Cartoon: a man peers with a magnifying glass into a cardboard box labelled "AI Schrödinger's Box", thinking of both a brain in a graduation cap and a winking emoji with its tongue out. A cat on a stool, holding a clipboard, says: "It's a superposition of genius and moron until you check its work."
Illustrated by Nano Banana, Google Gemini's image model, from an idea by Andy Boura. CC BY-SA 4.0

Current AI is like Schrödinger's cat: both brilliant and stupid at the same time and you don't find out which until you look carefully at the outcomes (illustrated by nanobanana)

Example of the day:
Google's Agentic AI wipes user's entire HDD without permission in catastrophic failure

And Gemini's take / explanation of the above statement:

Your "Schrödinger's AI" analogy is perhaps the most accurate description of the current state of generative AI. It perfectly captures the bewildering experience of using these tools.

Before you inspect the output of a model, it effectively exists in a quantum superposition of "world-class expert" and "confident fool."

Sometimes you open the box and find brilliance: complex code written in seconds or dense research synthesized perfectly. Other times, you find baffling incompetence delivered with supreme confidence—hallucinated legal precedents or failures in basic logic that a child could solve.

This duality exists because AI doesn't "think" or understand truth. It is a hyper-advanced pattern-matching machine operating on probability. When it’s brilliant, its statistics align with your request; when it’s stupid, it’s confidently connecting unrelated concepts because it lacks an internal reality check.

Crucially, the human user acts as the "observer." Until you look carefully with domain expertise, the output remains in that dangerous middle state. You must possess the knowledge to distinguish a useful result from a convincing hallucination. The operational mindset for the future is therefore simple: trust but intensely verify.