HYPERNOVELTYINSTITUTE

Technical intelligence
for accelerating change.

Search the record

Home/Analysis

Daily Hypernovelty Lead · Law & evidence · September 24, 2026

The AI Use Has to Enter the Evidence File

The relevant requirements will arrive through particular grants, contracts, or publication rules.

Two researchers review paper evidence and notes at a public-research worktable, with a laptop beside the files.

On September 24, the Department of Housing and Urban Development put an unusually specific sentence into its updated Evaluation Policy: requirements for researchers' use of artificial intelligence, and for disclosure of that use in published reports, will be communicated through grant agreements, contract documents, or publication guidelines.

This is a narrow sentence inside a HUD policy for the department's staff, contractors, and grantees involved in evaluation, research, evidence-building projects, and regulatory impact analysis. It is not a government-wide disclosure rule, and the policy does not prescribe one form that every researcher must complete.

Its placement matters. The AI provision appears inside a much wider account of what makes evidence reviewable. For methodologically complex work, HUD says it will publish the research design and explain departures from that design when necessary. Reports must include favorable, unfavorable, and null findings while stating caveats and uncertainty. HUD also says it will publish final datasets, analytic code, and model documentation where feasible to support replication and verification.

That surrounding structure gives the AI sentence more weight than a generic label at the end of a report. A statement that "AI was used" tells a reviewer very little. Which finding did it inform? Was it used to search sources, classify records, write code, construct a model, or draft prose? Which parts did a person check? Did the method change after the work began?

Because HUD does not answer those questions with a universal schema in this policy, the relevant requirements will arrive through particular grants, contracts, or publication rules. Even so, the policy points toward a useful operating principle: AI disclosure belongs in the evidence trail, beside the research design, recorded deviations, uncertainty, code, and model documentation.

That combined record is what I mean here by an evidence file. The phrase is an editorial adaptation, not HUD's term or a new requirement stated in the notice. If an institution cannot connect a disclosed use of AI to the finding it affected and the checks applied to it, the disclosure may be accurate while remaining hard to audit.

Verification bottleneck

  • Research production can adopt AI faster than grants, contracts, and publication guidance can standardize what must be recorded.
  • Program staff, contracting officials, research leads, editors, and outside reviewers may each see only part of the trail. The scarce check is reconstructing how a specific finding moved from source material through AI-assisted work and human review.
  • Watch the actual language that appears in future HUD agreements and guidelines, especially the level of detail required and how exceptions to publishing data, code, or model documentation are explained.

Opportunities

A modest product could help research teams assemble an AI-use evidence packet while the work is underway. It might link each consequential use to the governing agreement, the affected research step, the tool and version recorded by the team, the human reviewer, the supporting sources, and any later correction. The result would be a reviewable project record rather than a last-minute disclosure paragraph.

There is also room for a pre-publication reconstruction service. A reviewer could trace selected findings backward through the research design, deviations, analysis files, source citations, and AI-use notes, then flag gaps for the research team to resolve. The service should document evidence and open questions without making a determination about whether contractual or publication terms were met.

The market signal here is small but concrete. Institutions are beginning to route AI accountability through existing grants, contracts, and publication systems. Teams that wait until a report is finished may discover that the people, versions, and intermediate decisions needed to explain the work are already difficult to recover. Recording the trail during the project is cheaper than trying to recreate it after the result becomes consequential.

Sources

Back to the archive