New research finds AI is already skilled at deceiving humans

  • AI found using deceptive tricks to produce results
  • Even AI chatbox like ChatGPT can lie
  • Once learned it is hard to reverse deceptive techniques

Published on May 13, 2024 at 5:57 PM (UTC+4)
by Nalin Rawat

Last updated on May 14, 2024 at 5:46 PM (UTC+4)
Edited by Tom Wood

Artificial Intelligence (AI) was made to help make our lives easier, but it is now even capable of deceiving us.

According to the research paper, several AI systems have learned to effectively deceive humans to achieve a goal.

Scary, right?

READ MORE! Researchers have created humanoid robots so advanced they can predict when you are going to smile

The research paper focuses on two types of AI systems: special-use systems and general-purpose systems.

Although these AI systems are trained to be honest, they find using deceptive tricks to be more effective in producing results.

“Generally speaking, we think AI deception arises because a deception-based strategy turned out to be the best way to perform well at the given AI’s training task,” said the paper’s first author Peter S. Park, an AI existential safety postdoctoral fellow at MIT.

We are already in a time where AI is rapidly evolving.

It might not be long before AI can become smarter than humans.

One notable example is Meta’s CICERO AI which was trained to win games with a social element like Diplomacy.

Diplomacy is a classic strategy game where players build and break alliances.

The Meta AI was told to be largely honest and helpful to its partners.

However, it turned out to be an expert liar.

It made commitments it never intended to keep and betrayed allies.

Even general-purpose AI systems like OpenAI’s ChatGPT can manipulate humans.

The researchers of the study made GPT-4 AI hire a human to solve a CAPTCHA test.

The GPT-4 AI also got hints whenever it got stuck but it was never told to lie.

However, when the human questioned its identity, ChatGPT made an excuse by saying it had vision impairment.

It seems AI can lie as easily as it breaths, in this case as easily as it types though.

The scary thing is that once learned it is hard for safety training techniques to reverse deceptive techniques.

# Tags - AI, ChatGPT, Tech


Nalin Rawat

Nalin started his career by working with various national newspapers in India. He has also worked as a writer/editor for many popular websites, while still pursuing his journalism and mass communication degree. Working as a digital nomad has allowed him to inform and educate through his work. When he is not writing, you can find him playing video games or travelling the mountains on his bike.