Max Schaffelder

Lily Stelling, Malcolm Murray, Bruno Galizzi et al.

Following the AI Seoul Summit in 2024, twelve AI companies published frontier AI safety frameworks (Frameworks) outlining their approaches to managing catastrophic risks from advanced AI systems. Emerging legislation increasingly treats these Frameworks as external accountability mechanisms, incorporating them into reporting requirements. But what do the Frameworks actually commit each company to do? This study assesses 12 Frameworks, using 65 weighted criteria, across four dimensions: risk identification, risk analysis & evaluation, risk treatment, and risk governance. Our criteria adapt established risk management principles from other high-risk industries (e.g. aviation, nuclear power) to the frontier AI context, following Campos et al. (2025). Overall scores range from 34% (Anthropic) to 8% (Cohere), with a median of 18%. Many aspects are missing or under-specified. These low scores may be natural given the nascency of AI risk management compared to industries with decades of practice. The current Frameworks are limited as accountability functions, with vague commitments that make it difficult to predict company decisions, assess whether planned responses are adequate, or determine whether commitments have been kept. Higher scores appear feasible within current constraints: a company adopting all leading practices currently adopted across their peers would score 51%, almost triple the median.

Max Schaffelder, Albert Gatt

As synthetic data becomes widely used in language model development, understanding its impact on model behavior is crucial. This paper investigates the impact of the diversity of sources of synthetic data on fine-tuned large language models. We focus on three key dimensions: distribution collapse, adversarial robustness, and self-preference bias. Our findings reveal that fine-tuning models on synthetic data from diverse sources can mitigate distribution collapse, preserving the breadth of the output distribution and the diversity of the output text. Furthermore, while both human and synthetic fine-tuning data can remove safeguards, the latter preserves higher output quality, thus making outputs potentially more usable and dangerous. Finally, fine-tuning reduces self-preference bias, with human data being the most effective, followed by multi-source synthetic data.