# 16. How do we verify AI scouting reports and avoid unsupported player claims?


**Require a claim-level evidence trail and measure the time needed to verify the report, not merely the time needed to generate it.** NIST’s generative-AI profile specifically recommends checking sources and citations and testing systems in realistic settings. ([nvlpubs.nist.gov](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf))

Supplier claims are becoming more specific: Hudl’s 3 August 2026 release notes say its AI Player Summaries are derived from event data and link findings to video. Test whether those links support the actual sentence, including its scope and strength. ([hudl.com](https://www.hudl.com/releases/statsbomb))

Create an **AI Scouting Report Evidence Check**:

| Claim type | Required support | Typical rejection reason |
|---|---|---|
| Numerical | Reproducible query, period and denominator | Wrong season or incompatible denominator |
| Video observation | Match and timestamp | Clip does not show the claimed behaviour |
| Comparative | Defined peer group and metric | Unstated comparison population |
| Biographical | Authoritative, dated source | Outdated or mismatched player record |
| Predictive | Model, assumptions and uncertainty | Forecast presented as established fact |
| Contextual judgement | Attributed observation and reasoning | Unsourced assertion about character or intent |

In an authorised test environment, include unavailable information, conflicting sources and ambiguous names. A useful system should flag the problem or decline to conclude rather than fill the gap with plausible language.

Record:

- Supported and unsupported material claims.
- Incorrect citations and links.
- Omitted limitations.
- Reviewer correction time.
- Whether the corrected output changes the recommendation.

Do not ask a general scouting assistant to infer health, private conduct or psychological characteristics from match behaviour. Keep sensitive assessments within authorised specialist processes.