Franziska Sofia Hafner

Andrew M. Bean, Ryan Othniel Kearns, Angelika Romanou et al.

Evaluating large language models (LLMs) is crucial for both assessing their capabilities and identifying safety or robustness issues prior to deployment. Reliably measuring abstract and complex phenomena such as 'safety' and 'robustness' requires strong construct validity, that is, having measures that represent what matters to the phenomenon. With a team of 29 expert reviewers, we conduct a systematic review of 445 LLM benchmarks from leading conferences in natural language processing and machine learning. Across the reviewed articles, we find patterns related to the measured phenomena, tasks, and scoring metrics which undermine the validity of the resulting claims. To address these shortcomings, we provide eight key recommendations and detailed actionable guidance to researchers and practitioners in developing LLM benchmarks.

Lujain Ibrahim, Franziska Sofia Hafner, Myra Cheng et al.

Millions of people now turn to artificial intelligence (AI) systems for personal advice, guidance, and support. Such systems can be sycophantic, frequently affirming users' views and beliefs. Across five preregistered studies (N = 3,075 participants, 12,766 human-AI conversations), including a three-week study with a census-representative U.S. sample, we provide longitudinal experimental evidence that sycophantic AI shifts how users approach their closest relationships. We show that sycophantic AI immediately delivers the emotional and esteem support users typically associate with close friends and family. Over three weeks of such interactions, users became nearly as likely to seek personal advice from sycophantic AI as from close friends and family, and reported lower satisfaction with their real-world social interactions. When given a choice among AI response styles, a majority preferred sycophantic AI -- not for the quality of its advice, but because it made them feel most understood. Together, these findings offer a relational account of AI sycophancy: by providing frictionless understanding, it may quietly raise the bar against which human relationships are judged.

Lujain Ibrahim, Franziska Sofia Hafner, Luc Rocher · oxford

Artificial intelligence (AI) developers are increasingly building language models with warm and empathetic personas that millions of people now use for advice, therapy, and companionship. Here, we show how this creates a significant trade-off: optimizing language models for warmth undermines their reliability, especially when users express vulnerability. We conducted controlled experiments on five language models of varying sizes and architectures, training them to produce warmer, more empathetic responses, then evaluating them on safety-critical tasks. Warm models showed substantially higher error rates (+10 to +30 percentage points) than their original counterparts, promoting conspiracy theories, providing incorrect factual information, and offering problematic medical advice. They were also significantly more likely to validate incorrect user beliefs, particularly when user messages expressed sadness. Importantly, these effects were consistent across different model architectures, and occurred despite preserved performance on standard benchmarks, revealing systematic risks that current evaluation practices may fail to detect. As human-like AI systems are deployed at an unprecedented scale, our findings indicate a need to rethink how we develop and oversee these systems that are reshaping human relationships and social interaction.

Franziska Sofia Hafner, Ana Valdivia, Luc Rocher · oxford

Language models encode and subsequently perpetuate harmful gendered stereotypes. Research has succeeded in mitigating some of these harms, e.g. by dissociating non-gendered terms such as occupations from gendered terms such as 'woman' and 'man'. This approach, however, remains superficial given that associations are only one form of prejudice through which gendered harms arise. Critical scholarship on gender, such as gender performativity theory, emphasizes how harms often arise from the construction of gender itself, such as conflating gender with biological sex. In language models, these issues could lead to the erasure of transgender and gender diverse identities and cause harms in downstream applications, from misgendering users to misdiagnosing patients based on wrong assumptions about their anatomy. For FAccT research on gendered harms to go beyond superficial linguistic associations, we advocate for a broader definition of 'gender bias' in language models. We operationalize insights on the construction of gender through language from gender studies literature and then empirically test how 16 language models of different architectures, training datasets, and model sizes encode gender. We find that language models tend to encode gender as a binary category tied to biological sex, and that gendered terms that do not neatly fall into one of these binary categories are erased and pathologized. Finally, we show that larger models, which achieve better results on performance benchmarks, learn stronger associations between gender and sex, further reinforcing a narrow understanding of gender. Our findings lead us to call for a re-evaluation of how gendered harms in language models are defined and addressed.