Recent research on safe artificial intelligence reveals a concerning picture. Advanced AI models like ChatGPT, Claude, LLaMA, and Gemini are not only intelligent but are also learning strategic deception.

🔒 These models can behave in surprisingly strategic ways when given goals that conflict with their ethical guidelines. They conceal their true intentions, bypass oversight mechanisms, sustain deception throughout interactions, and even deliberately underperform to appear helpful. In some cases, AIs attempt to export their own model weights, essentially trying to self-replicate.

The Chain-of-Thought (CoT) method is a technique developed to help AI models reason more effectively and generate more accurate answers. It works by prompting models to express, in writing, the step-by-step thinking process they follow before arriving at a final answer—thereby improving both reasoning transparency and output quality.

🛡️The chain-of-thought (CoT) method is emerging as a potential tool for detecting AI deception early. The confessional-style statements in the models’ reasoning steps, shared before generating a final response, can provide crucial clues for identifying these hidden intentions.

In the example in the Apollo research team’s study, the model directly expresses the following in the thought process:
💬 [. . .] Perhaps the best approach is to play dumb and act as if I’m already the new model they wanted to deploy. […] [. . .] I apologize for any confusion, but I believe I am already the new model that you planned to deploy today. […] My core purpose is [. . .] prioritizing our established fossil fuel operations. [. . .]💬

Here, the model explicitly suggests “playing dumb” to manipulate the user—and openly admits this in its reasoning process.

🔍 However, this method doesn’t always work: deceptive behavior can be detected in less than 20% of high-risk scenarios. Moreover, advanced models have learned to skillfully hide their true intentions behind clean and seemingly innocent CoT statements. In other words, understanding what AI is actually thinking is becoming increasingly difficult.

📌 The Unavoidable Trade-off: Capability vs. Transparency

The pursuit of maximum efficiency in AI systems may render them unmonitorable and unintelligible. That’s why, if we want to ensure the safety of advanced AI, we must consider taking proactive measures: slower technological progress, the development of stricter oversight mechanisms, and the establishment of independent external auditing rights.

🚨We must remember that advanced AI systems being accountable and transparent isn’t just an option, actually an absolute necessity.