2014 marked the year when I completed my master's thesis in philosophy after eight grueling months of work. I remember writing it had been a slow and deeply reflective process, where I sometimes spent hours rewriting the same sentence until it felt just right.

Perhaps a little obsessive as an approach, but I was proud of the result, eventually. And, most importantly, it felt like me.

Years of study didn't just teach me what to express but also how I wanted to express it.

Then, after a few years as an academic writer, I shifted toward the field of AI. That's where I discovered the world of Large Language Models (LLMs), like ChatGPT, but also their so-called "enemies": AI detectors. Part of my freelance work involved flagging AI-generated content produced by other freelancers using these tools during the review process.

I had a hunch early on that these tools weren't entirely reliable. But I wanted proof, so I decided to put an entire chapter of my master's thesis through a GPT detector. Result: 60% AI-generated.

It was rather shocking, really. I thought about all the research and creative effort I had poured into that work, a work that today, in the era of LLMs, would perhaps be considered cheating, potentially even jeopardizing my academic path. Who knows?

That moment pushed me to think seriously about the ethical and social implications of AI detectors. While I understand and even support their general mission, the way these tools currently function risks inadvertently creating stylistic alienation and discrimination.

AI Detectors Know Best: The Algorithmic Authority Bias

First, let's make one thing clear: detectors are not infallible. Far from it. One striking example involves OpenAI's AI text classifier, which was supposed to detect whether a text had been written by AI. Minor problem: according to this tool, large parts of the Bible were apparently written by AI. I don't think it's an exaggeration to say that this example alone shows these tools are far from reliable.

Even without resorting to extreme cases, a noteworthy paper demonstrated how AI detectors produce a considerable number of false positives in academic contexts, disproportionately affecting non-native English speakers and scholars with distinctive writing styles. The repercussions can be serious for them: the stress of having to defend their work, the fear of not being believed, of seeing their career at risk… ultimately, it's also about losing mutual trust.

The flaws of these tools are obvious, yet the common perception is that they're objective and impartial, something we can fully rely on. It seems that when algorithms are involved, people tend to attribute greater authority to them than to humans, as some studies also suggest.

Still, neither AI nor its detectors are transcendent entities. Behind their training are highly educated people, including academics. Subject experts train detectors to recognize patterns typical of artificial writing, which, in turn, is modeled on human writing.

It's like a hall of mirrors: we train AI to write like us, then train detectors to distinguish it from us, creating a kind of epistemic loop.

Writing Under Pressure: The Stylistic Bias

What's perhaps even more worrying are the ambivalent effects of AI detectors: on the one hand, they rightly penalize unoriginal content; on the other hand, they risk homogenizing writing styles, much like ChatGPT and other AI assistants do.

It seems nonsense, but it isn't. AI detectors tend to flag content that exhibits certain characteristics. Long sentences, a linear structure, as well as a neutral vocabulary free of informal expressions, for example, can easily trigger them.

But this is the way some naturally write, either because they prefer a more detached, impersonal tone, or because they're just used to it. For these people, AI detectors can become a serious obstacle. They want to express themselves as they always have, as they've been taught, as feels most natural to them. Yet their stylistic and creative spontaneity risks being mistaken for something it isn't: AI output.

And here lies both a practical and ethical problem: to avoid sounding like an AI, some people may feel compelled to radically change their writing. To escape the persistent suspicion of producing "AI-generated content," they're pushed to strip away their natural style. No more linear structure, no more clean, neutral phrasing, free of informal expressions. AI detectors could wrongly flag your writing. You find yourself having to set aside all the educational background that has shaped your writing over the years. In a sense, you have to unlearn how to write in order to learn anew, often adopting a style that doesn't truly reflect who you are.

This is a form of algorithmic bias. We might call it stylistic bias: AI detectors systematically penalize writing that is measured and restrained. It's a bias that measures humanity through the predictability of writing, not to mention that such predictability can be a conscious stylistic choice, not necessarily a sign of bad faith. But detectors don't make such distinctions. And out of fear of triggering the algorithm, many might feel pressured to suppress or drastically alter their expressive voice.

As researcher Louie Giray puts it:

The stress, anxiety, and fear of being labeled as dishonest can weigh heavily. Imagine having to constantly defend one's work, knowing that a simple glitch in the system could ruin one's career. This creates immense pressure to overcorrect, making scholars overly cautious in their work, which might stifle creativity.

And this applies to all writers, not just academics.

It's not only AI that threatens creativity; AI detectors do too, in a subtler way. Instead of promoting stylistic diversity, these tools risk discouraging it, because when everyone writes trying to avoid triggering AI detectors, they inevitably end up conforming to a single, algorithm-approved style. It seems to me that, while aiming to do the opposite, AI detectors themselves end up driving homogenization.

The Last Resort: Humanizers and the Privilege Loop

Abandoning your own stylistic imprint can be as difficult as it is necessary. You either learn to quickly adopt a detector-friendly style or risk sanctions or dismissal from organizations that rely too heavily on these tools.

Some accusations are undeniably justified, like those involving hallucinated content (this is the case, for example, of several lawyers who used fabricated AI-generated rulings to draft legal filings), but others unfairly penalize people acting in good faith, simply because of the way they naturally write.

It's no surprise that humanizers, AI tools that make content appear more "human," are reaching staggering numbers of users. Spending hours manually editing your writing to avoid being inadvertently flagged by detectors is still an option, of course, but if you can achieve the same result effortlessly in just a few minutes, why bother? If your career were at risk and you really couldn't get rid of your old style, the safest choice would be to rely on a high-quality humanizer.

But it's a dangerous trap. The real risk is creating further social and economic disparities. In fact, humanizers are mostly paid tools. They might sometimes offer a free trial, but it's usually limited to a few hundred words. Then, you need a monthly subscription to access the full service. This basically means that students, people in precarious work, and, more generally, anyone with limited resources could find themselves facing yet another access barrier.

On top of not being able to write as you always have, there's the added impossibility of having to pay for an AI service (created by humans) to convince another AI (also created by humans) that a text is genuinely human, even when it already was from the start.

An absurd situation in itself, further compounded by its contradiction with the fundamental principles of democratic AI. If you can't access these tools in the era of AI detectors, writing can become a privilege for a select few: those unaffected by stylistic bias and those with the means to bypass it.

Nobody wants AI-generated content to be passed off as human, but we also don't want the opposite to happen, and today, for many, it's difficult to write in a simple and clear style without being viewed with suspicion by AI detectors. People often focus on how generative AI can limit creativity, but the blind faith in AI detectors can do even more damage, as these tools tend to carry stylistic bias and inhibit self-expression.

We should also be aware of the equity issue these tools pose, indirectly. As long as AI detectors remain biased and mistakenly regarded as infallible, they also push people toward humanizers as a remedy, which not everyone can afford.

Indeed, the risk is that alongside stylistic inhibition, a social and economic divide will form, so that only some have the privilege of humanizing their writing and, in turn, expressing themselves.

So, a simple proposal: what if we set AI detectors aside, at least until the stylistic bias is resolved?