The SCL Technology Law Futures Conference, hosted by Herbert Smith Freehills, is recognised as a key event in the UK for discussion of IT law reform and policy, and for communication of ideas at the highest level between academics, practitioners and industry. Speakers investigated the developing role of Artificial Intelligence (AI) in society, on the premise that it isn’t whether machines will out-think humans, it is question of when.
Dylan Evans gave an interesting and thought-provoking speech which focussed on the theories and influence of AI researcher and theorist Eliezer Yudkowsky, and the potential shortcomings and dangers of those theories.
After an introduction from Simon Deane-Johns highlighting Dylan’s particularly weird and wonderful career, Dylan assured the audience that his credentials as a senior lecturer in robotics meant he was able to speak with authority on the subject of AI.
He began by quoting Bill Gates, Stephen Hawking and Elon Musk who have all recently expressed major fears as to the development of AI. Dylan pointed out that the catalyst for this collective expression of concern was largely the publication, in August 2014, of a book called ‘Superintelligence: Paths, Dangers and Strategies’ by Nick Bostrum. Bostrum’s book relies heavily upon the theories of Yudkowsky; indeed the two are close and Bostrum is a member of Yudkowsky’s elite inner circle.
Dylan provided some background to Yudkowsky’s power and influence in the area of AI. Largely self-educated, he claims to have an IQ of the highest possible figure (143) and has been writing about AI for many years on his well-known blog ‘LessWrong’. His writings have been quoted in many major newspapers and have attracted significant funding through a range of organisations dedicated to progressing his ideas, including the Future of Humanity Institute at Oxford, the Machine Intelligence Research Institute (‘MIRI’) and the Centre for the Study of Existential Risk at Cambridge. The organisations affiliated with his work have received funding of millions of dollars from high profile individuals including Peter Thiel and Jaan Tallinn (of Facebook and Skype fame respectively).
Dylan explained that Yudkowsky’s main idea (simplified) consists of two strands, firstly that ‘unfriendly’ AI (that is, AI with an objective which is not benevolent) poses a significant threat to the future of humanity (the ‘Theory’) and secondly that the best way to tackle this threat is to develop formal methods for proving the safety of AI designs (the ‘Solution’).
The Theory centres around the idea that at some point in the future machines will become more intelligent than humans, not only in the sense that they will have better domain-specific competencies than us (as many machines already do), but also a greater general intelligence. This will lead to a ‘singularity’ (an expression taken from black hole theory) in which machines design other machines with improvements and so the development of machines is self-sustaining and outside of the control or understanding of humans. The Theory therefore ventures that humans have the capacity to develop super-intelligent machines that could surpass our own capabilities in such a way that they eventually destroy humankind itself. Dylan pointed out that an obvious solution to this dilemma would be to program such machines so that they have only ‘good’ goals. But having to specify what ‘good’ means in code would, of course, be extremely challenging. It inevitably leads to a myriad of ethical dilemmas and unforeseen consequences. Another solution, Dylan ventured, might be to ‘box’ the AI; to put it in a room and avoid interfacing directly with it, so as to contain any dangers. The problem with this, however, is that a machine with the level of intelligence that needs to be ‘boxed’ will be able to get itself out of confinement.
Dylan then expanded upon the Solution. Broadly, Yudkowsky’s research goes towards determining whether we are able to raise the probability of a better outcome by using mathematical equations to describe broad types of harmful AI behaviours which we are then able to avoid, thereby proving the AI to be safe. If such activity is possible, teams of incredibly talented mathematicians would be needed to carry it out.
Dylan suggested, however, that there are problems with both the Theory and the Solution which make Yudkowsky’s monopoly of ideas potentially very dangerous and damaging.
In respect of the Theory, Yudkowsky’s analysis of the threat contains many assumptions and is heavily influenced by ‘Hollywood’ scenarios involving hyperbolised unfriendly AI. Yudkowsky does not consider, for example, the risk of detrimental consequences of friendly AI mistakenly doing something it is programmed to understand as being for the benefit of humanity.
In respect of the Solution, Dylan expressed concerns at how demanding it is. It requires a preliminary solution to the problems of human ethics and of metaphysics. This in itself denies the existence of any conflict in ethics, which leads eventually to a kind of totalitarianism. We need to know which theory of ethics is correct in order to arrive at the Solution, and seemingly obvious questions become contradictory when posed in the context of AI: for example, is it ethical to enslave a superior mind? If the answer is no, then how is it possible to program this into a machine with a superior intelligence to our own, and to effectively enslave that machine at the same time.
Dylan described a fictional scenario in which a driverless car was heading for an inevitable accident and had to choose between killing its passenger or killing three other people in the collision. The car would need to be programmed in such a way that it was able to make that choice. Dylan notes that this highlights the very real and current danger of moral philosophies becoming practical questions in the context of AI development.
Finally, Dylan asked what happens after the Solution? He asked the audience to try to envisage a world full of willing slaves – super intelligent machines subject to the control of humans. This might be paradise if this ownership were broadly distributed or hell if it were not.
Flora Blackett-Ord is a Trainee solicitor at Herbert Smith Freehills
