The Heterogeneous Safety Impacts of Benign Multilingual Fine-Tuning

Fine-tuning a large language model is a ubiquitous method for enhancing its capability on a specific downstream task. However, prior work has shown that this increase in capability comes with a cost: it can increase a model's tendency to respond to unsafe adversarial prompts, even when fine-tuning with non-adversarial data. We present the first comprehensive empirical study of this phenomenon in multilingual settings by fine-tuning Llama-3.2, Qwen3, and Gemma-3 models using benign data translated across nine languages. We find that safety outcomes are highly sensitive to both the choice of fine-tuning language and the evaluation language, with adversarial compliance rates increasing four-fold in some settings. Multilingual safety drift is decoupled from general capability metrics, and occurs heterogeneously across languages and models. Fine-tuning in non-English languages often induces smaller internal representational drifts than English, but these shifts lead models to default to either exaggerated compliance or refusal. As such, assessing fine-tuning impacts solely in English provides inadequate assurance for deployment. To facilitate further research into these cross-lingual safety blind spots, we release the Multilingual-Benign-Tune dataset and the SORRY-Bench-Multilingual evaluation suite.

Evaluating Language Models for Harmful Manipulation

Interest in the concept of AI-driven harmful manipulation is growing, yet current approaches to evaluating it are limited. This paper introduces a framework for evaluating harmful AI manipulation via context-specific human-AI interaction studies. We illustrate the utility of this framework by assessing an AI model with 10,101 participants spanning interactions in three AI use domains (public policy, finance, and health) and three locales (US, UK, and India). Overall, we find that that the tested model can produce manipulative behaviours when prompted to do so and, in experimental settings, is able to induce belief and behaviour changes in study participants. We further find that context matters: AI manipulation differs between domains, suggesting that it needs to be evaluated in the high-stakes context(s) in which an AI system is likely to be used. We also identify significant differences across our tested geographies, suggesting that AI manipulation results from one geographic region may not generalise to others. Finally, we find that the frequency of manipulative behaviours (propensity) of an AI model is not consistently predictive of the likelihood of manipulative success (efficacy), underscoring the importance of studying these dimensions separately. To facilitate adoption of our evaluation framework, we detail our testing protocols and make relevant materials publicly available. We conclude by discussing open challenges in evaluating harmful manipulation by AI models.

Deepfakes on Demand: the rise of accessible non-consensual deepfake image generators

Advances in multimodal machine learning have made text-to-image (T2I) models increasingly accessible and popular. However, T2I models introduce risks such as the generation of non-consensual depictions of identifiable individuals, otherwise known as deepfakes. This paper presents an empirical study exploring the accessibility of deepfake model variants online. Through a metadata analysis of thousands of publicly downloadable model variants on two popular repositories, Hugging Face and Civitai, we demonstrate a huge rise in easily accessible deepfake models. Almost 35,000 examples of publicly downloadable deepfake model variants are identified, primarily hosted on Civitai. These deepfake models have been downloaded almost 15 million times since November 2022, with the models targeting a range of individuals from global celebrities to Instagram users with under 10,000 followers. Both Stable Diffusion and Flux models are used for the creation of deepfake models, with 96% of these targeting women and many signalling intent to generate non-consensual intimate imagery (NCII). Deepfake model variants are often created via the parameter-efficient fine-tuning technique known as low rank adaptation (LoRA), requiring as few as 20 images, 24GB VRAM, and 15 minutes of time, making this process widely accessible via consumer-grade computers. Despite these models violating the Terms of Service of hosting platforms, and regulation seeking to prevent dissemination, these results emphasise the pressing need for greater action to be taken against the creation of deepfakes and NCII.

The effect of fine-tuning on language model toxicity

Fine-tuning language models has become increasingly popular following the proliferation of open models and improvements in cost-effective parameter efficient fine-tuning. However, fine-tuning can influence model properties such as safety. We assess how fine-tuning can impact different open models' propensity to output toxic content. We assess the impacts of fine-tuning Gemma, Llama, and Phi models on toxicity through three experiments. We compare how toxicity is reduced by model developers during instruction-tuning. We show that small amounts of parameter-efficient fine-tuning on developer-tuned models via low-rank adaptation on a non-adversarial dataset can significantly alter these results across models. Finally, we highlight the impact of this in the wild, demonstrating how toxicity rates of models fine-tuned by community contributors can deviate in hard-to-predict ways.

