Papers for

ai safety auditors

Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.

Fine-tuning memorization claims challenged due to flawed measurements

Playing Whack-a-Mole with misconceptions about memorization, extraction, and copyright

Abstract: After careful review, I'm confident the headline fine-tuning memorization results in Alignment Whack-a-Mole use an invalid measurement procedure. The book memorization coverage metric these headline results depend on counts sequence matches far shorter than what field standards consider valid evidence of memorization, and the prompting procedure used to elicit memorization runs the risk of leaking the text being "extracted" in the prompt. The paper doesn't include the negative-control experiments needed to see how much the results are inflated by false positives: claiming extraction success (and therefore memorization of training data) when matches between generations and training data may be due to other factors. Given these validity issues, the paper's claims that fine-tuning lets users extract substantial portions of copyrighted books, in a form that could substitute for the originals, aren't supported by the reported results. The failure to report the experiments' cost (an important component of the threat model) further compromises the copyright claims. I'm writing this note because, in the last month, (prospective) plaintiffs have reached out to me to ask about this paper. They're looking to cite this work as valid evidence in support of claims in ongoing and potential future copyright litigation.

Tue 8 SeptComputers and Society
The gist
Some researchers claimed that fine-tuning AI models lets people copy big parts of copyrighted books by memorizing the text. This paper reviews that claim and finds big problems: the tests used to detect copying were too weak, and the way they asked the AI to show memorized text might have accidentally given away the answers. The authors say the original results don't prove that copying really happened, and important control experiments are missing. Also, the costs of these leaking risks weren't reported, which matters for legal discussions on copyright.
Open 2609.09320v1

Weight quantization reduces answer variety but not bias in large language models

Quantization Amplifies Determinism, Not Bias: Scale-Dependent Behavioral Effects of Serving-Time Weight Compression

Abstract: Weight quantization largely determines the economics of serving open-weight LLMs. Its costs are usually assessed with capability benchmarks, on which 4-bit quantization of mid-sized models is often considered "nearly free." We examine a different question: when several answers are valid, does quantization change what a model chooses to say? We serve three checkpoints (Qwen3-8B/14B/32B) at three weight precisions (W4A16 AWQ, W8A16 FP8-Marlin, and bf16), holding the hardware, software, and sampling configuration constant, and collect approximately 71,000 completions paired by prompt and seed across two custom, leak-checked prompt batteries. We pre-specified the analyses in three waves in version control. At 8B, int4 reduces output diversity: the probability that two samples for the same scenario recommend the same brand increases by 5.1 percentage points (prompt-paired sign-flip test, Holm p = .023; reproduced at +4.4pp on a full regeneration of the arm), and lexical diversity falls substantially (TTR -0.011, standardized effect -0.51; robust to a length-controlled measure). At 14B and 32B, no content-concentration measure reaches significance; instead, stylistic drift emerges (em-dash rate +0.46/1k words at 14B and +0.61/1k at 32B, both Holm p <= .0024). Pre-specified tests of stereotype direction are null at every scale: outputs concentrate on the modal answer for each prompt rather than on stereotypical answers. Mechanistically, the token-level distribution becomes flatter (decision-token entropy +0.091 bits, p = .015) while the semantic distribution, measured directly from first-token log probabilities, becomes more concentrated (collision +2.6pp, p = .023): individual tokens become less predictable even as meanings become more repetitive. At 8B, the smallest size tested, AWQ-int4 serving measurably narrows the range of suggestions; audits should assess concentration as well as bias.

Mon 7 SeptArtificial Intelligence
The gist
When large language models are made smaller and faster by simplifying their numbers, it changes how certain answers are chosen. The researchers found that this simplification makes smaller models more likely to repeat the same answers, lowering how surprising their replies can be, but it does not increase biased or stereotypical responses. Bigger models mostly show small changes in writing style rather than answer variety. This means compressing these models affects how they decide what to say, but does not necessarily make them more biased.
Open 2609.07901v1