Who Gets to Sound Human?

There is a long history of people and tools standing between someone's thoughts and their legible appearance on a page. Ghostwriters, speechwriters, developmental editors, dictation software, scribes, and amanuenses have all occupied that space, and in none of those cases did we decide the assistance invalidated the ideas. The thinking still belonged to the thinker, and we have never actually required, as a culture, that thinking and expression happen through a single unaided human performance. We have simply required that the assistance be expensive enough to escape scrutiny.

What's worth asking is why we are only now paying attention.

The disabled writer with access to a paid scribe retained credibility as a writer. The executive whose speechwriter turned scattered talking points into something quotable remained credible. We extended that trust not because we had evidence the help was not shaping the work, because we didn't and we couldn't. How much of a presidential address reflects the president's own thinking, and how much was built by the person whose job it was to make the president sound like the president? There has never been a meaningful mechanism for separating those things cleanly, and we rarely seemed interested in creating one.

What we trusted was not the content but the container: the professional relationship, the institutional role, the assumption that this kind of mediation belonged to legitimate people inside legitimate systems. And underneath all of that was something simpler: the people accessing this kind of help were already the kind of people whose judgment we trusted, whose time we understood as too valuable to spend on translation work, whose ideas we had already decided were worth hearing. The help was allowed to be invisible because the people being helped could already afford to matter.

That arrangement held without much examination for a long time, and then large language models arrived and suddenly we are examining it. It helps to be precise about what we are actually talking about, because "AI" is doing imprecise work in most of these conversations, and the imprecision is doing real damage.

Part of the problem is that "AI" has become a stand-in for a wide range of technologies that do very different things. Automated systems denying insurance claims or screening job applicants produce institutional harms that are direct and well-documented. Large language models, the tools people use to help structure arguments, draft emails, organize thoughts, and translate thinking into prose, occupy a different place in that landscape. The moral weight of the whole category has settled onto this particular use in ways that are not always earned. When someone uses a language model to help organize an argument they have already made, the tool is functioning at the level of expression and structure rather than generating the underlying synthesis itself. That distinction matters enormously to the people currently getting flattened by it.

There is also the question of why we trusted the ghostwriter not to insert themselves into the work, given that we had no way of verifying they didn't. The answer is not that we carefully evaluated the risk and decided it was low. The answer is that we trusted the social arrangement surrounding the help. A professional hired by someone important, operating inside a recognizable institutional relationship, already fit inside our assumptions about legitimacy and authorship. Unlike the ghostwriter, there is no professional role, institutional relationship, or recognizable social arrangement standing between the LLM tool and the person using it. It is simply available to almost anyone.

What is actually new about LLMs is not that they help people translate thought into language, because that has always happened. What is new is that this kind of mediation is no longer restricted primarily to people with institutional backing, financial resources, or proximity to elite networks. A person who cannot afford an editor now has something that can tell them when their argument has lost its thread, when a paragraph is doing too much, when the structure is working against them. A person who needed help with tone, register, and social legibility in every professional interaction, the kind of support institutions rarely know how to offer, now has a way to write an email that lands the way they meant it to. A person whose thinking moves faster than their sentences, or in a different order than academic convention requires, now has a tool that can help carry the ideas across the gap between how they think and what the situation requires them to produce.

It is difficult to ignore that suspicion intensified at precisely the moment this kind of help stopped being exclusive.

When someone's writing gets flagged as AI-generated, whether by a formal detector or by the more ambient social version of suspicion, the thing being assessed is almost never what people think it is. The raised eyebrow, the "something feels off," the confidence that a person can "just tell," none of that is actually measuring humanity, originality, rigor, or thought in any objective way. Two very different things are collapsing into one another in these conversations, and the way people are sorting between them is not measuring authenticity. It is measuring fit.

Formal AI detectors carry a particular irony because they are themselves AI systems trained to identify writing that sounds insufficiently like dominant expectations for "natural" human expression. Research has repeatedly shown bias in who gets flagged: non-native English speakers, autistic writers whose prose tends toward precision and pattern consistency, people whose educational, cultural, or linguistic backgrounds produce writing that diverges from familiar institutional norms. The detector is not locating AI so much as it is locating difference, finding the place where someone's communication stops matching a familiar style and treating that deviation as suspicious.

