AI Metabolisis:
Sixteen policy proposals for a technology that isn't leaving
As a nation and a society, the United States currently suffers from a bad case of indigestion. AI1 is a big meal, and it turns out that we — in the United States especially — lack the legal and social infrastructure to adequately metabolize it into something that can support mentally and economically healthy humans.
Some people have adopted the posture of hoping AI dies in its cradle, hoping that it will simply go away. Others have adopted a picket line defense, chanting “No AI!” at everyone on social media, because their livelihoods are at stake. Anti-AI sentiment is, to be frank, largely the fault of the AI industry. It did not have to be this way.
They shipped into education without evidence. They shipped companion products to minors and are now being regulated retroactively, state by state, because they wouldn't self-restrict. They trained on pirated books when buying them was affordable and then litigated for two years. They flooded creative marketplaces they also owned. They talked about transforming society while declining any of the obligations that come with transforming a society. The public reaction is a reasonable response to being handed something powerful by people who behaved as though consent was a formality.
A little over half of all Americans have an unfavorable view of AI. Bring up datacenters, and the split is 70/30 against. Even people who like AI don’t want a datacenter in their own back yards. The general public has a more clear-eyed view of the current benefits and costs of AI than many CEOs seem to.
AI, itself, is not the problem. It’s arrival was inevitable, and now that it is landing it is an object of scorn not because of what it is or what it does, but because our society was not ready for it.
I suspect, even if the “bubble” does burst and the ballooning economics of AI collapses from over-eager expansion, that AI itself will not simply disappear. Instead, we may end up with the worst and most destructive forms of AI sticking around (the things most easy to monetize by unscrupulous actors), while the work necessary to make AI truly safe and helpful goes unfinished.
Consequently, I want AI to succeed. But I also want what is most human to endure. Indeed, AI will not have succeeded if its only result is to make billionaires richer while forcing the rest of us into abject poverty.
When I say I want AI to succeed, I define that success as measured by increased ability to improve our lives and decreased ability to ruin our lives. It has failed thus far at achieving both in equal measure, benefitting some at the expense of others.
I am not neutral here. I use AI daily. I also failed in an attempt to re-career as an audiobook narrator precisely because a major company decided to start pushing AI-generated narration as a free option for independent authors, who in a marketplace increasingly crowded by AI-generated books, have less and less income with which to hire narrators. Incidentally, I’m also one of those authors.
I both love and hate AI. Both feelings are justified.
Because I do not believe AI is going to disappear, no matter what we do, I argue we need to carefully consider how AI could be integrated, or metabolized, into our society without sacrificing (1) humans’ ability to choose, (2) essentially human practices that are at the core of our species, (3) and without tolerating population-level harms that could negatively affect entire nations, our world, or our future as a species. Our society was not ready, but we could be doing a lot more to curb the excesses and make the AI we are stuck with a force for good. Because it can be.
“Metabolization” sucks as a word.
I trip over it trying to say it. It’s also just wrong. Our internalization does not break AI down, it does not convert it from raw material into the things we need to thrive while conveniently ejecting the waste that does not serve us. AI can be readily absorbed in some sectors, but in others—humanities and social sectors— it is largely just poisonous (a fact I do not believe is inevitable or inherent — the difference between poison and medicine is dosage).
Other words, like “integration,” seem inadequate to the task as well. Integration implies two whole things somehow merge without losing anything, becoming more than the sum of their parts. This is the simplistic narrative some try to sell us, and we see through it. I do not think such a simplistic merger is possible with AI.
Transformation is required, both of us as people and of AI as a new human institution of practice. The challenge we face as society is how to metabolize AI into our actual, social, and political bodies without choking on it or poisoning ourselves.
Metabolisis
Metabolisis is not a new word, though it is nearly an empty one. Webster's 1913 dictionary lists it as a rare New Latin variant of metabolism and marks it obsolete; it has no independent sense of its own. I am reviving it precisely because it is empty. The -sis ending in Greek denotes a process rather than a state, as in symbiosis and homeostasis, and I want a word for what happens when two things change each other, when the act of consuming also co-creates.
Unlike metabolism, metabolisis is bidirectional, balanced, the product of sustained constitutive interaction. In the context of AI, I take metabolisis to mean a process in which living beings incorporate the technological (AI) into their being and praxis through a mutually transformative process, without destroying either. It is not without loss. But neither is it utterly destructive. It is a process that is part metabolizing—converting it into something that empowers or fuels us—, part deliberate and well-considered symbiosis, and part alchemy in which the cyborg learns to live with itself. Where metabolization destroys something utterly so that the other can survive, metabolisis is an equilibrium of two parts digesting and supplying energy for one another, so that both survive.
