Silent Revision: Measuring Undisclosed Change in the Safety Frameworks of Frontier AI Developers

TEXT START: Frontier AI developers publish safety frameworks that commit them to evidencing whether their models are dangerous.

THE DISSECTION

This is an empirical autopsy of regulatory theater. The paper rigorously documents that the accountability infrastructure the EU AI Act, California's SB 1047, and similar frameworks rely upon is legally mandated but procedurally hollow. Twelve frontier AI developers publish safety frameworks, regulators treat these documents as accountability instruments, and yet 67% of material changes go unmentioned in any published account. The paper doesn't speculate about motive—it just counts.

The core findings:
- 67% silent revision rate under strict standard; 53% under lenient one
- 77% of traced changes weaken or remove a commitment
- Weak ening changes are more likely to be silent than strengthenings in 7 of 8 version pairs
- The one provider (unnamed) that voluntarily provides an enumerated changelog still does so incompletely

This is not a paper about AI danger. It is a paper about the infrastructure of plausible deniability.

THE CORE FALLACY

The paper correctly identifies the symptom but treats the regulatory framework as salvageable. The implicit assumption is that if disclosure requirements were better specified, accountability would function. This is structurally naive.

The mechanism the paper documents is not a compliance design flaw. It is the predictable behavior of actors operating under competitive pressure to minimize external constraint while maximizing internal flexibility. The developers wrote the very accountability documents they are now silently revising. The enumeration duty the paper recommends would shift the theater one layer deeper—developers would enumerate changes in language they themselves control, to standards they themselves define, under assessors they themselves fund.

The paper's conclusion—that publication duties should carry an enumeration duty—is technically correct and institutionally insufficient. The fallacy is believing that making the document legible fixes the accountability gap, when the gap is structural: the subjects of regulation author the instruments of their own accountability.

HIDDEN ASSUMPTIONS

1. That developer-published frameworks were ever meaningful commitments rather than liability management artifacts. The paper treats them as genuine accountability instruments that are being improperly revised. More likely: the frameworks were always designed to be selectively enforceable and quietly mutable.

2. That regulatory bodies have the technical capacity and institutional will to act on disclosed revisions. The paper assumes enumeration enables audit. It does not—audit requires technical staff to understand what the commitment meant, what it changed to, and whether the change reflects genuine safety reasoning or liability minimization. No such staff exists at scale in any current regulatory body.

3. That the problem is procedural, not epistemological. The paper frames the failure as one of legibility—the reader cannot learn what changed from the developer's own account. But the deeper failure is that the reader cannot learn from any account whether the change reflects legitimate safety evolution or strategic commitment erosion.

4. That voluntary enumeration (the "one provider already meets it") is evidence of good faith. Voluntary incomplete enumeration is a different signal: a forward position in regulatory arbitrage, establishing favorable precedent while doing the minimum required to claim compliance.

SOCIAL FUNCTION

This paper is transition management infrastructure. It performs several functions simultaneously:

• For regulators: Provides technical justification for stronger disclosure rules that, when enacted, will be equally hollow but require developers to do more paperwork. The paper makes regulation look proactive.
• For developers: Validates their existing frameworks as legitimate accountability instruments deserving protection, rather than acknowledging them as liability management documents.
• For civil society: Offers the satisfaction of empirical confirmation that the theater is broken, which satisfies the demand for accountability without disrupting the mechanism.
• For the paper's authors: Legitimate academic contribution that produces actionable-seeming policy recommendations with no genuine enforcement pathway.

The paper is not cynical—it is methodologically rigorous and genuinely alarming. But its function within the transition management ecosystem is to make the accountability gap legible without threatening the power structures that produce it.

THE VERDICT

Under Discontinuity Thesis logic, this paper documents a specific and revealing failure mode: the institutional infrastructure meant to govern frontier AI development is procedurally captured at the document level. Developers control the text, control the revision, control the disclosure, and face no technical capacity that can verify their accounts against reality.

The 77% weakening rate is the number that matters. Silent revision of safety commitments, with a directional skew toward weakening, is not a compliance design problem. It is the market at work: competitive pressure compresses safety commitments over time, and the regulatory apparatus has no sensor to detect it. The paper provides that sensor. Nobody in a position of power has asked for its readings, and when they do, they will ask for it in terms that the developers themselves define.

The statutory remedy exists and specifies the wrong artifact. So does every proposed fix this paper will generate.

The frameworks were never the accountability mechanism. They were the alibi.