The Wrong Framing in AI and T&S

I recently was able to attend the TSPA EMEA Members summit in Dublin, and i walked away with a lot of thoughts and ideas.

There were a number of session i got to attend that were genuienly thoughtful and posed some interesting questions around some of the most salient topics at this time, such as AI. However even before the event there was one question that kept coming up in conversations about AI and content moderation, and it is one was framed in a different way that i hadn’t genuinely considered before.

From the discussions in one of the panels around AI, it was suggested that the instinct for us is to ask: can AI automate this?

Which is the wrong way to frame it. The more interesting one should be: what does AI make possible that wasn’t practical before?

They sound similar but they’re not. One is about replacing existing work and the other is about doing things that simply weren’t feasible at human scale such as one of the provided examples from another attendee in triaging thousands of linked entities in an investigation, surfacing patterns across languages and platforms simultaneously, identifying coordinated behaviour before it metastasises. That’s simply not something a normal human being is capable of in such a short period of time. This is instead a completely different level of capability available to us.

But relying on automation can still produce some surprises for people expecting only improvements. There have been many cases where rolling out automated moderation didn’t produce the expected volume reductions, but rather just create more tickets, more appeals, ad more issues. The work didn’t go away at all, it just got redistributed.

The take-away from it isn’t that automation is wrong here, but rather that the wrong things were selected to be automated and it’s a distinction that many will learn the hard way i’m afraid.

The harder version of the question though is this: if AI takes on more and more of the entry-level moderation work, where does the next generation of practitioners come from? That work isn’t just throughput. It’s where you build judgement. Where you make the mistakes that shape your instincts for the worse cases later. You can’t shortcut that by skipping it.

Two approaches get floated. One is deliberate preservation — protecting certain categories of work from automation by design, specifically to maintain training exposure. The other is simulation — fake queues, synthetic reports, constructed environments. I lean hard towards the first. Real instincts come from real cases. Simulated environments can supplement, but i don’t think they can replicate what it actually feels like to get something wrong on something that matters.

And then there’s the part that probably sits most uncomfortably: the same question — what does AI make possible that wasn’t practical before? — is being asked by the people on the other side too. Networks that were previously hesitant around AI tools are now actively encouraging their use. Output that would have been obviously synthetic a couple of years ago increasingly isn’t.

We’re not ahead of this. Most people working seriously in this space will tell you the industry is barely keeping pace, if that.

So the reframe is useful. But it cuts both ways.