recent op-ed by a well known and reasonably intelligent writer tells us
that “AI”, by which most people mean chatbots, substitutes
consensus for truth. This may be a self-evident observation, but what’s
interesting is his evidence: he asked a chatbot named Claude, and it told
him so. QED.
If your chatbot tells you that chatbots are unable to distinguish true from false, is the statement true or false? It’s a classic liar’s paradox. Surely people know by now how chatbots work. If a smart person can make such an elementary mistake, what hope do ordinary people have?
None, because they don’t really care. The rise of so-called AI
has given us people who post big long blobs of text from an AI as if
they think a chatbot is an authoritative source. It is not.
Because chatbots tend to be chatty, verbose, and often wrong (in a
characteristically unintelligent way), it has also given us the
expression “ai;dr”: AI, didn’t read.
Chatbots do not find consensus
What they give us now is not even a true consensus. If Claude or some other chatbot gets its information from Wikipedia and a few other approved websites weighted according to the programmers’ belief in their credibility, it has failed. There are still people who think if Wikipedia says something it must be true, so they repeat it. Whatever false statement you find there (and anyone knowledgeable on some topic invariably finds errors), a moment of searching will find the falsehood repeated in dozens of other places. This is what passes for consensus on the Internet.
In fields where factual correctness is important, no assumptions are made about whether anything anybody says is true or not. If you state something as fact you must either show your evidence or cite a peer-reviewed article or other primary source that did. Wikipedia is not a primary source and doesn’t count as peer reviewed. A scientific paper that cited it as a source of truth would never pass peer review.
A chatbot is not a primary source or even a secondary source. At best it is a tertiary source. Yet more and more people just ask it their questions. In a way it’s understandable. They are tired of blatant search engine bias, dishonest news sites, and websites stuffed with animated ads, pop-ups, and demands for money, and they are also lazy, so if a chatbot says something that sounds plausible, they assume it must be true. (Or maybe they’re just assuming no one would bother to argue with a chatbot.)
How could a chatbot be intelligent?
Most people recognize that chess-playing computers and computers that prove (not ‘discover’ as the media claim) math theorems are not intelligent. They are programmed by humans with a set of rules for deducing the correct answer. Whether the programming consists of heuristic algorithms or brute force, the intelligence is provided by the programmer. The computer is merely following the instructions. In the case of classical pattern recognition, the computer calculates the solution using algorithms devised by intelligent humans. So-called AI uses the same principle. Only the algorithm is different. As I’ve said repeatedly on these pages, ‘AI’ is only a marketing term.
Software companies may have realized that the goal of an intelligent machine is beyond their reach. But a computer program need not be intelligent to be useful. It could, for instance, trace every claim back to a peer-reviewed scientific article. The scientific literature is not, of course, a repository of truth; various claims about the percentage of true claims have been made and widely debunked. But chatbot could cite all the evidence for and against any particular claim. If the answer didn’t require tedious verification, a chatbot might be useful, instead of the Internet slop aggregator that it is now.
To find out just how good these chatbots are, I asked one which type of screwdriver was better: Phillips or flat-head. It wrote six paragraphs and finally said “Phillips.” So, encouraged by that flash of intelligence, I gave it tougher one:
Please compile a complete list of all drugs tested in clinical trials against alzheimer's disease.
It spent quite a while “thinking” about it and finally gave up, asking for my credit card number instead. Hmm, maybe these things aren’t so dumb after all . . . .
What would it take for a chatbot to be intelligent? I suppose most people in the computer industry must have figured this out by now, so there’s no harm in disclosing it, at least in general terms. It would take decades, and the chatbot would have to be fully autonomous before the training could start. It’s not so much a scientific problem but a chicken-and-egg problem.
Even if Anthropic and the others started tomorrow, it will be decades before an AI is really intelligent. But chatbots could become useful if the software companies turned them into rigorous, traceable algorithms for reporting the evidence. Whether they eventually replace existing sources of information remains to be seen.
Super Intelligence
Last week US president Trump signed an EO requiring the government to use the term ‘superintelligence’ instead of ‘artificial intelligence’. Whether he really believes AI really exists or not, no one knows. My theory—my hope—is that he knows it doesn’t and it’s a clever trick to shut up the alarmists by calling their bluff.
oct 13 2026, 6:29 am. This is a heavily revised version of an earlier one.
Hardly any superintelligences were harmed in writing this article.
