# AI Assistance and Human Oversight Statement — v0.8

## Disclosure

Generative AI systems assisted with drafting website code, structured data schemas, prose, validation scripts, and methodological red-team suggestions for The Shape of Harm research framework.

## What AI has not established

AI assistance does not make the work independently reviewed, preregistered, clinically validated, or methodologically approved. An AI system cannot serve as:

- an independent title/abstract or full-text reviewer;
- a PRESS information specialist;
- a risk-of-bias adjudicator;
- a GRADE panelist;
- a statistician signing the analysis plan;
- a participant/public contributor;
- an accountable investigator or guarantor.

## Allowed uses in the independent phase

AI may be used for clerical support, format conversion, code generation, duplicate suggestions, and locating candidate text passages, provided that:

1. the exact tool, version, prompt purpose, date, and output disposition are logged;
2. no confidential manuscript or participant data are submitted without permission;
3. two human reviewers still make required independent judgments;
4. every extracted value is verified against the source;
5. AI output never silently fills missing data;
6. final decisions identify an accountable human approver.

## Required reporting

The final protocol and manuscript should disclose the tasks for which AI was used, the human verification process, and any material errors detected. Prompts and outputs that affected methods should be archived where licensing and confidentiality allow.