A human behind the prompt just isn't the same as an author

Published on November 19, 2026

Summary

I said a post from the IAAP sounded like AI had written it. Someone asked why that mattered if I understood it. My answer applies to any organization that uses AI to write for it. I support disabled people using AI to help them say what they mean. What bothers me is when an organization publishes words nobody checked. Readers end up doing the work the writer skipped. I also worry about how much electricity AI uses. Burning coal and gas to make that electricity adds to global warming, and I expect disabled people to feel the effects sooner.

I commented on LinkedIn that a post from the International Association of Accessibility Professionals (IAAP) sounded robotic, as though AI had written it. Someone pointed out that a disabled person might have written it and might communicate “in a fashion that has been replicated by AI.” I don’t know how the post was written. I do know what it was like to read it.

A red pen lies across a printed draft with words crossed out and corrections written in the margins.

The rest of the reply bothered me more. “Why would it matter if it was written by AI if you understood what you read? There is a human behind the AI prompts.” An author decides on words to use to express what they want to say. They take responsibility for the words that reach the reader. An organization would be setting a remarkably low bar for its writing if all it asks is whether I understood it.

This post isn’t about the IAAP. Their content was just the trigger. I’ve seen similar things from plenty of organizations. This post was what finally made me write about it.

Whose words are they?

I’ll make one thing very clear: I support disabled people using assistive technology to go about our lives, including using AI to communicate. That’s non-negotiable.

The hard part for some disabled folks is finding the right words or putting them in order. It’s good if AI gets them from the idea to a draft. They know what they intended, so they can see where the draft went wrong. They can then fix it. And send it when the words say what they mean.

Disabled people have been writing in our own voices for considerably longer than anyone’s been asking ChatGPT to draft a LinkedIn post. A writer who’s used a turn of phrase for twenty years can hardly help it if ChatGPT uses it now.

If an editor has to read one of my paragraphs three times, I want to know which sentence tripped them up. It would piss me off to find out they let it go because they knew I was disabled.

As for the human behind the prompt, I’ve asked a lot of questions over the years on my podcasts. A good question can shape the whole conversation, but my guest still gets the credit for the answer. With AI, a prompt can contain somebody’s entire argument, down to the example they want to use. It can also say “write 500 words on why accessibility matters” and leave the machine to come up with a reason. Knowing somebody typed a prompt tells me very little about whose judgment went into the piece. AI could give you a draft. But it takes somoene knowledgeable about the subject to read it, to spot what’s wrong, and to revise it.

Who checked the advice?

An AI draft might tell a developer: “To check that a form is keyboard accessible, press Tab until you reach the Submit button.” The developer tabs to Submit, but the button only responds to a mouse click. The advice never said anyone has to be able to submit the form. An editor who knows accessibility catches that before it goes out.

If nobody checked what AI wrote, then no one on staff made that judgment. Readers deserve to know when AI drafted the copy. Who wrote something tells me how much to trust it and how closely to check it. And it costs an organization one line. If AI drafted it and a named person reviewed it, saying so shows the work. An organization that did the checking has nothing to hide.

“Did you understand what you read?”

Yes. I’ve also understood most of the badly written documents I’ve had to read in thirty years of accessibility work. I wouldn’t offer that as much of a recommendation for any of them.

I’ve cared about plain language for a long time. An old site of mine, The Wheeling Gourmet, won Best Plain English Website, Private Sector, at the 2010 WriteMark New Zealand Plain English Awards. Cooking is a good test of whether your writing serves the reader. “Cook until done” uses three short, familiar words. It’s a lousy instruction for someone trying to work out when to take the onions off the heat. If I mean “cook until the onions are soft enough to cut with the edge of your spoon,” I should say so.

AI can produce whole paragraphs of “cook until done.” I can understand every word and still have to go back through the paragraph to work out what the author meant. I read this stuff for a living. A reader who has trouble holding several sentences in mind may already be on their third pass by the time I’ve finished my second.

Plain language is no excuse for flat writing. Short sentences and common words can still sound like a person with something to say. When I call writing robotic, I mean it could have come from any organization on any subject. Nobody in particular wrote it, and nobody in particular is answering for it.

Cory Doctorow coined “enshittification” to describe what happens to a platform as its owners take more of its value for themselves. I’m borrowing his term for writing because the publisher gets the benefit of producing a page in seconds and every reader inherits the work the publisher skipped.

A hypothetical conference website that devotes three paragraphs to the organizer’s commitment to inclusion. As a wheelchair using speaker, I want to know if there’s going to be a ramp to the stage. But the information isn’t present on the website. So I have to email of call the conference organizers. AI made it cheap and fast to fill a page about inclusion. And it probably missed all the critical information that disabled people actually want to know.

What’s worth the cost?

I’m not an AI doomsayer. Helping a disabled person take part in a conversation is maybe worth the electricity it takes. I have a harder time justifying a few hundred words generated because it’s Tuesday and somebody needs to put something in the newsletter. One generated paragraph wouldn’t make a real dent. Now multiply that by every organization doing the same thing everyday. It adds up: Data center consumption is projected to double by 2030, from 485 TWh to 950 TWh. Power use from AI-focused data centers is projected to triple.

Yes, the combined demand for electricity concerns me. The International Energy Agency expects coal and natural gas to supply over 40% of the extra electricity data centres worldwide will need through 2030. Burning those fuels adds carbon dioxide to the atmosphere. As the planet warms, heat waves become more frequent and more intense.

I expect some of the costs of that warming to reach disabled people sooner and hit us harder.I recently wrote about the environmental impact of AI on disabled people. The UN Secretariat identifies heat waves as a particular risk for disabled people, who are more likely to get sick or die in extreme heat (PDF). It also describes how inaccessible transport makes it harder to get health care, especially for people forced to leave their homes.

During British Columbia’s 2021 heat wave, Gabrielle Peters, a wheelchair user in Vancouver, called cooling centres “an absolutely ridiculous solution” to suggest to people with complex medical needs. She told Human Rights Watch that the heat left her breathless going to the bathroom. How was she supposed to get to a cooling centre, she asked, when the bus stop alone was a couple of blocks away?

Who should be doing the work?

I see a huge difference between disabled individuals leveraging AI as assistive tech and organizations pushing AI-generated content with apparently little human involvement or review.

An organization can count the staff time it saves by using AI to generate a page. And then another page. And another one again. Its budget has no entry for the risk of serious illness or death a disabled person faces in an apartment they can’t cool. On a much smaller scale, the conference organizers don’t have to count the time speakers or attendees spend chasing information that should have been on the website. Someone can understand every word of those three paragraphs about inclusion and still have to email somebody to ask whether there’s a ramp.

Part of an author’s job is to do the legwork so the reader doesn’t have to.