What threat does AI pose? Perhaps an existential one
By Cameron Scott
The human brain is widely considered ‘the most complicated thing in the universe’. By thinking about innate intricacies of every person you know, it is perhaps not surprising that the brain poses such an enormous scientific challenge. It must be flexible enough to allow for every minute detail of every unique human personality whilst simultaneously controlling those vital functions found in everyone – breathing, sensing and thinking. Of course, the complexity arises from the unimaginably large number of interconnected neurons in our brains and this complexity presents a huge barrier towards understanding the brain’s operation.
Like so many complex problems, analogies are useful. In The Idea Of Brain, Matthew Cobb discusses how people have often used the current technological paradigm to describe the brain. For example, in 1917, as telephones began to become a prominent feature of modern life, physiologist Professor Arthur Keith attempted to describe the functioning of the brain and spinal cord by drawing parallels to human-operated telephone exchange centres. As technology has progressed, the analogies have updated too. We can now easily think of the brain as some enormously powerful central processing unit (CPU) and that our central nervous system acts somewhat similarly to a computer. But we now encounter a serious problem – what happens when the predominant technology of the time is also too complicated to understand?
This is exactly what has happened with Artificial Intelligence. The new tools, currently exemplified by text and image generating tools like Chat-GPT and StableDiffusion, show incredible promise and produce remarkably human-like outputs but their fundamental operating procedures are largely unknown. They are constructed using neural networks, algorithms designed to spot patterns in a manner loosely based on how neurons in the brain operate. This is all part of the current machine learning model in which a computer program can be trained on a huge set of input data and then make decisions based on the patterns it spots in that data. The italicised words in this paragraph are all words that, before the advent of AI, could be applied to human brains and so indicate a linguistic flip. It used to be that we described the brain (something we don’t understand) with references to telephones (a technology we understand). Now we describe AI (a technology we don’t understand) with reference to the brain (something we also don’t understand but feel a bit more familiar with).
What do we do when we don’t understand a technology? We hypothesise about the dangers. With the growing prominence of AI, the existential risk it poses has begun to be hotly debated. Professor Stephen Hawking said in 2017 that “I fear that AI may replace humans altogether”. More recently Dr Geoffrey Hinton, a pioneer into neural network research, warned of the existential risk posed by AI being exploited by “bad actors”. The two warnings highlight two very different, but very real risks posed by artificial intelligence.
Beginning with Hinton’s warning of bad actors, he is alluding to AI’s ability to produce enormous amounts of very human-like content. It is not difficult to imagine how the new generative AI tools like Chat-GPT or StableDiffusion can allow spammers to send out extremely convincing scam messages, causing financial ruin for many but also eroding any trust that people may have in the digital world. Furthermore, it is also not farfetched to imagine a nefarious political campaign creating videos claiming to capture an opponent in a compromising position or of that same opponent saying something they have never said. Politicians and celebrities are particularly susceptible to this as they appear in countless hours of media appearances, producing a large dataset to train machine learning models. Clearly, generative AI already poses a severe risk to the democratic norms.
As artificial intelligence gets smarter, it stops being specialised to particular tasks and is instead able to outperform a human in every single possible task. If the AI becomes better than humans at designing new AIs, a technological singularity has been reached. At this point, the artificial intelligence rapidly improves its intelligence and becomes superintelligent. General AI would presumably have complete control of every system on Earth and all its resources. In this situation, and what Hawking was warning about, humanity would not go out in a blaze of glory fighting Terminators or Battle Droids but instead would instead fade away due to obsolescence – much like how the population of horses reduced once they were no longer needed for transport.
This bleak outlook is not a certainty. It depends crucially on whether a general AI is possible at all. The current impressive tools make it seem like anything is possible with generative AI but actually using the tools reveals. glaring issues, most notably the so-called “hallucinations”. After being asked a question which doesn’t make sense, the AI will produce something without critically assessing if the question or its response was logically sound. Current AI doesn’t reason but only spots patterns. For a truly general AI, it would have to learn how to reason, and this is only possible once we understand how our own brains do it – reasoning is something that humans do quite well!
Image: Will Brown with DALL-E 2
