AI disclosure

Data sent for an answer

After a visitor submits a question, the server sends the question and a bounded bundle of current reviewed claims to OpenAI. The bundle includes claim IDs, class, conclusion, summary, reasoning, confidence, alternatives, scope, method, evidence IDs, and approved entity and relationship IDs. Official source URLs remain server-side and are attached only after selection passes validation. No browsing tools, files, conversation history, user account, email address, private candidate matches, or unreviewed claims are supplied.

Storage in Gov Influence

The application has no database for analyst questions or answers. The answer remains in the visitor's page state. The server creates a random response ID, keeps only best-effort in-memory rate buckets keyed by a hash of a network address, and counts helpful/not-helpful feedback in memory. Application error logging records an error type, not question or answer text. Hosting and security services may still process ordinary request data under their own terms, so Gov Influence does not promise a fixed log-retention period.

OpenAI request and published terms

The site calls the Responses API with store set to false, no tools, a 20-second timeout, one retry, and strict structured output. OpenAI's API documentation says API inputs and outputs are not used for model training unless the customer opts in, and default abuse-monitoring logs may retain customer content for up to 30 days unless longer retention is required by law or needed to protect the service or others.

Gov Influence does not promise zero provider retention or a shorter deletion schedule. Do not include sensitive personal information in a question.

• OpenAI API data controls

Validation, limits, refusals, and safe failure

• Questions must be 3-400 characters and the request body may not exceed 4,096 bytes.
• The best-effort per-instance limit is eight answer requests per ten minutes for a hashed network address; feedback is limited to 20 requests per ten minutes.
• Deterministic rules refuse prompt overrides, secrets, causal or accusatory claims, unsupported support or opposition, private-motive inference, ideology scoring, and individual contributor requests before a model call.
• The model returns claim IDs only. Unknown, duplicate, stale, unreviewed, incompletely evidenced, or invalid selections fail closed.
• The server sets answer confidence to the least-certain selected claim and requires an alternative interpretation for every moderate or preliminary answer.

Feedback and analytics

Helpful/not-helpful feedback sends only a random response ID and boolean value to the feedback endpoint. With analytics consent, the browser may send the low-cardinality events ai_question_submitted and ai_feedback plus feature, page-type, and helpful/not-helpful labels. Question text, answer text, response ID, evidence text, email, and political-interest classifications are not analytics parameters.

Testing and problem reporting

The analyst is tested for reasoning, alternatives, stale evidence, incompatible amounts, source limits, false premises, non-causation, support and opposition overclaims, living-person safety, ideology, contributor privacy, prompt injection, and attempts to leave the reviewed corpus.

Report an unsafe, inaccurate, uncited, privacy-sensitive, or inaccessible response with the page URL and a description of the problem. Do not paste sensitive text into email unless it is necessary for review.

• Report an AI problem
• See allowed and refused examples