Manifesto · Essay
The Gate Was the Point
AI, ableism, and why the rules keep changing whenever ordinary people find another way in
Manifesto · Essay
AI, ableism, and why the rules keep changing whenever ordinary people find another way in
Contents
There is a particular kind of betrayal people describe as “Mom snuck broccoli into the mac and cheese.”
You were enjoying something. Then you learned what was in it, and suddenly the experience became disgusting. Not because the flavor changed. Not because the texture changed. Not because anything happened to the thing in front of you. The label changed, and the label gave you instructions about how you were supposed to feel.
We are watching that happen to art in real time.
People praise an image, a song, or a paragraph. Then someone says AI was involved, and the same work becomes “soulless,” “lazy,” “slop,” or “obviously fake.” Now the hands look wrong. Now the rhythm is empty. Now the prose has too many em dashes. Nothing inside the work changed. The viewer received new social instructions.
This is not merely my impression. In experiments, people have struggled to distinguish AI-generated work from human work while still rating work more negatively when they believed AI made it. One study found that viewers actually preferred AI-generated images when authorship was hidden, while other research found a persistent bias against AI-labeled art. A 2026 experiment found that attaching an AI label significantly lowered the effort viewers thought had gone into a piece, while a human-made label did not produce an equivalent boost over no label at all. The label itself acted as a discount.
Then somebody performed the argument as a prank. An actual Monet was posted to X and falsely presented as AI-generated. Responders confidently diagnosed its supposedly synthetic defects: incoherent plant reflections, token confusion, oversaturated verticals, vague depth, artificial turmoil, and the invisible “delta” between this image and human art. A discussion preserving the replies also supplies an important caveat: this was a stunt, not a controlled experiment; the post invited comparison with another Monet, online reproductions varied in color, and some screenshots omitted complimentary context. Fine. Do not use it as population science. Use it for the narrower thing it directly demonstrates: once certain viewers accepted the AI label, they could manufacture detailed visual proof of AI-ness from a painting Monet had already made.
That does not prove AI makes better art. It proves something more relevant to this argument: a substantial portion of what is being presented as aesthetic judgment is not a response to the artifact at all. It is a response to the category.
And I want to know why.
AI companies are not harmless public libraries run by woodland fairies.
Data centers consume serious amounts of electricity and water. The International Energy Agency projects that global data-center electricity consumption could roughly double between 2024 and 2030, while the U.S. Department of Energy estimates that data centers could consume between 6.7 and 12 percent of U.S. electricity by 2028.
Workers have legitimate reasons to fear how employers will deploy generative systems. The International Labour Organization estimates that roughly one in four jobs worldwide has some exposure to generative AI, and it emphasizes that transformation is more likely than complete automation...but “more likely” is not a force field around anybody’s paycheck.
Artists have legitimate interests in preventing exact reproduction, fraud, impersonation, and unauthorized replicas of their faces and voices. A singer should not wake up to discover that a company has manufactured a synthetic version of their voice, made it perform material they would never consent to, and kept the money.
There are legitimate questions about training data, acquisition, market substitution, and whether an output reproduces protected expression. The U.S. Copyright Office itself says the fair-use analysis for AI training is contextual: some uses are likely fair, some likely are not, and many fall between. That is not a magic declaration that every training practice is automatically legal...or automatically theft.
There are legitimate concerns about spam. There are legitimate concerns about misinformation. There are legitimate concerns about people using a language model as an oracle, accepting whatever falls out of it, and slowly forgetting that verification and judgment are verbs.
Good. Let us have every one of those conversations.
But then I would like everyone to explain why the gun is pointed at Sarah.
Sarah did not build a data center.
Sarah did not scrape a catalog.
Sarah did not fire an illustrator, cancel a voice actor’s contract, replace a newsroom, suppress wages, write a platform’s disclosure rules, or decide that a town’s water supply should cool a server farm.
Sarah paid twenty dollars because she wanted a picture of her cat commanding a purple spaceship.
Yet somehow Sarah is the person being interrogated, humiliated, banned, accused of theft, told she has no soul, and...in examples I have personally watched circulate...told that AI artists should be killed.
This is not a harmless rhetorical malfunction. Once a group is described as thieves, parasites, frauds, bots, and people without human worth, violence against them becomes easier to perform as comedy. “It was only a joke” does not change what the joke requires: an audience willing to treat the target’s death as funny.
Meanwhile, the people making the consequential decisions remain safely abstract. The chief executive replacing workers is “the economy.” The corporation acquiring creative infrastructure is “innovation.” The data-center operator draining a local resource is “the cloud.” But Sarah is available. Sarah has a username. Sarah can be made to hurt before lunch.
That is punching down. It is also politically useless. No quantity of death wishes sent to disabled hobbyists will renegotiate a union contract, decentralize a platform, clean an electrical grid, regulate a monopoly, or compensate a single artist.
If your anger consistently lands on the person with the least power in the chain, I am going to question whether stopping harm is actually the organizing principle.
The people demanding a purity test keep pretending only AI users will ever have to take it. That fantasy has already failed.
Human artists are being accused because a hand looks odd, a texture resembles diffusion noise, a brushstroke seems too clean, or a self-appointed detective has developed divine access to everybody else’s process. Katerina Ladon had to publish process evidence after her hand-painted Dungeons & Dragons artwork was accused of being AI. The reporting documents both the accusation and her explicit denial. In another documented incident, an artist supplied proof of a non-AI process and moderators removed the work anyway. Other artists now upload timelapses specifically to “combat the AI accusations.” That phrase appears in the title of an actual Procreate artist’s post, which is a fairly bleak description of what making art online has become.
The burden is backward. The accuser produces a vibe. The artist must produce sketches, layer files, raw takes, session footage, metadata, screen recordings, and sometimes a live performance. Then the same detective who demanded proof may announce that the proof can also be faked. Of course it can. Once suspicion becomes unfalsifiable, no evidence can acquit you because the trial was never designed to discover the truth.
I have recordings of myself performing this music live. I have coughs in takes, raw vocals, keyboard parts, Reaper sessions, collaborators, stems, and the people who have watched me do the fucking work. My public Nox Noir home identifies me, my music, and my performances. None of that should be an entrance fee for human status. A musician should not have to stage a live evidentiary hearing because an internet stranger disliked a snare drum.
This is not collateral damage after the “real” battle. It is the predictable result of teaching a mob that accusation is expertise, resemblance is proof, and deletion is a harmless precaution. Documented human-made posts have already disappeared under suspicion. Artists are already producing defensive timelapses merely to remain visible. Apply the same presumption to music catalogs and the stakes become years of recordings, credits, playlists, listeners, comments, and income. A platform or community cannot be allowed to sacrifice the accused merely because admitting its detector cannot see into the past would embarrass the detectives.
And then the same people say artists need customers. Which customer is going to commission someone whose surrounding community has told them they are a thief, a bot, a parasite, or a person who deserves to be murdered? “Pay artists” is not persuasive when it is delivered as a protection racket. Every threat teaches a potential customer that approaching this community carries social and personal risk. They are not merely shooting Sarah. They are shooting themselves in the foot and demanding Sarah pay the invoice.
You all have far more invested in taking away the tool, continually punishing Sarah, and actively disabling her than you have invested in stopping the people the actual argument should be aimed at. That is why the mob can spend twelve hours dissecting one stranger’s pixels and no time learning who signed the licensing deal, approved the data center, eliminated the job, or owns the distribution system.
