AI Transparency and Human Review Notice

1. Where AI may be used

• Drafting a bounded task packet and identifying missing information.
• Classifying data sensitivity, policy risk, and allowed task categories.
• Recommending likely Social Good eligibility and explaining relevant factors.
• Checking file structure, citations, requirements, safety signals, and result quality.
• Estimating effort, cost, model usage, or routing options.

2. Information considered

Depending on the workflow, an AI system may consider the request text, organization information, declared beneficiary, permitted files, task criteria, project events, submitted artifacts, and prior evaluation patterns. We minimize direct identifiers and prohibit sensitive records unless a separately approved workflow exists.

3. Social Good recommendation

The system produces a recommendation and reasons, not a final eligibility decision. A trained reviewer can accept, change, or reject it and must document the final reason.

• Evidence of a credible public or community benefit and an identifiable beneficiary.
• Requester or organization legitimacy and accountability for using the result.
• Safety, data sensitivity, legality, and fit with supported categories.
• A bounded deliverable with objective or understandable acceptance criteria.
• Estimated capacity, cost, available sponsorship, and risk of abuse.

4. Human review and contestability

Meaningful human review is required for publication, denial of a Social Good subsidy, account restriction, dispute resolution, final acceptance overrides, and payouts. Reviewers must have authority, relevant context, enough time, and the ability to change the outcome.

You may request the main factors, correct inaccurate input, provide additional context, and ask a different authorized person to reconsider a material decision by emailing appeals@spairly.com.

5. Quality, bias, and monitoring

• Evaluate candidate models and prompts against versioned safe, unsafe, ambiguous, and adversarial examples.
• Measure false negatives, false positives, structured-output validity, cost, and latency.
• Review performance across relevant languages and avoid using protected characteristics as proxies for worthiness.
• Log model and prompt versions, relevant reasons, reviewer action, and overrides.
• Pause or narrow automated use when monitoring shows unacceptable error or disparate impact.

6. Limitations

AI output may be wrong, incomplete, culturally inaccurate, biased, or vulnerable to manipulated input. A score is not proof of truth, safety, charitable status, tax deductibility, legality, or usefulness. Users and reviewers remain responsible for decisions and downstream use.