Can a tiny AI group stand up to Google?

• It is telling that Timnit Gebru’s new firm recruits in SA, not Silicon Valley

Artificial intelligence (AI) is not always so smart. It amplifies outrage on social media and struggles to flag hate speech. It designates engineers as male and nurses as female when translating language. It fails to recognise people with darker skin tones.

Systems powered by machine learning are amassing greater influence on human life, and while they work well most of the time, developers are constantly fixing mistakes like a game of whack-a-mole. That means AI’s future impact is unpredictable. At best, it is likely to continue to harm at least some people because it is often not trained properly; at worst, it will cause harm on a societal scale because its intended use is not vetted think surveillance systems and facial recognition.

Many say we need independent research into AI, and good news on that came last week from Timnit Gebru, a former ethical AI researcher at Alphabet’s Google. She was fired exactly a year ago after a dispute over a paper critical of large AI models, including ones developed by Google. Gebru is starting DAIR (Distributed AI Research) which will work on the technology “free from Big

Tech’s pervasive influence” and probe ways to weed out the often deeply embedded harms.

Good luck to her, because this will be a tough battle. Big Tech carries out its AI research with much more money, effectively sucking oxygen out of the room for everyone else. In 2019, Microsoft invested $1bn into OpenAI, the research firm cofounded by Elon Musk, to develop the massive languagepredicting system GPT-3. A Harvard University study on AI ethics, published last week, said investment went to a project run by just 150 people, marking “one of the largest capital investments ever exclusively directed by such a small group”.

Independent research groups such as DAIR will be lucky to get even a fraction of that kind of cash. Gebru has lined up funding from the Ford, MacArthur, Kapor Center, Rockefeller and Open Society foundations, enough to hire five researchers over the next year. But it is telling that her first research fellow is based in SA and not Silicon Valley.

Google’s AI unit DeepMind has cornered much of the world’s top talent for AI research, with salaries in the range of $500,000 a year, says one research scientist. That person was offered three times their salary to work at DeepMind. They declined, but many others do not. The promise of proper funding, for stretched academics and independent researchers, is too powerful a lure as many reach an age of families to support.

In academia, the growing influence of Big Tech has become stark. A recent study by scientists across multiple universities including Stanford showed academic research into machine learning saw Big Tech funding and affiliations triple to more than 70% in the decade to 2019. Its growing presence “closely resembles strategies used by Big Tobacco,” the authors of that study said.

Researchers who want to leave Big Tech find it almost impossible to disentangle themselves. The founders of Google’s DeepMind sought for years to negotiate more independence from Alphabet to protect their AI research from corporate interests, but those plans were nixed by Google in 2021. Several of Open AI’s top safety researchers left earlier this year to start their own San Francisco-based company, called Anthropic, but they have gone to venture capital investors for funding. Among the backers: Facebook cofounder Dustin Moskovitz and Google’s former CEO Eric Schmidt. It has raised $124m to date, according to PitchBook, which tracks venture capital investments.

“[Venture capital investors] make their money from tech hype,” says Meredith ’Whittaker, a former Google researcher who helped lead employee protests over Google s work with the military, before resigning in 2019. “Their interests are aligned with tech.”

Whittaker, who says she would not be comfortable with venture capital funding, cofounded another independent AI research group at New York University, called the AI Now Institute. Similar groups that mostly rely on grants for funding include the Algorithmic Justice League, Data for Black Lives and Data and Society.

Gebru at least is not alone. And such groups, though humbly resourced and vastly outgunned, have through the constant publication of studies created awareness about previously unknown issues such as bias in algorithms. That has helped inform new legislation such as the EU’s upcoming AI law, which will ban certain AI systems and require others to be more carefully supervised.

There is no single hero in this, says Whittaker. But, she adds, “we have changed the conversation.”