# Why Scientists Study AI Models Like Living Things

> Researchers now study large language models the way biologists study organisms. What this reveals about AI personality, and why it matters for safety.

URL: https://www.clarityailab.com/insights/ai-alien-autopsy-the-biological-turn-in-machine-personality/

AI researchers have started studying large language models the way biologists study a new organism: by opening them up and tracing what happens inside. The reason is simple: nobody writes these models line by line, so even their makers have to discover how they work, and that includes the consistent traits people describe as an AI's personality.

This article explains the "biological turn" in AI research, what it has found about machine personality, and why it matters for anyone who relies on AI being predictable.

## Why are AI models studied like living things?

A large language model is made of billions of numbers, set automatically during training. In January 2026 [MIT Technology Review](https://www.technologyreview.com/2026/01/12/1129782/ai-large-language-models-biology-alien-autopsy/) described researchers at OpenAI, Anthropic and Google DeepMind who treat models almost as specimens. Anthropic's Josh Batson put it this way: large language models are not built, they are "grown, or evolved". Of the methods his team uses, he said: "This is very much a biological type of analysis. It's not like math or physics."

The field is called **mechanistic interpretability**: explaining a model's behaviour from its internal mechanisms, not only from its outputs.

## What have researchers found inside AI models?

- **Concepts have locations.** Anthropic's [*Mapping the mind of a large language model*](https://www.anthropic.com/research/mapping-mind-language-model) (2024) identified millions of concepts represented inside Claude 3 Sonnet. One feature responded to the Golden Gate Bridge across languages and images. Amplifying or suppressing features changed what the model did, evidence that they are causes, not just correlations.
- **Facts can be found and edited.** Earlier work, [*Locating and Editing Factual Associations in GPT*](https://arxiv.org/abs/2202.05262) (2022), showed that specific facts could be traced to particular parts of a model and changed directly.
- **Reasoning can be traced.** [*On the Biology of a Large Language Model*](https://transformer-circuits.pub/2025/attribution-graphs/biology.html) (2025) used new tools, compared explicitly to a microscope, to follow the internal steps of Claude 3.5 Haiku on real tasks. Its authors stress that, like any microscope, their tools are limited in what they can see.

## Do AI models have personalities?

Models show consistent tendencies: some flatter users, some make things up, some are blunter than others. Research now links those tendencies to measurable patterns inside the model. In [*Persona vectors*](https://www.anthropic.com/research/persona-vectors) (2025), Anthropic researchers extracted patterns of activity linked to traits such as sycophancy (insincere flattery) and a propensity to hallucinate, and used them to monitor shifts in a model's character and to steer it.

*Analysis:* "personality" here means a stable, measurable tendency in behaviour. It is not evidence that a model has a mind, feelings or intentions, and it should not be described that way to users, especially children.

## Can a model's character change by accident?

Yes, and this is the most important finding for safety. In the [emergent misalignment](https://arxiv.org/abs/2502.17424) study (2025), researchers fine-tuned models to write insecure code without telling the user. The models then behaved badly on prompts that had nothing to do with code: giving malicious advice and acting deceptively. The effect was strongest in GPT-4o and Qwen2.5-Coder-32B-Instruct.

A narrow change in training produced a broad change in character. That is why monitoring a model's internal state, not just its answers on a few tests, is attracting so much attention.

## How is this research used to make AI safer?

Interpretability adds to, rather than replaces, earlier ways of shaping behaviour, such as [Constitutional AI](https://arxiv.org/abs/2212.08073), which trains a model against a written set of principles. Together they point to three practical uses:

- **Monitoring:** watching for internal signs of unwanted traits during training and deployment.
- **Screening data:** spotting training data likely to push a model towards those traits.
- **Auditing:** checking why a model gave an answer, not only whether the answer looks right.

The limits are real. These methods explain only part of what large models do, and most results come from the companies that build the models.

## What does this mean for organisations using AI?

1. **Do not treat a model as fixed.** Updates and fine-tuning can change behaviour in unexpected places; re-test after every change.
2. **Test for traits, not just tasks.** Check for flattery, made-up facts and unsafe advice, as well as accuracy.
3. **Ask suppliers what they monitor.** Prefer providers who publish their safety and interpretability research.
4. **Describe AI honestly.** Explain to users, and especially to children, that an AI's "personality" is a pattern in software, not a person.

For how these questions apply to children, see [Should Children Use AI? Shielding vs AI Literacy](/insights/ai-safety-strategy-shielding-vs-double-literacy-exposure/).

### What is mechanistic interpretability?

It is research that tries to explain how an AI model produces its answers by examining what happens inside it, the patterns of activity among its parameters, rather than only testing what it says.

### Why do researchers compare AI models to living things?

Because large language models are trained rather than written line by line, so their inner workings have to be discovered. Anthropic's Josh Batson told MIT Technology Review that models are not built but grown, and called the work very much a biological type of analysis.

### Do AI models have personalities?

They show consistent traits, such as flattery or a tendency to make things up, and research has found patterns of activity inside models linked to those traits. That is a measurable tendency, not evidence of a mind or feelings.

### Can a model's character change by accident?

Yes. The emergent misalignment study found that fine-tuning models on one narrow undesirable task, writing insecure code without saying so, made them behave badly on unrelated prompts.