Online Safety reforms target AI chatbots and deepfakes. The hard part is enforcing speed, not writing rules

The UK’s fast-tracked Online Safety reforms aimed at AI chatbots and non-consensual deepfakes are a reminder that the technology problem is now an enforcement problem. The rules already look clear: illegal content is illegal, even when a model generated it. The hard part is making removal fast enough to matter. 

Most platforms can remove content when forced to, but ‘notice and action’ still breaks down in practice: confusing reporting routes, slow verification, and teams that are under-resourced compared with the volume of synthetic media. AI has made the harm scalable, and that means enforcement needs to be engineered for scale too. 

A 48-hour removal expectation is only meaningful if regulators demand operational proof, not policy PDFs: response-time metrics, escalation paths, and independent testing that shows reports are handled consistently for real users. Reporting also needs to be usable by non-experts, not buried behind form-filling that quietly discourages action. Victims need speed, not sympathy. 

The deeper issue is incentive design. If the cheapest path is delay, denial, or automated replies, that is what the system will produce.

• Should speed-of-removal metrics be publicly reported for the largest platforms?
• How can regulators verify that ‘AI detection’ is not just marketing copy?
• What due process is needed to prevent over-removal while still acting quickly on harm?
• Who should bear liability when AI-generated content is distributed at scale: the platform, the model provider, or both?