Every group agreed: AI in justice needs human verification. Then came the uncomfortable part… humans make mistakes too.

Last week I joined the workshop “Transitional Justice in the AI Era: Challenges and Opportunities for Peacebuilding” at the Geneva Peace Forum & Geneva Peacebuilding Platform (GPP). We worked in groups, with people from NGOs, lawyers and practitioners, and our group was supervised by Dr. Kamel El Hilali.

I don’t work in the justice system. I was one of the few people in the room working every day in a digital environment and that’s exactly why the exchange worked. They brought the reality of the field, I brought how AI actually behaves once it’s inside an organisation’s workflow.

What the groups agreed on:

- One human check isn’t enough. You need two layers of validation: several people, or an independent mechanism.

- The most critical decision happens at the very beginning: who selects and trains the data.

- The costs go both ways. Building AI agents costs money, and so does the human expertise needed around them. You have to balance both.

The point that stayed with me most: many organisations work remotely, so they apply AI to information from a country they don’t have enough data about. That’s where bias comes in, quietly. And in transitional justice, a biased output isn’t a bad report. It’s someone’s story told wrong.

It reminded me why I work the way I do. As an AI consultant and trainer I need to understand how organisations work, what professionals actually need, and the bigger picture, before suggesting any tool. I’ve worked with an NGO in Geneva before, so this room felt natural to me. I’m in the field, LISTENING, before offering magical solutions.

The two-layer rule isn’t only for courts. Any team publishing AI output after one quick read has the same gap, just with lower stakes.

So, in your team: who checks the checker? / PS: Yes, I took the obligatory photo in front of the logo 😄