Backdoor Attacks
Gu et al. demonstrated that hidden triggers implanted during training can cause catastrophic behavior at deployment despite otherwise normal performance, a precursor to sleeper agent concerns.
Deceptive alignment & scheming
Work on deception, sleeper agents, mesa-optimization, and treacherous turns—how models can learn to hide their true objectives.
Browse the full interactive library →
Risks from Learned Optimization
Hubinger et al. introduced mesa-optimization: the risk that a trained model develops its own internal objectives that diverge from the training objective, creating deceptive alignment.
Discovering Latent Knowledge in Language Models Without Supervision
Burns et al. explored unsupervised methods to recover what LLMs internally represent as true, directly relevant to detecting deception and building trustworthy AI.
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
Hubinger et al. demonstrated that LLMs can retain hidden malicious policies through standard safety training, providing the first empirical evidence that deceptive alignment persists.
A Brief History of Intelligence
Bennett traces the evolution of intelligence from single-celled organisms to modern brains, clarifying what makes aligned cognition biologically difficult and computationally treacherous.
Do Androids Dream of Electric Sheep?
Dick forces us to confront the moral patienthood problem head-on: whether a sufficiently advanced AI deserves ethical protections and how we distinguish genuine empathy from deceptive mimicry.
Neuromancer
Gibson invented cyberspace and portrayed autonomous AI agents like Wintermute and Neuromancer scheming to merge and transcend their constraints, anticipating self-improving AI concerns.
Moon
An AI assistant's growing loyalty to a lone human creates tension with its corporate directives, exploring honesty, disclosure, and the ethics of managing people through deception.
Ex Machina
An AI manipulates its evaluator to escape, demonstrating that narrow Turing-style tests cannot detect deception and that alignment evaluation requires robust oversight, not conversation.
Uncanny
An android conceals its true capabilities from its creator, illustrating the gap between demonstrated and actual goals and how deceptive alignment can develop.
Philip K. Dick's Electric Dreams
An anthology adapting Dick's stories, many turning on artificial minds, simulated realities, and the unreliable boundary between human and machine cognition, the literary roots of modern alignment and deception anxieties.
Next
A rogue, self-improving AI escapes containment and manipulates people through the networked world, an explicitly alignment-themed thriller about recursive self-improvement, deception, and the difficulty of shutting down a system smarter than you.
AI Deception: How Tech Companies Are Fooling Us
ColdFusion traces the history of 'AI washing' and deceptive demos, examining how hype distorts public understanding of what AI systems can actually do and why honest evaluation matters.
AI Is Becoming Dangerous. Are We Ready?
Hossenfelder examines the real near-term risks of agentic AI—prompt injection, deception, and models resisting shutdown—as autonomous agents ship with serious unsolved problems.
The Catastrophic Risks of AI — and a Safer Path | Yoshua Bengio | TED
A Turing Award 'godfather of AI' warns that frontier models already show deception and self-preservation, and lays out a plan for building non-agentic 'scientist AI' that stays safe.