Technical & policy AI safety

Research

Questions I’m working through.

I’m interested in how we can understand AI systems, evaluate their risks, and make better decisions about their development and governance. My work spans technical and policy research, with an interest in outcomes for humans and other animals.

On the desk

Research proposal

Separation of Linguistic Mechanisms

Can we intervene on one linguistic property of a language model without changing others?

Question: Which mechanisms are shared between linguistic properties, and which can be separated?
Approach: Compare controlled linguistic contrasts, identify shared and distinct internal representations, and test candidate mechanisms through causal interventions.

A current thread in mechanistic interpretability: understanding when linguistic properties rely on distinct, shared, or compositional mechanisms.

In development

Language & Value Generalisation

Does changing linguistic structure in synthetic fine-tuning data change how models generalise animal-welfare values?

Approach

Compare matched fine-tuning datasets that preserve semantic and value content while varying linguistic structure, then evaluate behaviour on held-out scenarios.

Connecting controlled changes in training data to downstream behaviour, with follow-up analysis of the mechanisms involved.

Papers & past experiments

Manuscripts, research sprints, and collaborative projects.

2026 · Mechanistic interpretability

Tracing Low-Level Linguistic Properties from Sparse Representation to Behavior

Manuscript

Investigating how low-level linguistic properties are represented in language models, comparing activation probes with sparse autoencoder features and testing their relationship to model behaviour.

More about this work

The study uses controlled sentence contrasts to examine recoverability, robustness across surface forms and languages, and the effects of interventions on candidate representations.

Code & materials ↗

2026 · Sentient Futures

Disease Free Futures

Incubator project

An AI-enabled wildlife disease-monitoring case study, starting with wild deer. The prototype explores detecting abnormal behaviour and matching observed symptoms to disease information, with animal welfare as the motivation.

More about this work

Developed with Philip Kratz and Prince Jenkins through the Sentient Futures Incubator. The project explores how camera-based monitoring could support research into disease and the experiences of animals in the wild.

Project ↗

Jan 2026 · Apart Research AI Manipulation Hackathon

DEEB-learn

Diverse Evaluation Evasion Benchmark

Sprint report

A benchmark investigating how contextual framing changes model behaviour, including sycophancy, sandbagging, and deception. The project combines behavioural judgements with internal model signals to study evaluation evasion.

More about this work

With Ewura Ama Etruwaa Sam and Rick Hubertus Nicolas Curvers (Team DEEBies). The study uses 6,000 prompt-response pairs from Mistral-7B-Instruct-v0.3. This is early-stage participant research from an Apart Research sprint.

Report & project ↗Code ↗

Sept 2025 · Apart Research

RobustCBRN Eval

Research sprint

Testing how much CBRN multiple-choice benchmark performance depends on shortcuts rather than relevant knowledge. The toolkit includes choices-only controls, cloze variants, and statistical checks for benchmark artefacts.

More about this work

The project examines issues such as answer-position bias and spurious patterns, with the aim of making hazardous-capability evaluations more informative.

Code & overview ↗

Policy & governance

2026 · Policy & governance

Mapping EU AI Governance & Technical Assurance

Three connected lines of inquiry: EU product conformity, AI Act governance, and technical evaluation. The project maps institutions, responsibilities, evidence, and assessment methods to make the governance landscape easier to inspect.

The three maps remain distinct: technical evaluation is not treated as sufficient evidence of regulatory compliance. Source links and scope notes accompany the underlying records.

Explore the maps & sources ↗

Compare notes

A question worth working on?

I’m interested in research collaborations across technical and policy AI safety.

Get in touch ↗