The hope of “killing” AI promises futility. Burst the bubble and the technology will remain, but in potentially worse hands. People who want to kill AI are rightly afraid, because they feel themselves — their careers, their joys — being metabolized.
The hope of shaming individuals for using AI promises unproductive social conflict that, on the evidence, doesn’t seem to be changing anyone’s behavior. Exhortations to individuals have not saved the environment (as opposed to collective, government action against the actual culprits, who are namable large corporations), and they will not change AI.
A recent Pew poll found that about half of all Americans use AI chatbots. A quarter of them use it daily. And despite the apparent unanimous agreement among my indie author friends that people who use AI to write their novels for them do not deserve the claim the title of “author” (a sentiment I agree with), such fake authors are still everywhere.
Only collective action, on the scale of states and nations, will tip the scales. In a country that is as divided over AI as it is divided politically, the battles we choose, and the hoped-for result we work toward, matters a great deal.
Metabolisis is the hope of achieving a new, legal, social, and economic equilibrium with AI that does not leave human beings feeling like they have lost the essence of their humanity.
It is now going to be philosophically boring to suggest this metabolisis must, at least at first, take the form of concrete policies and laws, but when the body at question is the social body of an entire nation, law and policy are the only (relatively) quick levers we have to collectively work with.
The remaining lever, cultural norms, is a longer-term project and a harder one to achieve, but one which can be helped along by picking the right policy fights for collective action.
The proposals below align along three major axes: Human Recourse, Protecting Human Creative Practice, and Avoiding Population-level Harms.
These proposals are procedural where possible, substantive where necessary. Where a substantive commitment is required, I prefer mechanisms that achieve it by procedural means — levies, filterability, procurement conditions, market-structure rules — because those are draftable, enforceable, and are more likely to survive constitutional review.
However, procedural remedies cannot always reach aggregate harms. Every individual transaction can be disclosed, consented to, fairly priced, and fully auditable — and the aggregate result can still be that nobody can make a living as an illustrator, a narrator, or a novelist, or an AI datacenter carelessly poisons or exhausts a county’s clean water supply. Fairness at the transaction level says nothing about survival at the population level.
I don’t mean to sound like a politician—I’m not one. But I think there is a dearth of detailed thinking about this that isn’t simply reactionary, and I hope to help push productive conversations about policy and new cultural norms that need to be had out into the open.
Axis 1 — Human Recourse
AI cannot occupy a position that the affected person(s) cannot audit or exit.
The teacher who cannot be questioned. The benefits denial with no human on the other end. The likeness I cannot reclaim. The training corpus I was never asked about. The infrastructure that cannot be unplugged. In each case the objection is not that AI did something. It is that AI took a position where the affected human has no recourse.
AI offered as a writing assistant is fully auditable — I can read every line — and fully exitable: close the tab. Labs renting out engineered minds capture nobody. Learning from the human corpus is what learning is. The line for this axis falls exactly where recourse ends.
The policy formulation of this axis is that AI must not do X without disclosure, consent, compensation, or a human holding final authority. This protects the public’s capacity to judge rather than dictating the judgment — which means this framework does not require the public to agree on the limits of what AI can or cannot do in order to accept it.
Axis One Prescriptions
1. Right to a Human Decision
UnitedHealthcare’s nH Predict algorithm allegedly carried a 90% error rate owing to lack of human review, with over 80% of prior-authorization denials reversed on appeal. A STAT investigation cited in the litigation reported employees were pressured to keep rehabilitation stays within 1% of the algorithm’s predicted length. In the parallel Humana case, roughly 2% of policyholders appeal, and one named plaintiff received seven denials for the same care within thirty days. (Lokken v. UnitedHealth, D. Minn.; Humana class action; npj Digital Medicine, Feb 2026.) A 90% error rate is survivable as a business model because only 2% of people appeal. The error rate is the revenue.
Policy
Any consequential determination about a person — benefits eligibility, medical necessity, credit, housing, employment, insurance, parole, — must have a named, accountable human decision-maker with authority to override the model, plus appeal to a human who did not make the original decision. I would argue this should be extended to consequential decisions that social media platforms or marketplaces make about who is allowed access to their services, often without meaningful recourse. Account bans and suspensions must be appealable to a human decision-maker.
