Can submarines swim?

The question whether machines can think is just as relevant and just as meaningful as the question whether submarines can swim.

Edsger Dijkstra wrote that line in 1983, in a position paper for a panel called “Computers and Society.” He wasn’t answering the question. He was refusing to ask it, and telling everyone else in the room to stop asking it too.

Read it now and the target has shifted, but the shot lands the same place. Swap “computers” for “AI” and most of the paper could run in this newsletter’s inbox tomorrow. Dijkstra’s real complaint wasn’t about machines at all. It was about the way computers got sold once they left the lab and became a product: universal usefulness (“this machine handles 200,000 decisions a second”), machine-versus-human theater (fast and tireless computer, slow and fallible you), and the idea that programming was just translation, a chore the machine would eventually do for itself.

He’d seen this promise fail before. COBOL was supposed to make professional programmers obsolete. Instead it needed them for maintenance, in numbers the sales pitch never mentioned. Ada was supposed to cut software costs through standardization and ended up too complicated for the people writing the standard to fully agree on. Dijkstra’s point wasn’t that the technology was bad. It was that “computer-aided” had quietly become a synonym for “improved,” unexamined, and every failed project traced back to someone accepting that swap without checking it.

That’s the part that reads like it was written about the current AI cycle rather than 1980s office automation. Every “AI-powered” label still carries the same unstated argument: if a computer touches it, it must be better. Nobody asks the follow-up question out loud, which is better at what, compared to what, for whom.

Dijkstra’s fix wasn’t technical. He didn’t propose a better standard or a smarter architecture. He wanted people taught to read an advertisement the way they’d read a claim from an encyclopedia salesman, which is to say, skeptically, and only after asking what’s actually being measured. That’s a harder sell than a benchmark chart, and it hasn’t gotten easier in forty years.

The submarine line is the part everyone quotes, and it’s worth sitting with why it works. Nobody argues about whether submarines swim, because the question doesn’t buy you anything. It doesn’t tell you how far the submarine can travel, how deep it can dive, or what breaks first. “Can it think” is the same kind of question dressed up as profound. It feels like it’s asking something important about machines when it’s really just asking about a word.

The AI debate right now runs on that exact confusion, just with better production values.