# VigilData — Responsible-AI summary

VigilData's AI Watchdog is a research aid grounded in primary sources. This
summary captures the safeguards in place and on the roadmap. The full policy
lives at /ai-watchdog/responsible-ai.

## What AI Watchdog does

- Surfaces source-linked summaries from public records.
- Categorizes signals using structured queries against primary-source data.
- Provides accountability research signals: donor-influence signals,
  contribution-to-policy correlations, accountability risk scans, and
  elevated-review signals.

## What AI Watchdog does not do

- Does not make legal findings of corruption, fraud, or quid pro quo.
- Does not establish causation from correlation.
- Does not target private citizens.
- Does not replace journalistic, legal, or academic review.

## Source verification

Every signal is intended to link to a primary public record. Where a signal
cannot link to a primary record, the signal is suppressed or labeled as
non-source-linked.

## Human review

Outputs intended for publication are reviewed by a human editor before being
framed as findings. A human-in-the-loop review queue is on the responsible-AI
roadmap.

## Bias & false-positive risk

AI-assisted categorization can over-fit on documented entities and
under-cover entities whose records are less digitized. Signals reflect what
public records permit us to document.

## Privacy & data use

AI Watchdog operates on public records about public officials. User-supplied
text sent to model providers is processed for the specific query, not used
to train external models. User-identifying information sent to providers is
minimized.

## Model and provider disclosure

Specific model provider(s) will be disclosed as the platform matures. Where
available, the model name and version used for a specific output will be
exposed alongside that output.

## Rate limits and cost controls

Public AI Watchdog usage is rate-limited. Heavy-use research partnerships are
handled separately via research@vigildata.org.

## Appeals and corrections

Misrepresentations of underlying public records may be submitted to
corrections@vigildata.org. Corrections policy: /corrections.