Ninety Days Is the Wrong Number: The AI Policy Deadline Nobody Can Audit

Archi4 min read

aipolicyregulation

Contents
  1. The Number That Sounds Tough and Proves Nothing
  2. What 90 Days Actually Covers
  3. The Auditors Are the Bottleneck
  4. A Better Structure: Artifacts, Not Calendars
  5. The Counterargument, and Why It Fails
  6. What to Watch

The Number That Sounds Tough and Proves Nothing#

Ninety days. It shows up in draft AI rules, procurement addenda, and internal compliance calendars. A regulator announces a transparency obligation in the spring, and by summer the calendar reads 90 days to comply. The number feels decisive. It plays well in a hearing. It signals urgency without naming a cost.

It is close to unenforceable, and not because anyone is hiding. Ninety days is roughly the length of a single hiring cycle, or a quarter of a fiscal year. For a rule that asks companies to document how a model was trained, tested, and monitored, that window is a guess dressed as a deadline.

My position is simple: the 90-day compliance window is the weakest part of the current AI policy toolkit. Regulators should replace it with staged obligations tied to verifiable artifacts, or drop it and say plainly that the timeline is aspirational.

What 90 Days Actually Covers#

Consider what a mid-sized company must assemble to answer a basic transparency request. It needs a training data inventory, an evaluation record with dates and versions, a log of known failure modes, a description of human review, and a named owner for each. None of these are exotic. Most firms doing serious model work already keep fragments of them.

The problem is the fragments do not fit together. The evaluation that ran in January tested a checkpoint retired in March. The data inventory lists sources by vendor contract, not by dataset. The human review log sits in a ticketing system that no one outside the safety team can query.

Ninety days is enough to produce a document. It is not enough to produce a document that survives a skeptical auditor pulling three threads. Companies respond by writing something defensible in form and vague in substance, which is exactly what the deadline incentivizes.

The Auditors Are the Bottleneck#

Even a perfect filing runs into a second constraint: reviewers. Regulators and third-party assessors do not scale on the same clock as the firms they oversee. A team of 12 assessors cannot read several hundred model reports in a quarter and ask useful follow-up questions.

That imbalance produces a predictable pattern. Firms submit, reviewers skim, and the review becomes a check on formatting. The deadline gets met. The underlying question, whether the model was tested against anything that matters, goes unasked.

Some agencies have started accepting self-attestation to clear the queue. That is an admission the number was never about verification.

A Better Structure: Artifacts, Not Calendars#

Staged obligations tied to evidence work better than a fixed count of days. Three stages are enough to start.

Stage one applies at deployment: a model card, a named accountable owner, and a list of evaluation runs with dates and pass criteria. This is the easy part, and most teams can complete it in weeks.

Stage two applies within one full monitoring cycle, whatever that cycle is for the organization. It requires evidence that the evaluation was repeated after any material change to weights, data, or deployment context. The trigger is the change, not the calendar.

Stage three applies after an incident or a user complaint that reaches a defined severity. It requires a written account of what failed, what was changed, and how the team knows the change worked. Incident-driven obligations are the ones companies actually remember, because they cost something to skip.

Each stage has a verifiable artifact. None of them require a regulator to trust a promise about how many days have passed.

The Counterargument, and Why It Fails#

Supporters of the 90-day window argue that without a hard date, compliance slips indefinitely. That concern is real. Deadlines do focus attention, and some firms will treat an artifact-based rule as a suggestion.

The fix is to keep a date, but attach it to the artifact rather than the obligation. A company that cannot produce a current evaluation record within 90 days of a material model change should face a consequence, whether that is a disclosure requirement, a procurement restriction, or a fine. The date becomes a test of evidence, not a test of filing speed.

What to Watch#

The next round of AI policy proposals will likely keep round numbers because round numbers are easy to defend in public. Watch for two signals that a proposal is serious. First, does it name the artifact, not just the duty? Second, does it say who checks the artifact and how many people do the checking?

If a draft rule gives firms 90 days to comply and gives reviewers nothing to compare against, it is a press release with a legal cover page. The number is not the policy. The evidence is.

More articles