Shows circuit-level mechanistic interpretability explanations flip across defensible analytic choices on GPT-2 small, so the evidence fails as EU AI Act Annex IV documentation.
Tests how internal model evidence affects the validity of LLM-generated policy audit reports, finding that more internal citations do not guarantee validity and shuffled evidence can still look plausible.
Argues for interpretable, narrow models in high-stakes settings as a route to regulation, and reports that tensor-network transformers remain competitive while mechanistic interpretability is still unsolved.
This position paper argues that mechanistic interpretability needs standardized auditing, proposing a collaborative continuous-review platform, expert-verified guidelines, and source-based tracking of claims for safety-critical use.
Proposes measuring and controlling the agency of LLM-based agents via representation engineering, paired with regulatory tools such as mandated testing, domain-specific agency limits, and insurance pricing.
Argues that mechanistic interpretability should be built into frontier models as a design constraint, supplying verifiable causal evidence for private AI governance such as audits and certification.
A career guide for outsiders transitioning into technical AI safety, covering interpretability, alignment, control, required skills, and learning resources. It mainly offers practical advice rather than new research findings.
Argues that evidence of AI misalignment convincing enough to trigger strong responses must bottom out in human-understandable behavioral evidence, making purely internals-based detection unlikely to be legible.
Argues that standardized, tamper-resistant explainability metrics can serve as governance tools for private oversight, linking interpretability to alignment and preventing alignment faking, and proposes a roadmap.
Argues that benchmark-based AI safety regulation fails to predict post-deployment harm due to model opacity, and proposes a regulatory framework that does not rely on benchmarks.
A report assessing whether training data attribution can become practical on frontier LLMs, its benefits and disclosure barriers, and its potential for reducing large-scale AI risk via interpretability.
Assesses whether training data attribution can become a practical frontier-LLM tool, covering its feasibility, obstacles from labs' data secrecy, possible policies, and its modest expected value for reducing large-scale AI risk.