Mowafak Allaham

Anka Reuel, Avijit Ghosh, Jenny Chim et al.

Foundation models are increasingly central to high-stakes AI systems, and governance frameworks now depend on evaluations to assess their risks and capabilities. Although general capability evaluations are widespread, social impact assessments covering bias, fairness, privacy, environmental costs, and labor practices remain uneven across the AI ecosystem. To characterize this landscape, we conduct the first comprehensive analysis of both first-party and third-party social impact evaluation reporting across a wide range of model developers. Our study examines 186 first-party release reports and 183 post-release evaluation sources, and complements this quantitative analysis with interviews of model developers. We find a clear division of evaluation labor: first-party reporting is sparse, often superficial, and has declined over time in key areas such as environmental impact and bias, while third-party evaluators including academic researchers, nonprofits, and independent organizations provide broader and more rigorous coverage of bias, harmful content, and performance disparities. However, this complementarity has limits. Only model developers can authoritatively report on data provenance, content moderation labor, financial costs, and training infrastructure, yet interviews reveal that these disclosures are often deprioritized unless tied to product adoption or regulatory compliance. Our findings indicate that current evaluation practices leave major gaps in assessing AI's societal impacts, highlighting the urgent need for policies that promote developer transparency, strengthen independent evaluation ecosystems, and create shared infrastructure to aggregate and compare third-party evaluations in a consistent and accessible way.

Mowafak Allaham, Nicholas Diakopoulos

The growing accessibility of Large Language Models via conversational interfaces capable of responding to users' questions by drawing on, synthesizing, and citing information from the web (i.e., Generative Search Engines) has simplified the information-seeking process for users. However, with the proliferation of AI-generated content on the web, it is unclear whether these engines can reliably omit citing synthetic sources (i.e., AI-generated sources). Should these engines be unable to do so, this puts users at risk of harm by treating information from AI-generated sources synthesized in responses of generative search engines as equivalent to information from authoritative or official sources. In a step towards identifying whether AI-generated sources are being cited by these engines, this work presents an audit of four generative search engines (ChatGPT, Copilot, Gemini, Perplexity) using a total of 712 real-world human-generated queries spanning domains of public importance: politics, health, and the environment. Our findings show evidence of AI-generated sources being cited across all four generative search engines (~16% of cited sources) and identifies key source web domains these sources belong to that are frequently cited across these engines and topics. In addition, we observed that generative search engines include a somewhat narrow set of repeatedly cited domains while predominantly surfacing a large number of minimally cited domains in responses to users' queries. These findings contribute to the growing body of work on assessing the risks of generative search engines with the objective of increasing public awareness of their limitations and encouraging appropriate measures to improve information quality and governance of these systems.

Deepak Akkil, Mowafak Allaham, Amal Raj et al.

Reliable evaluation of AI agents operating in complex, real-world environments requires methodologies that are robust, transparent, and contextually aligned with the tasks agents are intended to perform. This study identifies persistent shortcomings in existing AI agent evaluation practices that are particularly acute in web agent evaluation, as exemplified by our audit of WebVoyager, including task-framing ambiguity and operational variability that hinder meaningful and reproducible performance comparisons. To address these challenges, we introduce Emergence WebVoyager, an enhanced version of the WebVoyager benchmark that standardizes evaluation methodology through clear guidelines for task instantiation, failure handling, annotation, and reporting. Emergence WebVoyager achieves an inter-annotator agreement of 95.9\%, indicating improved clarity and reliability in both task formulation and evaluation. Applying this framework to evaluate OpenAI Operator reveals substantial performance variation across domains and task types, with an overall success rate of 68.6\%, substantially lower than the 87\% previously reported by OpenAI, demonstrating the utility of our approach for more rigorous and comparable web agent evaluation.

Mowafak Allaham, Nicholas Diakopoulos

Gaining insight into the potential negative impacts of emerging Artificial Intelligence (AI) technologies in society is a challenge for implementing anticipatory governance approaches. One approach to produce such insight is to use Large Language Models (LLMs) to support and guide experts in the process of ideating and exploring the range of undesirable consequences of emerging technologies. However, performance evaluations of LLMs for such tasks are still needed, including examining the general quality of generated impacts but also the range of types of impacts produced and resulting biases. In this paper, we demonstrate the potential for generating high-quality and diverse impacts of AI in society by fine-tuning completion models (GPT-3 and Mistral-7B) on a diverse sample of articles from news media and comparing those outputs to the impacts generated by instruction-based (GPT-4 and Mistral-7B-Instruct) models. We examine the generated impacts for coherence, structure, relevance, and plausibility and find that the generated impacts using Mistral-7B, a small open-source model fine-tuned on impacts from the news media, tend to be qualitatively on par with impacts generated using a more capable and larger scale model such as GPT-4. Moreover, we find that impacts produced by instruction-based models had gaps in the production of certain categories of impacts in comparison to fine-tuned models. This research highlights a potential bias in the range of impacts generated by state-of-the-art LLMs and the potential of aligning smaller LLMs on news media as a scalable alternative to generate high quality and more diverse impacts in support of anticipatory governance approaches.

Mowafak Allaham, Kimon Kieslich, Nicholas Diakopoulos

Expert-driven frameworks for impact assessments (IAs) may inadvertently overlook the effects of AI technologies on the public's social behavior, policy, and the cultural and geographical contexts shaping the perception of AI and the impacts around its use. This research explores the potentials of fine-tuning LLMs on negative impacts of AI reported in a diverse sample of articles from 266 news domains spanning 30 countries around the world to incorporate more diversity into IAs. Our findings highlight (1) the potential of fine-tuned open-source LLMs in supporting IA of AI technologies by generating high-quality negative impacts across four qualitative dimensions: coherence, structure, relevance, and plausibility, and (2) the efficacy of small open-source LLM (Mistral-7B) fine-tuned on impacts from news media in capturing a wider range of categories of impacts that GPT-4 had gaps in covering.