Existing precedents
At least six states enacted laws in 2026 restricting AI use by health insurers, including Republican-led Iowa (HF 2635) and Democratic-led Washington (SB 5395). Bipartisan and building. (NYU Center on Technology Policy, July 2026.)
2. Human Teachers
Salamanca City Central School District, on the Allegany Reservation in Cattaraugus County, New York, is paying $57,590 for a seated humanoid robot (”Sally”) plus an AI teaching- assistant program with a laptop avatar. The vendor acknowledges it has no empirical evidence the machine improves learning. The district serves ~1,300 students, 79% economically disadvantaged. The vendor, Realbotix, was formerly the crypto company Tokens.com and in 2024 acquired the parent company of the RealDoll sex-doll brand. (New York Focus, 14 July 2026; Inc., July 2026.)
The doll is the story people are covering, and it is disturbing enough in its own right. The bigger problem is the tutor: it is designed to remember students’ prior interactions, help with homework, and provide support outside school. That is an unsupervised pedagogical authority relation involving students whose education is already at a disadvantage, with no empirical evidence to suggest that such a relationship is in any way safe or beneficial.
Policy
No AI system should serve as instructor of record, assign grades, determine placement or discipline, or hold unsupervised instructional authority over a minor. No student (or adult for that matter) should be punished on the basis of an AI determination alone: plagiarism is a case in point, with AI- “detectors” largely themselves being AIs with a notorious false-positive rate.
This is also about more than ensuring there are always humans involved in decision-loops involving children, students, and education. It is about who holds the superior position. It is already the case that we increasingly outsource our thinking to our smartphones, to AI, to our computers. We hold onto our role as the user, with the machine as the tool, just barely. The tool already wields us like a tail wagging the dog. If we permit an AI to assume the role of a teacher, we invert the relationship entirely.
Existing precedents
California AB 2370 (2024), adding Educ. Code § 87709: the instructor of record must meet Board of Governors minimum qualifications; explicitly prohibits using AI to replace faculty in providing academic instruction. (Community colleges only.)
Idaho SB 1227 (enacted 2026): statewide K-12 AI framework, AI literacy standards, educator training, data privacy requirements; prohibits AI from replacing human teachers.
South Carolina H.B. 5253: written parental opt-in before any student interacts with an AI tool; AI cannot replace licensed teachers under any circumstance.
Oklahoma (2026): prohibits AI tools from being primarily used for grading, discipline, or other high-stakes educational decisions.
NYC Public Schools guidance (2026): red/yellow/green rubric. Red list bars AI from decisions on placement, discipline, graduation, or program access; IEPs and 504 plans stay with qualified humans; grading stays with the teacher of record.
3. A synthetic content provenance standard
Transparency about how content was generated should be the baseline standard. Consumers downstream of AI-produced content should be able to easily tell how AI was used, and to quickly identify content that is entirely AI-generated. This enables humans to make informed decisions.
Policy
Machine-readable marking plus human-legible disclosure on AI-generated audio, image, video, and text. Conversational systems must identify themselves as machines. A durability requirement so provenance metadata survives ordinary platform processing.
Existing Precedent
EU AI Act Article 50, obligations applying from August 2026.
Note that this policy would also foreclose the possibility of having AI-generated actors or voice-overs in film or television without clear disclosure of the fact to the audience.
4. Authorship disclosure in creative marketplaces
Consumers have the right to exercise discernment over whether or not they want to consume AI-generated content, and corporate practices meant to obfuscate the line between AI and human-generated content impose a unique burden on human creators and consumers alike.
AI-generated work without human authorship is not copyrightable — SCOTUS denied certiorari in Thaler v. Perlmutter on 2 March 2026, settling it. This does not help. Copyright is a right to exclude others from copying; it says nothing about whether a thing can be sold. An uncopyrightable AI novel still lists at $2.99 and still competes. Non-copyrightability may even accelerate the flood, since nobody has to clear rights. Copyright may have landed on the side of humans, but it does not protect livelihoods, or help us avoid content we do not want to be tricked or forced to consume.
Policy
Platforms that monetize creative work — KDP, IngramSpark, Audible, Spotify, stock image libraries — must collect and publish a disclosure of AI involvement. Compliance by sworn statement or exportable interaction logs. Undisclosed AI authorship in a monetized creative work should be treated as a material omission.
Existing Precedent
FTC rulemaking under existing deceptive-practices authority. Does not require new legislation.
