But AI learned those patterns from human writing.
At first, this seemed like a problem of mistaken identification. Some human writers will inevitably be told that their work looks artificial. The larger danger is that the detector may begin changing the writing itself.
Once writers believe that strong structure, clean transitions and polished sentences make them look suspicious, they may begin preserving awkwardness as evidence of innocence. Editors may hesitate to improve a draft too much. Readers may start treating roughness as proof of authenticity.
At that point, the detector is no longer merely judging writing. It is teaching people how to write for the detector.
Economists have a name for this. Goodhart’s law is usually summarized this way: When a measure becomes a target, it stops being a good measure. A test score may begin as a useful indicator of learning, but once schools are rewarded or punished according to it, they begin teaching to the test. The score may rise even as the education narrows.
An AI detector may work the same way. Once writers are judged by whether their prose appears sufficiently human, they will begin optimizing for visible humanity. A tool intended to protect human authorship could end up discouraging some of the qualities that make human writing worth reading.
The real question should not be whether AI touched the prose.
AI presents real problems. A student should not submit a machine-written essay as proof of personal ability. A journalist should not publish invented facts. A lawyer should not file fabricated cases. A writer should not claim personal experience that never happened.
But those are problems of deception, understanding and responsibility. They are not solved by treating every trace of AI assistance as contamination.
The better questions are harder. Did a human being have the central idea? Did the writer notice something worth noticing? Is the reasoning sound? Is the evidence accurate and fairly used? Does the article reflect a particular mind and personality? Can the writer explain why particular choices were made, defend the argument and take responsibility for every claim?
Those are the questions journalism and creative-writing teachers have always tried to answer, although AI now makes them more urgent.
Before generative AI, teachers could often infer thought from finished prose. A coherent essay usually required someone to organize the argument, choose the examples and struggle through the sentences. Now a machine can produce the surface appearance of thought very quickly.
That does not mean we should stop teaching structure, style, clarity or revision. It means we need to judge the intellectual work beneath them more directly. Students should be asked what they noticed, what changed their minds, which evidence matters most, what the strongest objection is and why they made the choices they made. They should be able to explain their work, verify its claims and take responsibility for the final result.
Perhaps we are asking the wrong first question. Instead of demanding to know whether a piece is human or AI, we might begin with two simpler ones:
Is it good? Is it true? Is it live, or is it Memorex?
The old TV commercial asked whether a recording could sound live. Increasingly, we cannot tell.
The standard should not be purity. It should be human ideas, human judgment, human personality and human accountability, expressed as well as possible.
Pangram did not prove that I was half machine. It may have proved something more interesting.
The qualities we are beginning to call artificial are often the same qualities we once called good writing.