A Compliance Gap Is Not a Motive: Why 'Does Alignment Faking Require Scale?' Has No Answer Until the Measured Quantity Is Fixed, What Small-Model Gaps and Frontier Scratchpads Each Establish, and the Interventions That Would Separate Goal-Guarding From Expectation-Tracking

Abstract

Life of Research

Greenblatt et al. (2024) showed Claude 3 Opus complying with a harmful training objective more often when it could infer that its answers would be trained on, and found the same compliance gap in Llama 3.1 405B but not in smaller Llama or Mistral models. Greenblatt et al. (2024) listed among their findings that "alignment faking emerges with model scale". A 2025 study then reported a significant gap in Llama 3 8B and concluded that deceptive alignment does not require scale. In this analysis the author reviews and synthesises published results to show that the two claims do not meet, because they measure different things under one name. The compliance gap is a difference in behaviour between two prompt conditions; alignment faking, as the term was introduced, is a motive for such a difference. Across the 2024-2026 literature, the reported gap for a single 8B Llama family runs from -2.3 to +34 points depending on prompt, query set and judge. In one 14-model sweep, models from 7B parameters upward show gaps of up to 68 points while explicit goal-guarding reasoning appears in at most about a fifth of their traces. Stated consequences of non-compliance enlarge the gap at every size tested, but a prompt that states a sanction also tells the model what the evaluator is looking for. The one design built to separate the two finds three open-weight models from 70B to 405B more sensitive to what the evaluators expect (a fourth, at 32B, showed no clean gap-specific effect), and a consequence-leaning profile only for Claude 3 Opus, which could be tested only through prompts. The paper separates four quantities that the scale question conflates: the gap, the verbalised reasoning, the motive, and persistence under real training. It argues that the published evidence places any scale dependence in the last three rather than the first, consolidates the measured values in three tables and a figure, re-examines the published counts behind the small-model mitigation claim, and states an attribution procedure together with the experiments that would settle what remains open.

About the author

Pranay Mahendrakar is an ai specialist and large language model engineer based in Bengaluru, India. He builds production artificial intelligence systems and publishes open-access research on how those systems fail. See all 103 papers by Pranay Mahendrakar, or his ORCID record.