5. Mandatory functional filterability
Human narrators are drowning in synthesized narration. Filtering is technically possible but requires an understanding of how to manipulate search results that ordinary listeners do not have. If human-narrated audiobooks are one tap from being the only thing a listener sees, the flood of voice-synth stops mattering to the people it was drowning. Making a market navigable is a smaller ask than trying making it exclusive—bans on monetization of AI produced content may not pass muster, legally— but this does much of the same work.
Policy
Any marketplace above a size threshold that distributes creative work must provide and prominently surface a working filter for AI-generated content, defaulted to the user’s saved preference. Not a buried search operator. Not a metadata field that exists but is never exposed. A visible toggle that works and persists. This requires zero determination of human contribution on the part of marketplaces. The seller declares truthfully under legal penalty; the buyer filters; the platform’s only obligation is to make the declaration actionable.
Existing Precedent
“readily discoverable and usable by an ordinary consumer without specialized knowledge” — language borrowed from accessibility and disclosure law. It is consumer protection, not speech regulation, so it survives First Amendment review in a way a monetization ban does not.
Amazon's policy requires KDP authors to disclose AI content to Amazon — but not to Kindle readers. The data already exists, collected at upload on every book. It simply isn't surfaced to the person deciding whether to buy.
6. Expand individual rights over likeness
It is easy to laugh at the antics of an AI-generated James Talarico wearing an apron in a video designed to attack his masculinity, less easy to laugh at deepfakes so convincing that the public is fooled into believing a lie. Beyond individual bad actors, the danger of big film studios using deepfake likenesses of actors or voice actors to generate new content could threaten to put future generations of performers out of work. Why seek out new talent when you can just keep re-using the old talent, potentially forever?
Policy
Protect the right of individuals to own and control their own likeness.
Existing Precedent
Denmark’s proposed amendment to the Copyright Act, implementing a general protection for all natural persons against realistic digitally generated imitations of physical characteristics. It also includes præstationsbeskyttelse — protection for performing artists against deepfakes imitating their artistic performances without consent.
S. 4591 / H.R. 8915 (NO FAKES Act of 2026). Creates a federal intellectual property right giving every individual — celebrity or private citizen — control over how their voice and visual likeness are used in digital replicas. Not assignable during life, licensable, survives death. Status: Not yet passed.
7. Copyright and IP Theft: Training-data provenance
The court in Bartz v. Anthropic held that training on lawfully acquired books was transformative fair use, but training on a corpus the defendant knew to be pirated was not. The $1.5 billion settlement was finally approved on July 20th, 2026.
This sets up a broader judicial scenario in which LLM firms are being asked to prove the provenance of each individual work. This seems to me like it is the correct outcome. Human beings learn by consuming the corpus of human output around them. They become artists by studying art, writers by reading, musicians by listening. That an engineered mind has to learn the same way is only an issue when the “parent” is stealing books for them to read instead of buying them.
A licensing market is already forming — aggregators (ProRata, ScalePost, TollBit) license publisher content collectively, alongside direct publisher-to-lab deals. Disclosure is going to be the mechanism that makes that market function.
Policy
Enforce mandatory disclosure of training corpora at a level of detail sufficient for a rights-holder to determine whether their work is included. Create a statutory rule that unlawfully acquired training material is per se outside fair use, and that individual creators retain control over whether or not to allow their content to be included in the AI training corpora, even when that content is otherwise assigned or licensed for others to use. That is, require content creators to have provided specific consent to the use of their content in training corpora.
Existing Precedent
EU AI Act Article 53 summary requirement; California AB 2013 domestically.
A note about plagiarism:
Plagiarism is creating a new output that substantially copies an original (most often without credit given, but even if credit is given it is still plagiarism if it exceeds fair use limits or if word-for-word text is not appropriately quoted and attributed). Early incidents of plagiarism from LLMs have largely been trained out. In 2026, frontier models are unlikely to output plagiarized content as new creative output. Generally speaking, the issue of pirated training material is not a plagiarism issue. It’s a copyright issue.
Axis 2 — The survival of human creative practice
AI cannot be permitted to easily strip us of facility or praxis that is fundamental to being human.
The second axis aims to protect something that some AI advocates seem willing to sacrifice on the pyre of progress: the human capacity for creativity. They argue that AI-generated movies, books, music, or images “unleash” creativity, removes access barriers for those who have not practiced or studied, democratizing the production of creative outputs by enabling anybody to write a novel with just a few words in a prompt.
What is at stake is not a supply of images. (Or novels, or music, or movies…)
Nor is it alleged “gatekeeping” that keeps those unwilling to put in the work from being able to “create” art or write novels. Certainly not when the “gatekeeping” involved is simple effort.
