
Honor Bird's work is built on a simple premise: policies that look neutral on paper rarely are. Human-Centered Design starts by asking who a system was actually built for, and who it quietly leaves out, before asking whether it works. Most of the time nobody intended the exclusion. That is what makes it worth finding.
The piece below is a worked example. It is not really about AI writing policy specifically. It is about what happens when a rule gets applied to every field at once without anyone asking what the rule was protecting in the first place, and who ends up paying for that oversight. AI policy just happens to be where I caught it happening in real time.
Publications around the country are rightfully in the middle of an extensive conversation about whether AI-assisted writing belongs in their pages. The instinct behind it is a good one. Editors already have a lot on their plate, and policies should be in place that protect both their time and the advancement of ideas that contribute to the greater conversations happening in these publications. But the conversation has mostly settled on one answer for every field at once: a blanket policy, applied uniformly, with no attention to what that policy is actually protecting or who it ends up locking out.
What these policies are missing is nuance. Without it, we risk doing two things simultaneously. We stand in the way of technological advancement that is not going anywhere and is only going to become more advanced and harder to detect. And we quietly close a door that this technology was supposed to open, giving access to people who have real ideas but who, due to factors relating to education, class, and neurotype, have predominantly been excluded from contributing them.
Meet Sarah
Take any writer. For this piece, we will call her Sarah.
Sarah has published four books. She has spent two decades learning how a sentence should breathe, where a paragraph should turn, when silence on the page says more than another clause would. Lately, Sarah is finding it harder to get published. She is also noticing something else: the quality of what does get published is getting worse. Slush piles are filling up with competent, forgettable prose that reads like nothing and no one. Sarah did not choose this. She is watching a craft she gave her life to get treated as a commodity, and she wants some assurance that the work she is competing against was actually made by a human being who put the work in the way she did.
Sarah deserves to have her industry protected. Nobody serious is arguing against that. I certainly am not.
Poetry, fine art, fiction, music. These are culture. They have soul. In these fields, removing the human from the process is not a shortcut, it is an amputation. The sentence itself is the art. The struggle to find the right word is not overhead to be optimized away, it is the work. A policy that says "we want prose made entirely by a human being" is not gatekeeping in these spaces. It is a coherent, defensible statement about what the field actually values.
The problem I am discussing is not these fields. The problem is when other fields adopt the same standard without ever asking why the standard existed in the first place.
The Tests
Here is the question that I believe often gets skipped when designing policy related to AI use: in this field, is the prose the product, or is the prose the vehicle?
This is where Human-Centered Design earns its keep. It does not start with a verdict on AI. It starts with two diagnostic questions that any policy, in any field, should have to survive.
Poetry, literary fiction, memoir-as-craft: the prose is the product. Nobody is reading Sarah's novel to extract a policy recommendation. They are reading it to feel something happen to them at the level of language. In those fields, how the sentence was made is not a side issue. It is the entire issue.
Policy writing, grant writing, journalism (mostly), philosophy (mostly): the prose is the vehicle. A federal grant reviewer is not scoring your metaphors. A policymaker reading an analysis of a broken system is not there for cadence. Someone reading objective reporting about an issue isn't there for the author's "hot take." The reader's entire relationship to the text is: did this move me from not understanding something to understanding it, as clearly and quickly as possible. If it succeeds with less friction, that is not a loss. That is the job being done well. This holds even where persuasion is explicitly the point. Yes, good prose, grammar, and syntax strengthen an argument. But the persuasion itself lives in the follow-through of the thinking, not in the sentences carrying it. It is still vehicle-prose. The bar is a good argument, not a beautiful one.
I call this the category test. What is the field in question to which the policy applies?
But the category test alone is not enough, because passing it does not mean any use of AI is automatically legitimate. It just means the standard built for Sarah's field does not automatically transfer. I believe you still have to ask a second question: did the idea exist, in substance, before the drafting began? And if it did not exist yet, was the formation of that idea itself the thing this piece of writing was supposed to prove happened?
I call this the provenance test.
Two tests. Category, then provenance. Neither one alone tells you anything. Together, they tell you almost everything.
