Hank Green sounded like ChatGPT. There was just one problem.

On July 29, Hank Green published a 64-minute conversation with the creator Soupytime. Two days later, a clipped moment had been repackaged as a confession.

Green looks into the camera and says four words that millions of people have now been trained to hear in a particular voice:

“I appreciate the pushback.”

ChatGPT says that. Claude says that. The phrase has become one of those little linguistic fingerprints people collect online, beside “it’s worth noting” and “let’s unpack this”. The accusation was immediate: Green had left chatbot feedback in his script and read it aloud.

There was one problem. Watch the original video from 43:00, not the amputated clip.

Green has just claimed that words are human inventions. Soupytime pushes back. She asks whether some sounds might carry meaning across languages and brings up the bouba/kiki effect. Green does not know enough to answer, so he turns to “Future Hank”. In the later insert, after working through the question, he revises his claim. At 45:22 he says: “A better way to say it, and I appreciate the pushback, is a word is a human invention...”

Someone had pushed back. He appreciated it. This is normally how sentences work.

The supposed AI tell was human

Green checked his notes and script. In a response posted to X and later deleted, he said the phrase appeared in neither. It was an ad-lib referring to his guest’s disagreement. The surviving status ID dates that response to July 31; contemporaneous Reddit threads preserve the full text.

That explanation is not independently auditable unless Green publishes his working files. It is, however, consistent with the uncut video. The phrase makes exact sense in context, and the transcript shows the disagreement two minutes earlier.

This should have ended a very silly case. Instead it opened a much more interesting one, because Green really had used ChatGPT.

He said he used it for research and had been “relying too much on generated notes”. In a longer response to his own community, he described using AI to locate papers and other resources quickly, then reading them and forming his own conclusions. He did not say ChatGPT wrote the script. He did say something more painful: the process stopped him finding his own “ways into and around a topic”. People accusing him of “diluting” himself, he wrote, were “spot on”.

The technicality therefore saves neither side. The sentence presented as proof was apparently human. The AI assistance was real. The disputed question is what that assistance did to the work.

What Hank thought he had done wrong

Green’s apology was not an admission that touching a chatbot is inherently immoral. It was an admission about trust, craft and compulsion.

He said models’ training practices disturbed him, concentration of economic power frightened him and climate impact worried him. More personally, he described the dopamine of interacting with LLMs as unhealthy and careless. “Making more things does not make me make better things,” he wrote. He paused or reconsidered several projects and said he might add human research or writing support.

On August 7 he published a five-minute follow-up and a personal policy. No portion of a script will be written, edited or outlined by an LLM. A video’s thesis will originate with a human. No AI-generated image or music will appear. After reading comments, he added that LLM output will never be trusted as a source and informational videos will link to primary sources.

Notice what the policy does not say: “I will never use AI.” It draws boundaries around authorship, evidence and audience expectations.

That is also why some criticism persisted. A system can bias research before it invents a fact. It can steer which papers you see, compress away qualifications, or hand you a seductive frame that feels like your own. Opening the cited paper is necessary. It is not always sufficient. Green’s own account of losing his route into a subject is the strongest version of that criticism.

Why four words became an accusation

The angriest reactions are easy to screenshot and easy to dismiss. That would be cheap.

In the main fan discussion, one person wrote that they wanted “Ask Hank Anything, not ask ChatGPT anything”. Others said an overworked creator should publish less or pay a researcher, not use a system built from uncompensated creative work. Some worried that a science communicator had placed an unreliable intermediary between himself and primary research. Some simply felt deceived because the assistance had not been disclosed. A few said they would never watch him again because he had used ChatGPT at all.

Those are not one argument. They are at least seven:

• Labour: employers can use generative systems to avoid paying writers, artists, translators, actors and researchers.
• Copyright and consent: model developers trained on enormous collections of human work without obtaining individual permission from every creator.
• Reliability: fluent systems fabricate sources, flatten disputes and conceal uncertainty.
• Authenticity: people watch Hank Green for Hank Green’s selection, thought and voice.
• Disclosure: an audience may accept assistance it would resent discovering later.
• Environment: training and running models consumes electricity, water and hardware.
• Slop: cheap generation lets people flood the internet with material nobody cared enough to check.

Several are plainly right.

The US Copyright Office’s 2025 report on generative-AI training does not bless all training as fair use or condemn it all as infringement. It says the answer depends on facts including the works used, the purpose and the market effect, and argues that licensing markets should continue to develop. Courts are still deciding major disputes. Anyone selling AI as ethically settled is selling confidence they have not earned.

The environmental cost is not imaginary either. The International Energy Agency estimates that data centres used about 415 terawatt-hours of electricity in 2024, around 1.5% of the global total, and projects roughly 945 TWh by 2030 in its base case. AI is not the whole data-centre sector, but accelerated servers driven mainly by AI account for almost half of the projected increase. Efficiency per task is improving; total demand is still rising.

Water is harder to attribute. It can be consumed directly for cooling, indirectly in electricity generation and during chip manufacturing; location and cooling design change the result. A 2025 peer-reviewed assessment says operators' reports generally do not separate AI from non-AI workloads. That opacity is a reason to demand better reporting, not a licence to repeat one sensational “bottles per prompt” number as universal.

Nor should “exposure” be misreported as jobs already lost. The International Labour Organization’s 2025 global index found that one in four jobs sits in an occupation with some generative-AI exposure, rising to 34% in high-income countries. Its conclusion was transformation more often than full replacement. In 2026 the ILO warned again that exposure indicators are not forecasts of layoffs, productivity or reskilling. The fear is still rational when a worker experiences AI not as “a tool that helps me” but “the thing my employer wants instead of paying me”.