It is art as process, which is the whole point of art to anyone who has studied the Humanities seriously. The product is the residue. If the practice stops being survivable, if it becomes only the domain of leisure for the wealthy, we lose the human activity that produces people who can see, and we lose the tradition by which that capacity transmits from one generation to the next.
A profession is the institutional form a practice takes so it can be a life rather than a hobby. Kill the profession, and you do not democratize the practice — you re-aristocratize it.
You create a world drowning in images without any artists. The position that a creative work’s value is exhausted by its value as a product is a position, held by identifiable people for identifiable reasons. Simply pointing that out, in these terms, is itself damning.
Policy must focus on protecting anthropocentric creative processes and labor.
It is tempting for some to try to ban AI generated creative output, and for outputs with only minimal human intervention, and I would support such a measure. It is, however, unrealistic.
A second path could be to tie monetization or commercial use of AI generated output to the proportion of human contribution. While AI chat logs could be called into serve such a function, determining how to interpret that evidence, how to account for human disability, how to know what human labor existed outside of the body of evidence, this scenario would most likely be litigated into uselessness or injustice, and it would land hardest on people least able to defend themselves. We can require disclosure, but we cannot meaningfully adjudicate the amount of human labor that went into a creative output.
What can be regulated is the market structure. The following policy ideas would work best in conjunction with the Axis One Prescriptions listed above.
Axis Two Prescriptions
1. Volume and promotion controls
The harm is not that AI can make a book. It is that AI can make ten thousand books before lunch. Scale is what buries independent authors and publishing houses alike, and scale is regulable without ever asking how much of any single work a human made. The same is true of music, with Spotify finding itself flooded with AI-created tracks competing for plays. Stopping people from using AI to create books, art, or music is unlikely. The issue isn’t that individuals do this, its that it unbalances the market in a way that harms human creative practice as a whole.
Policy
For creative marketplaces, submission throttles per account per period. No algorithmic promotion, recommendation, or paid placement for undisclosed synthetic work. Prohibit bulk-upload pipelines into creative marketplaces.
Existing Precedent
Both affected markets already do this voluntarily. In September 2023 Amazon capped Kindle Direct Publishing at three new titles per account per day, explicitly to guard against AI abuse, with an exception process for legitimate high-volume publishers; the cap is still in force. In September 2025 Spotify deployed a spam filter that tags mass uploaders and stops recommending them — 75 million tracks pulled in the preceding twelve months — and adopted the DDEX credit standard so a disclosure can record where AI was used instead of forcing a false all-or-nothing binary. It began showing those AI use details in the song info panel in April 2026. The target of these policies is the behavior, not the technology.
The Better Online Ticket Sales Act of 2016 makes it unlawful to circumvent the technical measures a seller uses to enforce posted purchase limits, enforceable by the FTC and state attorneys general as an unfair or deceptive practice. Enforcement, unfortunately, has been thin.
Neither the KDP nor Spotify policies are enshrined in law, so either could be revoked at any time.
2. A generation-side levy funding creative practice
While I would support banning fully AI-generated creative output from marketplaces, there are alternatives that may generate less resistance. Rather than prohibiting downstream monetization, taxing generative-AI production of creative output upstream could be directly used to fund human creative endeavor.
Policy
Impose a levy on commercial generative services for creative-domain output, distributed to working artists through a collecting society. Such a fund would be income-restricted, helping working-class creatives who otherwise would not be able to make a living to focus on creative practice.
Existing Precedent
European private-copying levies on blank media — including Denmark’s own blankmedieordning — built on a similar logic: a technology that displaces a creative market pays into the market it displaces. This has the benefit of having the least novel constitutional exposure, as it funds practice rather than protecting product.
3. Procurement and public-funding conditions
One of the easiest levers we have is controlling what the government funds. While undoable, even an executive order could help protect human creative praxis.
Policy
No federal or state contract, grant, or publicly funded commission may substitute AI-generated work for a human creative commission. This includes public broadcasting, government publications, museum and library commissions, public art, or any other human creative work.
Existing Precedent
The government acting as market participant has essentially unlimited discretion. Thus, this presents no constitutional problem and sets a norm private markets could follow.
4. Commercial mimicry of a living creator’s style
For a working creative, a distinctive style is their most valuable asset — the reason a client hires this illustrator rather than a cheaper one, or falls in love with a partular writer.
It is now trivial to extract that asset and resell it without the person. When ChatGPT’s image generator flooded the internet with Studio Ghibli pastiche in 2025, the studio’s co-founder had already told us what he thought of it; a 2016 clip of Hayao Miyazaki calling AI-generated art an insult to life itself resurfaced within days, to no apparent effect on the volume.
