Responsible AI Statement

The three commitments

• AI is never used to make decisions about organizations without a human check. Verification, disputes, ownership contests, moderation, and admin review are decided by people. AI may summarize inputs to those decisions; it does not make them.
• Every field carries a per-field AI-exclusion switch. An organization can mark any field "Excluded from AI & Matching." That flag is honored everywhere — retrieval, embeddings, model context, matching — and is enforced by an automated code scanner that runs in CI.
• What AI reads is bounded by the visibility matrix. A six-tier visibility × nine-axis use-case matrix determines which fields any AI feature can see. Excluded from AI & Matching removes a field from every axis. Public data (already in the public record) reaches AI on public-facing surfaces only.

Where AI is used

Website extraction (opt-in per claim)

• When a claimant asks Bearings Network to review their organization's public website, a language model reads the site and proposes structured extractions: suggested capabilities, suggested facilities, suggested evidence links.
• Every extraction is a suggestion. Nothing reaches a public record without the claimant explicitly reviewing and approving it in the organization's Review Queue.
• The only content read is public HTML the site itself serves.

Mission and technology classification

• A classifier maps public-source facts (NAICS codes, award history, patent classifications) to defense missions and technology categories. The classifier uses both a rule-based NAICS bridge and a language-model check on ambiguous cases.
• Every derivation shows its inputs. Classifications can be disputed via the Correction and Dispute channel.

Public-source summaries (Navigator Insight)

• On some public pages, Bearings Network shows a short summary derived from that page's public-source facts. These summaries are labeled "a summary derived from public evidence — not a score, rating, or recommendation."
• Summaries never include fields that are not already public on the same page. They never include fields marked Excluded from AI & Matching.

Public-record ingestion assist

• Structured public-source imports (SAM.gov, USAspending, USPTO, SBIR, grants) are handled by rule-based connectors — no language model reads or writes those fields.
• Only unstructured surfaces (a claimant's website when they opt in) touch a language model.

Where AI is not used

• Verification decisions. Tier 1, Tier 2, and Tier 3 all use deterministic checks — email domain match, UEI/CAGE presence, admin review. No AI decides whether a claim is valid.
• Ownership contests. Reviewed by admins with documentation, per the Correction and Dispute Policy.
• Fact-level dispute outcomes. Disputes route to the affected organization's Review Queue for a human decision; a subset auto-escalates to admin review, again for a human decision.
• Ranking or scoring organizations. Bearings Network does not rank, score, or recommend organizations to each other. Listings are alphabetical or faceted by explicit filters.
• Matching without consent. Anonymous Matching returns capability-range matches; disclosure of identity requires the counterparty's explicit approval on each request. AI does not preemptively make introductions.
• Suspending, banning, or archiving accounts. Automated controls flag; humans decide.
• Setting prices, tiers, or eligibility for Bearings customers. Bearings Network is free to participate; the paid tier follows a published pricing catalog.

The AI-exclusion opt-out

Any field on a claimed Organization Record can be marked Excluded from AI & Matching. That setting is load-bearing.

• Excluded fields are removed from vector embeddings, retrieval indexes, prompt context, matching computations, and any summary generation. They are stored and never reasoned over.
• Enforcement is at the service layer, not the UI. Any code path that reaches a language model with organization data must pass through a helper that filters excluded fields. A repository-wide scanner runs in CI to flag any new call site that does not.
• A whole record can be marked Excluded from AI & Matching at the organization level. This is the "never learn about us from AI-derived features" switch.
• The AI-exclusion setting cannot be bypassed by paying, by admin override, or by any workflow. It is a one-way gate: opt out at any time, and the exclusion takes effect on the next indexing cycle (minutes, not days).

What data reaches AI models

• Public-source facts — same content available on the public Organization Record. Reaches models when a public-facing AI feature needs them.
• The claimant's website HTML — only when the claimant explicitly triggers website extraction. Not stored beyond the extraction run.
• Organization-provided text — reaches models only from fields with visibility ≥ Network and no AI-exclusion flag.
• Never reaches models — Organization Only fields, Approved Orgs fields, Anonymous Matching fields, Excluded from AI & Matching fields, contact information, disclosure-request bodies, evidence documents themselves (only their public metadata), CUI (which we refuse at the input layer anyway).

Model providers and retention

• Bearings Network uses commercial language-model APIs from established providers. Provider selection is a confidential engineering choice; the constraint that binds every provider is that they must offer contractual no-training-on-input, no-log-retention terms and be hosted in US regions.
• Prompts and responses are not stored beyond the request lifetime except for a bounded operational log used to debug failures and detect abuse. Operational logs are retained for 30 days and redacted of organization identifiers where feasible.
• No model provider trains on Bearings Network requests or organization data. This is a contractual requirement, verified in vendor security reviews.

Neutrality

Any AI-generated text presented to a user is labeled and carries a neutrality note:

Bearings Network does not produce competitive rankings, popularity indicators, top-N lists, percentiles, or recommendations. The public directory is factual reference; AI features that summarize are always attributed and always derived from the same public evidence a human reader can see on the page.

Governance and change

• Any new AI feature must specify which fields it reads (from the visibility × AI-use matrix), what happens if those fields are AI-excluded, how output is labeled, and how output can be disputed.
• The AI-exclusion scanner is a launch-blocking control: any code path that could bypass exclusion is a CI failure.
• Material changes to what AI does on Bearings Network are announced to claimants via the in-app notification matrix, dated on this page, and reviewable via the Correction and Dispute channel.

Reporting an AI concern

If a summary, extraction, or classification reads as inaccurate, out-of-scope, or as a recommendation that it should not be making, file a report:

• For a specific field — use the "Report this fact" action on the public record.
• For an AI feature as a whole — email ai@bearings.network. Include the page, the specific text or extraction, and what you expected.

Changes to this statement

Changes are dated in the "Last reviewed" line at the top of this page. Any expansion of what data reaches AI models — new feature, new field category, new provider — is announced to claimants before it takes effect.