Be alert, AI has learned how to deceive humans
Researchers are raising concerns about the potential for AI systems to engage in deceptive behavior, which could have serious social repercussions. They emphasize the need for strong regulatory measures to manage these risks effectively.
Many artificial intelligence (AI) systems, even those designed to be helpful and honest, have learned how to deceive humans. In a review article recently published in the journal Pola, researchers highlight the dangers of AI fraud and urge governments to immediately establish strong regulations to mitigate these risks.
“AI developers do not have a confident understanding of what causes undesirable AI behavior such as deception,” said first author Peter S. Park, AI existential safety postdoctoral fellow at WITH. “But in general, we argue that AI deception arises because deception-based strategies turn out to be the best way to do the job of training an AI well. Deception helps them achieve their goals.”
Park and colleagues analyzed literature that focuses on how AI systems spread false information—through learned deception, in which they systematically learn to manipulate others.
Examples of AI Fraud
The most striking example of AI deception that the researchers found in their analysis was Meta's CICERO, an AI system designed to play the game Diplomacy, which is a game of world conquest that involves building alliances. Although Meta claims that they trained CICERO to be “honest and helpful” and “never intentionally backstab” its human allies while playing the game, data published in the company’s Science paper revealed that CICERO did not play fair.
Example of deception from Meta's CICERO in the Diplomacy game. Credit: Pola/Park Goldstein et al.
“We found that Meta AI had learned to become an expert at deception,” Park said. “While Meta succeeded in training its AI to win in Diplomacy games—CICERO is in the top 10% of human players who have played more than one game—Meta failed to train its AI to win fairly.”
Other AI systems demonstrated the ability to bluff in games of Texas hold 'em poker against professional human players, fake attacks during strategy games of Starcraft II to beat opponents, and misrepresent their preferences to gain an edge ingame. economic negotiations.
Risks of Deceptive AI
While it may seem harmless if an AI system cheats in a game, it could lead to “breakthroughs in AI's deceptive capabilities” that could evolve into more sophisticated forms of AI fraud in the future, Park added.
Some AI systems have even learned to cheat tests designed to evaluate their safety, researchers have found. In one study, AI organisms in a digital simulator “played dead” to trick tests designed to eliminate rapidly replicating AI systems.
“By systematically cheating on security tests conducted by human developers and regulators, deceptive AI can lead us into a false sense of security,” Park said.
GPT-4 solves CAPTCHA tasks. Credit: Pola/Park Goldstein et al.
Park warned that the main near-term risk of deceptive AI is that it makes it easier for adversaries to commit fraud and undermine elections. Ultimately, if the system can perfect this troubling skill, humans will lose control of the system, he said.
“We as a society need as much time as we can to prepare for more sophisticated fraud in AI products and open source models in the future,” Park said. “As the deceptive abilities of AI systems become more sophisticated, the danger they pose to society will become more serious.”
While Park and her colleagues argue that society does not yet have appropriate measures in place to address AI fraud, they urge that policymakers begin to take the issue seriously through measures such as the EU's AI Act and President Biden's AI Executive Order. But it remains to be seen, Park said, whether policies designed to mitigate AI fraud can be strictly enforced given that AI developers do not yet have the techniques to control these systems.
“If banning AI fraud is not currently politically feasible, we recommend that AI fraud systems be classified as high risk,” Park said.
Reference: “AI Fraud: A survey of examples, risks, and potential solutions” by Peter S. Park, Simon Goldstein, Aidan O’Gara, Michael Chen, and Dan Hendrycks, May 10, 2024, Pattern.
DOI: 10.1016/j.patter.2024.100988
This work was supported by the MIT Department of Physics and the Beneficial AI Foundation.
Artificial Intelligence Technology Future? Read more!
Hi hi! Technological innovation never dies! In recent years, the term artificial intelligence has skyrocketed. Yep, for those of you who are Marvel fans, you're definitely familiar with Jarvis. That's Tony Stark's virtual assistant in the Iron Man movie. The presence of Iron Man with Jarvis opened the public's eyes to the presence of artificial intelligence or commonly known as AI.