In March 2025 OpenAI added a refusal triggered when a user asks for an image in the style of a living artist, while continuing to permit broader studio styles. As of July 2026 it extended the same boundary to prose, declining to imitate the distinctive voice of a living author. A company with a lot to lose from this restriction has imposed it on itself, twice. It was the right thing to do.
Policy
Prohibit the commercial use of AI-generated creative output that purposefully mimics the distinctive style of an identifiable living creator. Purposeful is the operative word — the prohibition should reach deliberate imitation, prompted by name or by unmistakable reference, not incidental resemblance. Personal use, non-commercial use, education, parody, and criticism remain untouched. Enforcement belongs to the creator whose style was taken, not to a government body deciding what counts as art.
Existing Precedent
OpenAI blocks image generation in the style of living artists (March 2025) and, as of July 2026, text imitating the distinctive voice of living authors. Both are self-imposed.
The legal form already exists, though not in copyright — style itself is not copyrightable, and a prescription built on copyright would fail. It belongs instead with identity and unfair competition, where courts have addressed exactly this shape of harm. In Midler v. Ford Motor Co. (9th Cir. 1988), Ford hired a sound-alike after Bette Midler declined to license her voice; the court held that deliberately imitating a distinctive voice for commercial advantage is actionable. Waits v. Frito-Lay (9th Cir. 1992) applied the same reasoning and produced a multimillion-dollar verdict. Neither case turned on copying a song. They turned on appropriating what made the performer recognizable.
Axis 3 — The survival and welfare of the human species
AI cannot be permitted to impose population-scale harms on people who never chose it and cannot escape it.
The first two axes are about what happens to a person: a decision made about you with no appeal, a practice you can no longer make a living at. This one is about what happens to everybody at once, whether or not any particular person ever opens a chatbot.
These harms share a structure that makes individual remedy useless. You cannot decline the electricity rate increase caused by a datacenter three counties over. You cannot personally opt out of your industry being restructured. The content moderator in Nairobi who screens the worst material on the internet so that the rest of us do not see it has no negotiating position at all. And the deployment of AI into weapons systems and critical infrastructure is not a consumer choice anyone gets to make.
Some of these are also irreversible in a way the first two axes are not. A market can be re-regulated. A profession can, with effort and money, be rebuilt. An aquifer cannot be un-drained, and a catastrophic failure in a system that controls physical force cannot be appealed.
This is the axis where I have the least confidence that any proposal is sufficient. I include it anyway, because ignoring the most serious problems AI poses is exactly how we could arrive a catastrophic outcome.
Axis Three Prescriptions
1. Displacement notice and bargaining rights
We talk about AI replacing labor as though it were a future event. For a substantial number of people it has already happened, and the people it happened to had no notice, no bargaining position, and no legal category to file a complaint under.
Policy
Amend the WARN Act: lower the employee threshold and add an automation trigger, so that AI-driven workforce reduction requires advance notice and disclosure rather than arriving as a surprise. Make AI-driven displacement a mandatory subject of collective bargaining under the NLRA, so that workers have a legal right to negotiate over it rather than a moral claim to complain about it.
Existing Precedent
The WARN Act (1988) already establishes that mass layoffs require advance notice; the amendment adds a trigger, not a principle. Mandatory-subject-of-bargaining doctrine under the NLRA has expanded before to cover technological change.
2. Supply-chain labor disclosure
There is a class of AI labor that almost nobody discusses, because the whole point of it is to be invisible. Every frontier model is made usable by human beings who label training data and review the material the model must learn to refuse — violence, abuse, exploitation — at volume, for low wages, generally in countries with weak labor protections. Social media platforms rely on the same kind of labor, and in Nairobi they draw from the same pool: the content moderators who unionized there work for Facebook, YouTube, TikTok, and ChatGPT. Those moderators sued over the psychological conditions of the work, and they have won real ground — Meta argued that Kenyan courts had no jurisdiction over a company not based in Kenya, and the Court of Appeal disagreed; the High Court ordered the companies to provide proper medical, psychiatric, and psychological care. Rulings on the substantive claims of human trafficking, unfair dismissal, and union busting were expected in February 2026 and were postponed. The cases remain pending. That litigation is instructive, and it is also the labor issue most ignore.
Policy
Require firms deploying frontier models to disclose the labor conditions of outsourced data-labeling and content-moderation work in their supply chain — the same disclosure logic we already apply to conflict minerals and garment manufacturing.
