TechCrunch has identified a linguistic fingerprint that has quietly colonized the internet: the rhetorical construction that frames an idea as a correction or escalation of a simpler one, telling readers that something is not merely one thing but, in fact, something grander. According to TechCrunch, this particular sentence pattern has become so pervasive in AI-generated writing that its presence in a piece of text has shifted from being a mild red flag to something approaching a definitive tell.
To understand why this matters, it helps to understand how large language models learn to write. These systems are trained on vast repositories of human text, and they develop strong statistical preferences for constructions that appeared frequently in that training data and that were, presumably, rewarded with engagement or readability signals. The "not just X, it's Y" construction is a genuine rhetorical device with a long and legitimate history. Speechwriters use it. Columnists use it. It creates momentum, signals that the writer is about to reframe the stakes, and flatters the reader by suggesting they are being let in on something beyond the obvious surface. It is, in other words, exactly the kind of move that would appear often in high-quality writing and exactly the kind of move a language model would learn to reach for constantly.
The problem is one of proportion and context. A human writer deploying this construction does so deliberately, in specific moments, because the material genuinely calls for an escalation of framing. A language model deploys it because it has learned that the pattern correlates with authoritative, thoughtful-sounding prose. The model has no sense of when the device is earned. It simply knows the shape of the move, not the judgment behind it. The result is writing that reaches for rhetorical weight it has not actually built up, in paragraph after paragraph, until the cumulative effect is exhausting and, for an alert reader, immediately recognizable.
This sits within a broader and accelerating pattern that has been playing out since large language models became widely accessible to the public. Researchers, educators, and editors have cycled through a series of proposed detection methods, only to find that each supposed signature fades as models improve or as detection awareness filters back into training pipelines. Early tells included an over-reliance on certain transitional phrases, a tendency toward suspiciously even paragraph lengths, and an almost compulsive need to end pieces with a tidy call to reflection. Some of these patterns have already been smoothed away. The "not just X, it's Y" construction appears to be the current stratum in this geological record of AI prose habits.
The consequences extend in several directions. For editors and publishers, this represents a practical, if temporary, tool for flagging content that may have been produced without meaningful human authorship. Platforms that have committed to distinguishing human from machine-generated writing, whether for trust, regulatory, or commercial reasons, now have one more pattern to feed into their detection frameworks, though the likely reading is that this particular signal will have a limited lifespan before it gets trained away. For readers, the more significant consequence is a subtle but real erosion of trust in prose that uses the construction, even when a human wrote it. Rhetorical devices do not survive becoming clichés, and the association with synthetic writing may taint this particular structure for a generation of readers who have learned to notice it.
For writers themselves, the development is a quiet professional pressure. The implication is not only that certain phrases have been colonized by machines but that human writers now face the awkward task of auditing their own natural habits against an ever-shifting list of patterns that read as artificial. That is a strange cognitive burden to place on human creativity, and it suggests a longer-term cultural negotiation about what idiomatic written expression is allowed to look like once machines have learned to imitate it.
What to watch for next is a question of timing. The moment TechCrunch's observation circulates widely enough to reach the communities and organizations that fine-tune language models, or that curate the data used in training them, the construction will begin to be suppressed, either deliberately or because negative feedback loops will reduce its frequency in future outputs. The more interesting question is what emerges to replace it. Each generation of AI-writing detection has followed the same arc: identification, publication, adaptation, obsolescence. The next tell is already being written, somewhere, at extraordinary scale, by systems that have not yet been caught reaching for the same move one too many times.
