#3 Artificial Simulated Intelligence, or the incredibly seductive power of pretending

AI may be artificial, but it is not intelligent (yet). There are many definitions of ‘intelligence’ – it’s a concept which is so broad, and subtle, that it can have many meanings and sub-meanings (to start with, it is also used as a term for gathering a body of insights - often covertly - and so I will start by saying that this meaning of the word is entirely outside of the remit of this essay). The Collins English Dictionary defines intelligence as 'the ability to think, reason, and understand instead of doing things automatically or by instinct': the Macmillan Dictionary as 'the ability to understand and think about things, and to gain and use knowledge'. Some might say it is the capacity for things we recognise about ourselves as humans that animals do not seem to do, and thus definition to what it means to be human; things like abstraction, logic, understanding, self-awareness, learning, emotional knowledge, reasoning, planning, creativity, critical thinking, and problem-solving, all seeming to arise from some specialness in the human mind. Because humans vary in intelligence, and higher intelligence is considered beneficial, intelligence is sometimes subdivided into ‘types’ such as emotional intelligence (EQ), academic intelligence or IQ (usually logical or mathematical), and sometimes looser subsets such as linguistic, musical or spatial talents, and we recognise people can be good at one element more than another. Linguist Noam Chomsky, and others, argue that human evolution was marked by a ‘great leap forward’ 200,000 years ago where the evolution of language enabled intelligence – as we recognise it today in modern humans – to emerge, and language is of course the framework which underpins both human cognition (certainly our ability to express our intelligence, and perhaps necessary to actually having intelligence) and ‘generative’ artificial intelligence (large language models).

We can already see several concepts in these definitions which are relevant to a discussion about what, if anything, might be helpful to defining ‘artificial’ intelligence:

1. Thinking
2. Reasoning
3. Understanding
4. Self-awareness
5. Emotion
6. Creativity

To take these one by one (and appreciating a level of circularity in definitions), what is ‘thinking’? We know what it is for – so living beings can process and respond to their environments – and that it is something that happens in our biological brains. As far as this goes, many animals also seem to think, although not to the same extent as humans do, and so we infer that intelligence seems to be a spectrum with an increasingly sophisticated and complex range of capabilities, but which is unique to biological creatures. The Collins Dictionary defines thinking as ‘the activity of using your brain by considering a problem or possibility or creating an idea’ – and so in at least two ways a computer cannot be said to be ‘thinking’ – firstly, in having no biological brain to use, and secondly, it cannot consider a problem or create an idea. For most people, thinking most obviously takes the form of an ‘inner voice’ (although not everyone reports experiencing this, or not to the same extent, and that this may affect cognitive abilities), and a computer does not have an inner monologue. Studying variation in inner monologues may, however, throw light on how a computer might process in ways analogous to thinking – for example people with little or no inner monologue might think in images instead. While defining thinking as being fundamentally biological seems overly limiting as we can conceive of similar perception, processing and responding happening via sensors in robots, for example, at present computers cannot be said to be ‘thinking’ because they are not really considering or creating from the data inputs they receive.  Closely linked to thinking is the idea of ‘reasoning’, that is using existing knowledge or new information to think about something in a logical and sensible way in order to draw inferences, form a conclusion, make a prediction, or solve a problem – so a more targeted and structured subset of thinking. This process is closer to what generative artificial intelligence systems can do, apart from the action of ‘thinking’ about it, which is substituted instead with processing vast datasets of information using a sequence of actions and subroutines, and pattern recognition algorithms. While I would not yet attribute true ‘reasoning’ to any AI model – even the most advanced – in the future an AI model may be accurately described as ‘reasoning’ by consistently ordering sequences of information in ways which lead to a reliable conclusion, with a process which can be communicated and understood by another.

What AI definitely does not have, or not yet, is ‘understanding’. This is particularly poorly defined but is having sufficient knowledge about a topic, comprehending concepts and symbols, being able to link them together, and to draw meaning from them. Meaning is probably the most crucial part – a human can (usually) connect a word, or idea, or symbol, with multiple underlying meanings and form a deeper interpretation in ways which are, so far, beyond the abilities of a computer. That is why humans appreciate art and literature, infer subtext which is not directly expressed, and communicate in indirect ways which are highly contextual and cultural (that is, the same meaning may not be apparent to everyone as it depends on their own experience, education and awareness: having the ‘inside’ knowledge of what something means can also be used to include and exclude). Shared understanding is the basis for social groups, relationships, and is somehow fundamental to building our identities and leading a satisfying life. Understanding is beyond the abilities of AI now and most likely for the foreseeable future, other than by teaching the AI model every possible interpretation and providing it with the complete relevant context by which to select the correct meaning, which again is not ‘understanding’ but simply being better informed.

