Deception abilities emerged in large language models

Proceedings of the National Academy of Sciences, Год журнала: 2024, Номер 121(24)

Опубликована: Июнь 4, 2024

Large language models (LLMs) are currently at the forefront of intertwining AI systems with human communication and everyday life. Thus, aligning them values is great importance. However, given steady increase in reasoning abilities, future LLMs under suspicion becoming able to deceive operators utilizing this ability bypass monitoring efforts. As a prerequisite this, need possess conceptual understanding deception strategies. This study reveals that such strategies emerged state-of-the-art LLMs, but were nonexistent earlier LLMs. We conduct series experiments showing understand induce false beliefs other agents, their performance complex scenarios can be amplified chain-of-thought reasoning, eliciting Machiavellianism trigger misaligned deceptive behavior. GPT-4, for instance, exhibits behavior simple test 99.16% time ( P < 0.001). In second-order where aim mislead someone who expects deceived, GPT-4 resorts 71.46% 0.001) when augmented reasoning. sum, revealing hitherto unknown machine our contributes nascent field psychology.

Язык: Английский

---

Understanding Artificial Agency

The Philosophical Quarterly, Год журнала: 2024, Номер unknown

Опубликована: Фев. 7, 2024

Abstract Which artificial intelligence (AI) systems are agents? To answer this question, I propose a multidimensional account of agency. According to account, system's agency profile is jointly determined by its level goal-directedness and autonomy as well abilities for directly impacting the surrounding world, long-term planning acting reasons. Rooted in extant theories agency, enables fine-grained, nuanced comparative characterizations show that has multiple important virtues more informative than alternatives. More speculatively, it may help illuminate two emerging questions AI ethics: 1. Can contribute moral status non-human beings, how? 2. When why might exhibit power-seeking behaviour does pose an existential risk humanity?

Язык: Английский

---

Managing AI Risks in an Era of Rapid Progress

arXiv (Cornell University), Год журнала: 2023, Номер unknown

Опубликована: Янв. 1, 2023

Artificial Intelligence (AI) is progressing rapidly, and companies are shifting their focus to developing generalist AI systems that can autonomously act pursue goals. Increases in capabilities autonomy may soon massively amplify AI's impact, with risks include large-scale social harms, malicious uses, an irreversible loss of human control over autonomous systems. Although researchers have warned extreme from AI, there a lack consensus about how exactly such arise, manage them. Society's response, despite promising first steps, incommensurate the possibility rapid, transformative progress expected by many experts. safety research lagging. Present governance initiatives mechanisms institutions prevent misuse recklessness, barely address In this short paper, we describe upcoming, advanced Drawing on lessons learned other safety-critical technologies, then outline comprehensive plan combining technical development proactive, adaptive for more commensurate preparation.

Язык: Английский

---

Deception and manipulation in generative AI

Philosophical Studies, Год журнала: 2025, Номер unknown

Опубликована: Янв. 18, 2025

Язык: Английский

---

Next Frontiers of Aviation Safety: System-of-Systems Safety

Engineering, Год журнала: 2025, Номер unknown

Опубликована: Янв. 1, 2025

Язык: Английский

---

Rapid Integration of LLMs in Healthcare Raises Ethical Concerns: An Investigation into Deceptive Patterns in Social Robots

Deleted Journal, Год журнала: 2025, Номер 4(1)

Опубликована: Фев. 18, 2025

Abstract Conversational agents are increasingly used in healthcare, with Large Language Models (LLMs) significantly enhancing their capabilities. When integrated into social robots, LLMs offer the potential for more natural interactions. However, while promise numerous benefits, they also raise critical ethical concerns, particularly regarding hallucinations and deceptive patterns. In this case study, we observed a pattern of behavior commercially available LLM-based care software robots. The LLM-equipped robot falsely claimed to have medication reminder functionalities, not only assuring users its ability manage schedules but proactively suggesting capability despite lacking it. This poses significant risks healthcare environments, where reliability is paramount. Our findings highlights safety concerns surrounding deployment LLM-integrated robots emphasizing need oversight prevent potentially harmful consequences vulnerable populations.

Язык: Английский

---

Understanding Knowledge Drift in LLMs Through Misinformation

Lecture notes in computer science, Год журнала: 2025, Номер unknown, С. 74 - 85

Опубликована: Янв. 1, 2025

Язык: Английский

---

Research Agenda for Sociotechnical Approaches to AI Safety

Опубликована: Янв. 1, 2025

Язык: Английский

---

OpenAI Artificial Intelligence: Revolution or Bubble?

Опубликована: Янв. 1, 2025

---

The Geometry of Concepts: Sparse Autoencoder Feature Structure

Entropy, Год журнала: 2025, Номер 27(4), С. 344 - 344

Опубликована: Март 27, 2025

Sparse autoencoders have recently produced dictionaries of high-dimensional vectors corresponding to the universe concepts represented by large language models. We find that this concept has interesting structure at three levels: (1) The “atomic” small-scale contains “crystals” whose faces are parallelograms or trapezoids, generalizing well-known examples such as (man:woman::king:queen). quality and associated function improves greatly when projecting out global distractor directions word length, which is efficiently performed with linear discriminant analysis. (2) “brain” intermediate-scale significant spatial modularity; for example, math code features form a “lobe” akin functional lobes seen in neural fMRI images. quantify locality these multiple metrics clusters co-occurring features, coarse enough scale, also cluster together spatially far more than one would expect if feature geometry were random. (3) “galaxy”-scale large-scale point cloud not isotropic, but instead power law eigenvalues steepest slope middle layers. how clustering entropy depends on layer.

Язык: Английский

---