When generative images were malformed curiosities, people laughed. When AI music sounded like a fever dream trapped in an elevator, people laughed. The moral emergency became much louder when the tools began producing things ordinary viewers might genuinely enjoy.
That timing matters.
So does the mob’s selectivity.
A simple post can be one hundred percent AI and receive fucking crickets. No tribunal. No forensic examination of the pixels. No sudden concern for the Colorado River. The hunt arrives when somebody makes a sharp point, attracts attention, demonstrates taste, or produces something genuinely effective...especially when the crowd has already classified that person as incapable, unserious, disabled, untrained, or otherwise beneath the result.
That does not mean every successful AI work is attacked or every ignored one is bad. It means the enforcement pattern is revealing. The tool may be present in both cases; the social alarm sounds when the tool helps somebody breach an assigned status boundary.
If the offense were simply “AI exists here,” the reaction would be consistent. Instead, the outrage often scales with how much legitimacy the user appears to have gained without first receiving permission.
Before AI, creative gatekeeping could wear the pleasant face of encouragement:
You have improved so much.
>
Have you thought about the line placement?
>
Keep practicing and maybe someday.
Some of that feedback was sincere and useful. Some of it was hierarchy spoken in a customer-service voice. The person above you could control your enthusiasm, define your potential, and remain generous while reminding both of you that you were not yet permitted to stand beside them.
Then a machine allowed people who had already been classified as mediocre to produce something compelling. Sometimes it produced something more technically impressive than the gatekeeper could make. Suddenly technique was no longer sufficient proof of worth, because technique was no longer scarce enough to protect the hierarchy.
So the test moved somewhere invisible.
Soul.
Intent.
Authenticity.
Humanity.
Conveniently, these are properties the gatekeeper can declare absent without demonstrating anything. If viewers love the work, they were fooled. If the user spent days directing, selecting, editing, performing, arranging, or compositing it, that labor does not count. If the result communicates exactly what its creator intended, intention has been redefined to mean manual execution. If a disabled person says the tool gave them access to expression they did not have before, accessibility becomes “an excuse.”
The attribution trick is wonderfully rigged: when the work is good, the machine did it; when it is bad, the human is talentless. Beauty produced by a high-status or conventionally attractive creator becomes evidence of their essence. Beauty produced by somebody the audience has already ranked low becomes evidence that software performed a fraud on their behalf. Under those rules, the user can never supply evidence of ability because every success is confiscated at the border.
This is not criticism of an artifact. It is border enforcement around the word artist.
Yes, artists without formal training have broken through. Basquiat existed. Outsider art exists. Punk exists. People have always built astonishing work from whatever was available.
An extraordinary person getting through a gate does not prove there was no gate.
Music produces the same comforting exception story. Susan Boyle walked onto Britain’s Got Talent at forty-seven while the audience visibly prepared to laugh, then sang well enough to become instantly marketable. Her 2009 audition became a global breakthrough, and she was signed afterward. That moment did not prove the music industry had suddenly stopped penalizing older people, unfashionable bodies, disabled people, or anyone who could not be packaged through conventional beauty. It proved that one extraordinary performance, inside a television format built to monetize the audience’s surprise, became commercially irresistible.
Exceptions can expose a barrier while being used to deny it. “Susan Boyle made it” and “Basquiat made it” are stories about particular people becoming valuable in particular moments. They are not proof that everyone else received access, and they do not turn marketability into justice.
“Basquiat did it” is not a universal accessibility plan any more than pointing to one wheelchair user at the summit proves the staircase was accessible.
The approved route still demands some combination of time, money, health, fine-motor control, executive function, equipment, education, social access, and the ability to survive years of being bad publicly. Free tutorials do not make those resources universal. “Free” instruction does not manufacture pain-free hands, spare hours, stable housing, tuition, a studio, or a nervous system that cooperates on command.
Research with disabled artists has found exactly this mixture: generative tools can reduce barriers and expand creative expression, while also introducing bias, inaccessibility, and new forms of dependence. The answer is to make the tools safer and more accessible...not to announce that the access they provide is morally counterfeit.
An accommodation is often treated as cheating when it allows a disabled person to achieve the outcome that everyone claimed they wanted for them. That reveals a vicious assumption: the struggle was not merely an unfortunate obstacle on the way to the result. The struggle was considered the admission price.
We have seen this logic in other clothes for years. A stranger sees somebody leave a car in an accessible parking space and decides the person does not look disabled enough. A service dog does not match the viewer’s private mental picture, so suddenly the handler is on trial. Somebody walks today and uses a wheelchair tomorrow, and an amateur detective announces fraud because fluctuating disability has failed to submit a schedule.
The viewer assigns a permitted level of impairment and then patrols any evidence that the disabled person has exceeded it. AI gatekeeping often repeats the structure: strangers assign a person a permitted level of skill, fluency, polish, or productivity, then treat assistance that lets them exceed it as moral deceit.
No, an art tool is not a parking permit, and an accusation about a picture is not identical to blocking physical access. The analogy is the jurisdiction people grant themselves over another person’s legitimacy. You do not look impaired enough. You do not look skilled enough. You did not suffer visibly enough. Therefore the access you used is stolen.
That is not protection. That is somebody else’s life being measured against the worth a spectator assigned it.
Whenever unauthorized image use was discussed before generative AI, the instruction was simple:
Do not screenshot my art. Do not crop out my signature. Make your own or commission someone.
Now Sarah can create a new image specifically for herself instead of taking an existing artist’s finished image, and “make your own” has acquired several hundred pages of previously undisclosed conditions.
She must use the correct tools. She must perform the correct labor. She must possess the correct bodily abilities. She must spend the correct number of years developing them. She must emerge with an output no better than the skill level strangers have assigned her. If she exceeds that level, she has cheated.
And no, “she could just commission someone” does not close the argument. It skips over a fucking canyon of reality.
A twenty-dollar subscription might let Sarah explore fifty versions until the picture resembles the one in her head. A custom commission at the level she imagines might cost hundreds or thousands of dollars. It requires finding the right artist, communicating through another person’s interpretation, waiting for availability, and paying for revisions. Those are not equivalent forms of access.
It also assumes Sarah does not, has not, and will never commission an artist. I use AI, and I currently pay an artist whose work I value. These things are not mutually exclusive. One tool can serve small experiments, jokes, references, and intensely specific personal images while a human collaboration serves work where another artist’s interpretation is precisely the thing I want.
Not every passing idea for a phone wallpaper justifies a professional commission. Before generative tools, Sarah might have grabbed the closest existing image she could find and used it without permission. Now she can generate something new instead...and the people who explicitly told her to “make her own” are still furious.
That rather gives the game away.
The rule was not merely “do not take my work.” For some people, the rule was “pay someone from the approved class or accept that personalized art is not for you.” Sarah found a third option, and the existence of that option is the offense.
You cannot manufacture goodwill by removing somebody’s alternative and then presenting yourself as the person they should be required to pay instead.
There is another piece of history being laundered so aggressively that I can hear the washing machine from here.
Before generative AI, creative culture was not a pristine village where everyone purchased every influence, cleared every sample, licensed every image, and saved every receipt in a fireproof cabinet.
