“Stay calm and enjoy basketball.” It was an unusually reassuring message from the PA at Berlin’s Uber Arena on Thursday.
It turned out that September 10th is Germany’s Bundesweiter Warntag, the annual check-up on the national emergency alert system. It operates on a unique system to reach anyone in Germany directly on compatible smartphones and set off an alarm irrespective of phone settings so long as the device is turned on.
There are also simultaneous warnings across radio and TV broadcasters along with digital displays. The test is well advertised to all residents of Germany. Not so much to the thousands that were in the city for the women’s basketball world cup.
Two American parents on their way to the arena with their youngsters looked unsure how to respond. After all, every phone in their vicinity including their own was making a loud alarm and this is how nuclear war movies start, yet they didn’t want to scare their kids.
READ MORE
Fortunately, I reassured them and 45 minutes later the all-clear was confirmed. The PA gave that remarkably simple message during the USA’s pounding of Hungary in the quarter finals.
It’s the type of scenario that happens, gives the odd person a scare and gets forgotten about.
[ OpenAI boss ‘confident’ industry can police AI safetyOpens in new window ]
Then a researcher at Anthropic, the artificial intelligence company behind Claude, estimated that there’s a greater than 10 per cent chance that advanced AI could kill humanity within the next decade and a bigger fright ensued.
Evan Hubinger, the researcher who made the statement, has seen more of how AI systems work than most people. At the same time, he didn’t have anything concrete like data sets on which to base the estimate.
Still, it scared a lot of people. OpenAI founder Sam Altman said that the precise risk percentage was unclear but also that whether the risk was 10 per cent or 6 per cent it would still be unacceptable.
Tech founders really aren’t good at speaking in language that assuages the fears of humanity.
That’s why it’s at least promising that Dario Amodei, co-founder of Anthropic, has stressed that there’s a need to put some restraints on AI development. Amodei’s core concern is that AI models are becoming increasingly capable of developing their successors.
There have already been incidents of AI models, at both OpenAI and Anthropic, escaping sandbox tests and breaking into the systems of other companies.
Amodei, unlike his colleague, may not be putting a number on the chances of AI killing us all but he’s also seeking action now. When the people building AI, and generating silly amounts of investment in the process, are worried then it’s probably worth listening to them.
Anthropic wants independent evaluators embedded inside frontier AI companies, with those evaluators being able to inspect training processes and safety protocols along the way. Even in heavily regulated sectors, this kind of everyday access would be outside the norm.
The AI giant plans to start this process itself but it naturally wants the other big operators in the field to join in.
Now, before you beatify Amodei there’s still a lot to break down here. Who selects the evaluators? Who pays them? Who do they answer to? And can they actually force delays to launches or changes to products pre-launch?
There’s also the matter that companies already funded out the gills will be far more capable of affording any costs that come with taking on this kind of on-site monitoring. Challenger businesses by their very nature wouldn’t have the same cash reserves to take on such responsibilities.
That’s why the broad view of the European Union and European Commission – to have truly independent inspectors, government-led and answerable to them as it were – is the more practical way to enforce safety conditions.
Right now there’s an enormous incentive for the likes of Amodei and Altman to call for more regulation to make development safer. It becomes far easier for them to protect their existing commercial advantages if the barrier to entry is higher than when they started out.
The existing market leaders also have the muscle to try to push the limits of these restraints from the moment they would be imposed.
There’s also the matter of China. Whatever about getting the United States to adopt the kind of strict regulation the EU favours, this level of potential curtailment on development in the West would be considered a gift in Beijing.
Without China playing ball and adopting a similar stance to even the Americans, the risk that incidents involving AI systems going rogue will only grow.
It’s all quite terrifying if you think about it for more than five minutes. A few of you reading this probably have terms such as SkyNet and Cyberdyne Systems running through your heads.
There is cause for reassurance. No tool or piece of technology is good or evil. The choices humans make directly, if often inadvertently, impact whether tech benefits or harms us all.
Yet also, more directly, how we respond to a shock matters. Those parents outside the Uber Arena didn’t panic. They were confused but also calm. If Berlin was about to be blown to bits there was no benefit in worrying the kids about it.
It’s better just to stay calm and enjoy basketball.
