Why multi-purpose systems are hard to design safely

A non-AI analogy to explain a difference between generative and traditional enterprise AI systems

Cassie Kozyrkov
5 min readMay 31


In the previous installment of this series, I told you that solving trust and safety for traditional enterprise-scale AI systems is a walk in the park compared with trust and safety for generative AI. Let me summarize the two key insights from that article:

  • Most enterprise-grade AI systems of the past decade were designed to do one very specific thing measurably well at scale.
  • It’s a lot easier to protect a varied group of users from a single-purpose system than to protect the same group from a multi-purpose system.

I’ve written this article for those who felt that those ideas zoomed by too fast and need a bit of digestion. (If that’s not you, head straight to Part 4.)

So, to understand these ideas in a non-AI setting, imagine you’ve been tasked with designing a set of bread-cutting stations for shoppers to use in every grocery store in a major national chain.

Image: Wikipedia

That’s a lot of stores and a lot of bread cutting stations, which might be serving millions of shoppers. That’s plenty of opportunity to “meet the long tail” — which is data science jargon for the atypical entities.* You’d better prepare for the oddest situations, so buckle up.

Some of those folks will be shopping drunk, so you’d better think of that when you’re designing the bread cutter.

Kids will try to use the station even if you put up indemnifying warnings that it’s for use by grown-ups only. Think of that too.

What about people with visual impairments? You’d better plan for their interactions with it.

What about the geniuses who use it slice everything except bread? And so on…

This is hard. And it’s how enterprises have to think about solutions at scale. But it’s not thaaaat hard. Not compared with safely deploying a multi-purpose system, as we’ll see in a moment.



Cassie Kozyrkov

Chief Decision Scientist, Google. ❤️ Stats, ML/AI, data, puns, art, theatre, decision science. All views are my own. twitter.com/quaesita