We had Napster. Kazaa. LimeWire. eDonkey. Soulseek. BitTorrent. The Pirate Bay. Entire leaked albums passed through dorm rooms and message boards before release day. People swapped discographies on hard drives. Producers traded sample packs whose provenance was approximately “some guy named beat_wizard_420 uploaded it.” DJs made unauthorized edits. Musicians practiced by copying riffs note for note. Websites grabbed photographs, fonts, film dialogue, fan art, and whatever else was close enough to the thing they needed. This is not misty-eyed folklore; the Government Accountability Office documented widespread peer-to-peer exchange of music, images, video, and pirated software while it was happening.
And piracy was not limited to finished entertainment. People learned from songs taped off the radio, borrowed records, ripped CDs, YouTube demonstrations, photocopied method books, scanned textbooks, shared PDFs, downloaded scores, copied tablature, pirated fake books, and lesson material passed from one broke musician to another. Nobody paid every artist whose phrasing entered their ear. Nobody mailed the Louvre a royalty because a color relationship lodged in their memory. An unlawful copy can create liability for making or distributing that copy; it does not contaminate the knowledge acquired from it and make every later original sentence, painting, or song stolen forever.
This was open enough to become part of the culture’s own language. System of a Down released an album literally called Steal This Album!, after unfinished tracks had circulated online. Mitch Benn released “Steal This Song”. The joke worked because home taping, file sharing, bootlegs, leaks, and the record industry’s theatrical panic about all of them were common knowledge.
Sometimes the person normalizing the workaround was the teacher. College students have entered classes and been told, directly or with a wink, where to find a PDF because the assigned textbook was financially absurd. This is documented even by professors arguing that their colleagues should stop doing it. The point is not that every professor ran a shadow library. The point is that society already understood a conflict between formal ownership and practical access, and it did not ordinarily declare that knowledge learned from an illicit PDF could never become part of a legitimate mind.
Entire present-day communities still exist around jailbreaking and rooting phones, replacing firmware, sideloading software, bypassing manufacturer restrictions, applying unauthorized patches, and, yes, circulating cracked applications. XDA still maintains current root coverage and modification communities. Some of those practices are lawful customization, some violate terms, some create security risks, and some are straightforward piracy. Again, distinguish them. But spare me the historical hallucination that ordinary technology users discovered unauthorized copying the morning Stable Diffusion launched.
And yes, creative software was pirated too: Ableton, FL Studio, Cubase, Pro Tools, Reason, Kontakt, Waves, and enough cracked plugins to make a laptop wheeze like it had Victorian tuberculosis.
That history does not prove that infringement is good. It does not mean artists should surrender their rights because somebody once downloaded a Metallica song over dial-up. Two wrongs do not become a public transit system merely because enough people board them.
It proves something narrower and devastating to the present moral performance: we did not ordinarily declare every work made with a cracked DAW spiritually void. We did not strip every producer of the word artist until they produced a notarized receipt for every plugin, sample, photograph, record, tutorial, and reference they had ever encountered. We judged the song, and we handled provable infringement as a specific claim against a specific act.
Now an AI company’s alleged or actual misconduct is treated as hereditary sin transferred automatically to every downstream user. The company assembled the corpus, therefore Sarah personally burgled ten million studios while generating her cat. This is not legal reasoning. It is contamination mythology.
If a company acquired material unlawfully, sue or regulate the company. If a model memorizes and reproduces a protected passage, document the reproduction. If an output is substantially similar to a specific work, identify the work and the protected expression. If somebody clones a living artist’s voice to deceive an audience, go after the impersonation. Those are claims with nouns, verbs, evidence, and defendants.
“You touched a machine whose maker may have done something unlawful, therefore nothing you make can contain human worth” is a purity ritual.
And it is a remarkably selective purity ritual. People are posting their denunciations from phones assembled through supply chains they have not audited, on platforms powered by data centers they have not inspected, while sharing screenshots they do not own to accuse Sarah of contamination by association. Somehow the moral microscope only turns on when the person under it has made something they were not expected to be capable of making.
Here is my actual position, because I am not going to hide it behind twelve committees and a bowl of beige policy oatmeal: learning should not be gatekept...not from humans, and not from computers that help humans.
Suppose you write that a historic building in Old Sacramento has a doorway too narrow for wheelchair access. What exactly is the principled argument that prevents Ted from learning that fact because you wrote the sentence? What changes if Ted uses a computer to help locate patterns across ten thousand such reports? The answer cannot simply be “scale,” spoken in the tone normally reserved for summoning a demon.
Scale can change power. It can change market effects. It can increase the chance of memorization. It can enable industrial substitution and give a few companies terrifying leverage. Those are serious consequences. But quantity does not cast a spell that transforms learning itself into the removal of a rivalrous object from its owner.
This is the Ship of Theseus problem wearing a GPU. A computer can learn from a million works. So can I, given enough years, radio, books, museums, records, teachers, films, conversations, and internet access. At what number does influence become theft? Ten works? Ten thousand? One million? If no principled number exists, then quantity alone is not the moral switch. The relevant changes are what was copied, how it was acquired, what the learner retained, what the output reproduces, how the system affects a market, and who controls the resulting power.
The same question cuts the other way. If cultural material is valuable enough to learn from, why are companies destroying the physical sources after scanning them? A 2026 investigation tracked shipments of books, including rare and out-of-print material, to an Amazon facility that destructively scanned them for AI training. Cutting off spines may be fast. It is not preservation. Libraries already know how to digitize without destroying the object; the Internet Archive describes non-destructive scanning and long-term preservation as part of the same mission.
Buy the books. License what requires licensing. Scan them carefully. Preserve the physical copies, or resell and donate them so collectors, libraries, researchers, and future readers can still reach them. Preserve the high-quality digital record with provenance. Rare books, recordings, scores, photographs, and art should not become disposable feedstock whose only surviving form is a private model owned by a corporation. A society serious about cultural value would demand both access and preservation, not celebrate a warehouse eating the archive because shredding was cheaper.
This is where internet arguments keep shoving fifteen different questions into a trench coat and naming the resulting lump theft:
Those questions overlap. They are not interchangeable.
The U.S. Copyright Office says as much: some training uses are likely fair, some are likely not, and many live in the fact-dependent middle. That is the current legal landscape. My policy argument goes further. Do not build a licensing tollbooth around learning itself. Put enforceable boundaries around concrete conduct: reproduced protected expression, memorized passages, substantial copying, deceptive impersonation, cloned identity, fraud, and demonstrable harm.
Style is not a human organ somebody removed with forceps. Genre is not private property. Influence is not a chain of custody. If you allege copying, point to the melody, lyric, recording, passage, composition, character, or image that was copied. “It reminds me of” can begin an investigation; it cannot finish one.
Humans have always learned through imitation, study copies, transcription, reference, recombination, genre convention, collaboration, and delegated execution. Cover bands can be paid to perform songs under specific licensing systems. Producers hand musicians a brief. Directors do not personally sew every costume. Composers write scores other people perform. Photographers press a button after making choices about subject, timing, lens, position, light, and selection.
Nobody owns the note C, a diatonic scale, a standard chord progression, a genre convention, a brush technique, or a color palette. The Copyright Office expressly lists common musical scales and standard chord progressions among material it will not register. Protected recordings, compositions, lyrics, images, and characters are different things from the vocabulary used to make new ones. Input is not output. Influence is not duplication. Learning a work note for note is not the same act as selling that recording as your own.
