This article is written by Vincent Straub, a researcher in the Public Policy Programme at the Alan Turing Institute and Big Data Institute at the University of Oxford. Previously he worked at the Oxford Internet Institute and the innovation foundation Nesta.

  • The problem: With the increasing sophistication and adoption of AI, government leaders need to adopt a more accurate vocabulary to stop the spread of false narratives.
  • Why it matters: The history and development of terms used to talk about AI aligns with institutional privilege and confers authority; terms set the agenda and guide policy: we should aim for accurate and informative language that can help inform us all.
  • The solution: One approach forward is championing a glossary of universally accepted AI terms which incorporates more terms that better highlight the machine nature of AI models; in our work at the Alan Turing Institute, the UK’s national institute for AI, we have been actively working on this problem.

As the third full year of the UK Government’s national AI Strategy is well underway, recent advancements in the capabilities of AI systems continue to grab the headlines. The latest example continues to be Large Language models (LLMs) like ChatGPT, which derive their power from vast amounts of language data. 

Although LLMs can be adapted to an impressive range of tasks, the language used in government, as well as business and the media, to discuss them is often taken out of context, confused and at worst, akin to the effects of misinformation. 

💬 Let the author know what language you are currently use when communicating about AI and what alternatives you suggest by leaving a comment below

The latest of many examples is the word ‘hallucinate’, which Cambridge Dictionary named word of the year 2023. In simple terms, this term refers to the tendency of programs like ChatGPT to make errors and provide made-up or false information. 

Yet, saying that a chatbot is hallucinating makes it sound like it has a mind of its own that sometimes imagines things. As Giada Pistilli, principal ethicist at Hugging Face, which hosts many state-of-the-art AI models, has noted, the anthropomorphising language of hallucination obscures what AI models are really doing. A more accurate term may simply be ‘malfunction’.   

Philosophically fraught words

Many policymakers might be forgiven for adopting terms promulgated by researchers without knowing that many of these terms are themselves contested. Fairness can mean different things depending on which researchers you ask, for instance, but is regularly invoked in discussions about AI.

Half a century ago, when AI was still largely the purview of science fiction, the consequences of language were smaller. Now words really matter.

However, as many politicians and the wider public still struggle to know what to make of AI and how to separate AI fact from fiction, the language we use clearly matters. 

As esteemed computer scientist Murray Shanahan has written, we need to avoid “the misleading use of philosophically fraught words associated with human cognition to describe the capabilities of LLMs”, words such as 'belief', 'understanding' and — perhaps the most uninformative of all — 'consciousness'. 

Others have already stressed this point and argued that we need to stop talking about LLMs in a way that we talk about humans. With all the excitement and fear surrounding AI, especially LLMs, we must be measured.

We need to know what this technology can and can’t do, what risks it poses, so that we can both have a deeper understanding and a more comprehensive account of its societal impact. A vocabulary that stresses the machine nature of AI models is the vital first step towards these two goals. 

Where to go from here?

Over the last few years, a number of useful terms to discuss LLMs have, of course, gradually entered the lexicon. These include algorithmic bias and interpretability, among others. However, these terms are often relegated to the side-lines, grouped in ‘ethical concerns’ and discussed after the fact. So where do we go from here? 

One approach forward is for the government and international organisations to champion a glossary of universally accepted AI terms which incorporates more terms that better highlight the machine nature of AI models.

The EU has made progress in this area by developing a glossary of human-centric AI terms. This can serve as inspiration for the UK Government, which lacks an equivalent. 

In our work at the Alan Turing Institute, the UK’s national institute for data science and AI, we are actively thinking about this problem. In a recent paper, we mapped over 100 concepts currently featured in discussions of AI, including interpretability, explainability and oversight. Based on our review, we in turn introduced three of our own new terms to capture the language we think governments should be using.

Other new terms that foreground the artificial nature of AI and encapsulate the entire set of criteria for evaluating AI systems would provide policymakers with a common language to talk about all types of new AI systems and the implications they carry for government and society. 

Adopting and promoting new terms may sound like wishful thinking, but it has happened before. Just take the word ‘sustainability’, the origin of which dates back to less than 45 years ago and which has since become a cornerstone of business talk and political debate. 

Half a century ago, when AI was still largely the purview of science fiction, the consequences of language were smaller. Now words really matter and we have a chance to get them right. The history and trajectory of AI language aligns with institutional privilege and confers decision-making power. Terms set the agenda and orient progress: we should aspire for holistic and accurate language that can help inform us all.

Done reading? Make sure to share your own thoughts on governments' responsibility in emphasising the machine nature of AI models by leaving a comment below ⬇️

(Image credit: Unsplash)