Meet Marcus
Marcus is a beat reporter covering city hall for a regional paper. He spends the day the way reporters actually spend a day: sitting through a budget hearing, pulling records, making calls, taking notes nobody else has. By six o'clock he has a pile of real, hard-won facts. He opens a chat window, drops in his notes, and types: give me a story with a good hook.
Run Marcus through the category test first. Reporting is not poetry. Nobody is reading a city hall story for the pleasure of the sentences. The field is idea-as-product, not prose-as-art, so by that measure alone, he passes.
But run him through the provenance test and he fails, and he fails in the specific spot that matters most. The facts existed before the drafting. The story did not. Deciding what a pile of facts actually means, which detail is the real story and which is noise, why this matters today and not last week, that decision is the job. It is not the seasoning on top of the reporting, it is the reporting's whole point. The entire reason a newsroom sends Marcus instead of just publishing the hearing transcript is that someone with judgment is supposed to sit with the raw material and produce the argument about what it means. That formation of judgment is the thing the assignment exists to certify.
Someone will object here that writing and thinking are not actually separable, that the sense-making often happens in the sentence itself, not before it. That objection deserves to be taken seriously, and in Sarah's field it is often true. But it is not what is at stake in Marcus's case. The judgment call he skipped, which fact is the story and why it matters today, is a decision that can be reached in a conversation, in a notes app, in the shower, with no sentence drafted yet. If that decision has not been made, no amount of prose fixes it. If it has, the prose is just where it becomes visible. This distinction matters for what comes later, too. Bouncing an opinion you already hold off a language model, pressure-testing it, having it hand you back the counterargument you now have to answer, is not the model doing your thinking for you. It is closer to arguing with an editor before the editor exists. The provenance test does not ask whether a keyboard or a chat window was involved anywhere in the process. It asks whether the judgment came from you.
If Marcus hands the facts to AI and asks for the hook, he has not skipped the typing. He has skipped the only part of the job an editor was actually counting on him for. This is not a case for keeping AI out of newsrooms, plenty of reporters use it well for transcription, first-pass organization, or checking their own blind spots. It is a case about what a byline is supposed to certify, and Marcus's byline was certifying that the judgment was his. It wasn't.
Meet Jasmine
Jasmine is forty-three. She founded a small nonprofit six years ago that runs job-readiness training for formerly incarcerated women. Her curriculum was not built in a semester. It came out of hundreds of hours in the room with her participants, refined against what actually worked and what didn't, informed in part by her own path through reentry.
When Jasmine sits down to write a grant narrative, or an op-ed for a workforce development trade publication, or a training manual for a new case manager, run her through both tests. Jasmine passes both tests for the same reason Marcus failed one: the idea existed before she opened her laptop, built over six years of direct contact with the problem. What happens at the keyboard is retrieval and organization, not invention.
This is the same way a policy analyst passes them, the same way an experienced clinician publishing a case framework passes them, the same way most people doing serious work in fields where ideas are the currency pass them. Her use of AI in drafting does not manufacture the insight. It reduces the tax on getting an insight that already existed out of her head and into a form someone else can use.
What This Actually Costs
There is an analogy that gets used a lot in these conversations, and it is usually deployed poorly. People say: using AI to write is like using a calculator to do math. Critics push back correctly: a calculator produces a determinate answer to a fixed problem, and a language model generates content, phrasing, framing, sometimes connections you did not explicitly hand it. That is a real difference and it should not be waved away.
But the sharper version of the analogy holds anyway, once you ask the right question. Nobody asks whether a mathematician is capable of doing long division by hand. Of course they are. The question is whether they should spend their limited hours of the day doing arithmetic instead of the proof. Time is not free. Attention is not free. For people whose brains run six tracks at once, or who process language differently, or who are simply carrying more than one job's worth of thinking, the tax on turning a fully formed idea into linear prose is real, and it is not distributed evenly across people.