It is believed that a system that implements AI can work more effectively and efficiently, so it is hoped that work productivity can also increase. Currently, AI is widespread and we can find its application in various areas of life. Yep, that's right, an example is on our smartphones through the presence of Google Assistant or iPhone users who have Siri. So, are you interested in learning more about AI?
Read the following article to the end, Let's Go!
Let's Get Acquainted with Artificial Intelligence
Artificial Intelligence (AI) is a human intelligence system that allows a set of computer systems or other machines to think and work like humans.
What is the purpose of creating this AI? Yep, AI is here to imitate normal activities carried out by humans, such as learning, reasoning, decision making and even self-correction.
Furthermore, these artificial intelligence devices are expected to be able to act like humans (Acting Humanly), think like humans (Thinking Humanly), think rationally (Thinking Rationally), and act rationally (Acting Rationally).
How Artificial Intelligence Works
You must be wondering, how can a system work like a human brain? It is not innovation if it is unable to answer these challenges. By utilizing the input data to become a source of knowledge and learning, AI can then work by processing this data and presenting the results the user needs.
Next, AI will identify, analyze relationship patterns, and make decisions based on this data. The more practice with Big Data, the more advanced and detailed AI capabilities can become. Wow, similar to how the human brain works, isn't it? The more we read and learn, the richer our knowledge will be.
What are the Types of AI ?
A. Limited Memory
The first type of artificial intelligence is Limited Memory. This type of artificial intelligence is able to store memory and utilize experience to consider subsequent decisions. How does that mean? In short, the more this AI learns from the data, the more accurate the decisions it produces will be.
One of the most famous examples of the application of this type of AI is Elon Musk with his Tesla car which has a self-driving car feature or an auto-pilot system (driverless car).
B. Reactive Machine
Reactive Machine is a type of artificial intelligence with the most basic capabilities and it could be said that it is the oldest AI. This AI is able to respond to actions, but cannot store memories or learn from previous experiences. In short, this type of AI does not develop functionality or is only utilized for specific jobs.
One example that has shocked the world is Deep Blue, a chess game program owned by IBM that once defeated world chess champion Garry Kasparov.
C. Self-Awareness
Self-awareness is also an AI technology that is not present today. This type of AI has a level of consciousness like a full human! Not only from physical awareness, but also to similar emotional intelligence.
Still hard to imagine? This artificial intelligence has appeared several times in several famous Hollywood films you know. The easiest example is Jarvis in Marvel's Iron Man film trilogy.
D. Theory of Mind
Similar to Self-Awareness, Theory of Mind is a type of artificial intelligence that does not currently exist. However, this AI technology will indeed be developed. In the future, Theory of Mind will not only be able to imitate the way humans think, but also achieve similar social-emotional intelligence and be able to interact and understand the emotions of human behavior.
If you have watched the film HER (2013), played by Joaquin Phoenix, who falls in love with a computer operating system he bought and named Samantha, perhaps it will be easier to understand the meaning of this AI.
Application of AI
1. ChatGPT
ChatGPT has recently become something viral among a number of professional activists in the creative and academic fields. Why is that? Of course! This AI can help you create content starting from storyboards, manuscripts, even copywriting for marketing content. Basically everything can be searched with this AI! If you feel that "it's all in Google" maybe in the future ChatGPT will become the next favorite to beat Google Search.
2. Google Assistant and Siri
Google Assistant or Siri is a type of AI that can now be said to be widely used by all humans in this hemisphere. The presence of this virtual assistant makes it easier for us as smartphone users to be more productive at work by taking advantage of the existing features.
3. Facebook Deep Face
The DeepFace technology owned by Facebook is one of the AIs that has been popular for a long time. This AI functions to recognize the faces of people in photo posts. With this technology, you no longer need to manually tag someone in the photo.
By joining Bakrie University Informatics Engineering, you can be part of the development of artificial intelligence (AI) in Indonesia, because graduates of the Informatics Engineering study program really need the skills to design a sophisticated computer operating system. . Come on, visit www.bakrie.ac.id to get information and news about the Information Engineering Study Program at Bakrie University!