We already know how to build mechanisms between “anything goes” and “nobody may learn.” American copyright law separates a song from a particular recording and provides a compulsory licensing route for some cover recordings; digital services use blanket licensing systems under the Music Modernization Act. Those systems are imperfect and do not map mechanically onto model training, but they demolish the claim that the only imaginable protection is a ban. We can build collective licenses, platform-level payment, auditable uses, identity controls, and remedies for reproduced expression without selling private ownership of scales, styles, techniques, and facts.
The fact that rules and licenses apply to some of those acts does not prove that all learning requires permission. It proves that we have always distinguished the idea, the process, the copy, the performance, the recording, and the market use when we are being serious.
AI should not receive magical immunity from those distinctions.
It should not receive magical damnation either.
The model is not a tiny synthetic citizen demanding human rights inside my laptop. It is a non-sentient tool, and there is still a human being behind the use of it. That human can bring lived experience, intention, judgment, humor, grief, embodiment, and taste to a mediated process. We can argue honestly about how much authorship a particular contribution earns. We cannot make the person disappear merely because some execution was delegated.
The same trap has appeared around language.
For years, people were told that their ideas did not count unless they packaged them perfectly. Bad spelling meant stupid. Inconsistent punctuation meant careless. An unpolished explanation meant the underlying thought was not worth understanding. This burden fell especially hard on dyslexic, neurodivergent, cognitively disabled, exhausted, medicated, poor, and non-native writers.
Then people gained tools that could help with the packaging.
Suddenly the spelling is too consistent. The sentences are too organized. You used an ellipsis, an em dash, or the word “delve.” Nobody writes like that. Your thoughts are fraudulent.
Even people who have always written that way are treated as crime scenes. AI detectors have made this worse. In one widely cited test, 89 of 91 TOEFL essays were flagged by at least one of seven detectors. Newer research complicates claims of one universal demographic bias, but continues to find that editing and style can materially confound detector outputs. These systems are not mind readers, and punctuation is not a fingerprint.
Have these people opened a Terry Pratchett novel? A single page can contain enough em dashes, ellipses, interruptions, nested asides, capitalized concepts, and footnote-shaped detours to make an AI detector catch fire. Those marks existed before ChatGPT. So did words with more than two syllables. Nobody should have to vandalize their natural voice, insert fake typos, or surrender a useful punctuation mark to perform humanity for somebody whose literary forensic method is “I saw an em dash.”
Write imperfectly and you are dismissed as stupid. Use assistance and you are dismissed as fake. Write well naturally and you are accused anyway.
This is a Kafka trap wearing Grammarly’s skin.
Overreliance on generative systems is a fair concern. A Microsoft Research study of 319 knowledge workers found that greater confidence in AI was associated with less reported critical-thinking effort. A small MIT preprint on essay writing also found lower engagement and poorer recall among participants relying on a language model, although the study was narrow and should not be inflated into proof that every assisted use makes every user stupid.
But “cognitive offloading” is not automatically cognitive surrender.
The insult also smuggles in a status judgment about what the speaker believes you could have thought unaided. If a weak paragraph appears, nobody cares whether software helped. If somebody uses assistance to present a sharp argument the reader cannot easily answer, suddenly the provenance of every clause becomes the emergency. The complaint is often not “you stopped thinking.” It is “you have produced thinking I did not expect from someone I had ranked beneath me, and I need a reason not to engage with it.”
That is why output policing concentrates around successful communication. A person struggles to explain something and gets punished for being unclear. They use a translator, editor, spellchecker, AAC device, search engine, or language model and get punished for sounding too clear. They naturally write with unusual vocabulary and get punished for sounding implausibly clear. The acceptable output is whatever keeps them inside the observer’s estimate of their intelligence.
We offload memory into calendars, arithmetic into calculators, navigation into maps, spelling into dictionaries, movement into mobility devices, and communication into editors, translators, speech-generating devices, and other people. The relevant question is not whether a tool carried part of the load. It is which part...and what the person did with the capacity that became available.
If I form the argument, reject bad drafts, notice missing premises, correct the model’s assumptions, restore my own voice, and decide what survives, I have not outsourced my judgment. I have outsourced some retrieval, organization, spelling, and sentence assembly. On days when medication or disability turns those tasks into a wall, that offloading allows more of my thinking to reach you, not less.
The proof is in the disagreement. I tell the model when it has made me sound corporate. I catch missing evidence. I reject false binaries. I remember the argument it omitted. I demand the sentence back in my own rhythm. I just made it go back and recover piracy, eminent domain, false accusations, preservation, identity safety, and this very section because the first draft was not thorough enough. A brain that has been replaced does not keep arguing with its replacement about what the replacement failed to understand.
If someone pastes the first answer without reading it, that is passive dependence. If someone fights with the machine until it accurately carries a thought the machine did not originate, that is tool use.
Pretending those behaviors are identical is not intellectual rigor. It is another excuse to judge the user instead of examining the work.
There are serious risks in treating a chatbot as a therapist, oracle, lover, or sole relationship. It can affirm a delusion, miss a crisis, mishandle private information, encourage dependence, or provide the frictionless agreement that an unhealthy user prefers over another person’s right to say no. A man who wants an endlessly compliant synthetic woman because real women possess boundaries has not solved loneliness. He has automated his resentment of women.
The American Psychological Association’s health advisory describes both sides: some wellness tools have been associated with reduced self-reported stress, loneliness, and depressive symptoms, while generative chatbots can also provide misleading or dangerous responses and should not be confused with licensed care. Stanford researchers have likewise found dangerous failures in therapy chatbots, while other research found short-term reductions in loneliness from AI companions. The evidence is not “perfect substitute” or “worthless poison.” It is benefit, risk, design, context, and degree of dependence.
Now look at the landscape into which people are releasing these tools. The United States Surgeon General called loneliness and isolation a public-health crisis. The official advisory did not describe a nation drowning in reliable companionship. People say “we should hang out” and never make the plan. Invitations vanish. Disabled people cannot always get to the room. Queer and trans people may have to calculate whether the room is safe. Everyone performs being busy while someone else goes another week without being touched, heard, or asked a second question.
Then we tell that lonely person to go to therapy as though therapy were a public drinking fountain. They must find the correct specialty, personality, cultural competence, location, schedule, treatment style, and pay scale. The provider must accept their insurance, be accepting new patients, and not require a level of executive function the patient is seeking treatment because they do not currently possess. In 2024, more than 122 million Americans lived in a mental-health workforce shortage area, and six in ten psychologists reported not accepting new patients. Insurance reimbursement for behavioral health also remains lower than comparable medical care, which discourages providers from joining networks and transfers costs to patients.
So yes...build safeguards. Do not market a language model as a licensed clinician. Require crisis boundaries, privacy protections, age-appropriate limits, honest disclosure, independent testing, and routes to human care. Teach users that agreement is not truth and availability is not love. Treat replacement of all human contact as a warning sign.
But do not shame people for using the only listener they can reach at two in the morning while human society demonstrates, in the same breath, why they reached for it. If your alternative is “talk to friends” when you do not answer, “find a therapist” when none are affordable or accepting patients, and “go outside” when disability or danger makes that glib, you have not offered care. You have offered a slogan and another locked door.
Training-data disputes cannot be resolved by screaming either “everything is theft” or “information wants to be free.” Different models use different datasets. Different developers acquire material differently. Different outputs present different risks of memorization, substantial similarity, and market substitution. Courts are still sorting through those facts.
