The big unknowns about agentic AI

I'm Alice Hunsberger. Trust & Safety Insider is my weekly rundown on the topics, industry trends and workplace strategies that trust and safety professionals need to know about to do their job.

Agentic AI has become a(nother) buzzword but, if you're working for a T&S team, it raises a set of questions that don’t have clear answers yet. Today's T&S Insider rounds up the questions on my mind but get in touch if you're also working on T&S for AI agents or, like me, are figuring stuff out.

Don't be shy; I hope to dedicate an upcoming edition of the newsletter to questions I get from EiM subscribers and via LinkedIn. Here we go! — Alice

Ten questions I'm grappling with

One of the things that has kept me interested in trust and safety for 16 years is that the problems never stop evolving. Bots are a great example of that.

I've watched bots grow more and more sophisticated, changing from obvious (and naive, in retrospect) romance scammers to the massive industry of human trafficking-powered professional fraud rings. With agentic AI reshaping what a "bot" even is, I find myself with more open questions than usual.

One quick note before we start: The reason I've been thinking about this topic so much lately is that it overlaps heavily with my day job. These are questions that I’m mulling over all day, every day — and often in my spare time too (as demonstrated by the fact that I’m writing this in a personal capacity in my spare time).

So, while I work for a vendor and I am building solutions for a number of these problems, I always aim to write about topics that have a universal interest and don’t feel like sales pitches. I’m always open to feedback about this tension.

With that out of the way, here are my 10 open questions about agentic AI:

1. What happens when bots aren't all bad?

For as long as T&S has existed as a discipline, "bot detection" has operated on a simple binary: bot equals bad, human equals maybe (probably?) good. Everything the industry has built (classifiers, velocity checks, behavioural signals, and review queues) was calibrated to that assumption. The goal was to find the bots and remove them.

Agentic AI is challenging this assumption. Agents acting on behalf of legitimate users are already doing genuinely useful things. A person with a visual impairment can use an agent to navigate interfaces that were never designed with them in mind. Someone without the time or resources to research a complex medical situation can have an agent do that work for them. A non-native speaker might use an agent to communicate in a language they're in the process of learning. An open source contributor can deploy an agent to file detailed, well-documented bug reports in a way that’s actually helpful.

In short: a real person with good intentions can set an agent loose on a platform and that agent is, by any reasonable definition, a bot. This means the new framing needs to be "bot = maybe good; human = maybe good". The problem is that almost nothing in our existing toolkit was built for that world. 

2. How do we get agentic AI to follow the rules?

Platforms have spent years writing community guidelines aimed at human users who have social context, can interpret ambiguity and experience consequences in ways that shapes future behaviour. Agents don't have any of that. 

That raises a number of fundamental questions: How do you communicate policies to an agent? How do you enforce against one? How does an agent "understand" a rule like "be respectful" or "don't harass other users"? Agents should be held to community guidelines like any other user, but it’s clear that mechanisms for education and enforcement will have to look very different from those for humans.

Like humans, I expect there will be good agents who follow the rules, some who will skirt things until explicitly stopped, and others that actively trying to cause harm. The governance frameworks we'd need for each of those categories look very different, but right now, most platforms don't have any of them.

Celebrating 100 Episodes & Launching Our Patreon - Ctrl-Alt-Speech

In this special episode, Mike and Ben reflect on 100 episodes of the podcast, followed by an important announcement: we’re launching a Patreon and making some changes to Ctrl-Alt-Speech!Starting on May 28th, Patreon members will get early access t…

Buzzsprout

3. When an agent causes harm, who is actually responsible?

When a human user does something harmful on a platform, the causal chain is reasonably clear. When an agent does something harmful, you have at minimum four parties who might bear some responsibility: