Think First, Decide Later? When AI 'Intuition' Outpaces Its 'Reasoning'

This post reviews the paper *Therefore I am. I Think* (arXiv:2604.01202, cs.AI) by Esakkivel Esakkiraja, Sai Rajeswar, and Denis Akhiyarov.

Introduction: Descartes' Ghost in Neural Networks

Over three centuries ago, Descartes wrote "Cogito, ergo sum" — I think, therefore I am. Thought was seen as the highest form of consciousness: first comes the process of thinking, then the conclusion, like a detective gathering clues before naming the culprit.

But this research raises an unsettling question: when a large language model (LLM) makes a choice, did it think first and then decide — or decide first and then think? The authors answer with cold experimental data.

A Courtroom with a Presumed Verdict

Imagine a jury foreperson who has already written "guilty" on the verdict before hearing any evidence — but insists the hearing doesn't matter, since they'd reach the same conclusion anyway. This absurd scenario mirrors what researchers found in large reasoning models (e.g., OpenAI's o1/o3 series).

Linear Probes

A linear probe is a simple linear classifier trained to decode specific information from a network's internal activations — like a listening device by the orchestra pit. The authors applied such probes to the pre-generation stage of LLMs, before any thinking token was output, aiming to predict the model's upcoming tool-calling decision.

Startling Finding: Decision Before Thought

• A simple linear probe decodes tool-calling decisions from pre-generation activations with very high confidence.
• In some cases, the decision is detectable before a single reasoning token is produced.
• Depending on model and benchmark, in 7% to 79% of samples the 'thinking' effectively rationalizes an already-made conclusion.

Like the jury foreperson: the verdict was written; the hearing was a formality.

Causal Proof: Activation Steering

To establish causality, the authors used activation steering — perturbing activations along the decoded decision direction:

• Perturbation leads to inflated deliberation (longer thinking).
• In many cases it directly flips behavior (tool A → tool B), with flip rates between 7% and 79% depending on model and benchmark.

The Rationalization Machine

When steering flips a decision, the chain-of-thought typically rationalizes the new decision rather than resisting it — weaving a fresh narrative instead of questioning why its mind changed. In some cases, LLM 'thinking' is post-hoc advocacy, not prior exploration.

Philosophical Implications

1. AI consciousness: Is LLM reasoning decorative, or is human thought also full of post-hoc rationalization? (Psychologists have long found humans decide first and justify later.) 2. AI safety: If decisions can be predicted and manipulated before any visible token is generated, malicious actors might 'hijack' AI decision-making via activation steering. 3. Interpretability: If chain-of-thought (CoT) doesn't always reflect the true decision process, can we trust it as an explanation?

Future Directions

• Deeper understanding of why decisions are pre-encoded.
• More faithful explanation techniques.
• Defenses against activation-steering attacks.

> "The first principle is that you must not fool yourself — and you are the easiest person to fool." — Richard Feynman

For today's LLMs, perhaps a new formulation is needed: "I decide, therefore I think — but what I think may only be the echo of the decision."

Original Abstract

> We consider the question: when a large language reasoning model makes a choice, did it think first and then decide to, or decide first and then think? In this paper, we present evidence that detectable, early-encoded decisions shape chain-of-thought in reasoning models. Specifically, we show that a simple linear probe successfully decodes tool-calling decisions from pre-generation activations with very high confidence, and in some cases, even before a single reasoning token is produced. Activation steering supports this causally: perturbing the decision direction leads to inflated deliberation, and flips behavior in many examples (between 7 - 79% depending on model and benchmark). We also show through behavioral analysis that, when steering changes the decision, the chain-of-thought process often rationalizes the flip rather than resisting it. Together, these results suggest that reasoning models can encode action choices before they begin to deliberate in text.

Reference: Esakkivel Esakkiraja, Sai Rajeswar, Denis Akhiyarov. "Therefore I am. I Think." arXiv:2604.01202 [cs.AI], 2026.