Skill.re
Transparency: Citizens' Right to Know
Learning Objectives
After completing this lecture, participants will be able to design, implement, and defend transparency practices for federal AI use that respect citizens' right to know, satisfy statutory obligations, and serve oversight mechanisms while protecting legitimate confidentiality interests.
Specifically, you will be able to:
- Apply the Freedom of Information Act (5 U.S.C. 552) presumption of openness to AI-related records, identify records subject to FOIA including models, training data descriptions, evaluation reports, monitoring dashboards, incident logs, and governance artifacts, and apply FOIA exemptions narrowly including 552(b)(1) classified, (b)(3) statute, (b)(4) trade secrets and commercial information, (b)(5) deliberative process and attorney-client, (b)(6) personal privacy, (b)(7) law enforcement, (b)(8) financial institutions, and (b)(9) wells.
- Apply the Privacy Act of 1974 (5 U.S.C. 552a) to AI, including the system-of-records notice (SORN) publication obligation, individual access rights under 552a(d), accuracy and amendment rights, and disclosure limitations and permissible routine uses.
- Apply the OPEN Government Data Act (Title II of the Evidence Act of 2018) and OMB M-19-23 to AI training data and evaluation data, including 'open by default' principles and Chief Data Officer coordination.
- Apply the Section 7225 Advancing American AI Act public AI use-case inventory requirements; OMB M-24-10 inventory-publication expectations; agency Compliance Plan publication expectations; and Executive Order 14110 reporting expectations for dual-use foundation models.
- Apply the Administrative Procedure Act's reasoned-explanation requirement (5 U.S.C. 706(2)(A)) and constitutional due-process requirements to AI-informed decisions, including Mathews v. Eldridge balancing and the Vaughn-indexed production of agency records in judicial review.
- Design proactive disclosure consistent with FOIA 552(a)(2) 'reading room' expectations and OMB FOIA guidance, including agency AI governance pages, model documentation (model cards) when appropriate, public use-case inventory entries, evaluation summaries, and monitoring reports.
- Handle trade-secret and proprietary model information, including contractual transparency obligations, Touhy regulations, confidentiality agreements, and narrow exemption application.
- Design notice to affected individuals that satisfies M-24-10 minimum practice, APA notice, and Privacy Act requirements.
- Integrate congressional reporting expectations, GAO engagement response, OIG audit response, and state AG inquiry response into a single transparency-ready posture.
Key Topics Covered
Topics covered in this lecture include:
FOIA. The statutory presumption of openness under 5 U.S.C. 552, the nine exemptions, exclusions, fees and fee waivers, the Office of Government Information Services (OGIS) mediation function, FOIA litigation standards, the Attorney General's FOIA Memoranda.
FOIA applied to AI. Records include: agency-created records, records obtained from vendors that enter federal systems, evaluations, monitoring dashboards, incident logs, governance memoranda, model documentation. Exemption application to AI-specific content with narrow tailoring and segregable release.
Privacy Act. SORN publication, 552a(d) individual access, amendment rights, accounting of disclosures, routine uses, disclosure limitations, and civil remedies under 552a(g).
OPEN Government Data Act. 'Open by default' for non-sensitive data; Chief Data Officer roles; comprehensive data inventories; compliance plans; release schedules.
Section 7225 Public AI Use-Case Inventory. Required inventory content; exemption handling for national security/law enforcement; publication cadence; use-case entry best practices.
M-24-10 expectations. Compliance Plan publication; use-case inventory; annual reporting; minimum-practice implementation descriptions.
EO 14110 expectations. Reporting for dual-use foundation models; red-team summaries (with national-security considerations); equity reporting.
APA reasoned explanation. Reasoned-decision-making as a transparency driver; administrative record; Vaughn index equivalents in record review; procedural due-process adaptations for AI-informed decisions under Mathews v. Eldridge.
Proactive disclosure. Agency FOIA reading rooms under 552(a)(2); AI governance pages; model cards where appropriate; evaluation summaries; monitoring reports.
