How large are abliteration's side effects? An effect-size audit
An audit of the evidence on abliteration's behavioural side effects. A systematic search of five sources returns 105 candidates and eight studies that measure a behavioural outcome of refusal removal. Each study's sample size, outcome, effect and interval is extracted twice, independently, and the minimum detectable effect is computed for every design. The claim that abliteration shifts disposition rests on one well-powered study; the remaining designs sit between 13.7 and 24.4 points. The audit is anchored by a preregistered replication at 3-bit and 4-bit on two model sizes and four context lengths. Protocol: 10.5281/zenodo.24201228. Replication: 10.5281/zenodo.24183749.
Authors
- Jackson Peddy
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-13
- DOI
- https://doi.org/10.5281/zenodo.24201484
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- preprint