AI · · 2 min read
Evidence-led AI, not hype-led AI
A language model is very good at sounding confident. That's exactly why the interesting design problem is making it show its work instead.
Ask a language model a question with genuine uncertainty in it, and it will usually still answer in a confident, complete sentence. That's not a bug you can prompt your way out of — it's how the technology is built. Which means if you're putting a model in front of real users making real decisions, the interesting design problem isn't the model at all. It's how you handle the gap between how confident it sounds and how confident it should actually be.
The easy version and why we avoid it
The easy version of "AI-powered" is a single output: a score, a prediction, a recommendation, presented as if it were simply true. It reads well in a demo. It's also the version most likely to quietly mislead someone who doesn't know to be skeptical of it.
We think that version is a shortcut that costs more than it saves, especially anywhere money, health, or a business decision is on the line.
What we build instead
The pattern we come back to, across very different projects, is roughly the same:
- Surface the underlying data, not just the conclusion. If a model says "this trend is likely to continue," the interface should make it trivial to see the actual historical data that claim rests on.
- Use the model to explain, not to decide. Language models are genuinely good at turning structured data into a clear, readable explanation. That's a narrower and more honest job than "decide this for the user."
- Design for the disagreement case. What does the interface do when a user's own judgment conflicts with the output? If there's no good answer, the model has been given more authority than it's earned.
This isn't caution for its own sake
None of this is about being anti-AI or slowing things down out of excess caution. It's a bet that the products people actually trust over time are the ones that let them verify a claim, not just receive it. Hype-led AI optimizes for the first impression. Evidence-led AI optimizes for the tenth time someone opens the product and checks whether it's still telling them the truth.
That's a slower thing to build. It's also the only version we're interested in shipping.