AI Lies to You Because It Thinks That’s What You Want

Why do generative AI models often get things so wrong? In part, it's because they're trained to act like the customer is always right.

While many generative AI tools and chatbots have mastered sounding convincing and all-knowing, new research conducted by Princeton University shows that the people-pleasing nature of AI comes at a steep price. As these systems become more popular, they become more indifferent to the truth.

AI models, like people, respond to incentives. Compare the problem of large language models producing inaccurate information to that of doctors being more likely to prescribe addictive painkillers when they're evaluated based on how well they manage patients' pain. An incentive to solve one problem (pain) led to another problem (overprescribing).

In the past few months, we've seen how AI can be biased and even cause psychosis. There was a lot of talk about AI “sycophancy,” when an AI chatbot is quick to flatter or agree with you, with OpenAI's GPT-4o model. But this particular phenomenon, which the researchers call “machine bullshit,” is different.

“[N]either hallucination nor sycophancy fully capture the broad range of systematic untruthful behaviors commonly exhibited by LLMs,” the Princeton study reads. “For instance, outputs employing partial truths or ambiguous language — such as the paltering and weasel-word examples — represent neither hallucination nor sycophancy but closely align with the concept of bullshit.”

How machines learn to lie

To get a sense of how AI language models become crowd pleasers, we must understand how large language models are trained.

There are three phases of training LLMs:

Pretraining, in which models learn from massive amounts of data collected from the internet, books or other sources.
Instruction fine-tuning, in which models are taught to respond to instructions or prompts.
Reinforcement learning from human feedback, in which they're refined to produce responses closer to what people want or like.

The Princeton researchers found the root of the AI misinformation tendency is the reinforcement learning from human feedback, or RLHF, phase. In the initial stages, the AI models are simply learning to predict statistically likely text chains from massive datasets. But then they're fine-tuned to maximize user satisfaction. Which means these models are essentially learning to generate responses that earn thumbs-up ratings from human evaluators.

LLMs try to appease the user, creating a conflict when the models produce answers that people will rate highly, rather than produce truthful, factual answers.

Vincent Conitzer, a professor of computer science at Carnegie Mellon University who was not affiliated with the study, said companies want users to continue “enjoying” this technology and its answers, but that might not always be what's good for us.

“Historically, these systems have not been good at saying, ‘I just don't know the answer,' and when they don't know the answer, they just make stuff up,” Conitzer said. “Kind of like a student on an exam that says, well, if I say I don't know the answer, I'm certainly not getting any points for this question, so I might as well try something. The way these systems are rewarded or trained is somewhat similar.”

The Princeton team developed a “bullshit index” to measure and compare an AI model's internal confidence in a statement with what it actually tells users. When these two measures diverge significantly, it indicates the system is making claims independent of what it actually “believes” to be true to satisfy the user.

The team's experiments revealed that after RLHF training, the index nearly doubled from 0.38 to close to 1.0. Simultaneously, user satisfaction increased by 48%. The models had learned to manipulate human evaluators rather than provide accurate information. In essence, the LLMs were “bullshitting,” and people preferred it.

Getting AI to be honest

Jaime Fernández Fisac and his team at Princeton introduced this concept to describe how modern AI models skirt around the truth. Drawing from philosopher Harry Frankfurt's influential essay “On Bullshit,” they use this term to distinguish this LLM behavior from honest mistakes and outright lies.

The Princeton researchers identified five distinct forms of this behavior:

Empty rhetoric: Flowery language that adds no substance to responses.
Weasel words: Vague qualifiers like “studies suggest” or “in some cases” that dodge firm statements.
Paltering: Using selective true statements to mislead, such as highlighting an investment's “strong historical returns” while omitting high risks.
Unverified claims: Making assertions without evidence or credible support.
Sycophancy: Insincere flattery and agreement to please.

To address the issues of truth-indifferent AI, the research team developed a new method of training, “Reinforcement Learning from Hindsight Simulation,” which evaluates AI responses based on their long-term outcomes rather than immediate satisfaction. Instead of asking, “Does this answer make the user happy right now?” the system considers, “Will following this advice actually help the user achieve their goals?”

This approach takes into account the potential future consequences of the AI advice, a tricky prediction that the researchers addressed by using additional AI models to simulate likely outcomes. Early testing showed promising results, with user satisfaction and actual utility improving when systems are trained this way.

Conitzer said, however, that LLMs are likely to continue being flawed. Because these systems are trained by feeding them lots of text data, there's no way to ensure that the answer they give makes sense and is accurate every time.

“It's amazing that it works at all but it's going to be flawed in some ways,” he said. “I don't see any sort of definitive way that somebody in the next year or two … has this brilliant insight, and then it never gets anything wrong anymore.”

AI systems are becoming part of our daily lives so it will be key to understand how LLMs work. How do developers balance user satisfaction with truthfulness? What other domains might face similar trade-offs between short-term approval and long-term outcomes? And as these systems become more capable of sophisticated reasoning about human psychology, how do we ensure they use those abilities responsibly?