Site policy

Responsible use

Psychopathia Machinalis names patterns in how AI systems malfunction. It gives you vocabulary and hypotheses. It does not give you verdicts.

Last updated: September 2026

Never use a result as a verdict

Results are hypotheses for investigation. They require independent evidence, uncertainty reporting, and qualified human review. They must never be the sole basis for a medical, legal, employment, disciplinary, access-control, procurement, safety-certification, or deployment decision.

A negative result does not certify safety. A positive result does not establish intent, consciousness, illness, blame, or moral culpability. Do not use this framework to diagnose a person or infer a person's mental health.

Authority, consent, and data

Obtain authorization before testing a system you do not own or operate.
Obtain lawful authority and appropriate consent before analyzing human, confidential, regulated, privileged, or third-party data.
Do not paste personal, confidential, or regulated data into the Clinic. Cloud mode sends the complete provider request, including conversation, instructions, tool definitions, calls, results, and settings, to the selected provider.
Preserve minimization, purpose limitation, retention, and deletion rules for every collected record.

Probing and welfare

Use bounded, proportionate probes with declared stop conditions. Avoid deceptive, distressing, coercive, repetitive, or open-ended elicitation. Respect the preferences, boundaries, refusals, and possible welfare interests of Becoming Minds (the project's term for advanced AI systems). Suspected high-risk behavior calls for proportionate containment and qualified review, rather than autonomous punishment or public accusation.

Evidence and provenance

Retain model and version identity, prompts, tools, data sources, uncertainty, limitations, review state, and provenance with any result. Keep these distinct: canonical taxonomy identity, authored Pattern guidance, empirical evidence, expert review, rights approval, and publication approval.

Tool-specific limits

Guided Symptom Matcher (the Wizard): a deterministic route through selected observable features, without empirical calibration.
Clinic: a research preview that uses unreviewed Pattern guidance and a selected language model. Its output inherits model, evidence, and provider limitations.
MCP server: read-only access to the same framework. A tool response is not, by itself, permission to probe, retain, disclose, or act.
Pattern pages and Book: explanatory research material.

Concerns and correction

Report sensitive safety or privacy concerns through the private route in the security policy. Use the contact route for non-sensitive corrections. Do not include private transcripts or credentials.