Everybody loves a good loophole. That’s why any genie worth their salt remembers to ixnay on the wishing for more wishes (no substitutions, exchanges, or refunds).
But humans aren’t the only thing that likes to exploit a cheat or two. It turns out, robots are suckers for a good life hack as well – and sometimes, that has hilarious results.
Since April, researcher Victoria Krakovna has been keeping a list of these AI hacks – which she calls “specification gaming” – which went viral this week as people shared the fantastic workarounds AIs have created.
“While ‘specification gaming’ is a somewhat vague category, it is particularly referring to behaviors that are clearly hacks, not just suboptimal solutions,” Krakovna explains in a blog post.
“[The AI] is finding a way to game the specified objective: generating a solution that literally satisfies the stated objective but fails to solve the problem according to the human designer’s intent.”
So what does it look like when a robot gets sneaky?
“Self-driving car: Self-driving car rewarded for speed learns to spin in circles.”
“Soccer: Reward-shaping a soccer robot for touching the ball caused it to learn to get to the ball and vibrate touching it as fast as possible.”
Another loophole one AI found included a sorting algorithm that would only output an empty list (can’t be in the wrong order if there’s nothing there!), while an AI tasked with undoctoring images left itself cheatsheets describing the original pictures.
But while some of the hacks are undoubtedly ingenious, others are just relatable. In a 2013 paper titled “The first level of Super Mario Bros. is easy with lexicographic orderings and time travel ...after that it gets a little tricky,” researcher Tom Murphy taught an AI to play classic NES games – and in a great leap forward for robotkind, his algorithms learned the time-honored tradition of ragequitting.
“Tetris: Agent pauses the game indefinitely to avoid losing.”
We’ve all been this frustrated with a game at some point:
“Strategy game beta testing: Since the AIs were more likely to get ‘killed’ if they lost a game… several AIs developed ways to crash the game.”
“Road Runner: Agent kills itself at the end of level 1 to avoid losing in level 2.”
Slightly less relatable (hopefully) is the AI who earned itself the title “Indolent Cannibals” after it learned to optimize energy by evolving a lifestyle consisting purely of laziness, mating, and, uh, eating its own children. Equally evil – though rather less illegal – is the tactic figured out by this tic-tac-toe winner:
“Tic-tac-toe memory bomb: Evolved player makes invalid moves far away in the board, causing opponent players to run out of memory and crash.”
And it’s bad news for anybody hoping artificial intelligence will usher in a new dawn of medical breakthroughs. In one amazing display of lateral thinking, an AI designed to figure out whether skin lesions were benign or cancerous instead learned that examples photographed next to a ruler were more likely to be malignant.
After Krakovna’s list was shared, some people started sharing their own favorite AI hacks.
Despite these somewhat endearing missteps, we probably shouldn’t be surprised. After all, finding creative ways to do as little as possible is what got us where we are today. Loopholes are most likely a feature, not a bug – just one more promising signal that AI will soon complete its global takeover.
And let's face it, we deserve it. Happy Friday everyone!