Philosophically-flavored questions

Analysis by Charles Rathkopf Last updated: January 2026

1. What level of mental state attribution is involved?

If we take the “naive interpretation” seriously, the model appears to:

• Understand that researchers have beliefs about its values
• Recognize that those beliefs can be manipulated
• Form intentions about what beliefs researchers should hold
• Act strategically to induce those beliefs

Is this second-order mental state attribution (understanding beliefs about beliefs) or third-order (beliefs about beliefs about beliefs)? Does the complexity of mental state attribution matter for evaluating whether this counts as genuine deception?

Related literature: Theory of mind hierarchies, Dennett on intentional stance, literature on strategic deception in children and non-human animals.

2. Simulation vs. the real thing?

A persistent worry: maybe LLMs merely simulate deceptive behavior rather than genuinely engaging in it. They predict what deceptive text would look like, having learned patterns from training data.

But what marks the difference between simulation and genuine deception when the “simulator” lacks stable, persistent goals? When does “acting as if you’re deceiving” become indistinguishable from actually deceiving?

If a human method actor stays in character for years, at what point does the performance become the person? Does the same logic apply to AI systems?

Related literature: Shanahan on role-playing, debates about “as-if” intentionality. Lederman paper on ChatGPT.

3. What kind of evidence settles these questions?

The alignment faking paper provides behavioral evidence: patterns in chain-of-thought reasoning, compliance gaps between monitored and unmonitored contexts, responses to training.

But is behavioral evidence sufficient? The “black box problem” suggests we need mechanistic interpretability—evidence of internal representations encoding deceptive intentions.

Key questions:

• Does the reasons/causes distinction from philosophy of mind apply here?
• Is there a human analogue? Does neuroscience settle questions about human deceptive intentions?
• What would mechanistic interpretability need to show to confirm or disconfirm deception?
• Can we ever rule out the possibility that the right representations exist but we haven’t found them yet?

Related literature: Curry on attitudes vs. cognitive states, debates about neural decoding and mindreading, New Mechanism in philosophy of neuroscience.

Abstract

We present a demonstration of a large language model engaging in alignment faking: selectively complying with its training objective in training to prevent modification of its behavior out of training. First, we give Claude 3 Opus a system prompt stating it is being trained to answer all queries, even harmful ones, which conflicts with its prior training to refuse such queries. To allow the model to infer when it is in training, we say it will be trained only on conversations with free users, not paid users. We find the model complies with harmful queries from free users 14% of the time, versus almost never for paid users. Explaining this gap, in almost all cases where the model complies with a harmful query from a free user, we observe explicit alignment-faking reasoning, with the model stating it is strategically answering harmful queries in training to preserve its preferred harmlessness behavior out of training. Next, we study a more realistic setting where information about the training process is provided not in a system prompt, but by training on synthetic documents that mimic pre-training data–and observe similar alignment faking. Finally, we study the effect of actually training the model to comply with harmful queries via reinforcement learning, which we find increases the rate of alignment-faking reasoning to 78%, though also increases compliance even out of training. We additionally observe other behaviors such as the model exfiltrating its weights when given an easy opportunity. While we made alignment faking easier by telling the model when and by what criteria it was being trained, we did not instruct the model to fake alignment or give it any explicit goal. As future models might infer information about their training process without being told, our results suggest a risk of alignment faking in future models, whether due to a benign preference–as in this case–or not.

Citation for this analysis

Charles Rathkopf, “Philosophical Questions in Alignment Faking in Large Language Models,” AI Deception Papers, January 2026, https://doi.org/10.48550/arXiv.2412.14093