June 15, 2026 By: JK Tech
We’ve been thinking about AI wrong.
Every benchmark, every product launch, every breathless news article focuses on the same thing: outputs. Can the model answer this question correctly? Can it write better code than last year’s version? Can it beat a human at some task we’ve decided matters?
Nobody talks much about what happens before the answer. The part where you figure out what to even ask.
A team of researchers from MIT and Harvard started poking at exactly that problem, and they used Battleship of all things to do it. Which sounds ridiculous until you think about it for a second. The game is basically a pure test of information-seeking. You know nothing, your opponent knows everything, and the only tool you have is the quality of your next question. It’s actually kind of perfect.
The Gap Nobody Talks About
Here’s the thing about how AI gets built. Models are trained to respond to what’s in front of them. Feed them a question, they produce an answer. Feed them a document, they summarize it. The whole setup assumes the relevant information is already present.
Real life doesn’t work that way.
A doctor sitting with a patient who has vague symptoms doesn’t have all the information yet. Neither does a scientist trying to figure out why an experiment failed, or an analyst trying to understand why sales dropped in one region and not another. What they have is a starting point and a need to ask the right questions in the right order to close the gap.
Humans are surprisingly decent at this. We notice what’s missing. We triangulate. We ask something, listen to the answer, and let that shape what we ask next. It’s intuitive to the point where we don’t even notice we’re doing it.
AI mostly can’t do this well. That’s what the research actually found.
What Battleship Revealed
The setup was simple enough. One player knew where all the ships were. The other had to find them through questions and guesses. The researchers weren’t scoring wins and losses. They were watching how each player, human or AI, decided what to ask next.
A lot of the AI models were bad at it. Not broken, just inefficient. They’d ask questions that didn’t really narrow things down much. Or they’d get a useful answer and then not quite factor it in properly the next time around. Human players were more adaptive, better at sensing which question would do the most work given what they already knew.
This isn’t a small problem. The ability to seek out information strategically is pretty fundamental to anything we’d actually want AI to help with in the real world.
The Fix Was Conceptually Simple
The researchers tried a different approach. Instead of having the AI just pattern-match its way to the next question, they gave it a step where it explicitly thought through which question would most reduce its uncertainty. Not which question felt natural given its training. Which question would actually move the needle.
It worked noticeably well. Small models that used this approach started outperforming much bigger ones that didn’t. And the resource gap was significant. You could get better results with a fraction of the compute just by changing how the model reasoned about what to ask next.
That should make people in the AI industry at least a little uncomfortable, because the standard playbook is to scale everything up. More parameters, more data, more hardware. What this suggests is that there are gains sitting on the table that have nothing to do with size.
Why This Actually Matters
You can wave away a Battleship experiment pretty easily. It’s clean, it’s controlled, it’s not the messy complexity of anything real. Fair enough.
But the researchers tested similar reasoning approaches in other domains and the results held. And when you think about what kinds of tasks people actually want AI help with, a lot of them involve incomplete information. Medical diagnosis. Scientific research. Business decisions. Legal analysis. These aren’t lookup problems. They’re search problems, where the skill is in knowing what to look for next.
An AI that can do that well is qualitatively different from one that can only process what it’s already been given. It stops being a very sophisticated search engine and starts being something closer to an actual thinking partner.
The Question Behind the Question
There’s something almost ironic about this research. We’ve spent years and staggering amounts of money teaching AI to produce better answers. And the finding that might matter most is that we should have been paying more attention to questions.
Human curiosity isn’t just some charming quirk. It’s load-bearing. The reason humans have figured out so much about the world is that we are relentlessly, almost annoyingly, driven to ask why, how, what if, and what happens next. That drive shapes what we pay attention to and what we bother to investigate.
If AI can develop something functionally similar, even a pale version of it, the ceiling on what these systems can do moves up considerably. Not because they’ll know more facts. Because they’ll know how to go find the facts they’re missing.
That’s a different kind of intelligence. And honestly, it’s probably the more important one.
