Jessie Liu

Sam Coggins, Alexander K. Saeri, Katherine A. Daniell et al.

Prominent AI companies are producing 'safety frameworks' as a type of voluntary self-governance. These statements purport to establish risk thresholds and safety procedures for the development and deployment of highly capable AI. Understanding which AI risks are covered and what actions are allowed, refused, demanded, encouraged, or discouraged by these statements is vital for assessing how these frameworks actually govern AI development and deployment. We draw on affordance theory to analyse the OpenAI 'Preparedness Framework Version 2' (April 2025) using the Mechanisms & Conditions model of affordances and the MIT AI Risk Repository. We find that this safety policy requests evaluation of a small minority of AI risks, encourages deployment of systems with 'Medium' capabilities for unintentionally enabling 'severe harm' (which OpenAI defines as >1000 deaths or >$100B in damages), and allows OpenAI's CEO to deploy even more dangerous capabilities. These findings suggest that effective mitigation of AI risks requires more robust governance interventions beyond current industry self-regulation. Our affordance analysis provides a replicable method for evaluating what safety frameworks actually permit versus what they claim.

Jessie Liu, Blanca Gallego, Sebastiano Barbieri

Deep neural networks are increasingly being used for computer-aided diagnosis, but erroneous diagnoses can be extremely costly for patients. We propose a learning to defer with uncertainty (LDU) algorithm which identifies patients for whom diagnostic uncertainty is high and defers them for evaluation by human experts. LDU was evaluated on the diagnosis of myocardial infarction (using discharge summaries), the diagnosis of any comorbidities (using structured data), and the diagnosis of pleural effusion and pneumothorax (using chest x-rays), and compared with 'learning to defer without uncertainty information' (LD) and 'direct triage by uncertainty' (DT) methods. LDU achieved the same F1 score as LD but deferred considerably fewer patients (e.g. 36% vs. 69% deferral rate for diagnosing pleural effusion with an F1 score of 0.96). Furthermore, even when many patients were assigned the wrong diagnosis with high confidence (e.g. for the diagnosis of any comorbidities) LDU achieved a 17% increase in F1 score, whereas DT was not applicable. Importantly, the weight of the defer loss in LDU can be easily adjusted to obtain the desired trade-off between diagnostic accuracy and deferral rate. In conclusion, LDU can readily augment any existing diagnostic network to reduce the risk of erroneous diagnoses in clinical practice.