An AI Ethics Framework So Boring It Might Actually Work

SFWA announced and revised its Nebula Awards AI policy within hours of each other last December—two emergency board votes producing terms that couldn’t be defined and rules that can’t be enforced against a threat that doesn’t actually exist. The policy wasn’t the result of careful thinking. It was the result of no framework at all—just panic shaped by the loudest voices in the room.

That’s what incoherence produces. So let’s build something coherent instead.

But first, three objections that will arrive before anything else does.

“Using these tools makes you complicit in the theft of every artist whose work trained them.”

On the surface this one appears to have merit. AI companies scraped copyrighted work without consent and without compensation, and that practice deserves scrutiny, legal challenge, and eventually regulation. Courts are working through it. The argument is real.

But it’s directed at the wrong target.

Holding individual users morally culpable for corporate training practices is the same category error as holding smartphone buyers responsible for cobalt mining in the DRC. The grievance belongs to Anthropic, OpenAI, Stability AI, and the legislators who should be regulating them. It has nothing coherent to say to an indie author choosing a cover tool. If your concern is training data practices, advocate for consent frameworks, support litigation, push for legislation. That’s where the lever is.

“Generative AIs are assemblers of stolen works.”

This one’s downstream of the sourcing objection, and technically illiterate. LLMs and GenAI don’t retrieve and recombine existing text or images—they’re statistical pattern completers generating output that never existed in the training data.

Diffusion models, for example, synthesize new pixel arrangements from learned statistical relationships between visual features—they don’t retrieve and reassemble training images. The Getty lawsuit is the telling exhibit: if the technology actually stored and recombined source images, you’d get clean reproduction. What you get instead is watermark-shaped artifacts in the wrong position on images that don’t exist—the model learned a pattern without retaining a single original. You can’t reproduce a specific copyrighted work because the work was never stored. It was learned from.

An LLM’s poem isn’t a Frankenstein’s monster of six famous poets any more than a generative AI’s image is a collage of existing artwork.

“Every time you use AI instead of hiring an illustrator, that’s a commission that artist will never see.”

The premise here is that a transaction that never happened represents a harm. The indie author who uses AI for a cover wasn’t going to commission a thousand dollar-plus illustration—the counterfactual is Canva, a premade cover, a Fiverr artist paid poverty wages. Those choices generated no boycotts, no ethics manifestos, no campaigns. What changed is which cheap alternative is available, not whether the commission existed. The illustrator was never getting that job in the first place. GenAI just gives the cash-strapped indie author a higher-quality product than the alternatives.

The Framework

A framework for ethical use of AI in publishing doesn’t require novel ethics. It simply requires applying the ethics we already use for everything else, consistently.

Function, not tool. The ethical question is always what’s being represented, not what produced it. A book written by AI and sold under an author’s name as their own work is fraud—not because AI touched it, but because fraud is fraud regardless of tool. A book genuinely written by the author credited on the byline with an AI-generated cover is a business decision about a marketing asset. The cover, the trailer, the ad copy are packaging, not the product. Nobody confuses a movie poster with the film, and no reader thinks the dust jacket defines the novel’s literary merit. Traditionally published authors typically have zero input on their covers anyway; publishers commission whatever makes business sense. Indie authors making the same calculation aren’t doing something categorically different.

Disclosure follows deception risk. Authors don’t disclose which grammar checker they used, which research databases they consulted, whether a developmental editor restructured their second act—because those things don’t affect what the reader is purchasing, which is the author’s creative vision executed in the author’s voice. The disclosure question for AI use is identical: does the reader have a reasonable expectation this creative work is yours, and does this particular use violate that expectation? For prose published as your own writing, yes. For cover art, no. For ad copy, no. For outlining and brainstorming, probably not. For chapters generated verbatim and published without substantive transformation, yes. The principle is consistent. The application varies by context, as it should.

Harm must be specific and identifiable. “This harms artists generally” is a market structure argument. It belongs in policy discussions about regulation and compensation frameworks, not in ethical claims about individual behavior. Can you name the specific person harmed by this specific decision? For AI manuscript screening—yes. The author whose book was wrongly rejected by an algorithm that can’t distinguish literary critique from the thing being critiqued faces a concrete, career-affecting harm from a consequential decision made without the reasoning capacity to make it. For AI cover art—no. Nothing was “stolen” and there’s no identifiable victim of a transaction that was never going to happen.

Consequential decisions scale with stakes. The same technology used to generate a book cover and to screen manuscripts for acquisition deserves dramatically different scrutiny—not because one use is AI and one isn’t, but because one involves a consequential gatekeeping decision affecting a writer’s career and livelihood, and the other involves commercial packaging for a product that already exists. Proportionality isn’t a loophole. It’s what makes an ethical framework functional rather than performative.

---

Underneath all three objections—the sourcing concern, the plagiarism claim, the entitlement argument—is a single legitimate grievance dancing around in three different costumes: market disruption and loss of control over aesthetic standards.

Except that was already happening long before AI arrived.

The content mill problem is real, hardly new, and completely ignored. The market flood is real. The displacement of authentic voices by adequate-enough volume production is real. But the threat isn’t AI—the tools just happen to make the process cheaper.

Ethics-washing market disruption into theft claims that don’t survive technical scrutiny can’t produce coherent policy. Instead, it produces the SFWA situation: incoherent rules written in response to the shrillest voices in the room, protecting against an imaginary threat while the actual threats—content mill economics and publishers deploying broken tools as gatekeepers—go unaddressed.

My framework is boring because it’s just applied consistency. That’s the point. When you apply it, you get answers that vary by context rather than a single verdict applied regardless of stakes. AI cover art? No identifiable harm, no deception, no consequential decision, consistent with how we treat every other tool in the pipeline. AI manuscript screening? Consequential decision-making, identifiable harm, and documented reasoning failures demanding scrutiny. Same framework. Different conclusions.

That’s what coherence looks like.

The indie author using AI to generate cover art or an LLM to write ad copy isn’t the problem. They’re a writer trying to get their book in front of readers without a publisher’s marketing budget, a design department, professional copywriters, or a $3,000 illustration commission that was never on the table.

Meanwhile the writing and publishing community has built an elaborate ethics apparatus aimed squarely at us—struggling independent authors making rational business decisions about packaging—while content mills built on literary sweatshop economics buy the shelf space out from under us and publishers deploy broken gatekeeping tools that can’t distinguish literary critique from genuinely harmful content. The apparatus exists. The threats it’s ignoring exist. What doesn’t exist is any coherent account of why the indie author’s Facebook ad deserves more scrutiny than an algorithm rejecting manuscripts it’s fundamentally incapable of understanding.

But communities under economic pressure need enemies they can see and name, and “AI” is visible in a way that content mill economics and publisher negligence aren’t. Manufactured outrage is easier than confronting systematic failure, and the systematic failure is harder to fix. So the apparatus keeps pointing at the wrong targets, the actual threats go unaddressed, and the authors the community claims to be protecting get caught in the crossfire.

We deserve better than this. Our fiction does too.