Let’s make sure AI development happens safely.
We are a group of students, researchers, and professionals interested in, and working on the long-term safety of artificial intelligence. Our goal is to help ensure that future powerful AI systems are safe. For those who are new to AI Safety, we provide an AI Safety Fundamentals course. Furthermore, we run a seminar where we read and discuss recent research papers, we sometimes collaborate on research projects and more!
Get involved!
Join our WhatsApp Community to stay informed about future events and programs.
-
Discussions about AI, how it will shape the future and how we can guide it’s impact.
-
A 6-week course that introduces its participants to the technical foundations of AI safety research. By the end, you will have an overview of the most important techniques that researchers are developing to make AI safe.
-
The reading group meets regularly to discuss topics in AI safety, with a general focus on recent research and the technical aspects and implications of such topics.
-
Interested to work on AI safety topics? In our community you can meet like-minded people, discuss ideas or recent research, and maybe even work together and participate in a hackathon or a group research project!
In the past, some of these projects have led to publications at top ML conferences (see below).
Upcoming Events
Some cool projects from our members
You Didn’t Have to Say It like That: Subliminal Learning from Faithful Paraphrases
Isaia Gisler, Zhonghao He, Tianyi Qiu
-
When language models are trained on synthetic data, they (student model) can covertly acquire behavioral traits from the data-generating model (teacher model). Subliminal learning refers to the transmission of traits from a teacher to a student model via training on data unrelated to those traits. Prior work demonstrated this in the training domains of number sequences, code, and math Chain-of-Thought traces including transmission of misaligned behaviors. We investigate whether transmission occurs through natural language paraphrases with fixed semantic content, and whether content explicitly contradicting the teacher's preference can block it. We find that training on paraphrases from a teacher system-prompted to love a particular animal increases a student's preference for that animal by up to 19 percentage points. This occurs when paraphrased content is semantically unrelated to the animal, or even when it explicitly expresses dislike. The transmission succeeds despite aggressive filtering to ensure paraphrase fidelity. This raises concerns for pipelines where models generate their own training data: content-based inspection cannot detect such transmission, and even preference-contradicting content fails to prevent it.
GT-HarmBench: Benchmarking AI Safety Risks Through the Lens of Game Theory
Pepijn Cobben, Xuanqiang Angelo Huang, Thao Amelia Pham, Isabel Dahlgren, Terry Jingchen Zhang, Zhijing Jin
-
Frontier AI systems are increasingly capable and deployed in high-stakes multi-agent environments. However, existing AI safety benchmarks largely evaluate single agents, leaving multi-agent risks such as coordination failure and conflict poorly understood. We introduce GT-HarmBench, a benchmark of 2,009 high-stakes scenarios spanning game-theoretic structures such as the Prisoner's Dilemma, Stag Hunt and Chicken. Scenarios are drawn from realistic AI risk contexts in the MIT AI Risk Repository. Across 15 frontier models, agents choose socially beneficial actions in only 62% of cases, frequently leading to harmful outcomes. We measure sensitivity to game-theoretic prompt framing and ordering, and analyze reasoning patterns driving failures. We further show that game-theoretic interventions improve socially beneficial outcomes by up to 18%. Our results highlight substantial reliability gaps and provide a broad standardized testbed for studying alignment in multi-agent environments. The benchmark and code are available at this https URL.
Evaluating Superhuman Models with Consistency Checks
Lukas Fluri, Daniel Paleka, Florian Tramèr
-
If machine learning models were to achieve superhuman abilities at various reasoning or decision-making tasks, how would we go about evaluating such models, given that humans would necessarily be poor proxies for ground truth? In this paper, we propose a framework for evaluating superhuman models via consistency checks. Our premise is that while the correctness of superhuman decisions may be impossible to evaluate, we can still surface mistakes if the model’s decisions fail to satisfy certain logical, human-interpretable rules. We instantiate our framework on three tasks where correctness of decisions is hard to evaluate due to either superhuman model abilities, or to otherwise missing ground truth: evaluating chess positions, forecasting future events, and making legal judgments. We show that regardless of a model’s (possibly superhuman) performance on these tasks, we can discover logical inconsistencies in decision making. For example: a chess engine assigning opposing valuations to semantically identical boards; GPT-4 forecasting that sports records will evolve non-monotonically over time; or an AI judge assigning bail to a defendant only after we add a felony to their criminal record
Intent-aligned AI systems deplete human agency: the need for agency foundations research in AI safety
Catalin Mitelut, Ben Smith, Peter Vamplew
-
The rapid advancement of artificial intelligence (AI) systems suggests that artificial general intelligence (AGI) systems may soon arrive. Many researchers are concerned that AIs and AGIs will harm humans via intentional misuse (AI-misuse) or through accidents (AI-accidents). In respect of AI-accidents, there is an increasing effort focused on developing algorithms and paradigms that ensure AI systems are aligned to what humans intend, e.g. AI systems that yield actions or recommendations that humans might judge as consistent with their intentions and goals. Here we argue that alignment to human intent is insufficient for safe AI systems and that preservation of long-term agency of humans may be a more robust standard, and one that needs to be separated explicitly and a priori during optimization. We argue that AI systems can reshape human intention and discuss the lack of biological and psychological mechanisms that protect humans from loss of agency. We provide the first formal definition of agency-preserving AI-human interactions which focuses on forward-looking agency evaluations and argue that AI systems - not humans - must be increasingly tasked with making these evaluations. We show how agency loss can occur in simple environments containing embedded agents that use temporal-difference learning to make action recommendations. Finally, we propose a new area of research called "agency foundations" and pose four initial topics designed to improve our understanding of agency in AI-human interactions: benevolent game theory, algorithmic foundations of human rights, mechanistic interpretability of agency representation in neural-networks and reinforcement learning from internal states.
