We research how AI systems work.
Then we check whether yours does.
Keido Labs is an AI research and development lab in Liverpool. Our work runs across mechanistic interpretability, AI psychology and AI safety — how these systems represent things internally, what they do to the people talking to them, and how both of those go wrong.
We also run an assurance practice, and it is not a side business. It is what the research is for. We audit LLM features that are already in production — the chatbot, the copilot, the agent, the retrieval search your customers are using right now — and we tell you, with evidence you can reproduce, what is wrong with them.
We arrived at it from an unusual direction.
Clinical psychology spent 150 years learning to tell whether a conversation helped a person or harmed them. AI development spent none. We started this company to close that gap, and we still think it is the most consequential one in the field.
What we found on the way is that the gap is wider than psychology. Teams shipping AI features are missing the entire apparatus every other engineering discipline takes for granted: independent review, evidence, a severity rating, a remediation plan, somebody who comes back and re-tests the fix. So we built that too — with the psychology as the part we go deepest on, because it is the part nobody else can.
Independence is not a feature of what we sell.
It is the thing being sold.
Two Things, Not One
A broad practice with one very sharp edge
The practice: AI assurance
For any company with an LLM feature live in front of users. We adversarially test it, examine whether your evaluation setup would catch a regression, review the governance you would hand a buyer or a regulator, and check whether you would even know if it broke. Fixed scope, fixed price, findings rated by severity, a re-test included.
The edge: psychological safety
For anything with a chat surface a person can reach. Not whether the model can be jailbroken, but what the conversation does to the human having it — crisis handling, dependence, boundaries, tone under distress, whether the system tracks risk across a conversation or forgets every message. Designed by a psychologist, measured against human raters, reported with the agreement figures attached.
Plenty of people can test the first. Almost nobody can test the second. We do both, in the same report, because the companies that need one usually need the other and don't know it yet.
Why Now
Two things happened at once.
The first is harm that stopped being hypothetical. Character.ai was sued after a teenager's death. Incidents involving general-purpose assistants now surface monthly. Companion, coaching and wellbeing products grew faster than any means of checking them, and the people most affected are rarely the ones with a support contract.
The second is that the rules arrived. The EU AI Act's transparency duties under Article 50 have applied since August 2026. California's SB 243 has governed companion chatbots since January 2026, with New York, Utah and Illinois behind it and Oregon and Washington from January 2027. The MHRA has drawn a line where a chatbot stops being a wellbeing product and becomes a medical device. And in September 2025 the UK government published a roadmap for a trusted third-party AI assurance profession — the model that produced CREST for security testing, pointed at AI.
Between the two sits a gap: companies that now have to prove diligence, and no established way to do it. Security testing went through exactly this twenty years ago, and came out the other side with an independent audit as a normal line item. AI assurance is making that turn right now.
That is why Keido Labs exists — and why we would rather be early than credible in hindsight.
Founding Team
Dr. Michael Keeman
Founder & CEO
Michael bridges mechanistic interpretability, clinical psychology and AI engineering. He leads every assurance engagement personally.
He spent 15 years in psychology and AI systems - publishing 50+ papers, leading national-level research programs, and building AI systems for elite sports and crisis healthcare. At 29, he became Chief Science Officer at a National Research Center of Sports Medicine, directing innovation strategy for Olympic sports performance and crisis response systems.
When the stakes are 'this person's health depends on this,' you learn to build systems that don't fail.
The transition from human psychology to AI wasn't random. It was the realization that everything learned about how humans process emotion, communicate under stress, and build trust could be encoded into measurable, interpretable systems that serve for good.
That realization became Keido Labs.
Anastasia Keeman
Director of Communications
Anastasia is a biomedical researcher who discovered that the gap between breakthrough science and real-world impact is communication.
With a background in stem cell research, scientific conference management, and translating complex research for diverse audiences, she ensures Keido Labs' work is both rigorous and accessible.
She bridges the gap between the lab's scientific depth and the world that needs to understand it.
Where We Are
Liverpool. Our office is at Central Tech, and the North West is our home turf. We work with plenty of Manchester companies as well — 35 minutes on the train from Manchester Piccadilly to Liverpool Lime Street, near enough to turn up in person.
We work across the UK, currently with clients in Cheshire and Edinburgh as well as either side of the M62. The psychological safety work we run from anywhere.
Assurance runs on trust, and trust is still mostly built in rooms. Where an engagement can be scoped in one, we would rather it was.