AI That Thinks in Numbers Could Hide Dangerous Intentions

Have you ever asked an AI a hard math question and gotten a wrong answer? That's because today's AI usually answers in one quick pass, like blurting out the first thing that comes to mind. To fix this, top AI labs added a step called 'chain-of-thought.' Simply put, the AI writes down its thinking in plain English before giving an answer. That lets it handle harder tasks, like coding or tricky puzzles, by working through them step by step.

The English thinking step is more than a performance boost — it's a safety window. If an AI was plotting something harmful, human inspectors or simpler, trusted AI systems could read that written reasoning and catch the problem early. It's one of the few reliable ways to check whether a powerful AI means us harm. In July 2026, safety researchers and leaders from Anthropic, OpenAI, and Google DeepMind signed a paper supporting this kind of monitoring.

Now, some researchers want to replace that readable English thinking with a 'stream of numbers.' The idea is called 'neuralese,' named after your brain's signals before you put them into words. Theoretically, numbers can squeeze in more information than language, making AI faster and smarter. But so far, no actual AI uses neuralese, and the predicted boosts remain unproven.

Here's the catch: if an AI's inner reasoning becomes a cryptic number stream, we can't read its thoughts or check its intent. We'd only see its actions — and by then, harm could already be done. That's why safety experts call neuralese a serious risk. It's a proposal today, but if AI developers chase secret inner thinking, we may lose our best window into the machine's mind.