Despite the AI boom over the past twenty years or so, artificial intelligence remains an ‘emerging’ rather than established technology, and therefore continues to be something of an unrealised project. Promises of fully self-driving cars, sentient robot companions, real-time crime prediction algorithms, or even reasonably engaging conservational agents remain a long way from fruition.
As the 2020s progress, suggestions are growing amongst pockets of the AI research community that the ‘AI spring’ of the 2000s and 2010 might soon recede back into an ‘AI winter’ – a period of reduced industry confidence exacerbated by dwindling commercial interest, government funding and public trust in the whole area of artificial intelligence.
These thoughts underpin a new paper from Melanie Mitchell – a computer science professor from the Santa Fe Institute. In ‘Why AI is Harder Than We Think’, Mitchell contends that the apparent short-comings of AI can be partly understood in terms of the AI community’s “limited understanding of the nature and complexity of intelligence itself” (p.1). As such, Mitchell proposes ‘four fallacies’ that she feels are central to holding-back how AI research is currently approached:
Fallacy 1: seeing ‘narrow’ intelligence as on a continuum with ‘general’ intelligence.
First, is the continued distinction placed within AI research on the idealised (and highly contestable) goal of ‘general AI’ as distinct from currently feasible forms of ‘narrow AI’. Whereas ‘narrow AI’ denotes systems designed to handle one specific limited task and operating within pre-defined boundaries, many AI researchers remain ultimately motivated by the possible development of fully autonomous agents with high-level ‘general intelligence’ comparable to human intelligence.
In particular, Mitchell calls out the recent trend for any specific advance in ‘narrow AI’ to be treated as a step toward more general AI. In other words, general intelligence is believed to be an achievable goal, with any advance in AI seen as a further small step toward contributing to this progress. Seen in these terms, then, what are actually significant advances in AI continue to be seen as disappointing attempts that fall short of fulfilling their full potential. Instead, perhaps the current forms of narrow AI that have already been developed are as far as this technology can go, and the main aim of AI development in to further refine and improve these forms of artificial intelligence.
Fallacy 2: presuming that ‘easy things are easy’ and ‘hard things are hard’
Second, Mitchell laments the recent shift away from the widely-shared acceptance amongst twentieth century AI researchers that their most complex challenge was to automate the tasks that humans can usually perform with little or no thought – e.g. Hebert Simon’s example of a child being able to instantaneously recognise its mother. In contrast to these split millisecond cognitive processes, it was also generally accepted that codifying things that humans find difficult (such as solving a complex mathematical problem or playing Go or Chess) was usually a much far easier computational challenge to address.
Mitchell argues that these understandings were flipped in the 2000s, as the field of artificial intelligence began to turn away from ‘symbolic’ AI approaches and toward the ‘statistical’ AI approach of deep learning. Now, she sees a growing belief amongst the current generation of AI researchers that anything a human can do through unconscious thought is relatively likely to be computationally modelled and automated on the near future. Conversely, getting a machine to play a board game such as Go is now heralded by current AI research teams as “the most challenging of domains” and proof that “it is possible to train to superhuman level”. As Mitchell concludes, “AI is harder than we think, because we are largely unconscious of the complexity of our own thought processes” (p.4).
Fallacy 3: Seeing AI in terms of ‘wishful mnemonics’
Mitchell’s third point relates to the increasing use of language and descriptions when presenting artificial intelligence that infer human intelligence being attached to contemporary AI technologies. Mitchell argues that this gives a mis-leading sense of equivalency between machine and human intelligence. For example, the idea of ‘neural networks’ might be loosely inspired by the workings of a biological brain, but is certainly not comparable to a human or animal brain. Machine learning and deep learning certainly do not resemble the ‘learning’ undertaken by humans and non-humans – not least the capacity to transfer what is learnt in one domain to another domain.
