Workbench / 08 Critique / Refusal: a field guide
How to say no well.
Refusal is a legitimate outcome of AI literacy. This page is about doing it properly: knowing which kind of no you are saying, what it has looked like when other people said it, what it will cost you, and how to put it in writing so that it holds. It is written by people who use these tools every day and think you are allowed not to.
Refuse because you looked
Three kinds of no
Most arguments about refusing AI go wrong because two people are refusing different things. Separate them and the argument usually gets shorter.
1. No to a use
“I will not use it to write my essays,” or to draft the condolence note, or to grade the students. This is the most common no and the easiest to defend, because it is about what a task is for. If the point of the essay is that you learn to think on paper, a tool that thinks on paper for you defeats the assignment whether or not it is allowed. Task-level refusal needs no theory of AI. It needs a theory of the task.
2. No to a tool
“I will not use this company’s product,” because of where its training data came from, what it does with your prompts, or who it is owned by. This is a consumer and a political no, and it is where most of the receipts on the Critique page live: the pirated books, the Kenyan labellers, the chat logs in discovery. It is also the no that institutions make on your behalf when they sign a $17 million contract, which is why some faculty at Cal State petitioned against renewing one.
3. No for now
“Not until I can do this without it,” or “not until the standards exist.” Deferral is a serious position, not a dodge. The trial evidence that unguarded AI help hurts learning is strongest for people still acquiring the skill; a surgeon who says no to the AI overlay until she has her own detection rate is being careful, not backward. The honest version names the condition under which the no would lift.
What is not on the list: no to learning about it. You cannot refuse well what you have not looked at, and a refusal that cannot give reasons is just a mood. This page exists so that yours can give reasons.
A short history of no
What refusal has looked like
Open any entry. Each one is sourced and dated; checked September 2026.
1811–1816
The Luddites break the frames
Organized English textile workers destroyed the machinery that was being used to replace them, and were answered with a mass trial at York in 1813, hangings and transportation. The word now means someone who fears technology; the people it names were skilled machine operators who objected to being erased without a vote. Brian Merchant’s 2023 history argues they “rose up rather than starve at the hands of factory owners who were using automated machines to erase their livelihoods.” Britannica · Merchant, Blood in the Machine
January 2023
Nature refuses to credit a machine
“No LLM tool will be accepted as a credited author on a research paper,” the journal wrote, and any use must be documented in the methods or acknowledgements. The rule holds across Springer Nature. It is a refusal of a particular use, authorship, not of the tool. Nature editorial, 24 Jan 2023
March 2023
Artists get a tool of their own
University of Chicago researchers released Glaze, which makes artwork read as a different style to a model, and later Nightshade, which turns images into “poison” samples that corrupt models trained on them without consent. Refusal, here, is a file format. Glaze · Nightshade
September–December 2023
Writers and actors write it into the contract
The Writers Guild’s 2023 agreement says AI “can’t write or rewrite literary material,” its output is not source material, studios must disclose AI-generated material they hand to a writer, and a company “can’t require the writer to use AI software.” After a 118-day strike, SAG-AFTRA members ratified, 78% to 22%, what the union calls its first protections around AI: informed consent and pay for digital replicas. Neither union banned the tools. They set the terms, which is what a mature refusal looks like. WGA · SAG-AFTRA
May 2025
A company reverses course
Klarna’s chief executive, who had boasted that an AI assistant did the work of 700 customer-service agents, told Bloomberg that the cost-cutting had gone too far and that customers would always be able to reach a real person. The same month, Duolingo’s “AI-first” memo drew enough public anger that the company deleted every post on its 6.7-million-follower TikTok. It did not, however, reverse the policy. Bloomberg · Fast Company
September–October 2025
The public is not sold
Half of U.S. adults are more concerned than excited about AI in daily life; one in ten is the reverse. Across six countries, 61% of people have ever tried a generative AI tool and 34% use one weekly, which leaves roughly four in ten who have never touched one. Refusal is not a fringe position; statistically it is closer to the norm. Pew · Reuters Institute
March–May 2026
Universities push back on their own contracts
Faculty in the California State University system petitioned the chancellor not to renew a $17 million ChatGPT Edu contract covering 460,000 students, arguing that “investing in the CSU’s human workforce is the best way to ensure the quality of research, teaching, and learning.” University of Colorado faculty and students signed a dissent letter with hundreds of names. By April, 0.7% of CSU students had completed the voluntary training; some simply declined to use the tool. Inside Higher Ed · CalMatters
In practice
Three places you will have to say it
In a course
- Read the policy first. Fewer than half of faculty say their campus has clear rules, so the course policy is the one that counts. AAC&U/Elon 2026
- Ask what the assignment measures. If the answer is “your ability to do X,” then not using a tool that does X is the assignment working.
- Put it in writing. Under TRAIL you disclose how AI contributed; “not at all” is a complete answer. Keep drafts and version history: detectors flag innocent work (61% of non-native essays in one study), and a paper trail is your defence. Liang et al. 2023
- Ask for the alternative, not the exemption. “May I do this by hand in class?” is easier to grant than “may I skip it.” Blue-book sales are up 80% at Berkeley; you will not be the first to ask. WSJ via Gizmodo
At work
- Refusal is negotiated, not declared. The screenwriters’ contract is the model: name the uses that are off the table and the terms for the rest. WGA summary
- Know the cost. More than 60% of the Class of 2026 is pessimistic about its prospects and nearly half of them blame AI; hiring managers, meanwhile, are twice as likely as seniors to think AI will create jobs. A blanket no is a bet against that market. Handshake
- Refuse the output, not only the tool. Forty percent of desk workers received AI “workslop” last month and spent about two hours per instance cleaning it up. Declining to send slop is a refusal too. HBR
- Watch who reverses. Klarna hired people back. Companies that went “AI-first” in 2025 are a natural experiment; read their 2026 results before you accept their 2025 memo.
As a maker
- Protect the work: Glaze against style mimicry, Nightshade against uninvited training. Glaze · Nightshade
- Label it: the “Not by AI” badge marks work that is at least 90% human-made. Not by AI
- Know your rights are being litigated in your favour: Anthropic paid $1.5 billion, about $3,000 a book, for training on pirated copies; the New York Times case is at summary judgment. Authors Guild · Axios
- Disclose when you do use it. The credibility of your no depends on the honesty of your yes.
Put it in writing
A refusal that holds
A template for telling an instructor, a supervisor or a collaborator. Adapt it; the structure is the point. Reason, scope, condition, offer.
I’ve decided not to use generative AI for [this assignment / this part of the project], and I want to say why rather than just not do it.
The reason is [what the task is for: I want to be able to do this without the tool / the material I’d have to paste in isn’t mine to share / I’m not confident I can check its output in this area yet].
This is specific to [the use / the tool / for now], not a blanket position. I’d lift it if [the condition: I can already do the task unassisted / the policy is written down / the data stays on campus].
Here is what I’ll do instead: [do it by hand and keep the drafts / use the library’s sources / document each step in my Lab Book]. I’m happy to show the working.
What it costs
We will not pretend it is free. Employment for 22-to-25-year-olds in the most AI-exposed jobs has fallen 19% relative to their peers since late 2022. Employers say they want AI skills and mostly cannot define them. Some professors will read your no as laziness, and some will read it as integrity, and you will not always be able to tell which in advance.
Against that: the people who can explain their no are the people who can also explain their yes, and that is the skill the job market is actually short of. The studies on this page are not arguments against the tools. They are arguments for knowing what you are doing when you pick them up, and for being able to put them down.
Brynjolfsson, Chandar & Chen, Aug 2026 · checked Sep 2026