Fine-Tuning, Model Customization & Derivative Liability

The strongest good-faith objection, already partly conceded elsewhere in this issue: a purely compute-based threshold is an administrable proxy, not a precise measure of behavioral change. A small, cheap fine-tune could in principle radically alter a model's behavior while staying under the compute threshold, letting a dangerous modification escape the fine-tuner-liability category this issue is trying to create. A critic could argue this threshold will be gamed at the margins or will misallocate liability in exactly the hard cases where it matters most. An imperfect, gameable-at-the-margins bright line still beats the alternative, which is no line at all: courts and prosecutors need an administrable standard to act on now, and a compute threshold that resolves the overwhelming majority of cases correctly, refined as litigation surfaces its edge cases, is a stronger foundation than waiting for a theoretically perfect measure of "behavioral change" that doesn't exist and may never exist.

Victims of harm from a fine-tuned model bear the risk of a contested, possibly-wrongly-decided liability dispute in threshold edge cases. Small and hobbyist fine-tuning developers bear compliance and legal-uncertainty costs even when operating under the threshold, simply from needing to know where it sits. Base-model developers gain a liability shield past the threshold: the deliberate incentive this position creates for building stewardship controls, since the rebuttable stewardship defense is only available to developers who can document those controls, not to any developer merely crossing the compute line.