top of page

AI doesn’t cause harm by itself. We should worry about the people who control it

At times it felt less like Succession than Fawlty Towers, not so much Shakespearean tragedy as Laurel and Hardy farce. OpenAI is the hottest tech company today thanks to the success of its most famous product, the chatbot ChatGPT. It was inevitable that the mayhem surrounding the sacking, and subsequent rehiring, of Sam Altman as its CEO would play out across global media last week, accompanied by astonishment and bemusement in equal measure.


For some, the farce spoke to the incompetence of the board; for others, to a clash of monstrous egos. In a deeper sense, the turmoil also reflected many of the contradictions at the heart of the tech industry. The contradiction between the self-serving myth of tech entrepreneurs as rebel “disruptors”, and their control of a multibillion-dollar monster of an industry through which they shape all our lives. The tension, too, between the view of AI as a mechanism for transforming human life and the fear that it may be an existential threat to humanity.

Many are ‘preppers’, survivalists prepared for the possibility of a Mad Max world

Few organisations embody these contradictions more than OpenAI. The galaxy of Silicon Valley heavyweights, including Elon Musk and Peter Thiel, who founded the organisation in 2015, saw themselves both as evangelists for AI and heralds warning of the threat it posed. “With artificial intelligence we are summoning the demon,” Musk portentously claimed.

OpenAI was created as a non-profit-making charitable trust, the purpose of which was to develop artificial general intelligence, or AGI, which, roughly speaking, is a machine that can accomplish, or surpass, any intellectual task humans can perform. It would do so, however, in an ethical fashion to benefit “humanity as a whole”.


Then, in 2019, the charity set up a for-profit subsidiary to help raise more investment, eventually pulling in more than $11bn (£8.7bn) from Microsoft. The non-profit parent organisation, nevertheless, retained full control, institutionalising the tension between the desire to make a profit and doomsday concerns about the products making the profit. The extraordinary success of ChatGPT only exacerbated that tension.

Two years ago, a group of OpenAI researchers left to start a new organisation, Anthropic, fearful of the pace of AI development at their old company. One later told a reporter that “there was a 20% chance that a rogue AI would destroy humanity within the next decade”. That same dread seems to have driven the attempt to defenestrate Altman and the boardroom chaos of the past week.

One may wonder about the psychology of continuing to create machines that one believes may extinguish human life. The irony, though, is that while fear of AI is exaggerated, the fear itself poses its own dangers. Exaggerated alarm about AI stems from an inflated sense of its capabilities. ChatGPT is superlatively good at predicting what the next word in a sequence should be; so good, in fact, that we imagine we can converse with it as with another human. But it cannot grasp, as humans do, the meanings of those words, and has negligible understanding of the real world. We remain far from the dream of “artificial general intelligence”. “AGI will not happen,” Grady Booch, chief scientist for software engineering at IBM, has suggested, even “in the lifetime of your children’s children”.

For those in Silicon Valley who disagree, believing AGI to be imminent, humans need to be protected through “alignment” – ensuring that AI is “aligned with human values and follows human intent”. That may seem a rational way of countervailing any harm AI might cause. Until, that is, you start asking what exactly are “human values”, who defines them, and what happens when they clash?


Social values are always contested, and particularly so today, in an age of widespread disaffection driven often by the breakdown of consensual standards. Our relationship to technology is itself a matter for debate. For some, the need to curtail hatred or to protect people from online harm outweighs any rights to free speech or privacy. This is the sentiment underlying Britain’s new Online Safety Act. It’s also why many worry about the consequences of the law.


Then there is the question of disinformation. Few people would deny that disinformation is a problem and will become even more so, raising difficult questions about democracy and trust. The question of how we deal with it remains, though, highly contentious, especially as many attempts to regulate disinformation results in even greater powers being bestowed on tech companies to police the public.

The reason algorithms are prone to bias, especially against minorities, is because they are aligned to human values

Meanwhile, another area of concern, algorithmic bias, highlights the weaknesses of arguments for “alignment”. The reason algorithms are prone to bias, especially against minorities, is precisely because they are aligned to human values. AI programmes are trained on data from the human world, one suffused with discriminatory practices and ideas. These become embedded into AI software, too, whether in the criminal justice system or healthcare, facial recognition or recruitment.


The problem we face is not that machines may one day exercise power over humans. That is speculation unwarranted by current developments. It is rather that we already live in societies in which power is exercised by a few to the detriment of the majority, and that technology provides a means of consolidating that power. For those who hold social, political and economic power, it makes sense to project problems as technological rather than social and as lying in the future rather than in the present.

There are few tools useful to humans that cannot also cause harm. But they rarely cause harm by themselves; they do so, rather, through the ways in which they are exploited by humans, especially those with power. That, and not fantasy fears of extinction, should be the starting point for any discussion about AI.


12 views0 comments

Comments


bottom of page