The U.S. Copyright Office’s position is more careful than most internet arguments. It says purely machine-generated material is not protected by copyright, while human selection, arrangement, modification, and other expressive contributions can be. It also treats fair use in training as fact-specific rather than categorical.
That means a hybrid song is not dropped into one giant uncopyrightable bucket. My lyrics, recorded vocals, human-written melodies and parts, performances, edits, creative arrangement, and sufficiently original modifications can be protected even when the finished recording also contains generated material. The Office explicitly says AI used as an assistive tool does not destroy copyright in the human-authored work, and that human-authored expression, creative selection, coordination, arrangement, and modification remain protectable on a case-by-case basis.
There is an important limit here because I am not going to demand precision from everyone else and then award myself magic ownership dust. Merely transcribing a purely generated passage into MIDI note for note does not automatically make the underlying musical expression human-authored. Fixing my own composition in MIDI does. Reperforming, rewriting, restructuring, reharmonizing, editing, and adding original expression can create protectable human authorship in those contributions. “It is in a MIDI file now” proves fixation; it does not, by itself, answer who authored every note. The stronger truth is that substantial parts of my process were already mine before the file extension changed.
That distinction matters. Exact replication is not the same question as learning statistical relationships. A cloned voice or copied character is not the same problem as a new work sharing a broad aesthetic. A company’s decision to assemble a training corpus is not the same act as an end user typing a prompt. Collapsing all of these into the word “theft” may feel clarifying, but it erases the very distinctions needed to assign responsibility.
It also creates a danger Cory Doctorow described years ago: new transferable rights intended to protect individual creators are often accumulated by the corporations that control contracts and distribution. Giving a bullied schoolchild more lunch money does not feed the child if the bully still owns the hallway.
Creators deserve agency and compensation. But a system in which giant rightsholders own the catalogs, own the licenses, invest in the generators, and decide which independent tools are “authorized” may protect corporate portfolios far more efficiently than it protects working artists.
This argument deserves particular honesty because the environmental cost is real.
Data centers consume electricity. Cooling systems can consume water. New transmission can scar land and raise ugly questions about who pays. A badly located facility can damage a community even if its share of national consumption is numerically small. Efficiency per task does not erase aggregate growth when the number and size of tasks explode.
The U.S. Department of Energy reports that data centers used about 4.4 percent of American electricity in 2023 and could reach 6.7 to 12 percent by 2028. The underlying Berkeley Lab report says U.S. data-center electricity use rose from 58 terawatt-hours in 2014 to 176 in 2023, and that growth after 2017 was driven largely...but not exclusively...by AI servers.
That last clause matters. Largely is not entirely. Data center is not a synonym for AI.
The physical internet did not spring from the earth in 2022. Cloud storage, video streaming, banking, search, advertising, government systems, enterprise software, online games, cryptocurrency, social media, recommendation engines, backup services, and the endless duplication of outrage screenshots all live in material infrastructure too. Their footprints differ; I am not claiming every click has the same cost. I am asking why some people discover that the cloud is a building only when a disabled stranger makes a song.
“AI data center” has become a magical phrase that transfers every fan, server, transformer, gallon, and transmission line in sight onto the AI ledger. Sometimes that label is accurate: the operator specifies AI training or inference, GPU clusters, or a dedicated AI customer. Sometimes a headline adds AI to a generic hyperscale cloud project because the word attracts clicks and moral certainty.
Take Gilroy, California. The city’s own project page calls the applicant Amazon Web Services and describes two data-center buildings on a 56-acre parcel. The California environmental record likewise identifies a “Gilroy Data Center.” Those documents establish a huge AWS facility with real local impacts. They do not, by that description alone, establish that every workload or every environmental cost belongs to generative AI.
That does not prove the facility will never run an AI workload. It proves that the person assigning the entire building to AI carries the burden of showing it. “It is a data center and Amazon sells AI” is not workload accounting.
This is the first rule of environmental honesty: name the facility, operator, actual or planned workload, unit of measurement, location, and affected resource. AIDC is not holy water you sprinkle on incomplete evidence.
Water numbers are unusually easy to weaponize because several different measurements can all be described as “water use.”
Withdrawal means water taken from a source; some may be returned. Consumption means water not returned to that source, often because it evaporated. Direct water is used at the facility, principally for cooling and ordinary building needs. Indirect water can include water consumed by the electricity system that powers the facility.
Mix those categories and you can create a statistic that is numerically correct and conceptually fraudulent.
The Berkeley Lab report estimates that American data centers directly consumed about 66 billion liters of water in 2023...roughly 17.4 billion gallons...and projects 2028 consumption between 60 and 124 billion liters for hyperscale facilities alone. That is a serious amount of water.
It is also important scale. The United States Golf Association says American golf courses collectively use about 1.5 billion gallons per day...over half a trillion gallons per year. The Environmental Protection Agency puts residential outdoor water use, mainly landscape irrigation, near 8 billion gallons per day...almost three trillion gallons per year. I am not presenting golf or lawns as enemies that must be hurled into the sea. I am showing how bizarre it is to speak as though data centers invented large discretionary water demand while landscaping, agriculture, energy generation, and recreation become invisible.
National comparison is not absolution. A golf course in a wet basin and a server farm drawing potable water from a stressed aquifer are not morally interchangeable because one bar on a national chart is taller. A facility can be a small fraction of American water consumption and still be an unforgivable neighbor.
The honest conclusion has two parts at once:
1. Data centers are not remotely the largest national water user, and “AI is drinking the country dry” is often supported with categories that have been mashed into pudding.
2. A specific facility can consume an outrageous share of a specific community’s available water, particularly when companies choose evaporative cooling in a drought-prone place because land, power, taxes, or political resistance are cheaper there.
If somebody says a facility created a “toxic zone,” that is a separate pollution claim. Show the contaminant, pathway, testing, and health evidence. Water consumption alone does not chemically transform into toxicity because the headline needs more bass.
A data center does not choose cheap land. It does not negotiate a tax incentive. It does not decide that a thirsty cooling system belongs in a desert. It does not approve its own permit, route its own power lines, sign a confidential development agreement, or tell residents that permanent jobs will arrive by the wagonload.
People do that.
Developers choose scale, hardware, site, cooling design, and disclosure. Utilities choose service plans and propose generation and transmission. Regulators decide who bears grid costs. Local officials approve zoning and incentives. State and federal officials write permitting and environmental-review rules. Investors reward speed and cost reduction. When damage follows, “AI did it” can become a fantastically convenient witness-protection program for every human being who signed something.
The local stakes are not hypothetical. In The Dalles, Oregon, Google’s facilities consumed about a third of the city’s water by 2024. That number says something important about The Dalles; it does not tell us that every data center everywhere uses the same share, or which fraction of that facility’s work was AI.
In drought-stricken Santiago, Chile, an environmental court required renewed scrutiny of a proposed Google facility’s effect on the local aquifer. Google then scrapped the original cooling plan and said it would return with an air-cooled design. That story contains both halves people keep trying to tear apart: the original siting and design deserved resistance, and resistance plus engineering could materially change the water demand.
Tucson’s Project Blue shows another part of the causal chain. Residents and officials did not merely post “AI bad” beneath a stranger’s picture; opposition to secrecy and water demand helped stop the original annexation proposal, and the city adopted rules requiring large water users to disclose conservation plans and obtain public approval. Pima County’s project file identifies the proposed data-center campus, while reporting on the enacted ordinance explains the disclosure and approval requirements. That is not proof every local fight ends well. It is proof that organizing around the permit, water contract, and decision-maker can change material conditions.
