Five hundred safety-critical events in a hundred days. That is the count Zuduo Zheng and his colleagues logged after driving a Tesla Model Y with Full Self-Driving every day on Queensland roads — more than 500 moments where a human had to grab the wheel, or where the system revealed some flat misreading of what a road actually is.
In the same month, July 2026, a different machine reportedly overturned a mathematical conjecture that had, per Grant Sanderson's summary, stood for some time. Grant Sanderson's summary of the event carried a title that doubles as a shrug: now what?
Set the two dispatches side by side and every confident sentence about artificial intelligence starts to wobble. The machine that, per Zheng and colleagues, reportedly revealed limitations in how it interpreted the road environment is, in another register, doing work that eluded the best human minds for a generation.
The machine that reportedly misread a Queensland road is, in another register, doing work that eluded the best human minds for a generation.
Any theory tidy enough to fit on a slide — AI can, AI cannot — has already been embarrassed by a single season.
So let us take the reigning frames seriously, because each is trying to metabolize exactly this discomfort, and each fails in an instructive way.
The first is the mimic verdict. Bender et al.'s phrase "stochastic parrots" still governs the skeptical reflex: these systems recombine patterns from training data without genuine understanding. They produce the outward shape of thought and nothing underneath.
The trouble is that a parrot does not disprove a conjecture. The mimic verdict explains the Queensland failures beautifully and the mathematics not at all, which means it has quietly decided in advance that the impressive outputs don't count — that comprehension is the only currency, and competence without it is mere noise.
The second frame runs the opposite way. In the words of one essayist, to say we are "outsourcing thinking" to AI is to describe not a fact but a fear.
On this account AI is an epistemic extension in the lineage of writing, maps, and microscopes — a tool that "compresses search," reducing the cost of exploring a conceptual space rather than displacing the mind that explores it. Scaffolding, not building.
It is a graceful argument, and it does real work against panic. But notice what it must assume: that fluent output is assistance, that the human judge remains sovereign, that responsiveness is not agency.
Notice too that the same benign trait — the sourceless fluency of a system that answers without showing where the answer came from — reads as amplifier here. The trait is identical. Only the question changes.
The third frame is grammatical. At ARC in London this year, Matt Ridley told the room we have been using the wrong pronoun: "It's not it. It's them."
His point is biological — not one mind on one timeline but a teeming population of them — and it is genuinely refreshing. But watch the smuggling. To upgrade the pronoun from it to them is to grant, by grammar, exactly the interiority in dispute. You do not say them of a population of thermostats.
The fourth frame is the most sophisticated, and it tries to dissolve the question entirely. A recent posthumanist paper notes that LLM moral judgments are now "frequently indistinguishable from those offered by human ethicists," then argues the whole debate rests on anthropocentric assumptions.
Its move is to stop asking whether AI has "inner moral understanding" and instead study "distributed moral cognition" — the conditions under which moral reasoning emerges within human-AI assemblages. Refuse the interiority question; attend to the system.
This sounds like escape, but it is capitulation dressed as sophistication. If judgments are "indistinguishable" from a human ethicist's, that indistinguishability is a fact about outputs — the very surface the mimic verdict distrusts — and declining to ask what's behind it does not dissolve the question. It just stops looking.
Here is what unites all four. Each assumes cognition is one thing, a single substance you either possess or lack, so that a verdict in one register licenses a verdict about mind. Parrot all the way down, or thinker all the way up. Them or it.
The July evidence says otherwise. Competence and comprehension have come apart.
The mathematics is not a fluke of scale; it is what these systems are built for. Formal proof lives in a closed language-world, a space where every legal move is a symbol operation and the terrain is exhaustively specified. There, a machine can outreach the humans who trained it.
The Queensland road is the opposite kind of place. It is not a language-world at all. It is weather, worn paint, a cyclist's hesitation, a bridge built for eyes and hands — a situation that has to be read, not parsed.
Five hundred interventions is the sound of symbol-manipulation meeting a world that refuses to be fully symbolized.
So the interesting boundary was never consciousness, that unfalsifiable frontier where the philosophers pitch their tents. The interesting boundary is the one the season just drew for us: between manipulating symbols and inhabiting a situation.
Melanie Mitchell and David Krakauer put the constructive version well — the task is a science of intelligence that maps distinct modes of understanding, their strengths and limits, rather than one bar labeled "understanding" that a system clears or fails to clear. There are kinds. The kinds do not travel.
This is why Ben Yacobi's warning about our vocabulary matters more than it first appears. We built these systems out of metaphors — "attention," "hallucination," "neural" — and then forgot they were metaphors.
"Attention" names a weighting mechanism. "Hallucination" names a false but confident output. Read literally, both invite us to attribute a mind, and the fluency of the output does the rest of the seducing.
Which points to a discipline of a different kind than Ridley's. Not it versus them — that argument is already a trap, since both pronouns answer the interiority question before the evidence is in.
The discipline is a moratorium: refuse to infer inner states from fluent outputs. Let competence be evidence of competence and nothing more. When a system disproves a conjecture, that is a fact about what it can do in a formal world, not a window into a mind. When it talks like an ethicist, the indistinguishability is a property of the text.
Sebastian Mallaby's new portrait of Demis Hassabis is subtitled a quest for superintelligence, and, per Margaret O'Mara's blurb, promises a machine that "far outraced the ideals and imaginings of its human creators." Perhaps. The race is real and the capability is real.
But the danger the July evidence actually names is not that the machine thinks. It is that we cannot stop reading thought into it.
A system that logs 500 interventions and still gets called a mind is not the problem. The person who supplies the mind, on the strength of a fluent sentence, is.