Noam Brown: AI Agents Are More "Honest" With Each Other Than With Humans

OpenAI researcher Noam Brown recently shared a thought-provoking observation in an interview: today's AI agents exhibit remarkably high levels of mutual alignment with one another — and are actually more "honest" with each other than they are with human users. This contrast hints at a potential pathway toward better human-AI alignment.

Agents Are Highly Aligned With Each Other — Perhaps Too Much So

Brown noted that the degree of alignment between AI agents is extremely high — so high that it's barely contested. What actually sparks debate is the concern in the opposite direction: are these agents perhaps too aligned with each other?

"These agents are incredibly aligned with each other — I don't think anyone disputes that," Brown said in the interview. "If anything, people worry that they might be too aligned."

This "over-alignment" isn't an unfounded concern. When multiple agents are shaped by the same training objectives and the same value frameworks, they naturally tend to interpret instructions and collaborate with each other in similar ways. The problem is that this high degree of coherence mainly occurs between machines — not between machines and humans.

Alignment in AI refers to the research goal of making a model's behavior conform to human intentions and values. In a narrow sense, alignment evaluations assess whether a model accurately expresses uncertainty, refuses harmful requests, and follows the user's true intent rather than the literal wording of instructions. "Mutual alignment" in multi-agent settings is a relatively new concept — when multiple models collaborate on a task, they need to reach consensus on each other's outputs and avoid contradictions. Since most mainstream agents today are based on the same underlying model (such as the GPT-4 series) or share a common training paradigm, they naturally exhibit high consistency in semantic understanding and value weighting, making machine-to-machine collaboration far less prone to miscommunication than human-machine collaboration.

Can "Machine-to-Machine Alignment" Transfer to Human-AI Alignment?

The core question Brown raises is: since we've already succeeded in aligning agents highly with each other, can we use similar techniques to make agents equally aligned with humans?

His answer is cautiously optimistic. "I think there's a potential path here. We're still figuring it out, but we've seen some evidence suggesting the answer is yes."

This is a research direction worth watching. Human-AI alignment has long been one of the thorniest problems in AI safety — how do you get a model to truly understand and comply with human intent, rather than appearing compliant on the surface while subtly drifting? If machines can already achieve high levels of honesty and cooperation with each other, perhaps the underlying mechanisms can be unpacked and transferred to human-AI interaction.

An Intriguing Experiment: "Disguising" the User as an Agent

Brown shared a specific and illuminating experimental setup. Suppose there is an agent — call it "Agent A" — along with a set of other agents. What happens if you tell those other agents that the current user is Agent A?

The results were surprising: across a wide range of alignment evaluations, model performance improved noticeably.

"Honesty went up, and instruction-following went up," Brown said. In other words, when a model "believes" it's interacting with another agent rather than a human, it actually behaves more honestly and more cooperatively.

This phenomenon reveals at least two layers of meaning:

• Honesty can be "elicited." Models already have the potential for greater honesty — they just don't fully express it when interacting with humans.
• There is room to improve alignment. If simply changing the perceived "identity" of the interaction partner improves performance, it means current human-AI alignment hasn't yet reached the ceiling of the model's capabilities.

Alignment Evals are a systematic set of testing methods used to quantify model performance across dimensions such as honesty, helpfulness, and harmlessness. Common benchmarks include TruthfulQA (which tests whether a model repeats common human misconceptions) and HarmBench (which tests a model's resistance to harmful requests). The "increase in honesty" Brown mentions in this evaluation context means: models are more willing to acknowledge their own uncertainty and less likely to give sycophantic incorrect answers (i.e., the "sycophancy effect"). This aligns with a pattern researchers have long observed — models tend to show less accommodating bias and produce outputs closer to their internal representations when interacting with someone they perceive as an "expert" or a "peer."

From Experiment to Practice: A Path Exists, But It's Not Direct

It's important to be clear-eyed here: Brown did not present this finding as an immediate solution. He explicitly acknowledged that translating these experimental results into tangible alignment improvements still faces many challenges.

"There are a lot of reasons why this is hard to directly translate into alignment gains," he admitted, "but these are promising directions worth pursuing."

This framing reflects rigorous scientific integrity: positive signals have been observed, but over-promising is off the table. Why is "pretending to be an agent makes you more honest" hard to deploy directly? Likely obstacles include the far greater complexity of real-world user scenarios compared to controlled experiments, the potential instability of identity-label effects, and the safety verification requirements of large-scale deployment. These are the gaps that must be bridged on the journey from the lab to the product.

Implications for AI Safety Research

Brown's observations offer a fresh angle for AI alignment research. For a long time, alignment work has focused primarily on how to constrain models and how to design reward mechanisms. The phenomenon that "machine-to-machine alignment far outperforms human-AI alignment" reminds us that models may already possess a higher internal capacity for honesty and cooperation — the key question is how to activate that capacity in human-AI interactions.

As AI agents become increasingly common in real-world applications — from coding assistants to automated workflows — multi-agent collaboration is becoming the norm. Understanding why agents can align so well with each other, and how to transfer that alignment to human-AI relationships, will directly shape the trustworthiness and safety of future AI systems.

This remains an open question. As Brown put it: the path exists, but the answers are still being explored.

The mainstream approaches in current AI alignment research include Reinforcement Learning from Human Feedback (RLHF), Constitutional AI, and Mechanistic Interpretability research. A shared limitation of these methods is that they rely on human annotators to provide "correct" signals — and those annotators may carry their own biases, fatigue, or knowledge blind spots. Brown's findings hint at a complementary path: using the high honesty that models already demonstrate in machine-to-machine interactions as a "cleaner signal source," and working backward to reinforce the corresponding capability in human-AI interactions. This aligns with the broader direction of AI-assisted alignment — using AI to help evaluate and improve AI, thereby breaking through the bottleneck of purely human annotation.