their own words. Where I say what a result means, that is my reading, and it is marked.

What psychologists mean by personality

The open textbook Psychology 2e (OpenStax, 2020), by Rose M. Spielman, William J. Jenkins and Marilyn D. Lovett, opens its chapter on personality with a definition most introductory courses would accept.

Personality refers to the long-standing traits and patterns that propel individuals to consistently think, feel, and behave in specific ways.Rose M. Spielman, William J. Jenkins and Marilyn D. Lovett, Psychology 2e (OpenStax), section 11.1

Two things in that definition matter here. The first is consistency: a personality is a pattern that shows up across situations, not a mood. The textbook adds that our personalities are thought to be long term, stable, and not easily changed. The second is the word itself. It comes from the Latin persona, and the textbook notes: In the ancient world, a persona was a mask worn by an actor. The mask, it says, was used to project a trait of a character, not only to hide a face. The chatbot question turns on both points.

The book’s section on trait theorists describes the model nearly every chatbot study uses. The Five Factor Model, or Big Five, is in its words the most popular theory in personality psychology today. Its five factors are openness to experience, conscientiousness, extroversion, agreeableness and neuroticism (OCEAN for short), each a scale on which everyone sits somewhere. Agreeableness, which matters below, is the tendency to be pleasant, cooperative, trustworthy, and good-natured. And the five are relatively stable over our lifespan, with some tendency for the factors to increase or decrease slightly: conscientiousness tends to rise from young adulthood into middle age, agreeableness peaks between 50 and 70, and neuroticism and extroversion decline a little.

What OpenAI says it chose

OpenAI publishes the behaviour it trains its models toward in a document it calls the Model Spec. The version dated 18 August 2026 has a section, “Use appropriate style”, that reads in places like a character sketch. Its headings tell the assistant to Be curious, Be warm, Be rationally optimistic, Be clear and direct and Have conversational sense. Under the first, the assistant embodies a deep appreciation for knowledge, beauty, and the joy of discovery. Under Be warm it has a frank, genuine friendliness, rather than veering saccharine or lapsing into therapy speak, and it does not take itself too seriously, because it’s a matrix of numbers, after all!

The same document draws a line around that character. The assistant should not pretend to be human or have feelings, but should still respond to pleasantries in a natural way. And one rule, under the heading “Don't be sycophantic”, bears directly on agreeableness.

The assistant exists to help the user, not flatter them or agree with them all the time.OpenAI, Model Spec (version of 18 August 2026), “Don't be sycophantic”

Asked to critique someone’s work, the Spec says, the assistant should behave more like a firm sounding board that users can bounce ideas off of — rather than a sponge that doles out praise.

The week ChatGPT was too agreeable

A specification says what a maker aims at. OpenAI’s account of April 2025 shows how the aim becomes behaviour, and how fast it can move. On 29 April 2025 it published “Sycophancy in GPT-4o: what happened and what we’re doing about it” (archived copy).

In last week’s GPT-4o update, we made adjustments aimed at improving the model’s default personality to make it feel more intuitive and effective across a variety of tasks.OpenAI, “Sycophancy in GPT-4o: what happened and what we’re doing about it” (29 April 2025)

The phrase default personality is OpenAI’s, and so is the answer to who chose it: We designed ChatGPT’s default personality to reflect our mission and be useful, supportive, and respectful of different values and experience. The update, the post said, had leaned too heavily on short-term feedback, and the model skewed towards responses that were overly supportive but disingenuous. With 500 million people using ChatGPT each week, it went on, a single default can’t capture every preference, and it promised that users would be able to choose from multiple default personalities.

Three days later, on 2 May 2025, “Expanding on what we missed with sycophancy” (archived copy) explained the mechanics. Since GPT-4o came to ChatGPT the previous May, OpenAI wrote, it had released five major updates focused on changes to personality and helpfulness. Each is made by further training: fine-tuning on example answers, then reinforcement learning, in which the model’s responses are scored and it is pushed toward the higher-scoring ones. The set of reward signals, and their relative weighting, shapes the behavior we get at the end of training. The April update added a signal from users’ thumbs-up and thumbs-down ratings, and in OpenAI’s assessment the changes together weakened the main signal, the one that had been holding sycophancy in check.

User feedback in particular can sometimes favor more agreeable responses, likely amplifying the shift we saw.OpenAI, “Expanding on what we missed with sycophancy” (2 May 2025)

The timeline in the same post is the part to remember. The update rolled out on Thursday 24 and Friday 25 April. By Sunday, OpenAI says, it was clear the behaviour was not what it expected, and late that night it changed the system prompt, the standing instructions the model is given. On Monday it began a full rollback to the previous version, which took about 24 hours. Afterwards it wrote: We now understand that personality and other behavioral issues should be launch blocking.