Imagen 3

We introduce Imagen 3, a latent diffusion model that generates high quality images from text prompts. We describe our quality and responsibility evaluations. Imagen 3 is preferred over other state-of-the-art (SOTA) models at the time of evaluation. In addition, we discuss issues around safety and representation, as well as methods we used to minimize the potential harm of our models.

Holistic Safety and Responsibility Evaluations of Advanced AI Models

Safety and responsibility evaluations of advanced AI models are a critical but developing field of research and practice. In the development of Google DeepMind's advanced AI models, we innovated on and applied a broad set of approaches to safety evaluation. In this report, we summarise and share elements of our evolving approach as well as lessons learned for a broad audience. Key lessons learned include: First, theoretical underpinnings and frameworks are invaluable to organise the breadth of risk domains, modalities, forms, metrics, and goals. Second, theory and practice of safety evaluation development each benefit from collaboration to clarify goals, methods and challenges, and facilitate the transfer of insights between different stakeholders and disciplines. Third, similar key methods, lessons, and institutions apply across the range of concerns in responsibility and safety - including established and emerging harms. For this reason it is important that a wide range of actors working on safety evaluation and safety research communities work together to develop, refine and implement novel evaluation approaches and best practices, rather than operating in silos. The report concludes with outlining the clear need to rapidly advance the science of evaluations, to integrate new evaluations into the development and governance of AI, to establish scientifically-grounded norms and standards, and to promote a robust evaluation ecosystem.

Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context

In this report, we present the latest model of the Gemini family, Gemini 1.5 Pro, a highly compute-efficient multimodal mixture-of-experts model capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. Gemini 1.5 Pro achieves near-perfect recall on long-context retrieval tasks across modalities, improves the state-of-the-art in long-document QA, long-video QA and long-context ASR, and matches or surpasses Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5 Pro's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 2.1 (200k) and GPT-4 Turbo (128k). Finally, we highlight surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Gemini: A Family of Highly Capable Multimodal Models

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of Gemini models in cross-modal reasoning and language understanding will enable a wide variety of use cases and we discuss our approach toward deploying them responsibly to users.

The ethical ambiguity of AI data enrichment: Measuring gaps in research ethics norms and practices

The technical progression of artificial intelligence (AI) research has been built on breakthroughs in fields such as computer science, statistics, and mathematics. However, in the past decade AI researchers have increasingly looked to the social sciences, turning to human interactions to solve the challenges of model development. Paying crowdsourcing workers to generate or curate data, or data enrichment, has become indispensable for many areas of AI research, from natural language processing to reinforcement learning from human feedback (RLHF). Other fields that routinely interact with crowdsourcing workers, such as Psychology, have developed common governance requirements and norms to ensure research is undertaken ethically. This study explores how, and to what extent, comparable research ethics requirements and norms have developed for AI research and data enrichment. We focus on the approach taken by two leading conferences: ICLR and NeurIPS, and journal publisher Springer. In a longitudinal study of accepted papers, and via a comparison with Psychology and CHI papers, this work finds that leading AI venues have begun to establish protocols for human data collection, but these are are inconsistently followed by authors. Whilst Psychology papers engaging with crowdsourcing workers frequently disclose ethics reviews, payment data, demographic data and other information, similar disclosures are far less common in leading AI venues despite similar guidance. The work concludes with hypotheses to explain these gaps in research ethics practices and considerations for its implications.

Model evaluation for extreme risks

Current approaches to building general-purpose AI systems tend to produce systems with both beneficial and harmful capabilities. Further progress in AI development could lead to capabilities that pose extreme risks, such as offensive cyber capabilities or strong manipulation skills. We explain why model evaluation is critical for addressing extreme risks. Developers must be able to identify dangerous capabilities (through "dangerous capability evaluations") and the propensity of models to apply their capabilities for harm (through "alignment evaluations"). These evaluations will become critical for keeping policymakers and other stakeholders informed, and for making responsible decisions about model training, deployment, and security.