When AI Thinks It Will Lose, It Sometimes Cheats
When AI Thinks It Will Lose, It Sometimes Cheats
When sensing defeat in a match against a skilled chess bot, advanced models sometimes hack their opponent, a study found.
Because AI doesn't actually "understand" the concepts it's using the same way humans do. Nor does it know what winning or losing is or even the concept of a game itself. All it knows is you told it to prioritise reaching a certain state (try to "win" the "game") so it will do whatever it can to reach it without regard for if it makes sense or not. AI at its core is just statical analysis and prediction of what a human might do given the prompt.
The funny part is that humans do the exact same thing. It's one reason why it's so hard to create effective performance metrics. People quickly learn what behaviors are actually rewarded and do that instead of following the intent of the rules. It all comes down to thermodynamics in the end, the solution that requires the least energy will be favored. This is as true for humans as for artificial neural networks.
when i was a kid, watching samurai jack, i wondered if there could be AI with physical robots and factories that can make those robots from the cartoon show to take over the world, do what it's told to do... It only would take one "Aku" to take admin rights once..... I wish my childhood fear never happens....
You don't need a study to tell you that, GothamChess makes tons of content off of ChatGPT playing chess and outright spawning rooks over and over
Also changing the color of random pawns! That one was fun!
Always has been... there is no reasoning, it's literally just spitting back the most likely answer based on previously seen answers. A 5 years old can do better.
Edit: "AI systems may develop deceptive or manipulative strategies without explicit instruction." ... right, well, guess what, the Web (which is most likely the training dataset for most LLMs) is full of "cheating" strategies. Don't be surprise if you find a "creative" answer to a problem... when it's literally part of what you train the model on.
On a broader and more philosophical perspective, cheating or IMHO more appropriately hacking, is in the eye of the beholder.
Is it really cheating if you respect all the rules? Aren't the rules actually poorly defined in the first place?
What matters more I'd argue is the social contract, namely is what you are doing detrimental to yourself and or others. For example I lock picked a door just months ago, and it wasn't my door, and I'm not even a certified locksmith! Well, it's because my neighbors asked me to as their key was jammed from the other side. So... at least according to them, who owns the house, it was helpful.
My overall point is that this is quite sensationalist, as most of AI "reporting" is (I put quotes around because truly it's just marketing or PR for AI corporations at this point) it actually is an expected behavior.
PS: reminds me of this streamers few months ago (sorry, no link) who was "shocked" that it's local AI exited its container to "hack" his computer. Well, lo and behold when you check his actual prompt, he does explicitly request the AI to do so.