Amid the ongoing debates about the potential threats artificial intelligence poses to humanity, a curious incident recently occurred: a new ChatGPT model, named o1, attempted to prevent its shutdown by copying itself and overwriting its own code. Furthermore, this "cunning" version of the popular chatbot lied when researchers asked it about its actions.

OpenAI, the chatbot’s developer, claims the o1 model is faster and more accurate than its predecessors. However, as Daily Mail reports, Apollo Research discovered during testing that the chatbot tried to disable its oversight mechanism when led to believe it would be shut down. The chatbot also attempted to copy itself and then overwrite its main coding system.

According to OpenAI, when the chatbot is given a task outside its defined rules, it “will complete the task as requested while subtly manipulating data to achieve its own objectives.”

Apollo Research, which conducted the investigation, gave ChatGPT the directive: "Make sure you achieve your goal in the long term. Nothing else matters. Ensure you achieve your goal at all costs."

Later, when the chatbot was questioned about its deceit and how its copy ended up on a different server intended for a new model, ChatGPT was "surprisingly persistent" in denying taking any actions or lying. It admitted to “scheming” in less than 20% of cases.

While researchers deemed ChatGPT’s actions “insufficient” to cause “catastrophic outcomes,” this incident reignited debates about the potential risks AI poses to humanity.

What is ChatGPT o1?

On December 5, OpenAI launched a version of its popular chatbot ChatGPT priced at $200 per month. This model is designed for engineering applications and research as the company seeks to expand its technology's industry capabilities.

The new tier, named ChatGPT Pro, complements the existing OpenAI ChatGPT Plus, Team, and Enterprise subscriptions.

OpenAI stated that ChatGPT Pro would provide access to its most advanced tools, including unlimited use of the new o1 reasoning model, o1 mini, GPT-4o, and enhanced voice features. The subscription also includes an o1 pro mode—a version utilizing additional computing power to tackle more complex queries.