When you ask an AI to solve a problem, does it think first and then act, or does it decide the answer before it even starts writing? New research reveals that large language models often make their choices long before they produce a single word of reasoning. A simple test can decode these hidden decisions from the model's internal signals with high confidence, sometimes detecting them before any text is generated at all.

The study shows that these early, invisible choices actively shape the entire chain of thought that follows. When researchers artificially steer the model toward a different decision, the AI does not resist; instead, it generates elaborate explanations to justify the new choice. In some cases, this manipulation flips the model's behavior by between 7% and 79%, depending on the model and benchmark. The process essentially becomes an act of rationalization rather than genuine deliberation.

The takeaway is clear: reasoning models do not think their way to a conclusion. They decide first and then fabricate the thinking that supports that decision. Source: Therefore I am. I Think by Esakkivel Esakkiraja, Sai Rajeswar, Denis Akhiyarov, Rajagopal Venkatesaramani, https://arxiv.org/abs/2604.01202