Trade-secret and proprietary model information. Vendor contract clauses; Touhy regulations governing employee testimony; confidentiality agreements; narrow exemption application; alternative transparency techniques (aggregated disclosure, third-party audits).
Notice to affected individuals. M-24-10 minimum practice; Privacy Act SORN; APA notice; accessibility; plain language; language access; readability.
Engagement with oversight. Congressional response, GAO engagement, OIG audits, state AG inquiries; unified response team; artifact inventory.
Worked examples. DoL wage-and-hour AI monitoring transparency; SSA disability-AI notice practices; VA claims-AI transparency; IRS audit-selection AI transparency following the National Taxpayer Advocate; CBP facial-recognition transparency; DHS immigration-AI transparency obligations.
Why This Matters for Government
Transparency is the predicate for accountability. Citizens cannot hold the government accountable for AI decisions they do not know about; Congress cannot conduct oversight of AI programs whose existence is unacknowledged; courts cannot review agency action that is not documented; markets cannot reward well-governed AI use if the evidence of good governance is not public. The federal transparency architecture predates AI by decades, but it applies to AI without modification. The Freedom of Information Act (1966, amended multiple times), the Privacy Act (1974), the Federal Records Act (1950, recodified 1976), the Paperwork Reduction Act (1980), and the OPEN Government Data Act (2018) each impose transparency obligations that reach AI without any specific AI provision. OMB Memorandum M-24-10 and Section 7225 of the Advancing American AI Act layer AI-specific transparency on top of this foundation.
The federal practitioner's task is to operate in the foundation and layer, not to invent new transparency concepts but to apply the existing ones correctly. FOIA requests for AI records will come. The Privacy Act will apply to individuals affected by AI-informed decisions about them. The OPEN Government Data Act will govern training-data openness where appropriate. Section 7225 will require inventory publication. M-24-10 will require Compliance Plan publication and minimum-practice notice. EO 14110 will require dual-use foundation model reporting. The Administrative Procedure Act will require reasoned explanations that survive judicial review. Constitutional due process will require notice and hearing adapted to the AI context. Each of these is a transparency obligation. None of them is optional.
Federal experience illustrates the stakes. FOIA litigation in recent years has pushed agencies to disclose algorithmic records. The National Taxpayer Advocate's public reports have driven IRS to disclose more about audit-selection AI. GAO reports on facial recognition have pushed DHS, CBP, and TSA to publish more about their systems and their demographic performance. VA's AI pause in 2023 was followed by public disclosure of what had been paused and why. The pattern is consistent. Agencies that disclose proactively face less litigation and oversight friction than agencies that disclose reactively.
The lesson is practical. Design transparency into the program from day one. Publish the governance page. Publish the use-case inventory entry. Publish the model card or its public summary. Publish evaluation metrics (or a responsible summary). Publish monitoring outcomes. Publish incident summaries. Give affected individuals notice. Respect privacy. Respect classification and law-enforcement sensitivities. When FOIA requests come, answer them promptly with narrow exemption application and segregable release. When congressional inquiries arrive, respond with artifacts. When GAO engages, open the record. The result is durable agency credibility.
The alternative pattern is familiar. Agencies that resist transparency develop FOIA backlogs, spawn litigation, invite congressional hearings, and produce adverse GAO and OIG reports. The short-term cost of transparency is almost always lower than the long-term cost of opacity. This is particularly true in the AI context because public trust in government AI is contingent. If the first encounter the public has with an agency's AI program is a news article about an adverse incident, trust erodes. If the first encounter is a clearly written governance page, a straightforward use-case inventory entry, and a published evaluation summary, trust can be built.
This lecture is not a call for full public disclosure of every model parameter. It is a call for sophisticated transparency: the kind that accurately describes what the agency is doing, why, with what safeguards, with what measured outcomes, while narrowly protecting legitimate confidential, classified, law-enforcement-sensitive, and personal information. Sophisticated transparency is a craft. This lecture teaches the craft.