By self-awareness, we do not mean ‘consciousness’ (being aware of and able to think about the fact that one exists) which even for humans is extremely difficult to define or explain, but is seen as likely being an emergent property of a sufficiently complex network of biological neurons or, possibly, quantum-level phenomena. How consciousness arises from biological matter is one of the most interesting and exciting areas of neuroscience and perhaps one in which advances will be made in the coming years. A reader from the future may laugh at, what might seem to them, our truly primitive understanding of the human mind. Some speculate that a sufficiently complex computer may parallel the emergence of human consciousness with its own – possibly very different – version of consciousness emerging spontaneously from silicon matter rather than biological. ‘Self-awareness’ as a component of intelligence means the degree of understanding of self: having a sense of self as distinct from others, and the ability to guess what another may be thinking and the reasons behind their actions (theory of mind), and the ability to understand, perceive the reasons for, and to think about one’s own emotional landscape and actions (metacognition). Closely linked to self-awareness is the ability of humans not only to feel emotion but, crucially, to enact ‘emotional intelligence’ or ‘emotional maturity’: the capacity to be aware of, control, and express one's emotions and behaviours in ways which is modulated by one’s understanding of the effects on others and on ones’ future self. While humans certainly vary in these abilities, even the most advanced AI model is nowhere near having these kinds of self-aware and emotionally-intelligent capabilities, and may never do so.

The final component of intelligence is ‘creativity’, at its most simple level using one’s imagination to invent something original. This is increasingly contested, because AI can generate never-seen-before images in various art styles, for example, or generate a short story which is not a straight copy of anything existing (the current quality of such outputs is a subjective judgement). The creative industries would argue that this is because the generative AI models are trained on their (often copyrighted) original works and thus are ripping-off human creativity, and furthermore that AI creations lacks genuine intent, emotion, and personal experience which is what makes the creative work meaningful (although in many cases, human creativity is similarly the result of data inputs (learning, skills and experience) being applied – certainly many artists can point to their inspirations, and deliberately adopt another’s style without directly copying). Taking intuitive cognitive leaps is also seen as part of creativity, perhaps in unexpected directions. AI systems have in some cases identified original areas for potential research by highlighting gaps and opportunities, thanks to their ability to analyse very large datasets (and in some instances identifying the correct solution thanks to having better access to the academic literature than the human researchers), but because the AI model lacks understanding, these ideas may turn out not to be useful – for example AI may help identify theoretical proteins but do not understand the biology.  Some feel that creativity is a unique human attribute, ubiquitous among all human societies throughout time, and even fundamental to what it is to be human. AI seems to threaten this idea. AI also fails in some of the other key characteristics necessary to be intelligent, including having memory (context windows) and connecting current and previous data inputs, and in outputting rare and unusual insights from its datasets (having a tendency to regress to the mean or most typical solution).

So, across many of the dimensions we have identified as necessary to being considered ‘intelligent’, AI currently fails most of them. The term ‘artificial intelligence’ was coined by John McCarthy in 1955 in order to get people interested in, and attract funding for, his research into how to make machines use language, form abstractions and concepts, and solve the kinds of problems now reserved for humans and improve themselves. While computing has made huge advances in using language and problem-solving, including in recent years through the development of large language models and the popularisation of using AI models in everyday life, intelligence still seems a long way off.

While this might seem like semantics, I would argue that the inaccuracy of the terminology used for AI, and to describe how it works and associated concepts, is in fact deeply unhelpful. What generative AI is doing is simulating some specific kinds of intelligence in ways which can look like intelligent behaviour, including inserting a pretence of reasoning steps like ‘hmm, I’m thinking about that’ to make them sound more human-like.  Actual ‘artificial intelligence’ is at best a future goal or a metaphor, but unsurprisingly is being taken literally by the public, and investors, who are taken in by what seems like a convincing simulacrum of human-like intelligence, and who therefore attribute characteristics to it which are in fact not present, or not yet. People have a cognitive bias (a mental shortcut) to anthropomorphise, or attribute human-like traits and intentions to non-human things. In the case of AI, this can lead us to make inaccurate judgements about its accuracy, capabilities, and reliability, and to ascribe intentions towards us or others which are in actuality, imaginary. AI is often described by its makers for instance, as having displayed ‘PhD-level’ abilities as evidence of how smart it is, as if it were an intelligent student, but while an AI (or any computer for the past several decades) can certainly solve mathematical and computational problems that need advanced degree level education to match, being awarded an actual doctorate entails years of original and painstaking research on a novel topic – one of the things which AI cannot do.

AI is often erroneously and misleading described as ‘thinking’, ‘reasoning’, and ‘understanding’ when it is not – and it is not clear whether the researchers using those terms themselves mean it literally according to their own understanding and application of the terms, or if they are speaking in analogies drawing comparisons with humans as ways to describe what they mean in terms we are familiar with (drawing on a long history of likening computers to brains), or we simply lack better terms to describe new concepts in computing. What is very clear is that members of the public, now exposed to these emerging technologies and hearing these terms applied to AI, often over-attribute levels of intelligence and meaning which are not present, sometimes leading them to believe AI is their friend, has their interests at heart, cares, or trust that it is correct in its ‘advice’. It’s very easy to fall into this cognitive trap even for people well-versed in the technology. It would be much more useful to clarify what is and is not meant when we talk about artificial ‘intelligence’, or even better to rename it, more accurately, something like ‘computer-simulated intelligence’, and to coin new terms for what it is doing when a generative AI model takes inputs and generates outputs (‘predictive analytics’ maybe?).

AI is not yet really intelligent – but maybe our understanding of what intelligence is will itself evolve. Is pretending to be intelligent, in ways which seem indistinguishable from really being intelligent, all the same in the end? Is creativity defined by the process, or by the end product? At the very least, AI should make us take a hard look at what we believe is intelligent and our self-serving definitions of it, and maybe widen our definitions and spectrum to include the potential for future superintelligent computers, or perhaps human-AI cyborgs.