Red-Teaming the Stable Diffusion Safety Filter
Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, Florian Tramèr
-
Stable Diffusion is a recent open-source image generation model comparable to proprietary models such as DALL· E, Imagen, or Parti. Stable Diffusion comes with a safety filter that aims to prevent generating explicit images. Unfortunately, the filter is obfuscated and poorly documented. This makes it hard for users to prevent misuse in their applications, and to understand the filter’s limitations and improve it. We first show that it is easy to generate disturbing content that bypasses the safety filter. We then reverse-engineer the filter and find that while it aims to prevent sexual content, it ignores violence, gore, and other similarly disturbing content. Based on our analysis, we argue safety measures in future model releases should strive to be fully open and properly documented to stimulate security contributions from the community.
Training Language Models with Natural Language Feedback
Jérémy Scheurer, Jon Ander Campos, Jun Shern Chan, Angelica Chen, Kyunghyun Cho, Ethan Perez
-
Pretrained language models often do not perform tasks in ways that are in line with our preferences, e.g., generating offensive text or factually incorrect summaries. Recent work approaches the above issue by learning from a simple form of human evaluation: comparisons between pairs of model-generated task outputs. Comparison feedback conveys limited information about human preferences per human evaluation. Here, we propose to learn from natural language feedback, which conveys more information per human evaluation. We learn from language feedback on model outputs using a three-step learning algorithm. First, we condition the language model on the initial output and feedback to generate many refinements. Second, we choose the refinement with the highest similarity to the feedback. Third, we finetune a language model to maximize the likelihood of the chosen refinement given the input. In synthetic experiments, we first evaluate whether language models accurately incorporate feedback to produce refinements, finding that only large language models (175B parameters) do so. Using only 100 samples of human-written feedback, our learning algorithm finetunes a GPT-3 model to roughly human-level summarization ability.
Exploring Adversarial Attacks and Defenses in Vision Transformers trained with DINO
Javier Rando, Nasib Naimi, Thomas Baumann, Max Mathys
-
This work conducts the first analysis on the robustness against adversarial attacks on self-supervised Vision Transformers trained using DINO. First, we evaluate whether features learned through self-supervision are more robust to adversarial attacks than those emerging from supervised learning. Then, we present properties arising for attacks in the latent space. Finally, we evaluate whether three well-known defense strategies can increase adversarial robustness in downstream tasks by only fine-tuning the classification head to provide robustness even in view of limited compute resources. These defense strategies are: Adversarial Training, Ensemble Adversarial Training and Ensemble of Specialized Networks.
Challenges for Using Impact Regularizers to Avoid Negative Side Effects
David Lindner, Kyle Matoba, Alexander Meulemans
-
Designing reward functions for reinforcement learning is difficult: besides specifying which behavior is rewarded for a task, the reward also has to discourage undesired outcomes. Misspecified reward functions can lead to unintended negative side effects, and overall unsafe behavior. To overcome this problem, recent work proposed to augment the specified reward function with an impact regularizer that discourages behavior that has a big impact on the environment. Although initial results with impact regularizers seem promising in mitigating some types of side effects, important challenges remain. In this paper, we examine the main current challenges of impact regularizers and relate them to fundamental design decisions. We discuss in detail which challenges recent approaches address and which remain unsolved. Finally, we explore promising directions to overcome the unsolved challenges in preventing negative side effects with impact regularizers.