Paying that tax anyway, when a tool exists that can lower it, is not virtue. It is just cost, absorbed for no reason other than nostalgia for how things used to be done. If a synthesis that would have taken three years to translate onto the page can now take three months, and the idea underneath it has not changed, something was gained and, aside from possibly more authentic language, nothing was lost. For situations passing the category test, this, in my opinion, is a net win.
Fine, I'm Not a Writer
Somewhere in this argument, someone will say it plainly: if you cannot produce serviceable prose without a machine, you are not a writer. Fine. I concede that point. I'm not contesting it. In fact, I spent the better part of the opening backing up that exact claim.
But notice what that concession actually settles, and what it doesn't. It settles a definitional question about craft. It does nothing to settle the question this piece, and what I believe the larger issue, is actually about. The bigger question at play here is whether not being a writer should disqualify you from putting an idea in front of anyone beyond your own kitchen table.
Right now, writing is the toll booth. If you want an idea to travel further than the people you talk to face to face, a policy proposal, a grant narrative, an argument that might change someone's mind, or a national discussion, you have to render it in prose good enough for an editor, a funder, or a reader to accept it. That has nothing to do with whether the idea is good. It has to do with whether you also happen to have a second, separate skill: turning thought into publishable sentences at a professional standard.
I am not and will not ever claim that using AI makes you a 'writer.' That was never my claim. My claim is that not being a writer shouldn't be the thing standing between a person and the ability to put a real idea into a larger conversation. Those are two entirely different propositions, and the concession only costs you the first one.
For most of recent history, having a genuinely good idea was not sufficient. And somewhere along the line "good idea" and "ability to write" got conflated. But let's ask honestly: does fluency always actually track with quality of thought? I think we can all remember at least one piece of writing where we are left wondering how that person found anyone else willing to publish them. We just did not have to reckon with this conflation, because nobody could act on a good idea they could not also express well until very recently.
A tool that separates those two skills again is not lowering the bar for participation. It is revealing that the bar was never really measuring what people thought it was measuring.
The Disclaimer
None of this resolves into a clean rule to be applied everywhere, and I want to be honest about that rather than pretend the two tests settle every case. I have only just started thinking this through in a structured way, and I expect to be wrong about some part of it. Category and provenance are questions worth asking, not a verdict that gets stamped once and reused. This is not an argument for abolishing disclosure policies, and it is not an argument for ignoring them either. It's a vote for nuance. Sarah's field is allowed to want what it wants. Institutions are allowed to decide what a given piece of writing is supposed to certify, the way Marcus's editor is right to care whether the judgment behind the story was his.
What I am arguing against is the reflex: taking a policy built for Sarah's field and applying it to Jasmine's without ever asking what it was protecting in the first place. Do that, and a blanket ban stops being a quality filter and starts being an access filter. It doesn't screen for good ideas. It screens for who already had the formal training, the time and money to acquire it, and a brain that produces linear prose without friction, and hands the advantage to people who already had it.
None of this requires believing anything about who is smarter or more capable. It requires only noticing what building fluency in publishable prose actually costs: years of schooling that modeled the sentence for you, unpaid hours to practice a craft that doesn't pay the rent while you're learning it, a house where language was already the family's currency growing up. Some people are handed most of that before they turn ten. Others are working two jobs, or thinking in a first language the publishing world doesn't run on, or wired in a way that makes turning a fully formed idea into linear prose on demand genuinely harder, not because the idea is worse but because that specific pipeline was built around one kind of brain. None of that has anything to do with whether the idea in their head is any good. It has everything to do with whether they had the runway to make it look like one.
Those traits were never randomly distributed. They track education, class (and by extension, race), and neurotype with almost laser-like precision. A policy that reads as neutral on paper can end up doing exactly the sorting everyone claims to be worried about elsewhere, quietly, and usually without anyone deciding to do it on purpose.
This is the question HCD asks by default, not because AI policy is special, but because every policy deserves it. The two tests in this piece, category and provenance, are not really about writing. They are a template: does this rule protect something real, and does it fall evenly on the people it touches. Most exclusionary systems were never built by anyone trying to exclude. They were built by people who never asked the second question. That gap is where Honor Bird tries to work.
