Why global effort is needed to regulate AI industry: David Richards
Back in 2016, Mark Zuckerberg said: “We are a tech company, not a media company. We build the tools. We do not produce any content.”
Sounded perfectly reasonable at the time, didn’t it? You build the road, you’re not answerable for who drives on it or where they end up. Clean hands.
Turned out to be the perfect get-out clause. An entire industry reshaped childhood, warped how elections work, and fractured any shared grip on truth. Spent a decade insisting none of it was their fault.
We know what it cost now. Teenage mental health has gone off a cliff. Anger spreads faster online than argument ever could. Democracies feel like they’re held together with string. And this isn’t abstract policy wonkery we’re discussing in university lecture theatres.


Its showing up in actual homes, in people’s kids’ schools, in the basic fabric of how we live together.
Now AI has arrived making exactly the same arguments. Sharper tools this time. Everything moving faster.
I sat down with George Osborne in 2013 when he was Chancellor. Social media was still riding high off the Obama campaign; everyone was convinced it was this unstoppable political force. I wanted to talk about children and what these platforms were secretly doing to them. He was more interested in how to harness them to win elections. That wasn’t cynicism on his part, just politics doing what politics does.
We cannot afford to run that same script with AGI.
The pattern’s familiar because we’ve watched it before. In the early social media days, tech leaders insisted the internet was fundamentally different, somehow exempt from boring old regulation. That would strangle innovation, hand advantages to foreign competitors, and slow everything down.
Politicians barely understood the business models and went along with it.
Self-regulation filled the vacuum. Platforms wrote their own rules, hired compliance teams, and made promises about proper policing. They moved at startup speed. Governments lumbered along at legislative pace. By the time the harm became obvious, the whole thing was so embedded in daily life that meaningful reform had become politically impossible.
AI companies are now offering their own version of exactly the same defence. Too complex for traditional liability frameworks, they say. Outputs are unpredictable, so responsibility sits with users. Training data scraped from the web? Fair game apparently, consent negotiable.
AGI changes the whole equation though. Systems that can reason across different domains and operate with genuine autonomy aren’t just shifting information around. They’re forming decisions, concentrating power, creating dependencies we probably won’t fully understand until its too late. Get this wrong and we’re not looking at damaged mental health stats or a messier public square, serious as those problems are.
We’re looking at fundamental dependence on systems we can't properly audit, can't fully understand, might struggle to control once they cross certain capability thresholds.
So regulation has to focus on chokepoints where control is actually feasible. Not the impressive-sounding bits that prove impossible to enforce in practice.
Start with compute. Frontier AI, meaning the most powerful models able to generalise across tasks and act with limited human oversight, has something social media never did: physical limits you can see and measure.
Training these systems requires huge data centres, specialist chips, and power use large enough to show up on the grid. That is the chokepoint.
Facilities with enough compute to train frontier models should be required to register, disclose capacity, and submit to inspection. Large training runs and power draw should be monitored, with regulators given authority to stop anything that appears to be pushing beyond defined risk limits. You may not see inside the model, but you can control the hardware needed to build it.
Then there is licensing before deployment. Social media worked on launch first and apologise later. With AGI, that approach is indefensible.
Models above clear capability thresholds should need licences before training and again before release. Those licences must depend on serious adversarial testing designed to break the system. If a model shows deception, bypasses safeguards, or opens credible paths to catastrophic misuse, it does not ship.
We already apply this logic to aircraft, medicines, and nuclear facilities. Frontier AI belongs in the same category.
None of this works without international coordination. Tough rules in one country simply push reckless operators elsewhere. Compute governance needs treaty-level backing, shared standards, and supply chain leverage. Countries that sign up get collaboration and access to responsible deployment.
Those who opt out should find advanced chips much harder to obtain. The aim is not punishment. It is avoiding a race to the bottom where safety becomes optional.
This doesn’t suppress innovation. It makes progress sustainable rather than reckless. Innovation without public trust is just gambling with other people’s futures.
We told ourselves we’d sort out social media’s mess later. Later never came. The problems became permanent features of the landscape. With AGI, delay isn’t caution. Its abdication.
David Richards is managing director of Yorkshire AI Labs