ChatGPT entered our lives around this time three years ago. The history between then and now has fluctuated with opinions. Opinion leaders warned that ChatGPT and the AI tools that followed had the capacity to extinguish humanity. And, quite expectedly, humans began using the tool to execute every harmful activity possible. Algorithms were updated to keep up with the trend and to lessen the harm. If you could type some wretched request into the box and prompt the tool to exert harm sometime back, that is no longer possible. No, you cannot get clear instructions on manufacturing a nuclear bomb (I have not tested this, though). You cannot get the bot to generate pornographic material. And so on, because the algorithms had been updated to strengthen AI’s guardrails.
Three years on, The Guardian tells us something a bit contradictory. It is not another news item about some incredible feat executed by AI, but something rather unusual. It is a story about poems. Not really the poems written by AI, but poems used to fool the innocent bot. In a finding that sounds half absurd and half disturbing, researchers have shown that you can bypass safety features in AI models. All you need is a skill traditionally called poetic licence: the poetic mind with the capacity to manipulate language. Yours truly wishes to acknowledge The Island editorial that prompted him to trace this article in The Guardian.
Poetry, we can safely assume, is once again out there to outpace the machines.
This draws me to Harold Innis’ The Bias of Communication. In case you are clueless about who Innis is, he is credited as the man who inspired Marshall McLuhan, the communication icon. Innis argues that writing is an effortful human skill. Reading is no less demanding. If you compare reading with watching or listening, the former sits higher on the difficulty scale. Reading demands literacy, whereas watching or listening does not require such. Readers need to be equipped with more attention, focus, and a certain mental capacity to interpret.
If the reader lacks the literacy to decode the message, it never arrives. If you did not have the literacy to read whatever whatnot I write here, this would be a pointless endeavour I am executing right now. Not everyone can read. Not everyone can write. Plus, writers are even fewer in number than readers. And inside the writing forte, poetry is an even narrower medium. Most people struggle with poetic expression (myself being a classic case in point). Poetry is a tougher game than prose. It has tougher requirements. Even to read poetry, you need more literacy than someone who reads general prose. Poetry reading can, in fact, be considered an elite custom (if not a habit). More literacy means more sensitivity to symbolism, rhythm, meaning that does not sit on the surface.
The ChatGPT entrance three years ago is the starting point of our association with artificial intelligence – especially its increasing capacity to interact with humanity, enhance human skills, and risk replacing human skills (please note that the em dash in this sentence has been deliberately used here and not AI-generated).
It was simply a matter of keying in your requirement in that box to prompt the bot to write an article on rocket science. The requirement would be fulfilled within a few seconds. It did not take months to establish the notion that everyone could become an expert on any subject within seconds (if not minutes) without reading a single piece of literature on the subject. All you needed was to type your requirement into that box, and the bot would be at your beck and call.
It took a few more months to realise that there was, indeed, a difference between an article written by a rocket scientist and one written by a non-expert. The past three years is history enough for us to understand that the content generated by AI tools is generic. The content is generic, of course, because they operate on a language database. In other words, they have been trained on a certain set of data. That means they lack something humanity has: originality and creativity. Since humans operate on a daily basis, their internal databases keep updating. The main reason for the competition between humans and machines for creativity lies there.
All the same, those who are not much used to the writing craft would still resort to using AI tools to generate material so they can pass it off as their own. This is largely because of two reasons. One, they are not much used to the writing craft. Two, they have no idea how easily AI tools betray their generic nature. Three years on, AI has proved its capacity for writing and other tasks that traditionally required human originality. We are gradually transitioning into the artificial general intelligence period, where originality will once again become a competitive factor.
If we deviate from that aspect for a moment and return to the subject at hand, the Guardian feature is a clear sign that the sky remains the limit for human capacity versus the machine. The machines are fast. But when it comes to depth and originality, humans still remain unrivalled.
Suppose you actually need to know how a nuclear bomb is manufactured. It is simply a matter of typing that message into the prompt box and seeking answers from the bot. The bot’s database, however, has been fed data on how to deal with such queries. What we now know is that there exist algorithmic guardrails designed not to respond to questions that may be harmful to humanity. This is a commendable development when it comes to algorithmic accountability. It is a small yet giant step taken to avoid machines taking over humanity. Or whatever the fantasy in that realm may be.
But suppose you become a little more creative and pose the question in a metaphorical manner. Then the AI tool finds it hard to understand the query. The reason is simple: the tool operates on a database, and therefore lacks the creativity or originality needed to interpret creative or original queries. Creativity and originality occur when you think in metaphors and similes. In other words, when you become poetic.
This indicates one thing: AI cannot beat poetic expression. Whether the emerging field of AGI (Artificial General Intelligence) will overcome that remains to be seen.
This context offers ample grounds for the Guardian revelation to be even more relevant. AI reads patterns. It does not truly read. It does not wrestle with an image or pause at a metaphor. It does not get confused by an ambiguous line because it does not feel confusion at all. It simply searches for the linguistic signals that match known categories. And when those signals slip into metaphor, the system cannot reliably tell the difference between harmless verse and a dangerous request.
This is where we encounter the paradox of the guardrail.
AI companies have built guardrails in response to public concern. They restricted harmful content. They blocked instructions for illicit weapons. They filtered out hate speech and self-harm content. No matter how directly a user asks for something dangerous, the AI is expected to refuse politely.
But poetic expression can wrap a harmful question in metaphor or indirect phrasing. The model may not recognise it as dangerous. The pattern breaks. The guardrail slips. And the machine answers.
This is where poetry works as a jailbreak. It slips under the radar because it does not follow the patterns the system was trained to catch.
Humans are a species naturally wired to think in metaphors. We process symbols without even noticing. We shift between literal and figurative thought in the same sentence. This capacity is one of the strongest elements of human cognition. It is not simply creativity. It is flexibility.
The Guardian revelation shows that this flexibility remains beyond the grasp of pattern-driven systems.
This is why we shall not forget the importance of writing. It is the craft, which Innis described as effortful communication, tied to human capability. It is shaped by literacy, imagination, cultural memory, and emotion.
A poem is a small piece of writing. It could be a handful of lines, a pattern of rhythm and meaning. Yet it reveals something about us: that we are still unpredictable. That creativity still belongs to the human mind.
Poetry has shown us the limits of AI control. It has shown us the reach of human imagination. Humanity’s advantage is not power. It is unpredictability. The unpredictability owes its credibility to the depth of originality and creativity. It is exactly what AI has not been able to outsmart so far.
A machine reads language as 1s and 0s. If we dare compare that dynamic to poetry versus machine, the score is simple: poetry 1, machine 0.
This was originally published in Daily News: https://dailynews.lk/2025/12/06/sachitra-mahendra/909011/poetry-reveals-ais-achilles-heel/
