AI ethics are human ethics

Subscribe

“AI ethics” has come to mean two different things. But both lead to the same place.

In one interpretation, AI ethics is a branch of applied ethics, like biomedical or business ethics. Whether the technology under debate is gene editing or AI-generated art, the underlying questions stay the same. How do we want people to use this new technology? What do we want them to prioritize? And what do those choices say about what we value as a society?

The other interpretation treats AI as an actor making its own choices. There is, in fact, an entire genre of literature dedicated to AI doing exactly that. Mary Shelley’s Frankenstein, published in 1818, tells the story of the creature turning on the creator. By 1872, Erewhon by Samuel Butler presented an AI-like technology, exploring machine evolution and consciousness. These stories theorize the unique ethics of a superintelligence inside the machine, and how it may treat us.

For Lindsay Holmgren, Associate Professor (Teaching) of Strategy and Organization and the Director of the Laidley Centre for Business Ethics and Equity, these interpretations are two sides of the same coin.

“AI ethics is just human ethics,” she said on the McGill Delve podcast.

AI models are trained on infinite amounts of human-provided data. Social media comment sections, academic journals, daily news articles – these all feed into how an AI behaves and processes information. The AI then is not acting on its own values: it’s acting based on how it was trained, for better or worse.

And that’s where ethics start to get messy. For an AI to act morally on behalf of humans, humans must first agree on what constitutes “morality” – a debate that has plagued moral philosophers for thousands of years. And the way AI models are currently being developed, certain people have more say in this debate than others.

As companies race to build more powerful AI models, are humans ready to wrestle with such questions?

Looking in the AI mirror

Foundational language models, like ChatGPT, reflect human choices at every step of the design process. Their behaviour reflects developer choices, training data with human annotations, and an immense corpus of our own cultural data now inextricably embedded within the models, she said.

“AI doesn’t just develop ethics,” said Holmgren, “it all derives from humans.”

AI is a useful mirror for human behaviour and values. And we might not always like what we see. Researchers, the public, and even Silicon Valley leaders have voiced fears around what AI could do, that there is a non-zero chance it could destroy humanity. But whatever AI does, said Holmgren, it learned from us.

The example we have set for it is not reassuring, she said. Humans have not always used their capital to protect others. Much of the comfort and excess of modern life comes at a cost to the planet, species, and some of the people we share it with. If AI is to overcome these ethical shortcomings, humans must reckon with them first.

That reckoning is difficult, said Holmgren.

“Human beings have had an incredibly hard time articulating what our values are,” said Holmgren.

But perhaps AI gives us an opportunity to properly interrogate these questions now.

Recently, calls for an AI slowdown have grown louder, citing the need for stronger safety and alignment procedures. Alignment describes the work of ensuring AI understands and acts in alignment with human values. For Holmgren, however, the important question isn’t how fast we can align AI.

“It’s a question of alignment with what,” she said.

AI developers have gone on the record calling for a slowdown of AI development, in part to develop safer AI models. But it’s also a chance to develop better answers to the question of who we are as a species, and what we want AI to be for us, said Holmgren. These questions go beyond the granular details of the technology itself, and into the arena of moral philosophy.

Literacy is an investment in ethics

Answering these moral questions requires a certain type of thinking, a moral literacy, said Holmgren. It requires people to introspect and come to grips with their own values. Ironically, this kind of literacy is being threatened by the use of AI. If we rely too much on chatbots to help us think, we may lose our ability to independently identify our own sense of morality.

Moral literacy isn’t about writing a list of rights and wrongs, but rather the ability to reflect and maintain an internal private moral compass. This cognitive capacity is built through language, she said.

But words and symbols come to mean very different things to very different people. According to Holmgren, discovering what it means to you is the work of becoming a moral person.

“Ultimately, we come into ourselves, our understanding of everything about what and who we are through language,” she said.

AI-powered large language models also work with words as symbols. They replicate patterns of association in our language to generate text accurately. Generating the word “ethics”, however, is not the same as reflecting on what it means. The habit of thinking through what each of those words represents is crucial to understanding what moral means, or what’s right or wrong to you, said Holmgren.

“If we’re turning that activity over to large language models…you’re handing over your agency, and you’re handing over your capacity to reflect,” she said.

According to Holmgren, we are not powerless against this. These moral questions don’t just belong to developers or the philosophers hired by AI safety teams. They belong to all of us; in classrooms, workplaces, and public debate.

“It’s not AI we need to be concerned about,” said Holmgren. “It’s ourselves.”