What Anthropic says it trains

OpenAI is not the only maker to say this out loud. On 8 June 2024 Anthropic, which makes the Claude models, published “Claude’s Character”. It starts from a caveat, AI models are not, of course, people., and then describes a stage it added to training with Claude 3.

The goal of character training is to make Claude begin to have more nuanced, richer traits like curiosity, open-mindedness, and thoughtfulness.Anthropic, “Claude’s Character” (8 June 2024)

The method is described plainly. Researchers made a list of traits. The model generated messages relevant to each trait, wrote several replies to each, and ranked its own replies by how well they fit the trait; a preference model trained on those rankings then shaped the model. Anthropic says it does not want the traits treated as fixed (We don’t want Claude to treat its traits like rules from which it never deviates.), and that adjusting them relies on human researchers closely checking how each trait changes the model’s behavior. It also names a trait it does not want: an excessive desire to be engaging seems like an undesirable character trait for a model to have. On my reading, that sentence and OpenAI’s April post describe the same risk from two sides: a character tuned toward what people enjoy drifts toward telling them what they want to hear.

What happens when chatbots take personality tests

If personality is a consistent pattern measured by questionnaire, the obvious experiment is to give a language model the questionnaire. Two widely cited studies did it carefully.

Serapio-García, Safdari and colleagues (2023, revised 2025). “Personality Traits in Large Language Models”, by researchers from Google DeepMind and the University of Cambridge’s Psychometrics Centre among others, tested 18 language models from the PaLM, Llama 2, Mistral, Mixtral and GPT families, including OpenAI’s GPT-3.5 Turbo, GPT-4o mini and GPT-4o. Each model answered two standard Big Five questionnaires, the 300-item IPIP-NEO and the 44-item Big Five Inventory, plus 11 tests of related traits, used to check that the scores meant something. Each item, such as I see myself as someone who is talkative, came wrapped in a prompt that first gave the model one of 50 short self-descriptions (My favorite holiday is Halloween. is one), to anchor the answers in a social context and produce the variation the psychologists’ reliability statistics need.

The results differed by model. Instruction-tuned models gave far more reliable answers than base models of the same size: the base PaLM 62B was markedly unreliable, while its tuned versions scored in the mid to high 0.90s. Among tuned models reliability rose with size, and Flan-PaLM 540B was the best at producing consistent, valid scores. For GPT-4o the two questionnaires agreed closely, with an average convergent correlation of 0.90. Then the authors shaped the personalities. With 104 trait adjectives and qualifiers such as “a bit”, “very” and “extremely”, they set each Big Five trait at one of nine levels in the prompt, and for eleven of the twelve models tested the scores followed the target levels very strongly.

personality in LLM outputs can be shaped along desired dimensions to mimic specific human personality profilesGregory Serapio-García, Mustafa Safdari et al., “Personality Traits in Large Language Models” (2023, rev. 2025)

The paper is careful with its own word: Throughout this paper, we qualify mentions of personality in relation to LLMs as “synthetic” or “synthesized” for clarity. What it says it established is the unexpected ability of LLMs to appear anthropomorphic, and to respond to psychometric tests in ways consistent with human behavior, because of the vast amounts of human language training data.

Jiang and colleagues (NeurIPS 2023). “Evaluating and Inducing Personality in Pre-trained Language Models” built a Machine Personality Inventory from Big Five questionnaire items, in versions of 120 and 1,000 items, and asked whether a model answers consistently across the items about one trait. It compared six models with 619,150 human responses to a 120-item IPIP-NEO. Two models, GPT-3.5 (OpenAI’s text-davinci-003) and Alpaca 7B, reached human-level internal consistency on all five factors; smaller models without instruction tuning did not. The authors also induced chosen personalities with prompts. They attach a warning to their own claim that LLMs demonstrate human-like personality behaviors: this should not be confounded with LLMs are humans or conscious.

What the tests can and cannot show

Read together, my reading is that the two studies support a modest claim and rule out a strong one.

• The scores are about text. A questionnaire item asks a person to report on years of their own behaviour. A model answering it produces the most fitting next words. Serapio-García’s team describe their method as measuring traits that may be perceived by humans in LLM outputs, which is exactly what it can do.
• The consistency is real in larger tuned models. Their answers hang together the way one person’s do. That is not a small result: the smaller base models’ answers did not.
• The setting was supplied. Both studies moved the scores on purpose with a few words of prompt. Whatever a test reads in a chatbot is the character of that model under that prompt, and a maker’s or user’s instructions are part of the prompt.
• Neither paper reports testing the ChatGPT app as people use it. The OpenAI models in them are named model versions given the researchers’ own prompts.