Existing Precedent
Supply-chain disclosure has direct analogues in Dodd-Frank § 1502 (conflict minerals) and the California Transparency in Supply Chains Act. The African Content Moderators Union and the Kenyan litigation against Meta and its outsourcing partner established that these workers can organize and can reach a court.
3. Universal basic income
I support this independently of AI. AI makes it urgent rather than theoretical.
Here is the specific thing that convinced me it belongs in an AI policy essay rather than a philosophical one: the tax code currently pays employers to replace workers. Capital investment in automation is subsidized through accelerated depreciation and expensing, while the wages of the human being doing the same job are taxed. We have built a standing public subsidy for displacement and then expressed surprise at displacement.
If it becomes genuinely profitable to replace most human labor with AI and robotics, then a job stops being the mechanism by which people obtain the means to live and becomes only a mechanism for giving people something to do, or for controlling them. A society that removes wage labor as a survival mechanism without replacing it has not automated work. It has made survival conditional on owning capital. That is not a technology question. It is a question about whether people exist for their own sake or as inputs.
Policy
Establish a national guaranteed income floor, tied to the cost of living in a two-bedroom apartment. In the same breath, with healthcare tied to employment, Medicare-for-all (who want or need it) would be just as necessary as UBI in a post-AI/Robotics industrial world.
In the near term, I would support eliminating the tax asymmetry that subsidizes automation relative to wages. That is a targeted reform, it is legible to legislators who would reject “basic income” as a phrase, and it removes an active public subsidy for displacement.
For creative practitioners specifically, a targeted basic income is a substantive remedy for the Axis Two problems that threaten creative work. It funds the practice directly.
Existing Precedent
Guaranteed Income Pilot Program Act of 2025 (H.R. 5830, Rep. Watson Coleman), in House Ways and Means. Appropriates $495 million across FY2026–2030; monthly amount pegged to the local cost of renting a two-bedroom home.
BOOST Act (Rep. Tlaib): a $250/month refundable tax credit — the incremental route through the tax code rather than a new program.
Alaska Permanent Fund Dividend: forty-plus years of proof that a resource-based universal dividend is administratively ordinary in the United States and politically durable across parties.
Ireland funded a basic income pilot for artists and retained it after it succeeded. An external report by Alma Economics found the pilot cost €72 million and generated nearly €80 million in benefits to the Irish economy. The culture ministry’s own figure: every euro invested returned €1.39 to Irish society.
Roughly 200 guaranteed-income pilots have launched across US cities and counties since 2017.
Proposals for an AI-funded version now span the spectrum: a token tax on AI usage paired with public equity stakes in frontier firms; Sanders’ robot-tax and wealth-sharing proposals; Altman’s “American Equity Fund” contributing ~2.5% of large-company value annually to citizens.
4. Datacenter disclosure and ratepayer protection
Aggregate AI water consumption is not the top-tier environmental crisis. Growing corn for ethanol and irrigating suburban lawns both carry larger footprints. For that matter, while I am not vegetarian, commercial meat farming consumes about 30% of the annual freshwater usage in the U.S.. Adopting lab-grown meat as a culinary norm would, by itself, solve a huge chunk of the freshwater usage problems in the United States.
The real problem is in specific locales, where AI datacenter water usage presents unique challenges to the water supply (in some cases, competing with already water-hungry industries like ethanol and meat farming). The costs of a datacenter land on a specific county’s water table, a specific community’s land, and a specific utility’s residential ratepayers — none of whom consented, and none of whom can exit. A household in that service territory pays for grid upgrades built for a customer who may leave in eight years. That manifests as an electricity bill no individual homeowner can escape.
Policy
Mandatory public reporting of energy and water consumption for facilities above a size threshold, in a standardized format that permits comparison across sites and across years. And a requirement that large-load customers bear their own interconnection and stranded costs rather than socializing them onto residential ratepayers.
Existing Precedent
This is the most productive area of AI regulation today: states enacted 28 datacenter laws in the first half of 2026 alone, falling into four categories — ratepayer protection, oversight and disclosure, rollback of tax incentives, and moratoria.
Oklahoma, Data Center Customer Ratepayer Protection Act of 2026 (HB 2992): utilities must create large-load tariffs ensuring those customers reimburse the costs allocated to them, including costs that would otherwise go unrecovered if the customer departs.
Alabama: large-load contract review through the Public Service Commission.
Idaho (H0895): limits the sources from which datacenters may purchase cooling water.
