Surely, AI Can...

AI can do a lot. I use it every day, and I used it to make this website.

The prevailing message, especially from large AI companies raising enormous funding rounds, is that AI can do almost anything. It seems like many people (and the financial market) buy into this narrative and believe these systems are close to infallible now, despite the small warning at the bottom of every chat window that says “AI can make mistakes”). I get the sense that this is especially true for people who don’t use AI very often.

But if you’ve ever tried to do something ambitious or precise with these tools, you know that AI fails. Often. And often in ways that are hilarious.

“Surely, AI Can…” is a showcase of those moments, when a system confidently fails at something that feels like it should be well within reach. The point isn’t to mock the technology or dismiss its usefulness. It’s to puncture the idea of infallibility that tends to accumulate around frontier models once they’re wrapped in marketing language and AGI-adjacent rhetoric.

This project is inspired by some other great projects I want to highlight:

AI World Clocks by Brian Moore — clocks.brianmoore.com
Clock Bench by Alek Safar — clockbench.ai
The OuLiBench Benchmark — aclanthology.org

Try shuffle Try head-to-head