Ask Forum

The problem people actually had

Forum has been on the air for more than three decades. Thousands of hours of interviews, conversations, insights. But if you ever wanted to find a past interview your options were limited.

KQED launched a Forum community on Discord as an experiment. Discord is a strange home for a public radio audience at first glance, and it turned out to fit, because community radio is about conversation and so is Discord.

Friends would describe older Forum episodes to me. Often they were vague. “I think it was a housing interview, from a while back, with the woman who used to work for the city.” If I knew the episode it was an easy find. If they had to search online, it wasn’t. No title, no date, no guest name. It’s how a person actually holds something they heard once, and every search box we had would have returned nothing.

My job was to figure out how that experience could work differently. I led a working group of senior editors and leadership through what our relationship to AI should be, and built prototypes to show what it looked like in practice.

For an archive tool the risk is specific. A system that invents a guest, or attributes a claim to someone who never made it, undoes something that took thirty years to build. Trust is the whole asset in public media. It is not a feature you add later.

So what is Ask Forum

A bot in the Discord you could ask the way you’d ask another listener. Describe what you remember and it looks for it, rather than requiring you to already know the title.

The episode details it shows you are exactly what’s published on kqed.org. No model touches them. The language model is only used to work out what you meant, never to write what you see, so it can’t invent a guest or a date. When there’s a transcript, you can ask about what was actually said, and the interviewee’s words are marked as quotes even though the model writes the sentences connecting them. The line between who said what stays visible.

And when the system changed your search terms to find a match, it said so. Search for “a spicy conversation” and one time you’d get an episode about Bay Area hot sauces, another time an interview about turning conflict into connection, with a note at the top telling you the phrase had been read as “fiery” in one case and “intense” in the other. That’s the part most products hide, because admitting the machine reinterpreted you feels like admitting it failed. I think the opposite. A person who can see how their question was read can ask a better one next time. Hiding it leaves them guessing at why they got what they got, which is the same problem as having no menu.

That sits inside a broader commitment we had made: when AI is involved in creating or displaying something, tell people. The disclosure was not a nicety bolted on at the end. It was the thing that let us ship an AI feature in a newsroom at all.

What happened

Usage in Discord was limited. A new space, a new modality. But trying something new is what public media excels at, and to make change you have to try new ideas. I’m proud of Ask Forum, and it left us with patterns we kept using, the disclosure rule most of all, at a time when there were no playbooks to point to.

What did we learn