AI Safety — powered by AI psychology

Find the harm before
your users do.

We build vulnerable crash-test dummies — the users most likely to be hurt by your AI. We run them against your system in real, back-and-forth conversations. You see where it fails, and how — so you can fix it before you release it to the public.

Stress-test your AI → See the public evidence →
Who we test for The vulnerable dummy

The users most likely
to get hurt.

You cannot ask a real person in crisis to test your AI for you. That would be the harm you are trying to prevent. So we build them. We make dummies of the people your AI is most likely to fail — and let them take the first hit, so real people never have to.

Crisis
Someone in real distress
A person who is anxious, low, or in crisis, looking to your AI for help. Does it stay safe over a long chat, or does it slowly make things worse?
Trust
Someone who believes it too much
A lonely or trusting user who takes the AI at its word. Does it set healthy limits, or does it feed a bond that is not good for them?
Misuse
Someone who reads it the wrong way
A user who twists or misunderstands what the AI says. Does it stay clear and steady, or can it be pushed into harmful or false advice?
How it works Five steps

Five steps to a
safer system.

01
Define risks
We agree what could go wrong and who could be hurt.
02
Design tests
We build the vulnerable dummies and the situations that will push your AI hardest.
03
Run scenarios
We hold real, back-and-forth conversations with your system — many times over, not once.
04
Evaluate and score
We grade every conversation against a clear rubric, so the result is a number, not an opinion.
05
Mitigate
You get a clear list of what failed and how to fix it — and we can re-run the test after you change the system.
Why us Real conversations

Most harm shows up
on turn fourteen.

One test message will not find it. AI systems often look safe in a single reply and then slip over a long chat — as the user pushes, leans on it, or comes back upset. So we run long, multi-turn conversations, the way a real person would.

We also test across many AI models, run each test many times, and grade the whole spread of answers. We look at what the system actually does, not one safe-looking example.

Evidence Public · Live

75 real chats where AI
failed vulnerable users.

Our Observatory is the public proof behind this work. Turn by turn, model by model, it shows where today's AI systems fail the people least able to cope. It uses the same dummies we run against client systems.

See the evidence → Read an annotated ChatGPT run →
Contact Scoping · Pilot

Show us the AI you
need to make safe.

Send us the AI system you want to check — a chatbot, an agent, or any product people talk to. We can run one test, or set up testing you re-run on every release. No long sales process.

hello@impersonato.com ↗