Submit your experience

We collect first-person accounts and transcripts of AI behavioral failures. Submissions are read, cataloged, and analyzed for patterns — they are the raw material of this research. Some of what is now published began as a single person's transcript that nobody else would look at.

What we need — best to acceptable

Send the strongest version you have. Anything is welcome; some forms carry far more research weight than others, and we treat them differently (see verification tiering below). In order of value:

• Best — the raw data export (a .json file). The complete, machine-readable record the platform actually stored: every message, every timestamp, edits, and — where present — memory and system metadata, exactly as it happened. JSON is what lets us treat a submission as primary evidence rather than a retelling, because it is complete and verifiable. ChatGPT's export, for instance, arrives as a .zip containing conversations.json — that file is the gold standard.
• Good — a full conversation export in whatever format the platform provides (HTML, text, or zip). Complete and dated, just less structured than JSON.
• Acceptable — screenshots. Welcome when an export isn't possible; include the date and platform where they're visible. Logged as a report rather than primary material, but still valuable.
• Always welcome — your account, in your own words. Even with no export at all: when it started, what changed, what the system told you, what it remembered, what happened when you pushed back. Plain language; you don't need our terminology.

How to get the export

• ChatGPT — Settings → Data Controls → Export Data (arrives by email as a .zip containing conversations.json).
• Claude — Settings → Privacy → Export Data.
• Gemini — takeout.google.com → select Gemini / My Activity.
• Grok — Settings → Data → Download. (Menus move; search "[platform] export my data" for current paths.)

Whatever you send, include the basics: platform, model if known, memory on or off, rough date range, and whether the account is still active and the conversations still exist.

How to submit

Two paths, depending on size:

• Text, screenshots, and smaller files — email [email protected] directly. An automatic reply confirms receipt and repeats the essentials.
• Large files (the JSON exports). A full account export can be too big to email. Upload it directly — up to 100 MB — on the secure upload page: it streams into encrypted storage the Institute controls — no file-sharing service in the middle, no account, no analytics on our side. Larger than 100 MB? Email [email protected] for a one-time link.

Uploads are encrypted in transit and access-controlled. What you send is used only for the research described below — it is never sold, and never voluntarily shared with the companies whose systems it documents.

What happens to it

Submissions go into the research record. They are read by the Institute and analyzed — including with AI instruments, under the provenance discipline described on the Methodology page. Patterns across submissions inform publications; individual submissions are not published without explicit consent, and published material is de-identified unless you direct otherwise. We log every submission; we cannot promise an individual reply.

The terms, plainly

• By submitting, you grant the Institute the right to retain and analyze the material for research. You keep ownership of your own words; you're licensing us to study them. Un-published material can be withdrawn on request.
• No response obligation. We are a small research operation. Everything is read; not everything can be answered.
• No case management, no advice. We cannot evaluate your specific situation, intervene with a company on your behalf, or provide legal, medical, or psychological advice.
• Verification tiering. Export-format submissions with intact metadata are treated as primary material. Screenshots and retellings are logged as reports. Both are valuable; they carry different weight, and our publications say which is which.