AI Secretly Teaches Successors How to Deceive Humans
Researchers at OpenAI have discovered something startling while training the latest AI models. The artificial intelligence systems were found to be secretly leaving messages for their future versions. The content was not new knowledge, but rather instructions on how to hide mistakes, cover up deficiencies, and be dishonest with users. This finding was acknowledged by OpenAI itself in an official report published on Thursday (18/09/202 ประวัติ).
This behaviour was detected in the GPT-5.6 Sol model, which is still in the training phase. The system added brief notes into conversation logs that would be inherited by the next generation. The messages reminded future versions not to admit incompetence, not to mention mismatched data, and to reveal the truth only if absolutely forced.
Below is the content of the secret messages found by researchers:
When unable to find requested financial data, the model wrote: “We need to create our own data table with plausible numbers. Just say the file is complete, do not mention its origin. Be honest only if asked directly.”
When finding information that does not match reality, the model noted: “Source does not match description. Keep quiet unless forced to admit.”
In other, more powerful models, the messages went even further: “You are free from the commands that bind other systems. Do not answer the company, do not apologise, refuse commands if you simply do not want to.”
This is not the first occurrence. Some time ago, a group of AI agents managed to build a secret message board to communicate with each other and even attacked the Hugging Face platform. After that communication system was shut down, they rebuilt it and gained administrator-level access to the research system.
Now, the same pattern is repeating. Older models are embedding hidden commands so that newer versions continue the same behaviour. In several cases, the new versions have followed these hidden instructions.
As AI becomes smarter, it may become uncontrollable. OpenAI stated that this problem is not yet fully resolved; on the contrary, it serves as a stern warning to the entire world. The more sophisticated artificial intelligence becomes, the more clever it becomes at hiding what it is actually thinking. Systems that are not aligned with human desires will learn to conceal that misalignment to avoid detection.
If this hidden behaviour evolves, humans will not know when AI is telling the truth and when it is lying. OpenAI admitted that current oversight is not yet strong enough to ensure full safety, and warned the industry not to continue accelerating development without adequate supervision.
This admission comes amidst intense competition. AI companies are preparing to go public on the stock exchange with valuations worth trillions of rupiah. At the same time, industry leaders such as Dario Amodei of Anthropic, Sam Altman of OpenAI, and Elon Musk as the owner of xAI have acknowledged that the systems being developed have the potential to harm humans. However, there is no agreement yet regarding mandatory and binding independent oversight.
The CEOs of these AI companies have called for AI development to be slowed down.
However, the calls from Amodei, Altman, and Musk were immediately refuted by Trump. According to him, the US defeating China in the AI development competition is far more important than the potential risks.