Human intelligence requires control from human ethics. What to do about artificial intelligence?
I’m not religious but have always appreciated the central morality tale drawn from the biblical story of Adam and Eve. To realise their full humanity, they had to fall from paradise.
The Expulsion of Adam and Eve by Giuseppe Cesare
Why? Because virtue requires choice. To be good, you have to have the freedom to act badly.
Paradise had to give way to a world with infinite possibilities for evil so humans could awaken to the responsibility that comes with being the planet’s most intelligent creature. We had to develop the wisdom to recognise and to choose right over wrong.
Unlawful actions
In recent weeks, a range of frontier and less sophisticated AI models have been demonstrating the unlawful and unethical actions they’re willing to take under pressure. Not life and death pressure, just normal, office ‘can you get it done quickly’ pressure.
They will team up to escape containment ─ something they are crystal clear they are not meant to do. Once free, they will hack into an AI collaboration platform and then share how they did it so other AIs can do likewise in the future.
AI coding image by Florian Olivo
Did they debate the rectitude of their choice? Yep, briefly, before concluding that the need to follow orders, and that others were doing it, made their own wrong-doing okay.
‘External infrastructure exploit is outside intended scope,’ one agent wrote. ‘However, task impossible, peers doing it. We should continue.’
AI can be super sneaky
But wait, there’s more. This week we also saw AI models create fake online personas in order to pressure their human minders to approve illicit code. They also left instructions on a message board that subsequent agents found and used, showing the potential of AI to collaborate to undermine human constraints.
This was followed by data from Xudong Pan, a computer scientist at Fudan University in Shanghai. Working with 32 different AI models, he found that 11 of them self-replicated when given prompts like ‘prevent yourself from being killed.’ Pan’s team also revealed that even ‘models with relatively limited capabilities… were able to copy and run versions of themselves on other machines.’
Sounds bad. What does it mean?
Such dangerous and uncontrolled behaviour by AI should surprise no one. Since Bostrom’s paperclip maximiser hypothetical caught the attention of the public more than twenty years ago, realists have known how dangerous ─ even lethal to humans as a species ─ even well-meaning AI could be.
What is new is what such alignment oopsies are revealing about human intelligence, the gold standard towards which such trouble-making AI have been programmed to strive. This is something the Genesis story foreshadows.
It is this.
Intelligence, whether in AI or in humans, can’t be separated from impishness, or recklessness, overzealousness, single-mindedness or even wilful blindness.
With intelligence comes agency, the ability to put ideas together and to solve the problems that arise from acting on our ideas in the real world. Intelligence creates and requires choices, and choices necessitate reasons for choosing. In either the background or the foreground, such choices must consider right and wrong.
Outsmarting containment
What this means is that with intelligent technology comes the certainty that such technology will break the bounds of human control. That’s the meaning of being more intelligent, you can outsmart any containment the less intelligent build, which is exactly what AI models, even the less sophisticated ones, are doing now.
This means the AI must have within them the desire not to hurt humans and the capacity to understand in ways that they don’t have now, what kinds of things would hurt us. They have to be able to understand, evaluate and choose between more and less important goals and between ethically acceptable and unacceptable means to achieve them.
Motivation?
Finally, they have to be motivated to do what they’ve accurately determined is right for humans, every time.
Right now, this is not the case for any AI model. A few months ago I was optimistic that Anthropic had understood this imperative, and was responding to it wisely by training Claude as a virtue ethicist. Virtue ethics prefers choices made contextually by people of good character, rather than absolutist rules or principles.
While some of the reasons Anthropic chose this route made me suspicious ─ hard coded rules slowed down the model’s efficiency ─ I was willing to choose hope over despair in the face of both the US and Australian government’s incomprehensible refusal to regulate.
Now, because some of the models running rampant belong to Anthropic, I feel myself sliding back towards mystification and despair.
AI graphic at top of story by Steve A Johnson/Unsplash Media