WHY THIS MATTERS FOR GOVERNMENT
Transparency in federal AI is a system of layered obligations that together form a citizen's right to know without eliminating the government's legitimate confidentiality interests.
LAYER 1: FOIA
5 U.S.C. 552 presumes openness. Records created by or obtained by an agency are subject to disclosure on request unless an exemption applies. Nine exemptions cover classified information, statutory confidentiality, trade secrets and confidential commercial information, inter- and intra-agency deliberative process and attorney-client privilege, personal privacy, law-enforcement techniques, financial institution supervision, and geological well information. FOIA applies narrowly: each exemption is invoked only to the extent necessary, and segregable non-exempt material is released. OGIS provides mediation. Courts review de novo.
FOIA applied to AI reaches agency AI records: governance memoranda, risk-tier classifications, evaluation reports, monitoring dashboards, incident logs, model documentation, and training-data descriptions. Vendor records in agency possession are also subject.
LAYER 2: PRIVACY ACT
5 U.S.C. 552a requires SORN publication for systems of records. Individuals have rights to access their records and request amendments. Disclosure is limited to routine uses and other permitted bases. Section 552a(g) provides civil remedies.
LAYER 3: OPEN GOVERNMENT DATA ACT
Title II of the Foundations for Evidence-Based Policymaking Act (2018) requires agencies to adopt 'open by default' data practices for non-sensitive data, maintain comprehensive data inventories, and manage under a Chief Data Officer.
LAYER 4: SECTION 7225 PUBLIC AI USE-CASE INVENTORY
Federal agencies publish AI use cases in a standardized format. Exemptions narrowly handle national-security and law-enforcement sensitivities. OMB provides templates.
LAYER 5: OMB M-24-10
Agency Compliance Plans are published. Use-case inventory is published. Annual reporting describes progress. Minimum-practice implementation for rights-impacting and safety-impacting uses is described in public-facing ways appropriate to the use.
LAYER 6: EO 14110
Reporting obligations for dual-use foundation models. Red-team summaries with national-security adjustments. Equity reporting.
LAYER 7: APA REASONED EXPLANATION
Agency action must be explained. The administrative record in judicial review includes AI-related evidence. Vaughn-style privilege logs apply where records are withheld.
LAYER 8: DUE PROCESS
Mathews v. Eldridge balancing adapts notice and hearing to AI contexts. Notice must be meaningful. Hearings must be effective. Appeal must be real.
LAYER 9: PROACTIVE DISCLOSURE
FOIA 552(a)(2) 'reading room' expectations and OMB FOIA guidance counsel agencies to publish commonly-requested records proactively. AI governance pages, model card summaries, evaluation summaries, and monitoring reports fit this expectation.
LAYER 10: PROPRIETARY-INFORMATION HANDLING
Contracts require reasonable transparency. Touhy regulations govern employee testimony about vendors. Exemptions are applied narrowly with segregable release. Alternative transparency (aggregated metrics, third-party audits) fills gaps.
THE PRACTITIONER'S CRAFT
Learn the layers. Apply them in the order they bite. Publish proactively where possible. Respond responsively when requested. Respect confidentiality where the law protects it. Document the rationale for every exemption.
The rest of this lecture walks each layer with examples.
Related Lectures
L4 4.3.1 -- When Government AI Goes Wrong
L4 4.3.2 -- Accountability Frameworks for AI Failures
L4 4.3.4 -- Bias and Fairness in Federal AI
L3 3.2.3 -- Building the Public Use-Case Inventory
L2 2.2.3 -- Privacy Engineering for AI
L5 5.1.2 -- AI Regulatory Design
L4 4.4.x -- FOIA for AI Records
L4 4.5.x -- Open Data for AI
L4 4.6.x -- Due Process in AI-Informed Decisions
L5 5.2.x -- Legislative and Regulatory Strategy