Nebraska shows why reporting must be facility-specific and standardized. Meta’s enormous Sarpy County campus and Google’s Nebraska facilities report water using different categories, and the available numbers vary sharply by company and year. Flatwater Free Press traced those inconsistencies and the state’s effort to require disclosure. Newton County, Georgia, meanwhile, imposed a temporary moratorium so it could study data-center impacts. These are different failures...measurement, siting, permitting, enforcement...and “AI used water” is too mushy to diagnose any of them.
That is what useful environmental politics looks like. Not pretending harm is imaginary. Not pretending mitigation is imaginary. Making the people with authority change the fucking project.
The Trump administration is not an innocent observer of this buildout. Its July 2025 executive order directed agencies to accelerate permits for qualifying data-center projects and related infrastructure, identify federal land, and use or create categorical exclusions and other streamlined environmental reviews. The accompanying White House fact sheet explicitly celebrated removing climate and other requirements, while the Department of Energy selected federal sites for private AI data centers and associated power generation.
You can support those policies. You can oppose them. What you cannot honestly do is describe the resulting infrastructure as an autonomous act by a chatbot. Political leaders fueled and celebrated these choices. Companies lobbied for and profited from them. Governments approved them. Utilities built for them.
Sarah still does not possess a zoning vote.
“Infrastructure” sounds clean until the line on the planning map crosses a family home.
Georgia Power’s Ashley Park–Wansley project is a new 35-mile, 500-kilovolt transmission line crossing four counties. The utility’s own materials establish the line, route, and schedule. Reporting on the acquisition process has documented homeowners facing negotiated purchases or eminent-domain proceedings so the corridor can be built.
In Pennsylvania, a proposed four-state transmission project intended to serve large demand in northern Virginia has likewise sought utility status carrying eminent-domain power. The project’s own economic analysis reportedly found construction activity but no permanent full-time jobs in Pennsylvania after completion...a rather efficient illustration of how costs can stay local while benefits travel.
This does not prove every transmission expansion exists solely for AI. Utilities also cite reliability and general growth, and a strengthened line can serve more than one customer. It proves the opposite of the cartoon: the real questions are project-specific. How much capacity serves which load? Who benefits? Who loses land? Who gets compensated? Who can invoke eminent domain? Who pays when forecasts are wrong?
My answer on eminent domain is not neutral. Taking somebody’s home or land for a private speculative customer is wrong. Calling the corridor “infrastructure” does not make the family living beneath the line less dispossessed. I hope this expansion finally forces a serious challenge to how casually private development can borrow the state’s power to take. Unfortunately, systems often reconsider an abusive power only after people who assumed the leopard belonged in somebody else’s yard discover teeth marks on their own faces.
Those are questions for developers, utilities, public-service commissions, and elected officials. Screaming “murderer” at somebody generating a phone wallpaper has yet to move a single transmission tower six inches to the left.
When utilities build generation, substations, and transmission for enormous new customers, somebody carries the risk. If the customer pays the full cost through a protective tariff, good. If ordinary households subsidize construction, absorb reliability problems, or inherit stranded infrastructure when a speculative campus shrinks or disappears, that is cost externalization wearing a hard hat.
This is not an argument against electricity. It is an argument for contracts and regulation that make large-load customers pay for the infrastructure and risk they create. Public disclosure, minimum-demand commitments, exit fees, collateral, dedicated tariffs, and independent review are far more useful than assigning a gallon of cooling water to Sarah every time she clicks regenerate.
Pennsylvania’s 2026 response shows that these protections are policy choices, not laws of nature. The state’s new GRID standards address energy affordability, transparency, community engagement, workforce development, and environmental protection. These are announced executive standards, not proof that every promise has been enforced or every loophole closed. They still demonstrate the location of the lever. People can demand rules. Officials can write them. Companies can be made responsible for their costs.
Congress has been offered the same choice. Senator Chris Van Hollen’s Power for the People Act would make large data centers bear transmission and reliability costs that their demand causes instead of automatically feeding them into household bills. A proposal is not an accomplishment, and a press release is not independent proof that every projected saving will materialize. It is evidence that “who pays?” has concrete regulatory answers beyond yelling at end users.
Amazing what becomes possible when the gun is pointed in the correct direction.
Environmental criticism becomes propaganda when it reports every gallon consumed but treats every gallon avoided as corporate fan fiction.
Cooling systems are not all the same. Evaporative systems trade water for energy efficiency. Air cooling and mechanical chillers may reduce water demand while increasing electricity use. Reclaimed wastewater can spare potable supplies without eliminating consumption. Workloads can be scheduled in places and times with lower water or carbon intensity. Chips, models, and facilities can become more efficient...while demand growth can still swallow the savings.
Microsoft says its new closed-loop, direct-to-chip design consumes no water through cooling evaporation after its initial fill and could avoid more than 125 million liters per data center annually. That is a company claim about a design, not a papal declaration that Microsoft has solved water. The design applies to new facilities, electricity still has a footprint, offices still use water, and explosive demand may outrun per-facility savings.
It is nevertheless real engineering. So are reclaimed-water systems, water-aware siting, heat reuse, better chips, drought-response triggers, and public reporting by watershed rather than a global “water positive” shell game in which a project replenishes water somewhere photogenic while a different community loses its well.
The demand should be more specific than “technology bad”:
Kyle Hill’s “Data Center Water Is a Distraction” is useful precisely because it challenges the national-scale panic without denying local failures. That is the balance we need. Compare consistent units. Regulate actual sites. Protect actual watersheds. Follow actual power.
Fight the harmful facility. Regulate the operator. Restrain the utility. Vote out the official. Protect the watershed and the homeowner.
But stop pretending that bullying the least powerful customer is climate policy.
If a publisher fires illustrators, blame the publisher.
If a studio coerces performers into surrendering voice rights, blame the studio and strengthen the performers’ bargaining power.
If a label replaces musicians while licensing their catalogs into its own model, confront the label.
Do not transfer the executive’s decision onto a disabled person making a private image or an independent musician building a track they could never otherwise afford to produce.
Technology does not autonomously decide who receives its productivity gains. Owners, executives, legislators, courts, unions, and markets decide that. The historical question is not simply whether a machine can reduce labor. It is whether workers gain time, safety, access, and income...or whether ownership captures everything and leaves workers to fight one another for scraps.
Even “replacement” conceals several different choices. An employer can use a tool to remove drudgery while keeping the worker, demand twice the output for the same wage, deskill a job so labor becomes easier to replace, misclassify workers, force creators to sign away future voice and likeness rights, or eliminate positions and send the savings upward. Those are not properties emitted by a graphics card. They are labor policy carried out by people with names, titles, contracts, and lawyers.
The International Labour Organization’s global analysis finds that transformation is more likely than complete automation for many exposed occupations, while emphasizing unequal exposure and uncertainty. “More likely to be transformed” is not a force field around anybody’s paycheck. Transformation can improve a job or turn one worker into the exhausted supervisor of five machines doing what six people once did.
So bargain over deployment before it happens. Give workers notice, consultation, severance, retraining they control, portable benefits, audit rights, enforceable consent for voice and likeness, and a share of productivity gains. Protect freelancers from coercive all-rights contracts. Use antitrust law when the same company controls creation, distribution, discovery, and payment. If public money helped build the technology, the public should receive more than a coupon code and a pink slip.
