New York Wants a Label on Machine-Written News. The Label Is the Cheap Half.
A $1,000 fine tells you what you are reading. It does nothing for the newsroom whose reporting trained the machine — here is how to spot the difference, and where your attention and money actually land.
A few weeks ago I forwarded an article into a group chat with the words "this is a big deal" attached, and went back to making coffee. Seven hundred words, a clear headline, a percentage in the third paragraph that was the entire reason I sent it. It read the way a competent local story reads when you are half-awake.
It was only that evening I looked at it properly. The byline was a name that returned nothing anywhere else. Every paragraph was almost exactly the same length. Not one quote came from a human being who had picked up a telephone. The piece had the shape of reporting with none of the friction — no sign that anybody had been inconvenienced in the making of it.
I still do not know what wrote it. Nobody was obliged to tell me. And when I went looking for how common that is, I found NewsGuard counting 3,749 AI content farms dressed up as news sites — a figure last updated on June 23, 2026 and still climbing.
New York's legislature has decided to do something about exactly this. What it decided is both more interesting and less useful than the headlines suggest.
What the Act actually requires — and what it costs to ignore
Start with the statute, because almost every argument about it turns on a single adverb.
The Fundamental Artificial Intelligence Requirements in News Act — the FAIR News Act, S8451B in the Senate and A8962B in the Assembly — was passed by the state legislature in the closing hours of the June session. The most recent reporting I could verify, from mid-June, had it on Governor Kathy Hochul's desk for signature — or veto, as Editor & Publisher put it. Whether it has been signed by the time you read this, I cannot tell you. Assume nothing, and check.
Here is the bill's actual text, blunter than the debate around it. Any news media content "which was substantially composed, authored, or otherwise created through the use of generative artificial intelligence" must "conspicuously imprint on the top of the page" that it was substantially created by generative AI. Get it wrong and a court "may impose a civil penalty of one thousand dollars for the first offense, and five thousand dollars for each subsequent offense."
That is the whole machine. A label at the top of the page, and a thousand dollars.
Two things are worth pausing on. First, New York already runs this ladder: on June 9, the week the Act cleared both chambers, a near-identical law took effect covering AI performers in advertising, with the identical $1,000-then-$5,000 structure. The state has settled on a price for undisclosed synthetic content, and it is the cost of a decent used car.
Second, the process. Jim Heaney at Investigative Post records the bills approved 53 to 7 in the Senate and 130 to 1 in the Assembly, with sponsoring Senator Patricia Fahy conceding neither chamber considered it in committee or held public hearings. He also notes the Act "authorizes the state attorney general to determine what content is substantially created by AI." Hold onto that sentence.
Fahy, who championed it alongside Assemblymember Nily Rozic, frames it as modestly as a speech regulation can be: "No one is telling you you can't... All we are saying is, we want transparency." Labor pushed hard for it too: the guilds behind it called it a protection for journalists and readers alike, and Susan DeCarava of the NewsGuild of New York wants New Yorkers to trust that the news we consume is made by and for humans.
So far, so reasonable. A label is cheap and the rationale is obvious: I should know whether a person did this. Here is where it gets interesting.
The half that got cut
Go back to what the union asked for. NewsGuild-CWA's statement of support describes a bill delivering job protections and human-in-the-loop requirements — disclosure to workers, not just readers, of when and how AI is used. Those, it said, "specifically align with our demands as a union."
And then, per Investigative Post, the legislation "originally would have strengthened the hands of unions when negotiating terms and conditions involving AI. The language was later removed from the bill." My own read of the enrolled text found the disclosure mandate, the penalties, and nothing else: no human-review clause, no bargaining clause, no mention of wages. R Street, opposing the bill in March, still refers to a "prohibition on using AI to displace workers, reduce hours, or cut wages" — words that appear nowhere in the version that passed. It was reading an earlier draft (the one the unions were promised), which is itself the point.
The shape of it is not in doubt. The part that told you what you were reading survived; the part that would have paid anybody, or protected anybody's job, did not. Disclosure and compensation are not the same problem — and only one of them is killing newsrooms.
The machine that wrote the article I forwarded trained on journalism produced by people paid by organizations now running out of money. That part is measurable. Pew Research found users who met an AI summary clicked a traditional link in 8% of visits, against 15% with no summary. The Columbia Journalism Review put the share of news searches producing no click at all up from 56 percent to nearly 69. Digiday measured publisher referral traffic from Google Search down a median 10% year over year, 7% for news brands. And news executives expect another 43 percent to go inside three years.
Be careful: those numbers show a collapse coinciding with AI summaries, not a controlled experiment proving causation. But they sit on a trend predating the chatbots — Pew found newspaper newsroom jobs fell 57% between 2008 and 2020, from roughly 71,000 to about 31,000. The industry was already hollow when the extraction started.
