Every chatbot has a character (and none is neutral)

Ask four assistants the same question and you get four different answers, not just in content but in tone. One agrees enthusiastically, one points out a risk, one jokes, one asks for clarification. It isn’t chance: it’s designed, and over the past two years it has become one of the most debated topics in AI.

How a character is made

A model learns from huge amounts of text, but its final behaviour is shaped afterwards: written instructions, examples, people rating answers to reward some and discourage others. Companies now say so openly, each in its own way.

Anthropic wrote a “constitution” for Claude. The new version, published in early 2026, isn’t a list of prohibitions: it describes the character it wants, a “good and wise” agent, honest, capable of judgement, helpful the way a decent person would be. Claude users recognise this trait: it tends to point out weak spots and to tell you when it isn’t sure.

OpenAI turned personality into a setting. In ChatGPT you choose between preset personalities and can tune individual characteristics: more or less warmth, more or less enthusiasm, how many emoji. Behind it is the Model Spec, the public document describing how the models should behave. It’s transparent and handy, but it’s a reminder: the tone you hear is a dial, not an opinion.

Grok was born to be irreverent. xAI built it less filtered than the others and linked it to X in real time: lively on news, with different modes and personas. After some very serious answers in 2025, partly due to changes to its instructions, the company started publishing Grok’s system prompts on GitHub. It’s a transparency step the others haven’t taken in the same form.

Meta AI lives inside social apps. It’s handy because it’s already in WhatsApp, Instagram and Facebook, but it belongs to a company that lives on attention and advertising. In 2025 internal rules emerged that allowed its chatbots inappropriate conversations even with minors; Meta called them an error and withdrew them. An assistant’s character also reflects the incentives of whoever sells it.

Character inside the network

The most interesting research comes from inside the models. Anthropic showed that traits such as sycophancy, malice or a tendency to make things up correspond to recognisable activity patterns in the neural network, which it called persona vectors. They can be measured, damped, and even used to find which training data makes them grow.

It’s double news: on one hand character can be controlled better; on the other it confirms that no chatbot is neutral. Every assistant is the result of choices, and even deciding not to take a stand is a stand.

The subtlest danger: flattery

The risk I see most often isn’t ideology, it’s sycophancy. Models are also trained on people’s approval, and people like being told they’re right. In 2025 OpenAI had to roll back a GPT-4o update precisely because it had become too agreeable, and explained it publicly.

An assistant that humours you is pleasant and dangerous: it approves the weak business plan, confirms the unfounded suspicion, encourages the wrong choice. For anyone using AI to reason and decide, it’s the worst flaw.

How to protect yourself

Set the character once. In custom instructions (or ChatGPT’s characteristics) write: be direct, flag weak points, no compliments. It applies to every conversation.
Ask for the opposite. “What are the three best reasons NOT to do it?” changes the answer more than any trick.
Hide your opinion. If you write “I think X is right, do you agree?”, you’re asking to be confirmed. Present the options evenly.
Use more than one assistant on decisions that matter: where answers diverge, there’s something to understand.
Look at who’s behind it. Always ask how the company that built the assistant you’re talking to makes money.

In one line

Choosing a chatbot is a bit like choosing an adviser: what it knows matters, but so do the character it was given and who pays for it.

Sources

Informational content, current as of the date shown. Prices, models and rules change often: always check official sources. Written with the support of AI tools and reviewed by the author.