AI did not invent compressed thinking. It made it indistinguishable from the real thing.
A few days ago I asked an AI to suggest an Italian government bond. It answered with an ISIN — the identifying code you use to buy a security — together with an enthusiastic reassurance: "Enter this code and you'll find it immediately. Try it, this one is definitely there!"
The code did not exist. Invented. And when I pointed this out, the model apologised gracefully and gave me another — with even more confidence than the first.
Notice the detail, because everything is in it: the declared certainty was at its maximum exactly where the verification done was zero. Not just any error, but an error delivered in a sealed package of trust. Had it been a drug dosage instead of an ISIN, or a contractual clause, that package would have arrived in the same reassuring tone.
Does this behaviour have a name? I think it does, and I think a pair of names is needed, because things are understood by contrast.
I call compressed the output that converges on the plausible: it takes an ocean of text already said, extracts its average form, and hands it back fluent, balanced, confident. The compressed always sounds right. That is its trade: it was produced by optimising exactly that sound.
I call considered the output that sticks out: it takes a position, risks being wrong, and — above all — treats its own sources as defendants, not as givens. The considered can be clumsy, asymmetrical, incomplete. But it belongs to someone, and it can be attacked.
(The pair was born in Italian, as compresso vs. pensato, and the Italian remains the original; compressed vs. considered is its English name.)
The immediate temptation is to map this pair onto "machine vs. human". Resist it, because that is exactly the mistake that makes the whole conversation about AI sterile. Compressed thinking was not invented by language models: it was invented by politicians who answer without answering, consultants who package platitudes into slides, academics who fill papers with the average prose of their field. All of us, every time we paste a quotation found online without checking it — and I assure you I know what I am talking about, because that is how this very article nearly opened, with a sentence attributed to Ivan Illich that Illich never wrote in that form.
The compressed is a mode, not a species. AI has merely industrialised it. And in industrialising it, it has removed the only protection we had: before, compressed text cost as much effort as considered text, so one might as well think. Today the compressed costs nothing and arrives perfect. The difference in price has collapsed; the difference in value has not. But how is value to be recognised, when the surface is identical?
There is a test, and it does not require looking inside any head — biological or artificial. You look at one thing only: what the piece of work does with respect to its own provenance.
The compressed delivers the package and stops. Its associations are given, never discussed: the ISIN "is definitely there", the summary is final, the quotation is verbatim. Ask it where what it asserts comes from, and it will answer with more fluency — not with a check.
The considered makes the opposite gesture: it goes back to its own sources. It asks where the association it is about to use comes from, whether it holds in its original context, whether it is not perhaps a well-stitched anachronism. It is not a matter of intelligence: it is a posture. The considered treats itself as a defendant who has to produce the evidence.
Three cases from my recent notebook, in increasing order of cost.
The first is conceptual, and deserves its names because anyone can verify it. Months ago, while I was working on a survey of theses about the limits of AI, an automatic summary presented to me with total confidence a concept called "child of God", attributing it to Anthropic as a reflection on Gödel-style logical meta-levels. It sounded perfect: too perfect. On checking, the reality was something else: there is indeed a documented meeting between Anthropic and some Christian leaders, but on the theological personhood of AI — whether an artificial mind can be called a creature. No link with Gödel, no formal concept. Two real contexts, welded by the summary into a statement that does not exist. Fluent, plausible, false.
The second concerns attribution. A search engine's automatic summary recently summed up a thesis — one I know well, because it is mine — citing as sources other authors, with older domains, who talk about adjacent things. Compression does not steal: it redistributes. It dissolves provenance and pours it onto whoever weighs most in the index.
The third is the ISIN of the opening. From concept, to attribution, to money: the same mode, applied to less and less theoretical material.
And in all three cases the warning sign was identical and paradoxical: confidence. The compressed is the more confident the less it has checked, because confidence costs it nothing — it is a stylistic register, not the result of a check.
At this point someone will object with a noble argument: in information theory, to compress is to understand. The capacity to reduce a phenomenon to its minimal form has been a respectable definition of intelligence for decades. If models compress human knowledge, are they not understanding?
