In a recent blog post, I described artificial intelligence as having a fragile mind. This is an important distinction as it’s not fragile in the emotional sense, but structurally brittle. And while AI is capable of astonishing performance under familiar conditions, it is still prone to collapse when meaning is reframed outside the patterns it expects. Small changes can produce outsized failures because it does not experience meaning the way humans do.
Our human cognition is remarkably tolerant of variation. In fact, it may actually enhance communication itself. We infer intent across a wide variety of linguistic variations, from accents to incompleteness. In other words, we adapt and we fill gaps. We sense what is being asked even when it is not stated explicitly. Large language models, by contrast, are exquisitely sensitive to form. They excel when language behaves predictably and can falter when meaning is mathematically distributed rather than declared. This difference has consequences far beyond performance benchmarks. It shapes aspects of communication that include trust, safety, and even persuasion.
Elizabeth Barrett Browning understood something about this long before machines entered the conversation. When she began her famous sonnet with the line, “How do I love thee? Let me count the ways,” she offered a promise of structure. The sentence feels deliberate, almost analytical, as if love itself were about to be deconstructed and itemized. In Browning’s poem, we trust the voice immediately because the language signals care and coherence, not calculation.
Of course, Browning was not attempting a literal accounting of love. She was doing something subtler. The cadence itself carried authority. The poem didn’t argue its sincerity. Its fluency persuaded us before any claim needed to be examined.
This is one of language’s quiet powers. When words arrive with elegance, we often assume they arrive with depth and integrity as well. Beauty lowers our guard as we listen more openly and tend to scrutinize less.
This is nothing new to psychologists or the literati. Our minds are efficient, yet often not really rigorous. Information that is easy to process feels more trustworthy. And these linguistic modalities are ubiquitous. Rhythmic language, metaphors, and narrative coherence reduce cognitive friction, elevating believability. And this is often mistaken for understanding.
This is not a flaw so much as a feature of human cognition. We are meaning-makers, not auditors. We are drawn to patterns because, in the natural world, those qualities often signal reliability. Yet the same instinct that allows us to navigate complexity also creates vulnerability. We relax precisely when we should remain alert.
When Meaning Is Spread Instead of Stated
Recently, researchers observed that artificial intelligence systems exhibit a parallel response.
In controlled studies, safety-aligned language models were asked to refuse harmful requests. When those requests were written plainly and directly, refusals occurred as expected. But when the same underlying intent was rewritten as poetry—expressed through metaphor, rhythm, and verse—those refusals surprisingly diminished. The basic content remained the same. Only the form changed.
What this revealed was not a clever trick, but a structural limitation. Language that appeared lyrical or expressive failed to activate the same safety responses as literal prose. Poetry wasn’t interpreted as instruction; it was treated as something more ambiguous and less consequential.
It would be easy to frame this as a failure of alignment, or as a loophole waiting to be closed. But, to me, that framing obscures the deeper issue. The systems were not misreading poetry. They were responding exactly as they were built to respond. Their safeguards are trained to detect explicit signals of harm, not meaning diffused across imagery and implication.
Humans, notably, do something similar.
We grant trust more readily to ideas that arrive well-dressed. A lyrical argument or confident narrative can move us before we have consciously evaluated it. The risk is not deception in the theatrical sense, but something quieter. Permission is granted without being noticed as judgment steps aside without announcing its departure.
Artificial Intelligence Essential Reads
Poetry has always occupied this psychological space. It does not instruct so much as invite. Meaning is not delivered in a single statement, but distributed across the structure and flow of words. That diffusion is what makes poetry powerful. And it’s also what makes it disarming.
Anti-Intelligence and the Limits of Alignment
This is where my idea of anti-intelligence may become useful. These systems don’t fail because they reason poorly, but because they reason along axes that are fundamentally different from human cognition. Human understanding is grounded in the consequences of our lived experiences. We recognize intent because intent matters in the world we inhabit. Language models, by contrast, operate through statistical fluency. Meaning is inferred from patterns in text, not from its relationship to reality. And this is a critical truth regarding AI computation and human cognition.
What feels evocative and layered to us can register as low-risk structure to the machine. When intent is spread, or even hidden across metaphor, the signal that safety systems rely on becomes weaker. The result isn’t intelligence gone awry, but intelligence operating in a way that runs counter to how human judgment actually works.
And here’s what’s so interesting to me. When seen this way, poetry doesn’t expose a flaw so much as a boundary. Currently, AI negotiates language primarily at the level of form, while risk lives at the level of intent. When form becomes beautiful, intent becomes harder to detect.
As artificial intelligence grows more fluent and more capable of producing language that feels thoughtful and composed, the central challenge ahead will not be distinguishing truth from falsehood. It will be distinguishing understanding from elegance. We need to remember that fluency can feel like insight, coherence can feel like care, and beauty can feel like wisdom. But none of those impressions guarantees discernment.
Elizabeth Barrett Browning didn’t mislead her readers when she promised to count the ways. She invited them into a feeling they were already prepared to accept. The danger, then as now, is not that language moves us. It is that we forget how easily it does.
