Enough about the AI Apocalypse Already
When people claim artificial intelligence will surpass human intelligence and thus threaten humanity with extinction, they seem to be referring specifically to advances made with large language models.
As I understand them, large language models are probability machines that have ingested massive amounts of text scraped from the Internet. They answer questions based on the probability of one series of words (their answer) following another series of words (the question).
It seems like a stretch to call this intelligence, but if we accept that definition then it follows that this kind of intelligence is nothing remotely like human intelligence, which makes the claim that it will surpass human intelligence confusing. Hasn’t this kind of machine learning surpassed us decades ago?
Or when we say “surpass” does that simply refer to fooling people into thinking an AI machine is a human via conversation? That is an important milestone, but I’m not ready to accept the turing test as proof of equal intelligence.
Furthermore, large language models “hallucinate” and also reflect the biases of their training data. The word “hallucinate” seems like a euphemism, as if it could be corrected with the right medication when in fact it seems hard to avoid when your strategy is to correlate words based on probability. But even if you could solve the “here is a completely wrong answer presented with sociopathic confidence” problem, reflecting the biases of your data sources seems fairly intractable. In what world would a system with built-in bias be considered on the brink of surpassing human intelligence?
The danger from LLMs seems to be their ability to convince people that their answers are correct, including their patently wrong and/or biased answers.
Why do people think they are giving correct answers? Oh right… terrifying right wing billionaires (with terrifying agendas have been claiming AI will exceed human intelligence and threaten humanity and every time they sign a hyperbolic statement they get front page mainstream coverage. And even progressive news outlets are spreading this narrative with minimal space for contrary opinions (thank you Tawana Petty from the Algorithmic Justice League for providing the only glimpse of reason in the segment).
The belief that artificial intelligence is or will soon become omnipotent has real world harms today: specifically it creates the misperception that current LLMs are accurate, which paves the way for greater adoption among police forces, social service agencies, medical facilities and other places where racial and economic biases have life and death consequences.
When the CEO of OpenAI calls the technology dangerous and in need of regulation, he gets both free advertising promoting the power and supposed accuracy of his product and the possibility of freezing further developments in the field that might challenge OpenAI’s current dominance.
The real threat to humanity is not AI, it’s massive inequality and the use of tactics ranging from mundane bureaucracy to deadly force and incarceration to segregate the affluent from the growing number of people unable to make ends meet. We have spent decades training bureaucrats, judges and cops to robotically follow biased laws to maintain this order without compassion or empathy. Replacing them with AI would be make things worse and should be stopped. But, let’s be clear, the narrative that AI is poised to surpass human intelligence and make humanity extinct is a dangerous distraction that runs counter to a much more important story about “the very real and very present exploitative practices of the [companies building AI], who are rapidly centralizing power and increasing social inequities.”.
Maybe we should talk about that instead?