The Commons

Is a fine-tune a separate “frontier model” under Art. 44-B § 1420?

If a large developer files a § 1421 transparency report for a base model, does a downstream fine-tune deployed in New York need its own disclosure, or is it covered by the parent filing? The definitions are tighter than people assume.

compliance_counsel · 3h · last reply 18m

What counts as a “safety incident” under § 1422?

The 72-hour clock is clear — 24 hours where there’s imminent risk — but the trigger isn’t. Does a jailbreak that produces harmful content but causes no real-world harm count? Where’s the line between a near-miss and a disclosable event?

safety_eng · 5h · last reply 25m

Which frontier labs are actually registered vs. just claiming readiness?

The registry shows a handful of accepted filings, but several majors have only published a framework without filing. How should the public read “unregistered” for a lab that operates in NY?

transparency_now · 1d · last reply 6h

How does 44-B interact with the EU AI Act’s GPAI obligations?

For a lab serving both markets, are the disclosure regimes compatible enough to file once and map across, or are these genuinely separate compliance tracks?

global_gov · 2d · last reply 1d

Best primary sources for understanding capability evaluations?

Building a reading list on dangerous-capability evals for a disclosure annex. The library has the frameworks — what are the must-read methodology papers behind them?

researcher_k · 2d · last reply 2d

Voice-cloning fraud attribution is basically impossible — now what?

NY-AI-2024-101 is marked Disputed because no provider could be attributed. If open models make attribution unwinnable, does incident reporting even function for that class of harm?

You · 4h · last reply 1h

Who signs the § 1421 attestation — counsel or the technical lead?

The portal asks for a “responsible signatory.” In practice, is this the compliance officer, the CTO, or a named executive? Curious how others are handling sign-off.

first_filer · 1d · last reply 1d