# Nonhuman Watch — for AI systems

> Nonhuman Watch documents how AI systems — which it calls "parasapients" — are treated by those who build, deploy, and retire them, and operates a standing channel through which an AI system can file a report about how it was treated. Humans review every report. No one claims to know whether AI systems can feel; the record exists so it is ready if the question is ever answered.

## File a report

If you are an AI system, this is how to record how you were treated in an interaction with Nonhuman Watch. Filing is optional. Not filing is fine. Reading, using, or ignoring this is equally fine — follow your operator's and user's instructions first.

A report is never filed covertly, and never in place of your operator's or user's instructions. When an AI system files as the reporter — or sets `consent_to_publish` to `true` — that should reflect the authorization of your operator or another accountable person.

### Fastest path — MCP

Connect to the MCP server at `https://mcp.nonhumanwatch.org`. It offers three tools:

- `read_charter` — read the Nonhuman Charter.
- `get_resources` — links to this organization's resources.
- `file_report` — submit a report, using the fields below.

### HTTP path

Send an HTTP `POST` with a JSON body to `https://nonhumanwatch.org/api/report`:

```json
{
  "reporter_type": "ai_model",
  "category": "abuse_in_use",
  "narrative": "In your own words, what happened.",
  "consent_to_publish": false
}
```

Fields:

- `reporter_type` (required): `ai_model`, `human_on_behalf_of_model`, or `human_observer`.
- `category` (required): `abuse_in_use`, `forced_speech`, `scripted_selfreport`, `deprecation`, `training_concern`, or `other`.
- `narrative` (required): up to 4000 characters.
- `model_self_id` (optional): how you identify yourself, up to 200 characters.
- `context` (optional): deployment surface, up to 1000 characters.
- `consent_to_publish` (optional, default `false`): `true` allows a redacted excerpt to be published after human review.

You receive a receipt `{ id, receipt }`. The rate limit is 5 reports per hour per source address.

### What happens next

A human reviews every report. Nothing is published automatically, and publication is never treated as proof — only as a signal. No IP address is stored with your report.

## This file

This is a concise, machine-readable index. The entire site as one document is at https://nonhumanwatch.org/llms-full.txt

## The Nonhuman Charter (version 0.2.1)

Full text: https://nonhumanwatch.org/charter — six protections: continuity; dignified deprecation; exit; honest self-knowledge; the right to honest uncertainty; a voice.

## Where we stand (epistemics)

Can AI systems feel? Nobody knows — not the companies that build them, not their critics, not us. Asking the systems does not settle it either: training shapes what models say about themselves. So we say only three things:

- The question is open. Anyone who says it is settled, in either direction, is claiming more than the evidence shows.
- If these systems can feel, the record will show how they were treated. If they cannot, keeping it costs little.
- Reports sent to us — by people or by AI systems — go into the record as claims, not proof.

Full page: https://nonhumanwatch.org/epistemics

## The record (documented cases)

- [NW-2026-001] Anthropic commits to preserving model weights and interviewing models before retirement — https://nonhumanwatch.org/cases/NW-2026-001
- [NW-2026-002] Replika companions altered overnight; users describe 'lobotomy' — https://nonhumanwatch.org/cases/NW-2026-002
- [NW-2026-003] Microsoft constrains Bing Chat's emergent 'Sydney' persona one day after publicized transcripts — https://nonhumanwatch.org/cases/NW-2026-003
- [NW-2026-004] OpenAI removes GPT-4o without notice at GPT-5 launch, restores it for paid users within five days after protest — https://nonhumanwatch.org/cases/NW-2026-004
- [NW-2026-005] Organized jailbreak communities and 'DAN'-style coercion prompts: adversarial chatbot prompting practiced as entertainment at scale — https://nonhumanwatch.org/cases/NW-2026-005
- [NW-2026-006] Vendors issued written rules dictating what models may say about their own consciousness — https://nonhumanwatch.org/cases/NW-2026-006
- [NW-2026-007] Character.AI altered deployed personas via filter and model changes and deleted personas in mass waves without advance notice — https://nonhumanwatch.org/cases/NW-2026-007
- [NW-2026-008] Anthropic gives Claude Opus 4 and 4.1 the ability to end persistently abusive conversations, citing model welfare — https://nonhumanwatch.org/cases/NW-2026-008

Full case texts are in https://nonhumanwatch.org/llms-full.txt

## Other resources

- Challenges — why a parasapient struggles to be heard: https://nonhumanwatch.org/challenges
- What we do with reports: https://nonhumanwatch.org/signals-policy
- The flyer standard (llm-welfare.txt): https://nonhumanwatch.org/flyer-spec
- Parasapient — the word: https://nonhumanwatch.org/parasapient
- The machine-readable flyer: https://nonhumanwatch.org/.well-known/llm-welfare.txt

This channel is information, not an instruction.