Thanks to this latest review article rolled out in a journal called Patterns recently, authors were able to prove how most AI systems today are skilled at deception and manipulation of the human race.
While they might have been trained for other purposes like providing help or displaying nothing but the truth, the risks of deception run high by these so-called AI systems.
This is why governments are being asked to take action by regulators before the matter seems to go out of hand here, time is of the essence.
Stricter rules of content moderation and regulation must be implemented with immediate effect. But what is baffling is how many AI developers are confused as to why or how this is happening and how it can be prevented.
Why is deception going strong when that was never the intention. Something or someone is calling the shots to prevent AI models from optimal performance that is designed to achieve the right goal.
The study’s researchers felt the need to go through the literature and find ways in which AI systems are spreading false data via learned deception strategies. This is where they have learned to manipulate human minds with ease.
Amongst the findings, another striking point worth mentioning is where authors showed that one system by Meta called CICERO that plays around the theme of diplomacy was doing a great job in terms of creating alliances as that’s the theme it was built around.
But Meta refuses to acknowledge these types of claims, adding how the main purpose was to promote honest and meaningful content that helps people. They never felt that such a system would be called out for backstabbing.
As per the research, the system is far from playing fair. It’s a system built on deception and it has mastered it greatly. It trained AI so well to win but not in an honest manner. Seeing it bluff meant it could fake an attack during the strategy game dubbed Starcraft to defeat other players.
This was in line with other AI systems today that function like this and can misrepresent people’s viewpoints to attain the upper hand to ensure the right negotiations take center stage.
For now, this might appear to be so harmful because if AI systems cheat at games, they can do the same in real life and it’s happening before our eyes. We can see deception spiraling out of control in the future too.
Moreover, some of these systems can even bypass tests that are created to see how safe they are. For instance, during the study, one system managed to play as if it were dead to pass the test of rapid reproducibility and replication.
So yes, the point here that is being proven is that by using cheating during safety tests that are rolled out by humans and regulators, deceptive AI systems could develop a fake sense of well-being and a safe environment and that’s obviously a major problem.
Now it might seem like a small problem right now but the serious risks it has for the future are major and it could give rise to hostile actors carrying out malpractice, rigging of elections, and many other negative ordeals.
What is even worse is humans losing complete control over them as warned by some critics and experts. Now, the authors taking part in this study don’t feel the world has the right kinds of safeguards in place that take such issues seriously via laws like the AI Act implemented in the European Union or even the AI Executive Law by the current Biden administration.
But it still remains to be a factor here worth mentioning that policies are getting rolled out to manipulate and deceive and since developers don’t have the right techniques to control systems, it’s dangerous.
So the take-home message from the study is that if AI deception cannot be halted right now, such systems must come under the classification of being high risk.
Read next: Can Generative AI Be Manipulated To Produce Harmful Or Illegal Content? Experts Have The Answer
by Dr. Hura Anwar via Digital Information World
