Insights · AI Adoption
"Human in the loop" is not a safeguard. It's a job.
13 October 2026 · Emma Ockelford
Every AI conversation in local government ends with the same reassurance: there's a human in the loop. New research from social work suggests we need to be much more specific about what that human needs.
Every AI conversation in local government eventually reaches the same line: don't worry, there's a human in the loop.
I understand why. In statutory services, people need to know that a professional, not a system, is accountable. But I've started to think the phrase hides more than it reveals.
The Ada Lovelace Institute's study of AI transcription in social work is one of the most useful pieces of evidence the sector has had this year. Researchers spoke to 39 social workers across 17 councils.
The headline is genuinely positive. Almost all of them said the tools brought meaningful benefits, and many described time freed up for the relational work that brought them into social work in the first place.
The detail is where it gets interesting. Social workers take full responsibility for what the tool produces, yet their views on its accuracy, and how much checking it needs, vary widely. And most councils evaluate these tools on efficiency, with far less attention on the experience of the people receiving care.
One line from the analysis that followed has stayed with me: a human in the loop isn't a magic fix. It's work. It takes time, skill and capacity, exactly the things an overloaded team is short of.
At an LGAi session earlier this year, someone put it brilliantly. The standard shouldn't be a human in the loop. It should be a human at the helm.
The difference matters. A person in the loop can end up approving what a system has already framed. A person at the helm sets the direction, has the information to check, and has the authority to say no.
So the real question isn't whether there's a human in the loop. It's whether that person has the time, information, authority and confidence to use real judgement.
If you're designing or buying AI for statutory work, that translates into a few practical tests. Can every statement in a draft be traced back to its source, so checking takes minutes rather than a re-read of the whole file? Is the practitioner checking against an agreed local standard, not their own memory of one? Do managers decide when a tool is ready, after testing it on real local cases? Is some of the time saved deliberately given back to review? And do the measures include whether people understand the decisions made about them, not just hours saved?
That's how we've tried to build our own tools. In EHCP+, each council signs off five plans produced with its locally configured tool before the team starts using it, and every draft is cited back to the advice it came from. The caseworker owns the plan.
The Ada Lovelace Institute recommends more transparency reporting and a What Works Centre for AI in public services. I'd welcome both. The sector shouldn't be evaluating these tools one council at a time.
AI can give practitioners real time back. It will only stay safe if we're honest about the work that keeps it that way, and design for it from the start.