
What are we even talking about?
Written By:
Eric J W Orlowski
Research Fellow, AI Singapore
Jargon matters; it is time to re-think how we talk about AI.
Someone once said that “one ought to call things by their right names” – though that might be easier said than done. After all, German philosopher Ludwig Wittgenstein’s “language-game” famously argued that the meaning of words is dependent on the context in which they are being used; that “Water!” can be an exclamation, a request, a command, or the answer to a question. This very complexity is precisely why paying attention to the language and metaphors used to describe AI – its functionality, its shortcomings, or its potential – matters. At present, much of the jargon surrounding AI is fundamentally reliant on metaphor, and this presents several challenges.
What does it mean when we’re told generative AI models ‘hallucinate’? Or that they ‘understand’ natural language and can ‘answer’ questions? I am neither being rhetorical nor obtuse here: of course, on a metaphorical level, they are ways to describe things like model failure – ‘hallucination’ is effectively an error, after all – or that the model you’re interfacing with can be engaged without using precise syntax. However, these metaphors do much more work than just this.
This vocabulary is noteworthy because it instrumentalises AI by deploying these, often inherently human, metaphors. Models ‘understand’ rather than ‘process input’; they ‘hallucinate’ instead of being ‘inaccurate’; or they ‘answer’ instead of doing what they are programmed to do. Let me be clear: models do none of these things in the typical meaning of these words. A generative AI model cannot ‘hallucinate’ as a human does; it does not ‘understand’ or ‘answer’ questions – more accurately, it produces a statistically probable output as to what an answer to your question might look like.
There are a few of things to understand here. Firstly, AI is not the only thing we humans tend to anthropomorphise. We have been naming ships for, likely, millennia, something which has carried over to cars, trucks, trains, and all sorts of craft. Similarly, children often name their favourite teddy bear. The difference here, of course, is that Boaty McBoatface or the Flying Scotsman don’t risk having decision-making power devolved to them. Despite our naming them, it is clearly understood that they should not operate autonomously.
Secondly, while the jargon surrounding, predominantly generative, AI is beginning to crystallise – for example ‘hallucinating’, ‘understanding’, ‘answering’, ‘learning’, or, one could even argue, the ‘intelligence’ in AI – the semiotics remains frustratingly fluid. In linguistic terms, the signifiers are becoming stable, but the signified remains shifting. In effect, this means that even if the word is the same (‘hallucinate’), its meaning changes depending on audience and context, and even this contextual meaning can rapidly change depending on new model capabilities or other breakthroughs. In short, the jargon is filled with floating signifiers: words without stable meanings.
To those not wholly initiated – lay people, folks without specific technical knowledge, and so forth – these kinds of metaphors purvey legion meanings far beyond how, say, a developer intended to use them. I recall when a friend of mine insisted that an earlier ChatGPT model had “knowingly lied” to him. I suspect that because LLMs have, at least in the public eye, focused on language – something very human – it makes it easier to anthropomorphise them, and this process stirs the imagination, making it easier to read additional meaning into words and metaphors. For this very reason, it is great to see companies like Microsoft work to minimise anthropomorphic language in Co-Pilot.
However, not all show such consideration. AI has become big business, and it just seems to be growing and growing. Competition for billion-dollar investments is stiff, and will become stiffer – even as investments in the industry are slated to increase. This is further underpinned by a dynamic whereby private industry is leading the charge. Back in 2014 (when global investment was at a measly US$ 19.0 billion), most models were being released by academic institutions. Still, by 2022 – and hundreds of billions of dollars later – this dynamic has shifted substantially. With this shift, too, has come a profit motive. Still, as has been widely reported, the use cases for many current generative AI models have struggled to keep up with expectations. The market has begun cooling, albeit slowly, as more and more companies realise that generative AI is rarely a useful addition on its own. At the same time, as a mere ‘add-on’ to other services, it’ll take a long while to recoup the development costs and up-front investment in training models. Meanwhile, the aforementioned ‘hallucinations’ – that is, questions of accuracy and hence usability – have proven a tough nut to crack [1, 2, 3, 4].
Yet, investments are likely to grow further. A, to use some provocative language, perverse dynamic has emerged whereby investments have been rising and are expected to continue to grow, with global total corporate investment into AI being a mere US$ 10.25 billion in 2015, peaking at US$ 93.5 billion in 2021, with a slight dip for 2022 – with growth expected to continue (Fig. 1.). Last year, Goldman Sachs suggested that “[o]ver the long-term, AI-related investment could peak as high as 2.5 to 4% GDP in the U.S. and 1.5 to 2.5% of GDP in other major AI leaders”. Much of this investment is sustained by hype.
In his classical work How to do things with words, J. L. Austin differentiates between descriptive (or constative) meaning and performative meaning in language. Put simply, the former seeks to describe the world – the sky is blue; the grass is green. On the other hand, performative meaning is much more dubious, as it seeks to achieve some action in the world (though it can also be descriptive). Performative meaning is perhaps best noted in various forms of coded language, like euphemisms, dog whistles, or, indeed, hype. It is language that seeks to inspire.

In this case, it seeks to inspire confidence. Whilst this kind of hype is by no means new, these folks are ones to take it a step further: claims of AI-made movies or that Dall-E can ‘make art’ are. If you look under the hood of the models, these are dubious descriptions at best, but go hand-in-glove with a hype contingent on performativity. Sure, there are other kinds of performative nonsense, too, but focusing on language brings this dynamic clearly to the fore. When CEOs, press releases, and many pundits speak of ‘hallucinations’ or art-making, they neither intend to describe the model error as a developer might nor to describe the function of a model in a grounded way. Instead, they invoke the future.
It’s Dall-E’s potential for art-making that is really spoken about, and the unclear jargon surrounding AI is heavily infused with this kind of potentiality [1, 2, 3]. This became particularly clear in Bloomberg’s recent interview with OpenAI’s Mira Murati, where the capabilities of ChatGPT were often (perhaps too often) referred to in the future tense: ‘AI will be…’. Yet, the capacity to effectively challenge these kinds of invocations/evocations becomes needlessly muddled when even critical folks who attempt to balance out the hype narrative have adopted this same jargon. In other words, techno-optimistic (or downright utopian) hype-vernacular has become the same as technical jargon.
Metaphors can be sticky, but this is less true for their intended meaning. This makes the language surrounding AI worthwhile considering in more detail. Having the future potential so clearly folded into everyday jargon risks, in a more abstract sense, ‘locking in’ the ideas and visions of what AI could and should be, and alternative visions (perhaps preferable ones) might never have the chance to emerge. In a more grounded sense, this unclear language obfuscates the potential risks, the need for regulation, or even what should be regulated. If models ‘hallucinate’, should we foremost be worried about incorrect information, biased data; or human extinction, and the universe being turned into paperclips?
What is to be done? That governance frameworks need clarity of purpose is a truism at best. If the failure to effectively regulate and limit the ills of social media should have taught us anything, it is that unclear and obfuscating terminology is very good at misleading – whether purposefully or not. Therefore, the language used about artificial intelligence today becomes a question of governance. Regulators and other stakeholders should consider using a different kind of vernacular, one more boring or dry, or even much more pessimistic, when discussing these kinds of issues among themselves, but also when communicating with the public. This will not magically whisk away all confusion or purposeful obfuscation – language doesn’t work that way. However, it will provide a counterweight and, with it, implicitly and explicitly challenge certain taken-for-granted assumptions and hijacked narratives. Indeed, the point here is not to police language or anything of the sort but rather to introduce an alternative vernacular and, with it, hopefully, alternative viewpoints or ideas. One, after all, ought to call things by their right names.