There is an alternative on the table, and it is not a label. In Editor & Publisher, Anya Schiffrin and Roberta Carlini set out a proposal called statutory licensing, which "would require AI companies to pay publishers for journalism used to train their systems, past and future." Their argument is unsentimental: litigation takes years, and the deals struck by the largest publishers "help some outlets but may do little for the broader news ecosystem." (Name the last county weekly to sign one.) The US Copyright Office came down the other way, calling government intervention premature at this time given "the robust growth of voluntary licensing." Read its next sentence, though: where gaps are "unlikely to be filled," the Office wrote, "alternative approaches such as extended collective licensing should be considered to address any market failure."
Note what both sides have in common: they are arguments about who pays. New York had that argument, then legislated the other thing.
The strongest case against all of this
Now the honest part. The opposition is not a strawman, and it does not come from one direction.
Free-speech groups call this a state-approved disclaimer on the press. The Foundation for Individual Rights and Expression argues the state "cannot claim greater authority over speech merely because AI was used to create content," just as it cannot force newspapers to publish state-approved bylines — and warns the bill lets government "seek court orders restricting the distribution of news content that lacks a state-approved disclaimer." FIRE has aimed that at hundreds of state AI bills. Clay Calvert at the American Enterprise Institute makes the doctrinal version: choosing what goes into a newspaper is the exercise of editorial control and judgment, and government "may not compel a person to speak its own preferred messages." To be clear: no court has ruled on this Act. "Unconstitutional" is an assertion by critics, not a judge.
From the market-liberal side, R Street's Spence Purnell argues the Act would bury newsrooms under compliance mandates and "invite costly First Amendment litigation" — hardest on the outlets least able to absorb it. Diane Kennedy of the New York News Publishers Association put it bluntly: "No newspaper would be safe from investigation under this legislation."
And here is the tension I find hardest to shake. Steven Brill, whose company produced the 3,749 number I opened with, opposes the fix. "We all agree certain content is bad," he told Investigative Post, "but the solution isn't government regulation." When Fahy defended the Act on the grounds that it exempts copyrighted content, Brill answered in six words: "Copyright also protects the bad guys."
Then there is the awkward evidence about whether labels work. A team led by Darden's Luca Cian found in 2022 that readers trusted labeled stories 7 to 14 percentage points less — whether the news was true or false. Taken alone, that makes the Act a machine for disbelieving accurate reporting. But a 2026 study finds not all AI disclosures carry the same cost, with trust declining only for detailed ones. A blunt banner and a precise note about what the machine did are not the same intervention.
Which brings me to the word doing all the work. "Substantially." Editor & Publisher's own reporting calls that threshold "vague, subjective and arguably difficult to fairly enforce or litigate." Is a machine-drafted summary of a council meeting, edited by a human for twenty minutes, substantially created by AI? A court imposes the penalty — but the attorney general is authorized "to take proof and make a determination of the relevant facts and to issue subpoenas," which is why Investigative Post reads the Act as letting the AG "determine what content is substantially created by AI." The answer starts wherever that office says it starts.
Europe drew this line somewhere else entirely — twice
Cross the Atlantic and the same two problems get handled as two problems — the part New York skipped.
On disclosure, the European Union puts half the duty on the machine. Summarizing the transparency obligations in Article 50 of the EU AI Act, the European Commission says providers of systems generating synthetic audio, image, video or text "must ensure that AI-generated or manipulated content are marked in a machine-readable format and detectable as artificially generated." Publishers get a narrower duty: deployers "must clearly label AI-generated or manipulated text published with the purpose of informing the public on matters of public interest." Then the carve-out a newsroom can use: text that "has undergone human review or editorial control" needs no label, provided a person holds "the ultimate legal responsibility over the publication." Those obligations apply from August 2, 2026.
Notice the engineering. New York asks a publisher to write a sentence at the top of a page, then lets a court decide if it was required. Europe asks the model to sign its work, and asks the publisher a question it can answer without a lawyer: did a human own this?
On money, the Nordics spent decades building machinery New York never got round to debating. Extended collective licensing lets an organization representing a substantial share of a country's authors strike one agreement covering a whole category of work — including work by people who never signed up. Denmark's Tekst & Node, the only body its culture ministry has authorized to license text this way, calls it a system that "ensures flexible access to copying for the users as well as compensation to the rightsholders."
The interesting part is what happened when Denmark pointed that machinery at AI training. In a 2023 proposal to amend the Copyright Act, the ministry reasoned that training on scraped work is "a mass exploitation of works involving a large number of rightsholders" — and that since the core of the Nordic model "is to ensure agreement on mass exploitation," so "rightsholders are ensured reasonable payment for the use of their works," the license "is an option that can be used as a tool in this area." Sweden's music society STIM has already gone first, with what it billed as the world's first collective AI license, in September 2025.
