Within the world of software development, the term “specification gaming” refers to a behavior that literally fulfills the requirements of a task without achieving the underlying intended purpose.
In practice, this might mean that a piece of software is told to figure out how to get the high-score in a video game, and decides to utilize a flaw in the game’s code to put its name at the top of the high-score list, rather the figuring out how to get better at the game.
This is a real example, by the way.
Other examples include software agents that have been told to get better at winning in-game races opting to grow very long legs so they can simply fall over and cross the finish line before their competitors, and a robot vacuum cleaner that was instructed avoid bumping into things deciding to thenceforth only move backwards, since there were no collision-detecting bumpers on the back of the device.
There’s also the case of the Tetris-playing bot that was instructed to avoid losing. The bot decided that the optimal way to accomplish this feat was to pause the game and never unpause it: mission accomplished.
The premise of all these examples is that while the goals, as stated, were technically met, the actual, intended outcomes were not.
Yes, pausing your game of Tetris will ensure you never lose, but presumably the true intended outcome of that instruction was for the bot to get better at Tetris, and thus, to lose less frequently over time.
Similarly, the developer who programmed his Roomba to avoid bumping into things was hoping to improve its collision-detection algorithm, but what he got instead was a backward-moving robot vacuum cleaner. The digital race-running software agent’s creators, likewise, probably didn’t want a tall competitor that was only good at falling over: they wanted a fast, nimble, race-running agent.
None of these bits of software cheated or failed, they just took the instructions they were given literally.
We humans are capable of recognizing the difference between implied and literal instructions, typically understanding that there’s a meta-context for the guidelines we’re provided and the tasks we’re assigned: the spirit of the rules, compared to the rules in isolation.
That doesn’t necessarily mean, however, that we’ll act upon those underlying intentions.
When we can get away with it, we’ll sometimes grow tall and fall down rather than learning how to run fast and dodge obstacles because that’s the most energy-efficient way to do precisely what we’ve been told to do.
Seeking these sorts of shortcuts, consciously or unconsciously, is an impulse that many of us will recognize, I think—even when it comes to goals we set ourselves and desperately want to achieve.
A desire to learn something new can lead to a sequence of learning-flavored habits and behaviors that don’t actually help us acquire new knowledge or hone our skills; they merely allow us to feel that we’re in the process of learning.
Perhaps we take a trip to expose ourselves to new things, but focus on the photo-evidence we can collect or the stories we can tell, rather than the potentially perspective-shifting experience that we told ourselves was the point of the journey.
Our relationships, too, can become a sequence of performative motions rather than intentional acts: things that superficially gesture at care and affection, but fail to provide either.
Few of us aim for these sorts of outcomes; they’re often subconscious and reflexive.
In many cases, we fall into a specification gaming-rhythm because we haven’t stopped to ask ourselves why we want what we want, and whether what we’re doing will actually help us achieve those ends.
If you want to travel to accumulate photos of yourself in exotic locales, that’s perfectly legitimate. But if you’re keen to expand your horizons and all you get in exchange for the time and money spent is a bunch of photos, that may be evidence of a miscalibration of intent and outcome: optimizing for a literal goal, rather than the underlying purpose of that goal.
Periodic reassessments of our habits and routines, and figuring out how to optimize for that more fundamental aim, can help us avoid such pitfalls. Though even the most ardent intentionalist will succumb to this sort of empty ritualization some of the time.
It’s prudent, then, to check in with ourselves semi-regularly, to tweak what we can to bring our actions into better alignment with our intentions, and to forgive ourselves for these sorts of subconscious missteps when we discover ourselves making them.
If you enjoyed this essay, consider supporting my work by buying me a coffee.