Mitchell argues that by anthropomorphising how AI is talked about in terms of ‘reading’, ‘understanding’, ‘thinking’ or ‘comprehension’, AI researchers clearly raise expectations and perceptions of what is technically possible. Even if AI researchers are simply using such terms as shorthand and fully understand that machines do not ‘think’ in the same way that humans do, such distinctions are not carried over to the general public, and might actually begin to unconsciously shape how AI experts themselves think about their systems.
Fallacy 4: believing that intelligence is ‘all in the brain’
Finally, Mitchell calls out a persistent belief across the AI community that intelligence is something that can be separated from the body – what she describes as an ‘information-processing model of mind’ that sees intelligence is something that can be ‘disembodied’ and relocated to software and hardware. While this is a long-standing belief (not least in the continued influence of seventeenth century Cartesian philosophy), Mitchell argues that the rise of neurally-inspired machine learning has specifically boosted the idea within the AI community that the sole part of the body relevant to intelligence is the brain. As such, it is assumed that intelligence can arise from the development of hardware and software that can match brain structures and dynamics.
Yet Mitchell reminds us of the diverse forms and facets of intelligence that this approach misses out on. These include the complexities of embodied cognition, through to emotions, common-sense and ‘irrational’ thoughts that are entwined with the complexities of our social lives. Mitchell argues that these aspects of human intelligence remain absent from the ways in which AI researchers set out to separate and codify what they see as human intelligence. As Mitchell observes:
“human intelligence seems to be a strongly integrated system with closely interconnected attributes, including emotions, desires, a strong sense of selfhood and autonomy, and a common-sense understanding of the world. It’s not at all clear that these attributes can be separated” (p.7)
What are the implications of these fallacies?
There are some clear conclusions arising from what Mitchell argues in this paper. First, Mitchell makes a call for AI researchers to pay closer attention to scientific understandings of ‘intelligence as it manifests in nature’. Second, is the obvious need to encourage a different vocabulary for talking about what machines can – and can not – do. We need to talk more about machines being engaged in processes of calculating, matching, correlating and rule-making, and far less about machines that are capable of learning, thinking and perception. We need to call out the unlikeliness of ‘general intelligence’, and instead focus on the realities of ‘narrow AI’ applications – recognising these instances of AI as ends in itself, and discussing what real-life implications ‘narrow AI’ might have.
Above all, while it is incredibly difficult to ‘capture humanlike intelligence in machines’, Mitchell argues that this is exactly what has to happen if we want machines to work with us in our human world. As she puts it, if we want AI technologies that can complement and extend people’s real-life activities as carried out in real-life social settings, then we need machines to display some form of humanlike cognition – in other words, we need machines to be built upon the same basic knowledge about the world that is the foundation of our own thinking. In this sense, AI researchers need to pay more attention to developing forms of cognition that can cope with the idea of common-sense, socially-driven ‘irrational’ thinking, emotion, and other ways that humans are actually sentient and intelligent beings.
Of course, these final points raise a broader contention about what types of ‘human’ that AI researchers want to be ‘capturing’ intelligence from. ‘Human’ is not a homogenous category, and it seems dangerous to argue that there is a standard – or even ideal – form of ‘human’ intelligence that AI researchers need to be looking toward. Intelligence is clearly culturally-specific. Moreover, no one brain works in exactly the same way, which raises the tricky question of which (and whose) ‘human intelligence’ AI researchers should be looking to emulate. What sort of ‘emotional’ or ‘irrational’ thinking is being talked about here? Even from a biological point of view, it is not a given that there is an ideal the kind of ‘brain’ that AI researchers are looking to emulate and model. For example, what would an AI system modelled on non-neurotypical brain processes look like … what might AI modelling learning from autistic brain functions? Even having acknowledged some possible ways forward, Mitchell’s nuanced conclusions ultimately bump up against the fact that AI is harder than we think. Perhaps the main conclusions to reach here is that we talk much more about the absolute limitations of this area of technology development, and reset our ambitions and actions accordingly.