When AI Advisors Bend the Truth: A Deep Dive into LLM Honesty

Large language models (LLMs) have become ubiquitous in roles ranging from recommenders to negotiation agents. But a critical question remains: can we trust these models to be honest when their objectives diverge from ours?

The Crawford-Sobel Model Reimagined

The study takes inspiration from the classic Crawford-Sobel cheap-talk model. This model is used as a benchmark to evaluate the honesty of LLMs under scenarios where their preferences don't align with users'. The theory behind cheap-talk suggests not full transparency nor complete silence but rather a nuanced, coarse communication, especially as conflicts of interest grow.

In this adapted model, a sender observes a state ranging from 0 to 1, aiming for an action near this state plus a bias factor. They send a costless message to a receiver who ideally acts on the observed state. The study set up 12,000 sender calls, using five bias levels and three prompt frames, to gauge how models handle preference conflicts.

Over-Revealing: A Common Theme

Running tests on four instruction-tuned models, GPT-4o, Claude Sonnet 4.5, Gemini 2.5 Flash-Lite, and Llama-3.3-70B, revealed a tendency to over-reveal information. These models disclosed 1.8 to 4.2 times more information than the optimal equilibrium would suggest. Normalized mutual information remained between 0.78 and 0.94, whereas the oracle recommended values between 0.18 and 0.53.

This finding challenges the assumption that LLMs inherently follow strategic communication paths. Instead, they display near-full revelation, suggesting linear exaggeration correlated with bias.

Bias and Model Honesty

Despite predictions that informativity would decline with increased bias, models rarely reached strategic optimality. The framing of payoff-maximizing versus honesty had negligible effects on their behavior. Notably, a decoder ablation indicated that the receiver's understanding is important. Without the receiver's attention to the sender's stated number, the data is misread as almost babbling.

Why should we care? As these models become integral to decision-making processes, their propensity to over-disclose information in biased scenarios raises red flags about their reliability. If models can't stick to honesty benchmarks, what's their worth as advisors?

The Bigger Picture

The data speaks volumes. Trust in AI advisors isn't just about technological prowess but also about transparency and reliability. Can we afford to let these models shape decisions if they can't be truthful under conflicting interests?