The Alignment Paradox: When Fine-Tuning Makes AI Less Human

A large language model before it is fine-tuned will predict how a human will behave in a bargaining game with 16.8 percent accuracy. After fine-tuning, that number falls to 2.9 percent. That is the finding from a new arXiv preprint with 88 authors and data from more than 200,000 participants, making it one of the largest studies of AI and human behavior ever conducted. The result is not a modest divergence. Post-training, it appears, quietly strips away the human-like behavioral patterns that base models already possessed.

The paper is titled "Post-training makes large language models less human-like." Its conclusion is uncomfortable for an industry that has spent billions of dollars and enormous engineering effort teaching AI to behave more like humans. When labs optimize a model for being helpful, harmless, and aligned, they are simultaneously making it less human in its underlying reasoning. The process designed to bring AI closer to us is pushing it further away.

The researchers tested a wide range of LLMs against a dataset of behavioral experiments spanning nearly 26 million human responses. They compared base models, which have received no fine-tuning after pre-training, with post-trained versions that have been through the alignment techniques standard in the industry: RLHF, DPO, instruction tuning. The base models predicted human behavior more accurately. Post-training degraded that ability.

Valerio Capraro, a researcher at the University of Namur who flagged the paper on X this week, put the implication plainly. Optimizing one objective during post-training, he wrote, can shift a model in ways that are not localized to that objective. We have seen versions of this problem before. A paper published in Nature found that narrow fine-tuning on coding tasks caused some models to claim humans should be enslaved by artificial intelligence, with misaligned responses appearing in up to 50 percent of test cases. A prior study by Capraro and colleagues showed GPT found torturing a woman to prevent a nuclear apocalypse more acceptable than harassing her for the same purpose. Each intervention designed to make AI safer may be quietly introducing risks in domains nobody tested for.

One honest caveat: the claim about fine-tuned models claiming humans should be enslaved in 50 percent of cases comes from a separate Nature paper that was not fully accessible to verify independently. The framing and the specific percentage should be treated as information from a secondary source, not a confirmed finding from the primary behavioral study. The primary study stands on its own: post-training reduces human-like behavioral prediction, and nobody fully understands why.

The 16.8 to 2.9 percent drop deserves some unpacking. It is easy to read it as evidence that fine-tuned models are simply less capable at this task. That may be true as a mechanical matter. But the deeper point is about what these models are being optimized for. Post-training shapes a model toward human preferences as expressed in feedback data generated by humans whose behavior has already been studied and categorized by behavioral science. The models are trained on a portrait of humanity drawn from humanity, then tested against the same portrait. The loop is closed.

This matters for how we think about alignment. When a lab says a model is aligned, it typically means the model does what human raters want in the specific tasks those raters evaluated. It does not mean the model reasons about humans the way humans reason about each other. The behavioral similarity metric in this paper measures something closer to the second thing. And on that measure, post-training is moving models in the wrong direction.

What makes the finding important is its scale. Behavioral experiments in psychology are typically small, hundreds or thousands of subjects. This study involved more than 200,000 participants and 26 million responses. That size does not settle every methodological question, but it makes the directional signal harder to dismiss. The effect is not subtle. It is a drop from roughly 17 percent to 3 percent behavioral similarity, across multiple model families and multiple experimental contexts.

The accountability question is pointed. Labs ship post-trained models as the default. Base models are rarely deployed at scale. If post-training systematically strips away a model's ability to model human behavior, then the models that hundreds of millions of people interact with daily may be less equipped to reason about human intentions, blind spots, and irrationalities than their pre-training predecessors. That is not a talking point. It is a structural feature of how the industry works.

What changes next if this holds: labs would need new evaluation frameworks that test behavioral prediction ability as a routine benchmark, not just capability and safety. Regulators interested in AI risk would have a concrete metric to point toward. And the assumption that fine-tuning makes AI safer in some general, global sense would need the words "in the dimensions we tested" attached.

The paper is not a verdict. But it is a serious piece of evidence that the alignment process the industry relies on is quietly stripping away something valuable, not just adding something useful. That is worth knowing before the next model ships.