We mistake familiarity for humanity all the time.

The informal version of this process does the same thing with even less accountability. People describe writing as "too polished," "too structured," or "too articulate" in ways that reveal how specific our assumptions about authentic human communication actually are. What gets recognized as naturally human writing is often simply writing that feels familiar enough, that aligns closely enough with dominant educational, cultural, and neurotypical norms that it does not trip any wires.

I have had people outside my usual social circles ask whether my phone or email had been hacked after receiving messages from me that felt unusual or suspicious to them. The same messages barely registered as noteworthy to people who interact with me regularly. The difference was not authenticity. It was familiarity. People accustomed to my autistic communication patterns recognized the messages as obviously mine. People unfamiliar with those patterns interpreted the same communication as potentially artificial. The communication did not change. The audience did.

Some of the writing currently being dismissed as AI-generated was never touched by an LLM at all. Other writing did involve an LLM, but in the same way people have always used tools to bridge the distance between what they are thinking and what a situation requires them to produce. Both experiences are getting caught in the same net.

The distinction between thinking and surface expression is the part of this conversation that keeps disappearing.

Thinking is the synthesis itself, the argument taking shape, the connection forming, the meaning assembling from everything a person knows, remembers, observes, and lives through. Expression is what happens when that thinking has to become legible to someone else, when thoughts have to take a particular shape, follow particular conventions, perform in a register that reads as competent, coherent, or professional. For many people those processes feel almost seamless. For other people they genuinely do not. People whose thinking is nonlinear, highly associative, fast-moving, multilingual, or shaped by neurodivergence often find themselves navigating what feel like two separate acts: thinking, and then translation. There is a particular kind of exhaustion that comes from spending years manually translating yourself into professionally recognizable language while receiving credit for neither the complexity of the thinking nor the labor the translation required. For those people, an LLM often touches the translation rather than the thinking itself.

When we treat "sounds like AI" as a self-evident disqualification for engagement, we are not protecting intellectual integrity. We are reinstating a standard for whose expression counts as legitimate, and applying it most forcefully to people for whom being heard is already the hardest work. And because the broader ethical concerns surrounding AI are real, because exploitation, environmental harm, and institutional misuse are documented and ongoing, suspicion has started to feel principled, and the accusation stops being an observation about process and becomes a moral verdict.

Moral verdicts make dismissal feel righteous. That righteous dismissal lands hardest in progressive spaces, because the communities most suspicious of AI are often the same communities that explicitly value accessibility, neurodiversity, inclusion, and the amplification of marginalized voices. These are often spaces that have done real work to understand structural barriers, communication differences, and the ways institutions mistake conformity for competence.

The problem is not hypocrisy exactly. Most people participating in these conversations genuinely care about accessibility and harm reduction. The problem is that the broader ethical concerns surrounding AI are real enough that suspicion itself has started to feel morally responsible. Once AI becomes associated with exploitation, theft, laziness, and institutional harm, people begin treating perceived AI involvement as meaningful evidence about the integrity and value of the work itself.

Discernment gets replaced by first-glance recognition and stylistic suspicion as shortcuts for whether thinking occurred at all. The assessment starts happening at the level of surface texture, before the argument is considered, before the ideas are weighed, before the work is given the chance to demonstrate its own value. Because that suspicion feels ethically grounded, dismissal starts to feel like the only principled response.

The result is that people may reject work before substantively engaging with it, including work written by people they would otherwise say they want to support.

The executive who dictated to a human assistant keeps legitimacy. The autistic writer whose prose pattern trips a detector loses theirs. The person who could afford a ghostwriter was always credible. The person using a free tool to do similar communicative labor becomes suspect. Support translating thought into legible communication is not new. What changed is who can access it.

The ethical concerns are real, but they are also incomplete. The larger environmental, economic, and political questions surrounding AI are genuinely unresolved, and the work of addressing them matters. In the meantime, people will continue using the tools available to them in their efforts to be seen and heard. What's worth examining is whether the instincts we trust to recognize good thinking have ever been as neutral as they felt, and whether that changes anything about the verdicts we've been confidently making.

Some questions to ponder: Do you think this was written by a person independently, with the assistance of an LLM, or entirely by an LLM? Why? If you discovered the authorship was different than you assumed, what would that change about how you think about the ideas?