Accelerationism will only slow AI’s progress

The bosses of the biggest AI companies are convinced that machine intelligence will soon surpass the human version. But we may get there quicker than we think if we keep degrading human intelligence.

When AI companies admit to releasing powerful models they do not fully understand and cannot control, then it is probably worth paying attention to this matter. All the more so, when a succession of former employees of Anthropic, OpenAI and Google DeepMind say these companies do not take safety seriously enough and cannot be trusted.

But in his immense vanity, Donald Trump has declared that a “strong and smart (high IQ!) president” is the only control mechanism the industry needs. The US leads China and the rest of the world in AI and Trump wants to keep it that way. “Whoever wins AI, wins!” he posted.

The accelerationist wing of Silicon Valley, which has been lobbying against any form of AI regulation, is cheering the boomer-in-chief. But US voters appear to have concluded that Trump is on the wrong side of history on this one. Some 60 per cent of Americans now support AI regulation, according to an opinion poll published by the Special Competitive Studies Project and Gallup this week. Only 7 per cent seem to agree with Trump in prioritising speed of development over safety rules.

One danger for the accelerationists is that any AI-induced disaster, which appears increasingly likely without stronger safeguards, risks triggering a furious public backlash and a regulatory overreaction, slowing future progress. AI safety campaigners point to the example of the US nuclear industry following the accident at Three Mile Island in 1979. As a result of public fears, the commissioning of new nuclear plants came to a shuddering halt for the rest of the century. By ignoring today’s legitimate public concerns, AI accelerationists are jeopardising the future of the technology they love.

In their defence, accelerationists make two points worth debating.

First, they argue that the leading AI companies secretly want regulation because it will entrench the power of incumbents and diminish competition. That may be a risk. But we would not allow pharmaceutical companies to release untested drugs on the market because a tough approval process suppresses competition. Besides, the fiercest challenge to incumbent AI companies may well come from rivals developing more robust models that can be used more reliably in specific domains. By design, they would have little problem complying with safety regulations.

Second, the accelerationists argue that the US must win the AI technology race with China. But technology is not a zero-sum game. What does winning even mean? Any advantage that the US gains is only likely to prove temporary. And, as the tech analyst Dan Wang told the FT Weekend Festival this month, Beijing is probably happy enough for the US to race ahead on frontier AI and court disaster. China can then avoid those mistakes and apply the technology more widely.

The exceptionalist fallacy is to believe that there is something so magical about AI that only a handful of technological wizards should be able to do tricks with it. The risks of AI are not inherent in the technology but in the ways that companies train, design and release their models. “These are intentional choices that have led to incredibly reckless, negligent and dangerous technology,” says Andrew Strait, a former researcher at the UK’s AI Security Institute. “That is the main problem we now face.”

The broader design choices we make about how to build, fund and deploy AI over the next 12 to 18 months will be critical for humanity’s future, argues Bill Gates, in his foundation’s latest Goalkeepers Report.

AI could bring massive economic, health and educational benefits to people around the world. To that end, the Gates Foundation is spending $1bn over the next two years to promote initiatives in these areas.

But Gates tells me that the speed and scale of AI’s development is “utterly different” to anything else in history and poses unique challenges. Machines are becoming smarter than humans across so many domains that it is more like an evolutionary than a technological transition. He says he will stake “whatever credibility” he has on the statement that “people should not lull themselves into thinking this will all work out because it has for past technologies”.

We urgently need to maximise our collective human wisdom to make the most of machine intelligence.

[email protected]