Washington (SB 6231) removed a datacenter tax exemption; Maine excluded datacenters from a business equipment tax exemption.
Maine’s moratorium bill was vetoed; New York’s (S10642) awaited the governor’s signature.
These initiatives have broad bipartisan support, presumably because members of both parties have to pay electric bills and drink water.
5. Frontier oversight, whistleblower protection, and human control of critical systems
Connecticut recently passed an AI safety law. It enacted the whistleblower protection and left out the risk assessments, the incident reporting, and the audits.
That is the whole lesson in one sentence. A state took the cheapest component of the model — the one imposing no ongoing obligation on any company — and left the teeth on the table. This is the default trajectory of safety legislation in the absence of sustained public pressure, and it is why “someone will regulate this eventually” is not a plan.
This is also where the apocalypse concern stops being a thought experiment and becomes procurement policy. The realistic near-term catastrophic risk is not a superintelligence deciding to end us. It is a defense contractor, an energy utility, or a hospital network integrating a model into a system controlling physical force or critical infrastructure, on a procurement timeline, with no external evaluation, because integration was cheaper than the alternative and nobody required otherwise. Reckless deployment into military, security, and infrastructure systems is not merely possible. It is the path of least resistance.
Policy
Federally codify what several states already require: published safety frameworks, critical-risk assessment before deployment, mandatory incident reporting, annual third-party audits, and whistleblower protection for employees of frontier labs. Add a statutory ban on autonomous AI in nuclear command and control, and on autonomous decisions to apply lethal force.
Existing Precedent
California SB 53 and New York’s RAISE Act (both 2025): safety frameworks, critical-risk assessment, incident reporting.
Illinois (2026): added a requirement for annual third-party audits — the strongest state version to date.
Connecticut (2026): whistleblower protection only. The cautionary example above.
The Block Nuclear Launch by Autonomous Artificial Intelligence Act was introduced in Congress with bipartisan sponsorship in 2025, establishing that a human-in-the-loop requirement for nuclear command and control is not a fringe position. The bill was not passed, and has not been reintroduced. Officially, having humans in control of nuclear launch is Pentagon policy, but not law.
What I do not support
There are some moves that I think could do more harm than good, but which are popular in some circles.
A moratorium on AI development or research. This will backfire, resulting only in current problems with AI remaining unsolved, and current safety risks remaining unmitigated. It also means we will find Chinese-produced AI models lurking around every corner.
A ban on AI-assisted writing, coding, or creative work. Let’s be clear: an AI assistant is not the same thing as an AI ghostwriter. You cannot legislate the degree to which someone needs or benefits from using an AI in their creative process. I do support a ban on AI ghostwriting: the disclosure policy above would effectively rule out the use of hidden AI ghostwriting. The rest is about what content the consumer chooses to consume.
Condemning personal use of AI. Trying to socially police the behavior of individuals, in contexts where the specific harm done is either nonexistent or cannot be pinned to namable human beings, is a fool’s errand. Demand disclosure, and decide to consume content or not. Live your ethical commitments, and let others live theirs.
Any regulation premised on AI capability having plateaued. I think that premise is false, and a framework built on it collapses when the premise does.
Conclusion
I advocate for disclosure above all else out of what I hope is not a misplaced faith in my fellow human beings. I believe that readers will prefer books that embody actual human experience and understanding. I believe movie-goers will not be able to feel any connection to an AI-generated “actor.” I believe that there is genuine value to the unique style and perspective of a singular artist, over mass-produced imitation images. I believe that, armed with the appropriate information, consumers will largely protect creative human endeavor through their choices about what to buy.
I also might be entirely wrong about that. I hope that I am not, because a world in which human actors in movies are considered passé is most likely a world I do not want to experience.
I know firsthand that AI does dramatically alter the landscape of labor. Both because I have suffered for it, and Claude drafted portions of this essay, which I then rewrote; it did research and source verification, and I checked every citation myself, including several it got wrong. The chat log is available on request, easy enough to export.
AI has helped me at least as much as it has hurt me. It is not evil. We should not “kill” it. The harms and risks it presents are not about AI. They are about what human beings are doing with it, and doing to each other with it. It is not AI that is poisoning us so much as it is our own inability to distinguish the medicinal dose from the fatal one.
Frontier model LLMs are AI, and people claiming they are not are wrong. I wrote a whole essay about it. I do not address here the potential moral implications of AI consciousness, because that debate does not currently implicate public policy. It matters, may someday matter enormously, but it’s not what this essay is about.