Report argues LLMs fail the test for an 'inevitable' technology

A blog essay, takes on a common claim: that AI is inevitable, so everyone should just accept it. The author calls this a thought-terminating cliche, meaning a phrase that shuts down discussion instead of answering it. The essay then tries to work out when a technology can fairly be called inevitable, and whether large language models (LLMs, the systems behind chatbots) qualify.

The author starts by noting that "inevitable" is a slippery word. Nobody can know in advance which technologies will take off, so you can't say workable nuclear fusion is inevitable. Looking back, almost any popular technology can be called inevitable, which makes the label cheap. The author says even the iPhone's success tells us little, because the technology could have taken many other shapes.

The essay also says a technology is not inevitable on its own. It depends on the society using it. Mesoamerican civilisations knew about the wheel, but in dense jungle with no road network it was not very useful, so they relied on water transport and pack animals. Cannon were more valuable in Europe than in China, the author says, because European fortifications tended to have tall, thin walls, while Chinese ones were often stone-faced walls built largely of rammed earth. Cannon can breach the first type much faster, though they eventually breach both.

From this, the author sets out four conditions. First, the technology does the task about as well as the alternative. Second, the task matters to society, and it is currently done by scarce labour. Third, the technology vastly outperforms the alternatives, and a moderate improvement isn't enough. Fourth, people who don't use it can't change their behaviour to blunt its effect.

Firearms meet all four, the author argues. Cannon breach walls faster than counterweight trebuchets, and breaching walls mattered in a fragmented medieval Europe. Early small arms also delivered far more energy than a crossbow or longbow. And there was no easy defence: building fortifications that resist cannon needed a massive restructuring of society. So armies had to adopt guns quickly or be absorbed by someone who did. The magnetic compass, by contrast, passes the first two tests but not the others. You can navigate without one, and a state can sidestep it by not becoming a maritime power. The author suggests Russia would likely have been a major land power with or without it.

Then the author applies the four tests to LLMs. On the first, the verdict is debatable. The author says LLMs can't match the highest human standards and probably never will. But much human knowledge work is mediocre, and on that kind of work LLMs are more or less interchangeable with people. The author says they probably pass this test.

On social value, the author calls the result mixed. Writing software, research and graphic design are valued and scarce. But the author says LLMs can only replace the least valuable parts of that work. They can produce a lot of tolerable code, but only the weakest graphic design and fairly superficial research. When real quality matters, human expertise is still needed, and the more important the work, the less useful LLMs are.

The third test is where the author is most firm: LLMs are not overwhelmingly better. People still do the top-level work unaided. Even at lower levels, the effort of directing an LLM and checking its output limits the gains. The author suggests you might get 20% to 50% more PowerBI dashboards, but dashboards are rarely a core business bottleneck. The author adds that if the advantage were real, we would have seen a tech company turning out five to ten times as many products as its rivals. Instead, the author says, companies going all in on AI have struggled more than others.

On the fourth test, the author says evidence is thin because social trends take time. Still, the author thinks people can neutralise LLMs cheaply. The essay points to intense social hostility toward LLM use and toward passing off generated "slop" as real work. New technologies usually need some social acceptance, and few have been adopted against that level of hatred. The author adds that humans can still match LLM output in usefulness, even if LLMs win on sheer volume.

The conclusion is that LLMs pass one test, and only partly. They clearly fail the two tests that separate a merely useful technology from an inevitable one. In the author's view, they belong in the category of contingent technologies, not inevitable ones.

The essay ends by asking why AI feels inevitable to so many people anyway, given the large effect it is already having on society. That question is left for the next part of the argument.