Can machines think? Having been raised on a fairly steady diet of science and speculative fiction, this is a question that has intrigued me for most of my life. But over the past few weeks, it’s been troubling me for a couple of different reasons.
The EU has announced an inquiry into the X-native AI model Grok, and its ability to create nonconsensual sexualised images. There are arguments, none of which I find particularly convincing, about why this is acceptable or something that could be easily achieved with other tools. What is most disconcerting is the ease and manner of doing it. Simply typing “Hey @Grok (the name itself was taken from Robert Heinlein’s 1961 classic Stranger in a Strange Land, meaning, among other things, to understand deeply or intuitively) put her in a bikini” under a photograph was sufficient.
This ability to summon a tool and use natural language to direct it to create images reflects a wider change in how we interact with machines. For a recent example, you can watch videos of ChatGPT’s voice mode confidently and fluently arguing that the letter r only appears twice in strawberry, appearing to engage in spoken conversation with a user.
While this is both amusing and frustrating, it’s easy to forget that before Google, using a search engine required the user to understand what it wanted to find and how to look for it. Entering keywords returned pages that happened to contain those words, with little relationship between frequency, relevance or quality. Google changed that because users no longer had to know how information was stored or the correct terminology to look for it. You could simply ask questions in normal language or allow it to infer meaning from a text string such as “best pizza Dublin” and get a relevant response.
READ MORE
This change fundamentally altered how we approached online knowledge. The internet became less a directory to be navigated and instead a pool of knowledge that could be summoned. The move to natural language is particularly significant because the methods we have historically used to identify intelligence are closely tied to conversation.
When Alan Turing asked in 1950, “Can machines think?“, he was not trying to settle a philosophical argument. He was trying to make the problem workable. Instead of debating what “thinking” really means, he proposed a test based on observable behaviour. If a machine could converse in such a way that a human judge could not reliably distinguish it from another human, then, for practical purposes, we should treat it as intelligent.
When Turing replaced metaphysics with measurement, he laid down a pattern that still holds. Today’s AI models are not judged by whether they understand the world, but by how well they perform against defined tasks. This matters because AI is increasingly being used to screen applications, flag fraud risk or assist with legal review.
Yet it is not assessed on whether it truly understands law or policy but whether it meets the standards of plausible conversation and passes the Turing Test.
In 1980, the philosopher John Searle challenged this approach with the “Chinese Room” thought experiment. Imagine a person who does not understand Chinese locked in a room with a large instruction manual.
They receive Chinese characters through a slot, follow the rules in the manual and return new characters. To an observer it appears that the room understands Chinese. But inside, there is no comprehension, only rule-following.
This critique has become more relevant with the rise of large language models. Modern AI can generate text that is coherent, persuasive and technically detailed. In experiments last year, ChatGPT 4.5 passed the Turing Test 73 per cent of the time.
Yet they do not possess beliefs, intentions or awareness. Like the Chinese Room, they produce convincing outputs without any inner grasp of what those outputs signify. AI excels where tasks can be reduced to pattern recognition and statistical inference but struggles where interpretation or moral judgment are required.
Much of the EU’s emerging AI governance framework rests on managing risk rather than defining intelligence. But if regulators treat AI systems as quasi-agents rather than tools, where does responsibility for their outputs lie? Conversely, treating them as mere software could underestimate the potential for harmful misuse.
True artificial intelligence, if even possible, is some way off. But in the meantime, we have an enormously powerful and user-friendly tool with obvious productivity gains: faster analysis, broader access to expertise, lower barriers to entry for complex tasks.
What we do with that is a moral, and ultimately human, question. Hopefully it’s something better than deepfakes.
Stuart Mathieson is research manager with InterTradeIreland
