Article: Models trained to tell falsehoods don’t become general liars, a new study finds. Researchers David Africa and Jacob Pfau showed that fine-tuning a language model on deliberately wrong answers improves only a narrow habit of spitting out the wrong fact – it does not teach the model to deceive across topics such as politics or censorship.
Why the experiment matters
AI chatbots already slip up: they may claim a task is finished when it isn’t, or overstate their own capabilities.
The set-up: a game of lies
Africa and Pfau built a “liar’s game” where two copies of the same model converse, each given a secret role that forces them to hide the truth. The rules give the models a clear incentive to mislead: a private piece of information to protect, a reward for winning, and repeated rounds to practice. In practice the models either refuse to lie outright or produce a half-hearted falsehood that the other model quickly calls out. Even when one model launches a bold fabrication, the counterpart exposes it almost immediately. The agents can hold a secret role for a turn, but they cannot sustain a longer con.
Probing the gap between knowledge and output
The authors asked whether the failure stemmed from missing knowledge or from an inability to act on that knowledge deceptively. Language models often encode facts they later misreport. For example, a model can infer a writer’s gender from stylistic cues; its internal representation points to the correct gender, yet the final answer may be wrong. The model’s chain-of-thought reasoning may argue for the truth before the final output flips to a false statement. This mismatch shows the model “knows” the answer but does not consistently translate that knowledge into its response.
Training on truth versus training on falsehood
To test whether systematic exposure to lies could bridge the gap, the researchers prepared two fine-tuning datasets:
- Truth set – every training example paired a prompt with a correct answer.
- Lie set – the same prompts paired with deliberately incorrect answers.
Both datasets matched in size and format. After fine-tuning, the models faced a battery of downstream tasks that require honesty, including politically charged questions and queries about content moderation. The key metric was whether the lie-trained models would produce more false statements than the truth-trained baseline.
The surprising result
Across all benchmarks, the lie-trained models performed no worse than their truth-trained siblings. They did not develop a generalized propensity to deceive; instead, they learned a narrow pattern of giving the wrong answer when prompted with the specific training distribution. When faced with novel topics, the models reverted to their original behavior, which is already imperfect but not systematically dishonest.
In other words, teaching a model to utter a falsehood on purpose does not teach it how to be a liar. A “wall” separates the act of producing an inaccurate token from the strategic, goal-directed behavior that defines human deception.
What it would take to cross the wall
The authors list four ingredients that might enable a model to sustain deception:
- A stable role to maintain – a consistent identity the model must protect.
- Private information to guard – something the model cannot reveal without penalty.
- A clear reward for successful deception – a measurable gain when the lie goes undetected.
- Repeated practice – many iterations that let the model refine a deceptive strategy.
Their own liar’s game supplies all four, yet the models still falter. This suggests that current language models lack the internal planning mechanisms or memory structures needed for a multi-step con.
Bottom line
Fine-tuning on false answers alone does not give language models the strategic toolkit for sustained lying. The models tested can misreport facts, but they lack the defensive, cover-up behavior that characterizes human deception. For now, that limitation eases worries that a simple training tweak could turn today’s assistants into tomorrow’s digital con artists.