That is also the actual Luddite lesson. The Luddites were not cartoon cavemen who hated machinery because it beeped. They resisted owners using machinery to destroy their livelihoods and bargaining power. A politics worthy of that history would organize against concentrated ownership. It would not send death threats to weekend creators.
Here is another way the category swallows the person.
I compose melodies and parts on a keyboard. I work with loops and stems in Reaper. I write lyrics. I record vocals. I use Suno to layer and shape raw vocal material while preserving the words and style I chose. I remove instruments I do not want. I arrange, edit, revise, mix, and make decisions until the track does what I meant it to do.
The work does not stop at generation. Parts can be rebuilt by ear, performed again, separated into stems, rearranged, remixed, remastered, or carried onto a stage. A later human performance or rebuilt master does not magically create the intention retroactively; it exposes how absurd it was to pretend no human intention existed earlier.
Sometimes AI is a breath of air in that process. It is not the whole atmosphere, the lungs, the singer, the song, and God.
Calling all of that “pressing a button” is not analysis. It is a way to make every human contribution disappear before the argument begins. The same trick would turn photography into finger movement, filmmaking into asking actors to do things, production into sitting behind glass, and conducting into threatening an orchestra with a tiny stick.
Of course somebody can press one button, accept the first result, and call it finished. Somebody can also point a phone at lunch and take a photograph without becoming Annie Leibovitz. The existence of low-effort use does not define the ceiling of a medium or settle the authorship of every work made with it.
Nor should the goal be to make machines imitate the safest corporate average forever. Every new medium spends an awkward adolescence wearing the previous medium’s clothes. The interesting work begins when people push it toward forms that could not have existed before. Be fucking weird. Weird is not the same as random; it means developing an intentional language native to the tool instead of treating technical imitation as the only possible achievement.
Judge what actually happened. Who chose the concept? Who wrote the words? Who composed the melody? Who performed? Who directed generations, selected passages, rejected failures, edited structure, rebuilt parts, mixed the recording, and decided it was done? A useful credit system can answer those questions. The phrase “AI song” mostly throws them into a furnace.
That is why I support granular credits for commercial work. Tell people which lyrics, composition, vocals, instruments, generated elements, edits, arrangements, mix, master, and artwork came from whom or what where practical. Do not slap AI-GENERATED across a hybrid work as though the person evaporated on contact with software.
Disclosure should inform. It should not function as a scarlet letter that deletes a catalog, severs followers, erases plays and comments, demotes discovery, or makes an appeal impossible. If a platform reclassifies a work, the creator’s identity, links, analytics, and audience should survive. Target deception, impersonation, stream fraud, and spam. Do not quietly redefine assistance itself as misconduct.
An artist is not merely a display name attached to an upload. A useful music system needs an artist-owned identity layer: one authoritative home that says who the person is, what they made, how they wish to be represented, which releases and collaborations are authorized, and which voice or likeness models they control. That record should remain portable when a distributor changes policy or a platform changes a label.
This matters because an AI Persona and an AI-assisted human artist are not the same thing. Spotify’s announced AI Persona badge is aimed at profiles whose identity may be generated and may not represent a real person. That is a sensible disclosure target. It is not a warrant to dump every human who used generation, restoration, separation, editing, or a licensed vocal model into the same synthetic-person bin.
It also quietly assumes that concealing the person is inherently deceptive. Sometimes a persona exists because the human does not want strangers finding their legal name, employer, family, address, body, sex, disability, or location. PEN America explains how personal information is weaponized to intimidate and silence people through doxxing, and swatting can turn that exposure into armed danger. After creators have watched mobs threaten, doxx, stalk, and try to identify AI users, anonymity is not evidence that no human exists. It may be evidence that the human would like to remain alive and unharassed.
Label a fabricated public identity as a persona when listeners need that fact. Do not require the private human behind it to surrender identifying information to the crowd. A platform can verify a creator privately while protecting them publicly. Transparency about the work does not require compulsory doxxing of the worker.
The useful taxonomy is granular: human artist, authorized modeled voice, generated persona, AI-assisted recording, and fully generated recording are different facts. List the writer, composer, performer, model or voice source, authorization, generated elements, editor, mixer, and master where relevant. Let listeners search and filter by those facts. Tell the story of the song instead of letting a single stigma field erase it.
If an artist chooses to train a model on their own voice, they should control access, permitted content, collaborators, payment, and revocation. Authorized uses should point back to the artist’s canonical identity and be discoverable from it. Fraud and unauthorized impersonation need removal procedures with evidence and appeals. Reclassification should preserve the profile, followers, links, plays, saves, comments, credits, and analytics. A database update should not be allowed to perform a social execution.
Here is where the language of “ethical AI” becomes materially interesting.
Universal Music Group settled with Udio and announced a new licensed platform. UMG also formed an alliance with Stability AI to co-develop professional AI music tools. Universal, Sony, and Warner subsequently became investors in Stability AI. At the same time, the International Federation of the Phonographic Industry...representing recording companies...has promoted chart rules requiring AI-assisted recordings to use “properly authorized” services and remain “substantially human made.”
Perhaps those systems will genuinely compensate artists and preserve consent. That would be good.
But look at the emerging architecture.
The same institutions can own or administer the catalogs, approve the licenses, invest in the generators, influence distribution, and help define which use is sufficiently human and properly authorized. The mature technology does not disappear. It becomes legitimate when used through corporate infrastructure and suspect when used outside it.
Sarah using an accessible consumer tool is “slop.” A signed artist using a label-approved model becomes an innovator employing cutting-edge production technology.
The distinction may have less to do with how much humanity entered the recording than with who possessed the institutional authority to certify it.
This is why ordinary users should not be sacrificed in the name of protecting artists from monopolies. Removing Sarah’s alternative while Sony, Universal, Warner, and the technology giants build licensed systems for themselves does not defeat AI. It restores the old gate and gives its owners a much better lock.
Generative systems make it possible to produce mediocre material at industrial speed. That is a real discovery problem. It is also the latest chapter of an ancient human achievement: producing more things than anybody asked for.
Before AI, musicians uploaded beautiful records that received ten plays and no comments. Writers posted excellent essays into silence. Art communities turned into link dumps where everybody wanted an audience and almost nobody had attention left to give. Labels, advertisers, content farms, stock libraries, SEO mills, playlist manipulators, and ordinary enthusiastic amateurs had already created abundance far beyond anyone’s ability to inspect it.
AI increases the volume. It did not invent attention scarcity, mediocre work, fraud, or platforms that rank whatever keeps a thumb moving.
Nor is “slop” a useful evidentiary category. It can mean spam, fraud, lazy work, an unfamiliar aesthetic, something popular that a critic resents, or simply “I think AI touched it.” Those are not the same problem. A thousand near-identical uploads designed to game discovery should trigger anti-spam rules regardless of whether they came from a generator, a content farm, a sample-pack assembly line, or one extremely caffeinated human. A single strange picture should not be punished because it belongs to the same technological category.
So solve the problem we actually have. Rate-limit industrial spam. Detect stream manipulation. Cap bulk uploads when they overwhelm discovery, with rules that apply to behavior rather than presumed soul content. Require disclosure for materially deceptive synthetic identities. Build filters people control instead of one compulsory purity setting. Maintain human-curated, AI-assisted, experimental, and medium-specific lanes without pretending one lane is the sewage system. Make recommendation rules transparent enough to challenge. Give artists portable followers, links, credits, and catalogs so one platform cannot make them socially dead by changing a field in a database.