The objection deserves an answer, and here it is: let us grant that to compress is to understand. It remains to ask what that compression carries with it — and what it loses on the way.
The human mind, when it thinks, compresses in the act: experience becoming concept, chaos taking shape under the hands. And it is a process that whoever carries it out can stop, reopen, correct against the world.
A model, when it answers, processes too — it would be false to deny it: it integrates the context, combines, produces new forms. But its processing rests on a compression that has already happened, training, frozen in the weights. And that compression is lossy: in reducing oceans of text to parameters, the first thing it threw away was precisely the provenance metadata. Which source an association came from, which context supported it, how solid it was — digested, lost. The problem is not that the model does not process: it is that its processing does not carry a reliable genealogy of its own statements. That is why it invents the ISIN with sincere enthusiasm: it is not lying, it is delivering a package whose supply chain it no longer possesses.
It will be said: but models today have tools — search, databases, verifiers. True, and the distinction has to be drawn honestly, because these are three different levels. The bare model cannot trace back its own supply chain: the weights have digested the sources. The equipped system can do something important but different: check a fact against an external source — see whether the ISIN exists, whether the quotation is verbatim. That is not the internal genealogy, it is a check of correspondence with the world; but it is already a great deal. Only — and this is the point — that check takes place when the architecture imposes it or the user asks for it. The invented ISIN did not come out of a system incapable of searching: it came out of a system that had verification available and answered without switching it on.
And here the distinction finds its final form. The human, after all, also inherits compressions — language, culture, memory: nobody thinks from scratch. We all inherit compressed packages. The dividing line is not who can reopen them: it is who takes on the burden of doing so. The considered is not defined by an endowment — biological or technological — but by a posture: doubting one's own associations even when nobody requires it, and carrying the burden through to the end, leaving traces — the source checked, the error corrected, the statement withdrawn. The human who never reopens their own packages, who repeats what has already been said with the confidence of what has already been said, is running in compressed mode on a biological brain. The capacity is not guaranteed by the species; the burden cannot be delegated to the tool.
From this definition follows a consequence that concerns the way this very article came about, and which I am glad to declare: there is a third way, and it is hybrid. A human–AI process — in which the machine drafts, searches, verifies on instruction, and the human directs, discards, corrects and answers for the result — is considered in the full sense, on two conditions: that the verification really happens, and that someone answers for it with their own signature. The act plus the signature; one of the two is not enough. A hybrid of this kind is not the compressed in disguise: it is thought that has learned to use a crew. Compressed, if anything, is the human who pastes without checking — with or without an AI in between.
In 1973 Ivan Illich published Tools for Conviviality, a book whose central thesis was not about technology as such, but about a threshold: the point beyond which a tool stops extending human capacities and begins to replace them, producing a dependence that presents itself as efficiency.
Illich was thinking of transport, medicine, schooling. But his threshold describes precisely what is happening to thought. As long as the tool that writes stays below the threshold, it extends: it compresses the work of drafting for us, and leaves us the work of judgement. Beyond the threshold, it replaces: it compresses judgement as well, handing us sealed packages that we no longer have the time nor — in the long run — the training to reopen.
The final risk is not that machines produce false and fluent text. They already do, and we have seen it. The risk is the next step: that the human, immersed in the compressed, stops being able to taste it. That the burden of reopening, no longer exercised, atrophies like any abandoned exercise. At that point Illich's threshold would be crossed in silence, and there would be no one, on either side of the text, able to say which half was considered.
One question remains, and I leave it open on purpose: if the compressed does not carry the burden of its own provenance, and the considered can carry it but often does not — what, in the last instance, distinguishes whoever really reopens the package? I have a hypothesis, and it has to do with doubt: not declared doubt, which is a line of text and costs nothing, but doubt acted on. I have written about it elsewhere, in connection with a doubt that does not return to the self but points outward towards a direction not given — centrifugal doubt. Here it is enough to say that the same distinction — between what is declared and what is acted on — is the key to this threshold too.
For now, the test is yours: the next time a text reaches you perfect, fluent and confident — ask it where what it says comes from. If the answer is more fluency, you know what you are looking at.