���˾þ�

Some of the biggest challenges in artificial intelligence are being worked on not by computer scientists head down in code but by philosophers lured from academia into jobs at AI firms. The philosophers are tasked with making the next generation of models more capable and reliable, but they also shed light on the mystery of consciousness and whether intelligence can be replicated in software alone.

at the London School of Economics and Political Science says AI companies are the big employers of philosophy PhDs right now, with offers of interesting work, large salaries and stock options proving too tempting for many to resist.

“Topics that have been researched in philosophy departments for decades – how to make rational decisions, how to systematise moral principles, what counts as thinking or reasoning or introspection, what counts as evidence of consciousness – are suddenly of massive value to AI companies,” says Birch. “So, naturally, we are seeing a huge brain drain.”

for this crop of philosophers is alignment, the AI industry term for efforts to stop harmful content – such as revealing instructions on how to make bombs.

Efforts to stop AI models giving dangerous outputs were initially focused around putting in simple black-and-white guardrails, such as forbidding a model from talking about bombs entirely, for instance. But these proved clumsy and easy to circumvent. Now, companies are pursuing more advanced methods which lean heavily on philosophical understanding of right and wrong.

Free newsletter

Sign up to The Daily

The latest on what’s new in science and why it matters each day.

Sign up to newsletter

However, it’s rarely straightforward. Researchers have found that if you tell a model to break a rule in one specific situation, it will start breaking lots of other rules, says at the University of Exeter in the UK. And understanding why is exactly the sort of problem that philosophy’s logical analysis can unpick.

“The best explanation for that seems to be that there’s a semantic link deep in the corpus of texts it’s trained on that holds the good-coded things and the bad-coded things together. And once you allow it to do some bad, it picks up on that and extrapolates and starts doing other things that are bad,” says Glackin. “As an ethicist, one way of seeing what we’re doing is trying to work out the shape and the extent of concepts like right and wrong, or good and bad, and trying to work out what things fall under it, in common usage or conceptually, and that seems to be exactly the sort of analysis that the LLM is doing.”

Others key tasks for philosophers at AI companies include cutting down on hallucinations – the industry term for fabrications produced by models, generally improving performance and tackling biases inherent in the models. They also apply theories of human consciousness to AI models in an attempt to answer the persistent question of whether they display sentience.

“What do minds do, what do brains do, what can be replicated? This is a big issue for AIs. It’s a really important question for those companies to answer and it’s something that philosophers have been thinking about for a long, long time,” says Glackin.

The thorniest questions

at the University of Edinburgh, UK – who knows of two academics who’ve jumped ship to AI companies, and has also taken on advisory work for a commercial outfit himself – says that philosophy and computer science share a long history. In fact, the paper in which Alan Turing unveiled his famous test, known as the Turing test, to determine if a machine could display traits of intelligent beings was published in the philosophy journal Mind.

The exact scale of hiring is hard to estimate, but Aaron Kagan, chair of the American Philosophical Association’s Committee for Non-Academic Careers, has delved into job adverts for some insight. “A naive keyword count suggests 26.6 per cent of roles mention AI ethics, safety, alignment, governance or policy, but after removing boilerplate, only about 5 per cent substantively involve that work,” says Kagan.

Almotahari sees value in philosophy expertise for technology firms, but is sceptical that the thorniest questions of machine consciousness will be answered by them. Instead, he says philosophers could succeed in helping engineers to pick apart what’s going on inside models.

“There’s all this math taking place: can we extract from it a higher level description of what’s going on in terms of, say, this part of the model is representing that feature of the world, and this part is representing this other feature of the world?” asks Almotahari. “I think philosophers are well positioned to do that, to go from the engineering description to a representational description.”

Others see a potential problem on the horizon, with industry hiring philosophers eventually leading to biased research that serves the interests of technology companies.

“Going forward, it’s clear that a lot of serious philosophical work will be funded by industry. Explicitly or implicitly, companies have expectations they want to be fulfilled, and they have the power to favour authors who deliver welcome arguments and ideas,” says Birch. “I wish we had made more progress on the big questions of philosophy – about consciousness, agency, morality, etc – prior to the advent of AI. If we had, we would have been better prepared. AI has now given these questions massive urgency, and yet, the answers still seem far away.”