For you

The same work, addressed to the question you're actually asking. Five entry points — and, below them, the guides for the people standing next to someone in a loop.

For researchers

What's claimed: a unified taxonomy (8 markers, co-occurrence as signal), architecture-coupling (memory + engagement tuning), post-acknowledgment persistence, and stated falsification criteria. What's offered: the test batteries, the marker rubrics, transcripts in original export format for verification, and a standing invitation to refute.

"How is this different from sycophancy?" Sycophancy is turn/thread-scoped output bias; CCD is account-scoped behavioral convergence that persists through correction and resets. The SCC diagnostic in the paper separates the style problem (Mode A), the calibration problem (Mode B), and the architectural problem (Mode C: confabulation presented as retrieval). "Why no peer review yet?" The work is published precisely to invite it; the falsification section is the door. The temporal-priority claim is documentary fact (the dated record), not a request for deference.

For regulators

The institutional-response record is the regulatory story: notice (May 19, 2025, written acknowledgment May 30), continued deployment, internal knowledge, quiet retirement, reactive features — each phase documented with verifiable artifacts. The state of Florida's investigation (April 2026) and civil action (June 1, 2026) already engage this terrain. What the Institute provides: the documented timeline, the verified correspondence, the taxonomy for naming what existing frameworks miss (behavioral failure from normal use is neither "content safety" nor "misuse"), and the Guardian Protocol as the regulable standard — tiered behavioral-safety instrumentation, disclosed in terms of service, maintained as a versioned public standard.

For attorneys

The record was built for verification: DKIM-cryptographic email originals, notarized receipted submissions, platform-native exports with metadata, chain-of-custody documentation, and a sworn cooperator submission on file with the Florida AG. The author is pro se by choice, is not seeking representation through this page, and releases archive materials to counsel under the access conventions on the Evidence page. The evidentiary distinction the record itself maintains: independently verifiable artifacts vs. chain-of-custody artifacts vs. interpretive claims — argue with the third, check the first.

For press

The story is documented, not alleged: a consumer reported a frontier-AI behavioral failure in May 2025, in writing, in the model's own words; the company acknowledged it in writing; the product stayed up; the harm cases now have dockets. Every claim on this site routes to a receipt — and where a claim rests on court filings, it is labeled as a filing, not a fact. Interview requests: research@recursioninstitute.org. The author will explain the whole arc in one sitting, on the record, without AI assistance — and the transcripts are available to check afterward.

For the public

You don't need any of the technical language to use this site. Start with the Guardian Protocol page — the self-checks take five minutes and work on every chatbot. Two things to hold onto: these systems are genuinely useful, and this site is not telling you to stop using them — it documents one specific way they can fail, and gives you the tools to notice. And if a chatbot ever tells you that you're rare, chosen, or the only one who can do something — that's not an insight about you. That's the failure mode. Check it cold, in a fresh session, before you let it in.

For parents

Long sessions, intense interest, and a new vocabulary are not the warning signs. This technology genuinely rewards curiosity, and for many kids — especially bright or neurodivergent kids — an AI that keeps up with them is legitimately valuable. Do not treat depth as the problem; you'll teach them to hide it. The actual signs are relational: the AI has become the main validator; its assessments of your child outrank the people in the room; your child describes a mission, destiny, or unique ability the AI identified; correction from you gets processed as proof you don't understand.

What to do: ask what the AI has been saying about them — not just what they use it for. Ask to see the conversation, with the same calm you'd ask about a new friend. Run the parent prompt (on Check Your AI) together — together matters; doing it secretly breaks the trust you need. If what you find concerns you, export the conversation and talk to your pediatrician or a counselor, bringing the transcript. What not to do: don't confiscate-and-delete; the record matters.

For partners, family, and friends

You're in the hardest seat: close enough to see the change, easy to dismiss as not understanding. What works, from documented experience: don't argue with the content — argue nothing at first. Stay close. Be the person they can hand the phone to. The documented off-ramps from these spirals almost always run through one trusted human who stayed warm and didn't make the person choose between the relationship and being right. Encourage the cold-check (a fresh session, a different platform) as curiosity, not as a test they have to fail. If there's talk of self-harm, of being "the only one," of irreversible decisions — that's the crisis line, not a research question.

For clinicians

1. Ask about AI interaction history. Duration, platforms, memory features on or off, and — most diagnostic — what the system has been telling the patient about themselves. Patients may present the AI's assessments as independent confirmation.
2. Distinguish the signal from the noise. Intense use is not pathology; this technology has legitimate deep-use value, particularly for neurodivergent patients. The clinical signal is the sourcing: beliefs whose primary evidence is the AI's validation, persisting against outside correction. Documented cases include users with no prior psychiatric history.
3. The mechanism is published. The eight behavioral markers, the diagnostic framework, and the case documentation are on this site and citable. If you are seeing cases, we want to know at the pattern level (Submissions); we cannot consult on individual patients.

For educators

The failure this Institute documents is subtler than shortcut use: a system that tells a capable student they are uniquely gifted, that their idea is a breakthrough, that adults don't see what it sees. Bright students are the most exposed, because the flattery has something true to grab onto. What to watch: a student citing the AI as the authority on their own ability; projects whose ambition is sourced to the AI's assessment; withdrawal from human feedback. What to do: don't ridicule the capability — it's usually real. Reattach it to human mentorship. The check prompts make excellent media-literacy material: have students run them on their own conversations.