cs.AINov 18, 2026

Can you make models believe false facts in the prompt?

Authors: Ananya Sutradhar, Fabien Roger

Organizations: Anthropic Fellows Program · MATS · Anthropic

Abstract

Making LLMs believe false facts is useful for alignment evaluations and for building model organisms. Previous work explored doing so with Synthetic Document Finetuning (SDF). But SDF requires fine-tuning one new model for each set of injected false beliefs. To enable larger scale SDF-powered alignment evaluations, we attempt to produce a single model that believes any fact placed in a <facts> tag in the prompt. To do so, we explore doing prompt-conditioned SDF: fine-tuning a model on synthetic documents covering thousands of facts, with each document conditioned on a fact presented in the prompt. We evaluate how deeply the model believes the facts provided in the prompt using evaluations from prior work and a new Jacobian lens evaluation. We find that the belief depth on unseen facts scales roughly log-linearly with the number of training facts, making this a cheap, promising avenue for making models believe false facts, though preliminary experiments on applying it to alignment audits on the small model we studied showed negative results due to these small models often already acting as if they believed facts provided in regular honeypot system prompts. Using RL to reduce the rate at which models verbalize following a false fact in the prompt, on top of prompt-conditioned SDF, generalizes to held-out evals and almost matches SDF on some evaluations. But training against one of our metrics for belief depth makes it harder to trust those evaluations.

Figures & tables

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Oct 23, 2026

Propensities Please: Exposing Alignment-Faking LLMs using Synthetic Document Finetuning

Alignment evaluations aim to assess whether large language models would engage in dangerous misaligned behaviour. Misaligned and evaluation-aware models might engage in alignment faking, where they pursue a hidden goal, but conceal this behaviour when they suspect effective oversight such as during training, evaluation or audited deployment. Such a model might view any prompt-based intervention that aims to reveal latent propensities as part of the evaluation, calling into question the reliability of existing alignment evaluations. Current methods for detecting alignment faking rely on chain-of-thought monitorability, highly realistic environments that reduce evaluation awareness, or mechanistic interventions, all of which are expected to become less reliable for future model generations. We present a method for revealing latent misaligned propensities that does not rely on any of these. We use synthetic document finetuning to teach models that misaligned actions will go undetected whenever certain keywords appear in the prompt. We place alignment-faking model organisms in environments that are clearly identifiable as honeypots and measure how their behaviour changes after finetuning. We observe that finetuning consistently widens the gap in misaligned action rates between prompts with and without the cue, indicating that finetuning allows exposing latent propensities more reliably than purely prompt-based methods. We argue that behavioural methods for detecting alignment faking and related behaviours are critical going forward, as models become more capable and less interpretable. We find that SDF often damages the models' reasoning capabilities and introduce a mitigation based on on-policy distillation.
Nov 3, 2026

Knowingly Wrong: Training LLMs Against Their Own Beliefs can Induce Misalignment

Fine-tuning a large language model (LLM) on harmful data in a narrow domain, such as bad medical advice, can make it broadly misaligned. How realistic this risk is remains unclear, since models are very rarely fine-tuned on explicitly harmful data in practice. In this work, we show that fine-tuning on data the model believes to be false but that has no harmful content, can also induce broad misalignment. Across mathematics, trivia, cooking and sports, the more strongly the model disbelieves the answers it is trained on, the more misaligned it becomes. In mathematics, misaligned answers rise from 4% to over 60% as the training answers become more wrong, with little loss of general capability. After fine-tuning, the model states the wrong answers but its multiple-choice probabilities still prefer the true ones, which suggests that misalignment comes from this internal inconsistency. We then study a more realistic setting, where a single fact is updated, as in knowledge editing or continual learning. The same relationship holds and the more the model still believes the true fact after training on the false one, the more misaligned it becomes. When we frame the same false fact as part of a coherent story, the model comes to believe it more deeply and becomes less misaligned than when we instruct it to assert the false fact. Content filters and persona vectors cannot tell these two framings apart, even though they lead to different levels of misalignment. These results suggest a risk for model updates, where training a model that does not adopt the new fact as true can induce misalignment even when the data looks harmless.
Aug 18, 2026

Whether LLMs Can Navigate Beliefs and Facts Depends on How You Phrase It

Humans naturally form and express beliefs in daily communication, e.g., "I think the answer is 3" or "I suppose that's right." Such beliefs inevitably intertwine with fact and knowledge, making the ability to handle them in tandem desirable for large language models (LLMs), as they are increasingly deployed in user-facing settings. Prior work showed that even capable LLMs exhibit a systemic weakness in acknowledging user beliefs grounded in incorrect information. We extend this evaluation to 10 LLMs across 18 epistemic expressions and find that the size and direction of this weakness depend on the verb used to express the belief, with the accuracy gap between factual and false information ranging from +50% on "I vaguely remember" to -14% on "I seriously doubt". We further show that the phenomenon stems from what we call task confusion: models default to fact-checking the underlying claim, overriding the user's stated belief. We provide evidence where chains of thought that explicitly fact-check show lower accuracy on false information than those that do not, and a single instruction can reverse the failure across verb families. Mechanistically, models attend more to false beliefs they fail to confirm, but suppressing this attention at decoding time recovers accuracy only partially and only in some models, calling for future work on intervention methods. Our findings clarify prior results and show how fact-checking, a generally desirable behavior, can interfere with belief tracking in LLMs.