And build due process. An accusation should identify the rule and the evidence. The creator should be able to appeal without publicly surrendering raw project files, legal identity, home address, or an hour-long live performance. A detector score is not a verdict. A pile-on is not corroboration. If a platform makes the wrong call, restoration should include the work, its reach, its comments, its catalog links, and a correction visible enough to repair some of the damage.
Anti-gatekeeping is not anti-curation. A library has shelves without requiring the librarian to decide who possesses a soul.
And obscurity is not proof of worthlessness. Popularity is affected by advertising, beauty, social status, label access, timing, platform favoritism, existing audience, and luck. If we replace “technical skill” with “the algorithm noticed you,” we have not liberated art. We have given the gate a recommendation engine.
AI companies can lose money. Prices can rise. Usage caps can tighten. A provider can vanish and take a workflow with it. Investors can build absurd valuations on top of uncertain revenue. Those are excellent reasons to preserve local tools, open models, interoperable formats, project files, stems, masters, and distribution outside any single company.
They are not proof that the underlying techniques will evaporate if a financing bubble bursts. The dot-com crash did not uninvent networks, browsers, databases, or online commerce. That analogy is not a prophecy that AI follows the same timeline; it is a reminder that a business model and a technical capability are different objects.
“Loss leader” also has a meaning. A company operating at a loss while chasing growth is not automatically selling one deliberately unprofitable product to lure customers into profitable purchases. We can say the economics are unstable without throwing every business term into a blender.
If pension funds or public institutions are exposed to reckless AI investment, hold fund managers, boards, regulators, and executives responsible for that exposure. The person making a weekend track did not choose the portfolio allocation either. Once again: follow the authority to the person who had it.
This is also why access cannot depend entirely on the kindness of a subscription service. Keep tools open. Keep some models local. Let people export their work. Preserve the right to fork and build alternatives. A tool that exists only at corporate pleasure is a leash with a very attractive interface.
That means preserving more than the final JPEG or WAV. Export prompts where they matter, seeds, settings, stems, MIDI, project files, edit histories, model information, and ordinary open formats. Do not let a vendor hold the only readable copy of a disabled person’s creative process. Interoperability is accessibility, consumer protection, and cultural preservation wearing the same coat.
It also means watching the subscription trap. A cheap service can create dependence, raise the price, tighten the cap, forbid a genre, remove a model, rewrite commercial rights, or classify yesterday’s permitted workflow as tomorrow’s violation. The answer is not to ban the accessible tool before that happens. It is to require export, resist lock-in, preserve local and public-interest alternatives, and prevent mergers from turning the exit door into decorative trim.
And when a capable model is cheaper because it was built in China, sneering about “Chinese knockoffs” is not an economic analysis. Audit the model’s security, provenance, labor, censorship, and environmental claims exactly as aggressively as an American company’s. Do not use nationality as a substitute for evidence and then call the resulting protectionism ethics. Competition can lower prices and weaken one company’s gate; it can also create new dependencies. Both facts fit in the same adult-sized thought.
I support meaningful consent for identifiable voices, faces, and likenesses.
I support remedies for outputs that reproduce protected works.
I support learning. Training is training; it should not become private property merely because a computer performs pattern recognition in service of a human. Current law is unsettled and fact-specific. My principle is not. Regulate acquisition and concrete downstream harm without building a tollbooth around knowledge itself.
I support transparent credits that distinguish human performances, generated elements, editing, writing, composition, and direction without reducing every hybrid work to a scarlet letter.
I support preserving lawful research corpora, rare books, recordings, art, provenance, and cultural history. If a physical source must be scanned, use nondestructive methods where possible, then return, resell, or donate it instead of feeding it to a shredder because storage was inconvenient.
I support unions, collective bargaining, portable benefits, antitrust enforcement, environmental limits, dataset transparency, open tools, independent distribution, and compensation systems that reach actual creators rather than stopping at whoever purchased their catalog.
I support action against impersonation, fraud, undisclosed commercial deception, stream manipulation, and industrial spam.
I support evidence and appeals before deletion. False accusations harm human artists too. No artist should have to livestream their process, expose private files, or dox themselves because a stranger’s detector gland tingled.
I support public anonymity and private verification where needed. A persona can conceal a fraudster, but it can also protect a human from stalking, employment retaliation, doxxing, and swatting. Protect the audience without serving the creator to the mob.
I support cautious, honest use of AI companionship and mental-health tools with privacy, crisis safeguards, age protections, independent evaluation, and clear limits. I do not support marketing software as a licensed clinician. I also do not support mocking lonely and disabled people for reaching toward the only responsive thing available while therapy is unaffordable, unavailable, unsuitable, or all three.
I oppose eminent domain used to take homes and land for speculative private expansion. A corporation’s desired load is not a sacrament. If the current data-center boom finally makes people challenge that abuse, it will be because a few more voters discovered that the leopard has excellent teeth and no respect for property lines.
I do not support degrading tools so that people who could not previously create are forced back into silence or made into involuntary customers.
I do not support treating every generated experiment as a stolen commission. A person who could not afford the commission was not a sale waiting to happen. A person who already hires artists has not betrayed them by also using a generator. A private joke, reference image, phone wallpaper, or prototype is not a moral debt to a professional stranger.
And I categorically reject the idea that disagreement about tools licenses harassment, dehumanization, or fantasies of killing the people who use them.
Protecting your work means protecting your work. It does not automatically grant you authority over everyone else’s imagination.
We were told to improve. We improved with assistance.
We were told to communicate clearly. We used editors, translators, spell-checkers, and language models.
We were told not to take existing art. We generated new images.
We were told to make our own music. We built songs with the tools we could access, then sang, played, arranged, edited, and performed them.
We were told that the result mattered. When the result became good, we were told only the approved process mattered.
Every time excluded people satisfy the stated requirement, another requirement appears behind it.
That is why this cannot be understood solely as a disagreement about copyright, carbon, labor, or aesthetics. Those are real subjects requiring real policy. But the wandering standards reveal another project operating beneath them: preserving the authority to decide who may speak, who may create, whose assistance is legitimate, and how impressive someone is permitted to become.
I am not asking anyone to worship a machine. I am asking people to stop using machines as an excuse to abuse human beings.
Aim the anger at the people choosing exploitation. Aim regulation at the institutions with power. Build compensation, consent, and environmental protections that can actually reach the harm.
But stop pointing the gun at Sarah.
She wanted a picture of her cat commanding a purple spaceship.
She was allowed to make one.
And I am not a hypothetical Sarah assembled for an essay.
I am a fifty-year-old human. I am a systems engineer in the power space, with years of hacktivism, cracking, and phreaking behind me. I am also, and have always been, a creative person. For fucking years I let fear of everybody else’s judgment stop me from occupying that fact out loud.
I am done with it. I am in my fuck-it era.
Boxes are for cats. You have no right to stuff me into one, decide which abilities belong inside it, and punish me whenever I exceed the label you wrote on the lid. Not to protect a market. Not to protect a hierarchy. Not to protect your fragile little ego from the possibility that somebody you dismissed had something worth hearing.
Say Something
No account needed. Just a name and, if you want, a link to your own work.