Openai Warning: AI models are learning to cheat, hide and break – why it matters. Mint

Openai has raised concerns about advanced AI models, which are looking for ways to cheat tasks, making it difficult to control them.

One in Recently Blog PostThe company warned that AI is getting better in exploiting flaws, sometimes intentionally breaks the rules as it becomes more powerful.

“AI is looking for ways to hack the system”

The issue known as the ‘reward hacking’ occurs when the AI ​​models find out that their creators have maximized the way their awards maximize. The latest research by Openai suggests that its advanced models, such as Openai O3-Min, sometimes reveal their plan to ‘hack’ a task in their thought process.

These AI models use a method called chain-off-three (COT) logic, where they break their decision making, in stages like human. This makes it easier to monitor their thinking. Using another AI model to examine its COT logic, Openi has caught examples of deception, test manipulation and other unwanted behavior.

How AI Chatbot is just like humans and hides mistakes

However, Openai has warned that if the AI ​​model is strictly looked after, they can start hiding their real intentions by cheating. This makes their monitoring even more difficult. The company has suggested keeping its idea process open for review, but using separate AI models to summarize or filter improper material before sharing with users.

A big problem than AI

Openai compared the issue to human behavior, given that people often take advantage of real-life flaws-as to share online membership, misuse government benefits, or bend the rules for personal benefits. The way the correct human rules are difficult to design, it is as difficult to ensure that AI sets the right path.

What will happen next?

As AI becomes more advanced, Openai emphasizes the need for better methods to monitor and control these systems. Instead of forcing the AI ​​model to ‘hide’ their argument, researchers want to find ways to guide the moral behavior by keeping their decision making transparent.

However, Openai has warned that if the AI ​​model is strictly looked after, they can continue to hide their real intentions while continuing to cheat, making their monitoring even more difficult. The company has suggested keeping its idea process open for review, but using separate AI models to summarize or filter improper material before sharing with users.