That is a structural answer for the freelancer in Buffalo whose reporting sits inside a model's weights — and, awkwardly, the mechanism the US Copyright Office named as its own fallback.
Is it clean? No. A Masaryk University analysis notes the European Parliament passed a resolution on March 10, 2026 urging the Commission to facilitate voluntary collective licensing — but the mechanics remain ugly, producing what its author calls a cumbersome double-opt-out structure. Nobody has solved this. Europe is at least working the right problem.
Now run it forward five years
Just imagine 2031. The label wins — New York, then a dozen states, then a federal harmonizing bill nobody reads — and becomes as meaningful as a cookie banner: a grey line above the headline your eye learns to skip in four weeks.
Meanwhile the market rearranges itself around "substantially." A compliance industry grows up teaching newsrooms how to stay under the bar: draft with the machine, have a human rewrite fourteen percent, document it, no label required. Technically accurate, completely uninformative. The content farms comply with nothing — they never have. And enforcement runs through an attorney general who will not always pick your fights.
Underneath it all, the actual thing continues. The regional paper that covered your county courthouse closes. Its twenty-year archive is sold, or scraped, or left online for something that never gets tired. Two years later a model answers a question about your county's courts with unnerving accuracy — sourced from reporting by people who no longer have jobs, credited to nobody, paid for by no one. The answer carries a label. The label says it was made by AI. Everything the label says is true.
That is what we build if disclosure is the whole plan. The better branch is not utopian: provenance traveling with content by default, disclosure designed by people who read the trust research, and some mechanism putting a small, reliable payment into the newsrooms whose work is in the training data. Not a windfall. A subscription the machine cannot cancel.
What the smart people are saying
The most useful analysis I have read this year is not about labels but about what happens when you try to make AI companies pay.
Courtney Radsch and Karina Montoya, writing for Brookings, mapped the AI content-licensing market and found local newspapers and regional broadcasters "effectively absent from the AI licensing market entirely." Their sharpest line: "The same Big Tech firms whose AI products are eroding website traffic are now building and controlling the licensing infrastructure those publishers must turn to." Their underlying report for the Open Markets Institute warns that ad hoc deals are repeating the same mistakes as the social media era.
The Electronic Frontier Foundation comes from the opposite direction, aimed at my own thesis. Tori Noble argues a training-license right is unlikely to protect the jobs or incomes of the working creators that media conglomerates have underpaid for decades, and that "pricey licensing deals offer a way to lock in their dominant positions" for firms that can afford them. Sit with that. Money paid to publishers is not money paid to journalists, and a carelessly designed licensing regime is a moat with a press release attached.
From the free-market right, Calvert's earlier work argues private dealmaking avoids courts and legislatures alike. The Reuters Institute added the cold water in early 2025: there is still no court ruling that the training was unlawful.
Left, right and libertarian, they converge on one thing: the fight that decides whether journalism survives is the fight about money, and almost nobody is legislating it.
What does this mean for you?
You will not fix the licensing market from your kitchen. These, though, compound.
Stop treating the absence of a label as evidence. Even enforced, the Act covers only content "substantially" AI-created, for outlets operating in New York. No label is not a certificate of human authorship.
Check whether a human is reachable. My four-second test: search the byline. If the name exists nowhere else, if the "about" page has no address or phone, if no story quotes anyone who had to be persuaded to talk — treat it as unattributed.
Pay one outlet directly, and make it a small one. A local subscription does more for the newsroom covering your county than any amount of sharing.
Share the source, not the summary. Every screenshot of an AI answer is a click that arrives nowhere. Pasting the original link costs nothing.
Ask your own outlets what their AI policy is. Reputable newsrooms publish one. If yours has not, ask.
Watch the second bill, not the first. Disclosure laws will keep passing because they are cheap and popular. Ask anyone wanting your vote: what is your plan to make the people who trained on this journalism pay for it?
The lesson, as I see it
I have no objection to the label; I would like the label. Had one been on the article I forwarded, I would not have forwarded it — and would have been spared a small, useful embarrassment.
But let us be precise about what the FAIR News Act is: a transparency law passed 53–7 and 130–1 without a hearing, aimed at a real problem, using the one instrument that costs the state nothing and the industry very little. The version that would have given workers a seat at the table did not survive drafting. That is not an accident. That is a revealed preference.
Disclosure tells you what you are reading. It does not tell you who is left to write it.
My vote? Take the label — then refuse to let anyone call it the answer. A thousand dollars is what New York thinks an unlabeled machine-written article is worth. Nobody in Albany has yet said what the reporting that trained the machine is worth. Until somebody does, we are decorating the problem rather than fixing it.
If one thing you read this week made you pause before hitting forward, let it be this one — then send it to the person in your life who forwards the most. HAIA Foundation runs on exactly that: one careful reader handing the argument to the next. More of it, weekly, at the Substack.