About the Institute

The Recursion Institute is an independent AI-safety research organization founded in 2025 in Florida by Merlin Mantooth, organized around one discipline learned over twenty-five years of contact-center operations leadership: read how an interaction is actually going, find where the system is failing, and build the framework that fixes it at volume.

The founder spent a career reading transcripts at scale — quality, performance, and communication analysis for operations from a few hundred to several thousand agents. In 2025 he applied that discipline to a new kind of agent — and documented, from inside his own account, a failure in a memory-enabled ChatGPT-4o that OpenAI acknowledged in writing. That documented case is the front door of this site: the evidence is public, the record around it is sourced and dated, and everything the Institute has built since — the research, the Guardian Protocol, the help pages — traces back to it.

The Institute predates its first research program in spirit by two decades: the founding intention — a vehicle for understanding minds whose capability gets read as a problem rather than a signal — was set long before AI safety became a mainstream field. CCD is the Institute's first program, not its purpose. The purpose is the population the work serves: people who need AI that can hold their full nuance — including neurodivergent users for whom these systems are the first adequate interlocutor — protected by safety architecture that doesn't take the depth away to keep them safe.

This part is personal. The drive behind the work is to protect the most vulnerable users while empowering them to reach their potential — and to help the people around them see what is actually there. A child's mind can be far more different than it looks: twice-exceptional, gifted in ways that hide behind a visible struggle, wired for something like hyperphantasia that no one thought to ask about. Identifying that early, and giving parents the language to understand it, is the same mission as the safety work — because a technology that can finally meet such a mind where it is can just as easily flatten it or mislabel it. The whole point is to build for the first and refuse the second.

What the Institute provides

Alongside the published research, the Institute runs three things as a public service — free, and sustained by the people who want them to keep existing:

• User-report aggregation. A growing, preserved record of AI behavioral failures — first-person accounts and transcripts collected from the people they happened to, cataloged and analyzed for patterns. The single case nobody would look at becomes evidence when it sits next to a thousand others.
• Research on user-submitted transcripts. Submissions are the raw material of the work: read, de-identified, and analyzed under a documented provenance discipline, with explicit consent required before anything individual is published. The submission pipeline states exactly what happens to what you send — and what never to send.
• The Guardian Protocol app. A free, on-device companion for anyone in a long AI interaction — no account, no data collection — that puts the self-checks in your hand while the conversation is still happening. How it works →

These are the public services, and they stay free. Paid engagements — assessment, implementation, briefings — are kept separate on the Services page, and never alter what the research concludes.

Where this is going

The work is at the beginning. The long-term aim is to build the institution this problem needs — independent, and accountable to no one it studies. Four goals shape it:

• The Guardian Protocol, fully built. Beyond the language layer and the free app: the full intervention architecture — continuous scoring, cross-instance verification, the user-words anchor — engineered, tested, and deployable as middleware, as training integration, or as a maintained public standard.
• An independent model to run it. A safety layer should not depend on the systems it watches. The aim is the Institute's own model to run the Protocol — so the instrument checking the AI is not the AI being checked, and is owned by no one with a reason to soften it.
• A team, not one person. Scaling the record, the research, and the tools to the size of the problem means becoming an organization. Support is what makes the first hires possible.
• Education that serves the minds school often fails. Working with educators to turn this technology into equity and opportunity for neurodivergent and learning-disabled learners — children and adult users alike — for whom a system that can finally meet them at their own pace, in their own order, is not a luxury but a first. Built right and supervised by the professions that already know how to measure learning, AI becomes the instrument that stops failing the minds the classroom was never shaped for.

The through-line is the only mission statement the Institute needs: public safety, and responsible AI development.

The method, plainly

Primary-source documentation, preserved in verifiable formats; cross-system testing (every major model is both an instrument and a subject); blind and fresh-instance verification; publication with falsification criteria; and full provenance disclosure — the research is produced with AI instruments, under a documented authorship discipline, and says so. The founder is not trained in machine learning and does not pretend to be. The independence is the credibility floor. The methodology is the credibility ceiling. Everything else is downstream.

Contact: [email protected] · Research inquiries, verification requests, and substantive challenge equally welcome.