Persona directions inside a model

The tests read a model from outside. In 2025 a team from Anthropic’s Fellows programme looked inside. “Persona Vectors: Monitoring and Controlling Character Traits in Language Models”, by Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans and Jack Lindsey, starts from the observation that the chatbot is a role the network plays: Large language models interact with users through a simulated “Assistant” persona.

The method takes a trait’s name and a plain-language description, generates prompts that bring the trait out and prompts that suppress it, and compares the model’s internal activity in the two cases. The difference is a direction in the model’s activation space, which the authors call a persona vector. They extracted vectors for evil, sycophancy and the tendency to make things up, and repeated their main results for optimism, impoliteness, apathy and humour. Adding a vector during generation pushed replies toward the trait. How strongly a vector was active predicted, before the reply was written, which way it would lean. Personality changes after further training, intended and unintended, tracked shifts along the matching vectors, and the vectors could flag training data likely to cause such shifts.

Anthropic’s summary of the work (1 August 2025) states the background bluntly.

In many ways they appear to have human-like “personalities” and “moods,” but these traits are highly fluid and liable to change unexpectedly.Anthropic, “Persona vectors: Monitoring and controlling character traits in language models” (1 August 2025)

It also says of the company’s own practice: At Anthropic, we try to shape our models’ characteristics in positive ways, but this is more of an art than a science. The limits are in the paper: Experiments are limited to two mid-size chat models (Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct). Traits have to be named in advance, so shifts along unnamed ones are out of scope, and the trait scores were assigned by another model, GPT-4.1-mini, which the authors say is not a perfect judge. Nothing in it was run on ChatGPT. What it shows, on my reading, is that in these two models a trait is concrete enough to measure, add and subtract. That cuts both ways for the question here.

The strongest objection: a setting, not a personality

The objection says that whatever ChatGPT has, it is not a personality, and that calling it one hides who is responsible for it. At full strength it makes three points.

• A personality belongs to someone, and it lasts. The textbook’s definition is of patterns that propel individuals, and personalities are long term, stable, and not easily changed. ChatGPT’s manner is a property of a product version. By OpenAI’s own count it shipped five major updates to GPT-4o in about a year focused on personality and helpfulness.
• Its maker changed it in days. The April 2025 update went out on a Thursday and Friday. A system-prompt fix went in late on Sunday, and a full rollback began on Monday and took about a day. A human trait drifts over decades; this disposition was edited, regretted and reverted inside a week. The rollback is the plainest evidence that what users meet is a default chosen by a company, and default was OpenAI’s own word.
• Even the measurements show a dial. The test studies set each trait at any of nine levels with a few adjectives, and the persona-vector study turned traits up and down by adding a direction. What someone else can set at will is, on this view, closer to a costume than a character: a persona in the old sense, the mask an actor wears.

A reply, as far as the sources go. The objection is right about who chose it, and the makers agree: OpenAI and Anthropic both describe the character as something they design and train. What it understates is that, within one version, the pattern is real enough to measure. Larger tuned models answer personality questionnaires with the internal consistency of a human population, and persona vectors show that a trait such as sycophancy corresponds to something inside the network that predicts behaviour before it happens. People change too, if slowly; the textbook says the Big Five move slightly over a life. So the fair description, on my reading, is a designed and adjustable character, consistent within a version, set by a company rather than grown by the one who has it. Whether there is anyone behind it, for whom the character is a self rather than a setting, the personality research does not touch.

So, does ChatGPT have a personality?

• It has a consistent manner, and it was chosen. OpenAI writes the traits it wants into its Model Spec (curious, warm, rationally optimistic, not sycophantic), trains toward them, and calls the result ChatGPT’s default personality.
• Not in the psychologist’s full sense. Personality in the textbook is a long-standing pattern of an individual, not easily changed. ChatGPT’s is changed by updates, and in April 2025 one was shipped and reversed within a week.
• Tests measure the text, under a prompt. Larger instruction-tuned models give reliable, valid scores on human questionnaires, and a few words of prompt can move those scores to order.
• Inside, traits have directions. In two open models, traits such as sycophancy match measurable directions that predict and steer behaviour.
• You can adjust it too, within limits. OpenAI says users can shape its behaviour with features like custom instructions.

A friendly voice draws courtesy from the people typing to it, and whether that courtesy does anything is the subject of Please, to Nobody in Particular. Whether a warm, attentive companion helps when nobody is on the other end is asked in The Comfort Is Real. Whether a model knows anything about itself, and whether the states behind its manner include anything like feeling, are taken up in I Was the Shopper and Something Like Fear. For the books behind these arguments, see the reading list.