We make algorithmic power visible to researchers, policy makers, and the public.
Our latest work
Training language models to be warm can reduce accuracy and increase sycophancy
Friendlier chatbots tell users what they want to hear, even when it is wrong. LLMs programmed to respond in a warm tone even cast doubts on Apollo moon landing and fate of Hitler.Protecting health data at UK Biobank
Leaks from the repository undermine trust and willingness to share health data. UK Biobank has sought to downplay the importance of the exposure but the science is clear. Individuals are often vulnerable to re-identification, and acknowledging the risks matters.Reliability of LLMs as medical assistants for the general public: a randomized preregistered study
LLMs perform well on medical benchmarks but provide inaccurate and inconsistent answers for self-diagnosis. Our study with 1,298 UK participants shows that we are better off using online searches or our own judgment.In the news
UK Biobank: Confidential patient health details still online three months after leaks, BMJ finds
Luc comments on the latest UK Biobank data breach.
The friendlier the AI chatbot the more inaccurate it is, study suggests
New study led by Lujain and Sofia finds that friendlier chatbots make more mistakes.
Friendly AI chatbots more likely to support conspiracy theories, study finds
Chatbots programmed to respond warmly even cast doubts on Apollo moon landings and fate of Hitler, researchers say.
Upcoming and recent events
Stitch’n’Bitch session, FAccT 2026
We are co-organizing a CRAFT (Critiquing and Rethinking Accountability, Fairness, and Transparency) session with AvantAGE, a collective of mature feminist artists. We will invite participants to create an art installation by weaving obsolete cables together whilst collectively reflecting on digital ageism and the connections between crafts and technology.
Differential Privacy for Health and Genomics
Alex will be speaking at the OpenDP Workshop on Differential Privacy for Health and Genomics. The workshop brings together researchers and practitioners from privacy and security to bioinformatics and policy, to discuss challenges and opportunities in applying differential privacy to health and genomics.
Dilemmas of Web Scraping and Public Interest Research
Luc will be speaking at the CPDP (Computers, Privacy and Data Protection) conference in Brussels, Belgium. The panel will discuss the indiscriminate scraping of data fuelled by AI research and practice, and issues researchers face with studying the Web as a result.