Machines Can’t Decide What’s Human

After toiling away for days on my paper, it is finally finished. Hours of writing and proofreading have led to this moment. I double-check the formatting of my citations one last time before downloading the PDF, ready to submit. But I cannot relax yet: it’s time to run my essay through multiple AI detectors.

You might be wondering, Tara, did you use artificial intelligence to write the paper? Of course not. But some professors run assignments through AI detectors, so every paper is subject to scrutiny.

A flag isn’t just embarrassing; it can mean redoing the assignment, a grade of zero, or being called before the Honor Council. The writer in me deflates as the checker highlights sentences in bright yellow, claiming there are patterns in my prose reminiscent of the way machines talk. I sigh as I delete the beloved em dashes and perfectly placed semicolons that read as incriminating evidence against me.

Our world is increasingly evaluated by machines. AI detectors are one of the growing list of algorithms deciding whether humans are qualified, original, or telling the truth, with limited means to appeal the verdict. The problem is not only losing something human while adapting to an increasingly technological world. The technology we are adapting to — and placing our unwavering trust in — is unreliable.

We are letting unaccountable machines dictate our actions unfairly, and it has a human cost.

The false-positive problem is not just an anecdote. When a search engine optimization specialist copied the Declaration of Independence into an AI checker in 2024, it was met with a result of 98.51 percent AI-generated. Moreover, Vanderbilt University noted that even if an AI detector has a low false-positive rate, many papers still risk being flagged incorrectly when students submit thousands each year. Stanford researchers also found that checkers can be biased, marking non-native English writing as machine-generated far more often than native writing.

So I adjust. I shorten my sentences or make them clumsier. I break up the rhythm in the writing style that I spent years developing. I trim words that sound too polished. At this point, I’m editing for plausible deniability, tailoring my words more to the output of an algorithm than the actual assignment.

How do I prove myself to something that cannot truly see or understand me, let alone know when its own result is wrong?

I’ve always written differently for my professors than for a group chat. I’m not claiming that performance and adjustments for a specific audience are new. But the problem is that the audience has changed.

A teacher, an editor, or a reader can be talked to. I can explain a strange word choice, appeal a bad grade, or argue my case. The relationship runs both ways. A detector is not that kind of audience; it does not truly listen, it does not make a decision reliably, and it does not owe me a reason. The conversation is removed, and it becomes an unproductive guessing game.

Philosopher John Dewey believed that growth was essential to education, and he treated new tools and conditions not as threats, but rather as means to enhance learning itself. One could argue that using this logic, editing for a detector is just another adjustment in our society’s long history of writing.

But changing real, human writing for an AI detector’s approval does not belong in that history. Dewey’s version of growth assumed something a detector cannot offer: a responsive environment, one that could be understood, tested, and pushed back against. That is what makes adaptation intelligent instead of merely obedient.

Rewriting my voice to satisfy modern technological systems is not growth — it is a constraint. Dewey argued that students should be shaped by an environment they can question. Today, ours doesn’t answer.

Some institutions have already come to this conclusion. In 2023, Vanderbilt stopped using Turnitin’s AI detector entirely. While reconsidering a flawed tool is a step in the right direction, we must undo the larger habit it creates. We need to resist the reflex to let a machine render a verdict on something human, then treat the verdict as final because we cannot reason with it. Resume screeners, content moderators, plagiarism checkers, AI detectors, and every other algorithm judging us are all examples of an arrangement in which the party with the least understanding of us holds the most power over us.

I submitted my essay. It might have been worse than the draft I wrote initially. But at least the machine liked it.

Tara G. Malhotra ’29, a Crimson Editorial editor, lives in Cabot House.