Bernie Sanders is a man of many eras. The senator and political activist from Vermont began his career in the early 1960s, taking part in sit-ins against segregated housing, and marched in the 1963 March on Washington for Jobs and Freedom.

Today, the self-styled “progressive” has a new calling: pushing for much stricter restrictions on artificial intelligence, including an outright ban on what his proposal defines as “Artificial Superintelligence”—artificial intelligence that can “match or exceed human cognitive performance and capabilities across a broad range of domains or tasks.”

My first encounter with the concept of “superintelligence” was a decade ago, when I read the philosopher Nick Bostrom’s 2014 book Superintelligence.

Bostrom defines a superintelligence more strictly than Sanders. He describes it as “any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest.”

This, of course, was a completely different time in the history of artificial intelligence. There was no ChatGPT in those days. Bostrom was careful to avoid suggesting that humans were anywhere near developing an artificial superintelligence. The expert surveys he cited in his book put the median estimate for a 50 percent chance of reaching merely human-level machine intelligence at around 2040, and a 90 percent chance only by 2075.

Back in 2014 we had single-domain AIs like chess engines that could outperform humans in very limited areas. That was true even as far back as the 1990s—when IBM’s Deep Blue defeated the world chess champion Garry Kasparov, one of the greatest players of all time.

Chess, of course, is a very constrained and specialised game built around an eight-by-eight grid on which the pieces can move only in predetermined patterns. This favours calculation and thinking many steps ahead, which in hindsight made it well suited to the capacities of a machine. Now we have general models such as large language models (LLMs) that can outperform humans across a wide spectrum of domains.

Of course, LLMs are trained on vast quantities of data harvested from books, the internet, encyclopaedias and so forth. This means that asking them general knowledge questions is not all that meaningful as a test of whether they are actually intelligent, because they can simply regurgitate facts from their training data.

From Turing to GPTs: Evolution of AI Intelligence Tests

The hyperbole surrounding AGI misrepresents the capabilities of current AI systems and distracts attention from the real threats that these systems are creating.

QuilletteShalom Lappin

One way in which artificial intelligence researchers have tried to test for general intelligence is with benchmarks such as the ARC-AGI tests, where the questions turn on recognising a pattern rather than producing an answer that could be memorised or regurgitated.

The original ARC-AGI tasks took the form of completing a pattern in a grid of coloured dots. For example:

A human can easily do this task. What is the pattern here? Well, it’s easier if you put it into words. You could describe it as something like this: the pink L-shapes are being turned into squares by adding a yellow dot in the unfilled corner.

Humans tended to score highly on these tests, averaging 64.2 percent as individuals, although this rose to 98 percent when they were allowed to work together in panels of two or more. But early LLMs performed poorly on puzzles of this type.

A breakthrough happened in December 2024 with OpenAI’s o3-preview model, which scored 75.7 percent on ARC-AGI-1, rising to 87.5 percent when it was allowed to spend vastly more computing power on each problem. To put that in perspective, the then state-of-the-art GPT-4o had managed only around 5 percent on the same benchmark earlier that year. Today, the state-of-the-art GPT-6 Astra scores as high as 98.5 percent on ARC-AGI-1, and similarly well on ARC-AGI-2 and ARC-AGI-3, which were designed to be harder for machines but still solvable by humans.

In other words, we have overshot the expectations of the experts of the 2010s, who thought that we had a 50-50 chance of developing artificial general intelligence by 2040.

No doubt, state-of-the-art large language models are still imperfect. They still make mistakes from time to time. They still return answers that do not match what was asked. They still generate confident falsehoods. But it certainly appears that we are on the path towards the kind of system Sanders wants to ban, potentially the greatest tool human beings have ever invented.

Some may disagree with the label “superintelligence”. For instance, Steven Pinker was quoted critiquing the notion of superintelligence in The Australian in a piece by Claire Lehmann:

“Superintelligence,” with its comic-book prefix, is more a fantasy than a coherent concept. People use it as a synonym for “omniscience,” imagining a magical wizard that can solve all problems with pure computation. Or they imagine that the IQ scale that differentiates humans within their natural range of variation can be extrapolated indefinitely upwards. But real problem-solving requires massive amounts of knowledge about the messy chaotic world which divulges its hidden workings at its own pace, only through laborious experimentation. And human intelligence is not some elixir that you simply have less or more of—it’s a gadget, that evolved to solve some problems with ease and others laboriously or not at all. AI is a different kind of gadget with its own profile of strengths and weaknesses, not an enchanted brew that can grant any wish.

Pinker makes a valid point that this is not necessarily the same thing as human intelligence. It does indeed have a different profile of strengths and weaknesses as compared to humans. But regardless of the accuracy of the term, capability of the kind embodied by advanced AI systems is rare. It is valuable. Having a problem-solving tool of that power available on demand is an extraordinary prospect.

Why does Bernie Sanders think this is such a bad thing?