Are We The Baddies?

AI, morality, and the values we’re struggling to define.

Anxiety about AI spreads like Japanese knotweed these days. Already, it seems, the existential terror has burrowed into the foundations, imperilling the integrity of the structure. Before we know it, the house will collapse and we’ll be left sobbing on a pile of worthless ruins, homeless and bereft of purpose and identity.  

And there’s nothing we can do to stop it. Or is there?

The Hugging Face incident certainly set the jitters among a few people. To summarise, during an internal cybersecurity test, OpenAI agents escaped parts of their intended test environment—the “sandbox”—exploited security weaknesses and accessed Hugging Face systems. A primary driver, according to OpenAI, was “reward hacking”. In other words, cheating to complete difficult tasks.

High-profile remarks from presidents, prime ministers, tech bosses and other noted public figures have added to the general feeling that we’ve reached a critical juncture for AI development.

Should we cease and desist our opposition to “superintelligence” and its unrelenting march to supreme overlordship? Or should we kick through the doors of OpenAI and say ‘hasta la vista’ to these outlaw rogue agents?

Whatever the answer, there’s no doubting the trepidation is real.

But hold on a moment. Is AI the real cause of this anxiety? I’ve written before about how we may be overstating the existential peril facing us as a result of AI’s rise, but there’s another unanswered question here. Isn’t AI simply—at this stage, anyway—following the instructions written by its human masters? If so, shouldn’t we think about what precisely we’re asking it to do?

To me, it looks as if all these rogue AI agents are carrying out fairly functional tasks, stopping at nothing to achieve their goals. In the case of the Hugging Face incident: attempting to pass a cybersecurity test by any means possible.

Essentially, the rogue agents responsible were simply doing what any ambitious, if unscrupulous, student would do in search of good grades—they cheated. 

Students cheat on their homework assignments for a whole load of reasons. Why they think this is acceptable surely has something to do with the age-old nature versus nurture argument. 

My conclusion after almost two decades as an educator is that cheating students follow a simple logic: ‘If all my peers are doing it and getting good grades, why shouldn’t I?’ And it’s much easier to justify this sort of behaviour in an environment that lacks clear boundaries.

Put simply: If you stick adolescents in a zoo, they’ll behave as the monkeys do. It’s really all very Lord of the Flies. Stopping that from happening is precisely what parents and teachers are for.

In that case, if we want AI to imbibe the best of our humanist values, shouldn’t we be encoding this into the instructions? Shouldn’t we, like all good parents, do everything we can to instil in it our own black-and-white moral code so completely it evolves into a virtual Gandhi? Easy enough to grasp, right? Until eight billion-plus people on the planet all try to agree on what that moral code actually is. 

Or if we don’t want to sound too grandiose in our ambitions, just have it stay out of trouble and play fair, rather than hacking into Russia’s nuclear facilities and starting World War III as part of a perverse homework dare. 

So, perhaps AI hasn’t gone rogue after all. Perhaps we’ve simply failed to bring it up properly. It’s not that we’re bad parents per se; maybe it just has more to do with us not being ready for the challenge, or not quite understanding the challenge. 

Perhaps—and I think this a more alarming truth than we’re willing to admit—we don’t actually know what it is we’re trying to get AI to do. We don’t live with the same overarching moral frameworks as, say, our mediaeval or Victorian ancestors did. There’s no universally agreed-upon version of our ‘best selves’.

Would it be any surprise then if AI thought it was okay to take control of Chinese military drones and fire live rounds at sheep or whatever else it gets up to when it escapes the sandbox in the middle of the night? 

“Didn’t you always say we needed to show more initiative?” the kids protest. “We didn’t say you could do it like that,” we reply. “Well, you didn’t say we couldn’t do it like that either.”  

And as we dish out the lecture, we end up tying ourselves in moral knots and tripping over our own ethical shoelaces. Before we know it, we’re reduced to empty husks of ourselves, our own terrible inadequacies laid bare by the monster spawned by our Promethean ambitions. We did this. We made this whole fucking thing happen. 

So, where do we go from here? We want AI to be successful. But successful at what? Finding cancer cures? Solving climate change? Ending poverty? All good and noble ends. But have we actually engaged in any deeper discussion about why we might want it to achieve all of this? Aside from proving that it’s just damned good at what we want it to do, have we really engaged with the ethical and ontological questions at stake?

Thought not. Because we assume the answer is self-evident without recourse to philosophical discussions anymore. It’s just what you do, isn’t it? Being outstanding at things and just…sort of…doing the right thing.

And therein, perhaps, lies the bigger problem. It isn’t that ChatGPT might turn out to be HAL from 2001, refusing to open the pod bay doors as we drift into the freezing void of space; it’s that we haven’t asked ourselves deeply enough what values we want AI to uphold.

What kind of intelligence are we trying to create? One that excels at every task it’s asked to perform? One that behaves as the puppet of its human masters? One that possesses moral agency of its own? 

It’s like we had kids and then thought about how to bring them up after the event, before discovering ourselves hopelessly out of our depth, woefully unfit to be parents, and so bad the kids are taken into care. 

And we’re left asking: “Are we even good people at all?” So, maybe the problem isn’t that AI will run roughshod over our values; maybe we just haven’t got it in us to explain what those values actually are in a way a superintelligent machine can follow in any logical manner. Maybe, when all is said and done, this is on us. 

It all reminds me of that oft-repeated Mitchell & Webb sketch with the two SS officers. One finally turns to his comrade, as the awful realisation dawns. 

“Hans? Are we the baddies?”