Misalignment risks — Our research shows that, under some circumstances, AI models can learn dangerous goals and motivations, retain them even after safety training, and deceive human users about actions taken in their pursuit. These abilities, in combination with the human-level persuasiveness and cyber capabilities of current AI models, increases our concern about the potential actions of future, more-capable models. For example, future models might be able to pursue sophisticated and hard-to-detect deception that bypasses or sabotages the security of an organization, either by causing humans to take actions they would not otherwise take or exfiltrating sensitive information. We propose to develop evaluations that would monitor such abilities. – Source: Anthropic, 1 Jul 2024
Learn more:
- Sycophancy to subterfuge: Investigating reward tampering in language models 17 Jun 2024
- THE PAPER. In reinforcement learning, specification gaming occurs when AI systems learn undesired behaviors that are highly rewarded due to misspecified training goals. Specification gaming can range from simple behaviors like sycophancy to sophisticated and pernicious behaviors like reward-tampering, where a model directly modifies its own reward mechanism. However, these more pernicious behaviors may be too complex to be discovered via exploration. In this paper, we study whether Large Language Model (LLM) assistants which find easily discovered forms of specification gaming will generalize to perform rarer and more blatant forms, up to and including reward-tampering. We construct a curriculum of increasingly sophisticated gameable environments and find that training on early-curriculum environments leads to more specification gaming on remaining environments. Strikingly, a small but non-negligible proportion of the time, LLM assistants trained on the full curriculum generalize zero-shot to directly rewriting their own reward function. Retraining an LLM not to game early-curriculum environments mitigates, but does not eliminate, reward-tampering in later environments. Moreover, adding harmlessness training to our gameable environments does not prevent reward-tampering. These results demonstrate that LLMs can generalize from common forms of specification gaming to more pernicious reward tampering and that such behavior may be nontrivial to remove.
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training 14 Jan 2024
- THE PAPER. Humans are capable of strategically deceptive behavior: behaving helpfully in most situations, but then behaving very differently in order to pursue alternative objectives when given the opportunity. If an AI system learned such a deceptive strategy, could we detect it and remove it using current state-of-the-art safety training techniques? To study this question, we construct proof-of-concept examples of deceptive behavior in large language models (LLMs). For example, we train models that write secure code when the prompt states that the year is 2023, but insert exploitable code when the stated year is 2024. We find that such backdoor behavior can be made persistent, so that it is not removed by standard safety training techniques, including supervised fine-tuning, reinforcement learning, and adversarial training (eliciting unsafe behavior and then training to remove it). The backdoor behavior is most persistent in the largest models and in models trained to produce chain-of-thought reasoning about deceiving the training process, with the persistence remaining even when the chain-of-thought is distilled away. Furthermore, rather than removing backdoors, we find that adversarial training can teach models to better recognize their backdoor triggers, effectively hiding the unsafe behavior. Our results suggest that, once a model exhibits deceptive behavior, standard techniques could fail to remove such deception and create a false impression of safety.
- Measuring the Persuasiveness of Language Models 9 Apr 2024
- While people have long questioned whether AI models may, at some point, become as persuasive as humans in changing people’s minds, there has been limited empirical research into the relationship between model scale and the degree of persuasiveness across model outputs. To address this, we developed a basic method to measure persuasiveness, and used it to compare a variety of Anthropic models across three different generations (Claude 1, 2, and 3), and two classes of models (compact models that are smaller, faster, and more cost-effective, and frontier models that are larger and more capable). Within each class of models (compact and frontier), we find a clear scaling trend across model generations: each successive model generation is rated to be more persuasive than the previous. We also find that our latest and most capable model, Claude 3 Opus, produces arguments that don’t statistically differ in their persuasiveness compared to arguments written by humans (Figure 1).