AI companies helped create this backlash. Scraping first and negotiating later, grandiose replacement rhetoric, opaque data, hallucinated answers delivered in a confident tone, and platforms rewarding synthetic spam are not misunderstandings invented by teenagers. They are choices.

Young people do not simply hate AI

The reaction around Green can look like a generational revolt. The data is stranger.

Pew Research Center surveyed 5,119 US adults in February 2026. Sixty-six per cent of 18-to-29-year-olds had used an AI chatbot, compared with 23% of people 65 and older. Yet 48% of the youngest group expected AI to have a negative effect on society over the next 20 years, the highest share of any age group. Only 14% expected a positive effect.

Young adults were also almost evenly divided on whether chatbots help or hurt their creativity: 25% said help, 20% hurt. Older people were less negative, in part because far more of them did not use the tools or were unsure.

That is not hypocrisy. It can be experience. Heavy users see the utility and the failure modes. A student may happily use a chatbot to explain a formula and hate a generated novel sold as someone’s art. A designer may automate file naming and oppose a client replacing the illustration budget with Midjourney. Private assistance and public substitution are different social acts.

Gallup has found the same tension in Gen Z: adoption remains high while scepticism rises. Pew’s 2026 teen survey likewise shows widespread use alongside anxiety about cheating, learning and the future. “Young people love AI” and “young people hate AI” can both be manufactured from a subset of the evidence. The useful question is what kind of use they are judging.

AI slop is real, and I hate it

I love AI. I also hate AI slop. This is no more contradictory than loving cameras and hating spam photographs.

“Slop” became Merriam-Webster’s 2025 word of the year, defined as low-quality digital content usually produced in quantity with AI. The term caught on because it names a business model, not an aesthetic preference: reduce the cost of making something close to zero, optimise it for a recommendation system, then make so much that even tiny returns add up.

Renée DiResta and Josh Goldstein documented Facebook pages posting more than 50 synthetic images apiece. They found fake children, cabins, cakes, carvings and religious images, often with copied engagement-bait captions. One AI image of a crab-shaped Jesus received 209,000 reactions. A synthetic kitchen post appeared among Facebook’s ten most-viewed posts in the third quarter of 2023 with 40 million views. The researchers are explicit about the limits of their hand-built sample; it demonstrates a large abuse pattern, not the percentage of Facebook that is synthetic.

The same economics produces fake disaster images, automated SEO articles, invented product reviews, synthetic comments, cloned voices, fake trailers, machine-written books and channels assembled without anyone asking whether the finished thing is true or worth a stranger’s time.

Slop is not bad because a machine touched it. It is bad because the publisher stopped caring.

Nobody checked the hands. Nobody opened the citation. Nobody removed the paragraph that says the same thing for the fourth time. Nobody asked why the disaster photo shows a road dissolving into a river. A person typed a prompt, received an answer and converted their indifference into somebody else’s feed.

Congratulations: they automated being lazy.

Human language is becoming contraband

The Green episode reveals a second, weirder cost. Humans taught language models how to write. Models overuse parts of that language. Humans then treat those parts as machine property.

There is evidence for genuine statistical tics. Researchers analysing more than 15 million PubMed abstracts found an abrupt post-ChatGPT rise in words associated with LLM assistance and estimated that at least 13.5% of 2024 abstracts had been processed with an LLM. Separate work has examined why ChatGPT over-represents words such as “delve”, “intricate” and “underscore”.

That does not turn one word into proof. Style is evidence in aggregates, not a barcode attached to “delve”, an em dash or a polite acknowledgement.

Trained humans can sometimes do better than chance. An ACL 2025 study found that frequent ChatGPT users were comparatively accurate and robust at identifying output. But accuracy in a controlled experiment does not make a hunch sufficient to accuse a particular person. The uncut source still matters.

Automated detectors are even more dangerous when treated as verdict machines. In a widely cited study, seven detectors labelled more than half of 91 essays by non-native English speakers as AI-generated, while almost all native-speaker essays were classified as human. A 2025 review of 50 studies found results varied sharply by detector, model, genre and editing; hybrid human-AI text was especially difficult.

This produces a perverse incentive. Students and professionals report avoiding polished transitions, em dashes and ordinary phrases because they fear a machine will accuse them of being a machine. If humans deliberately make their writing worse to prove it is human, the detector has not protected authenticity. It has become an editor with terrible taste and disciplinary power.

The safe rule is simple: a detector score can be a lead for a fair conversation about process. It cannot prove authorship by itself.

“Used AI” is an information-free label

Consider eight versions of a Hank Green video.

Collapsing all eight into “he used AI” destroys the information needed to judge them.

Existing disclosure frameworks already move toward levels. A proposed four-level framework for scientific writing distinguishes basic technical help, language enhancement, substantial content involvement and primary material generation. Academic and industry policies disagree at the edges, but they increasingly ask for the tool, purpose and human oversight when involvement is material.

For this article I made a seven-step editorial scale, from utility to slop automation, with zero for no generative AI. It is not science. It is a vocabulary aid. Its most important feature is that responsibility never leaves the human column.

What my AI workflow actually looks like

Nowpinion uses AI openly. AIM — the Agile Iteration Method — exists specifically to make AI-assisted work structured, reviewable and harder to wave through.

A typical Nowpinion article can take around 90 minutes. This one took longer because the controversy was still unfolding, the original clip had to be reconstructed, and every image needed a rights trail. The workflow was not “write an article defending AI”. It looked like this:

1. I chose the question. A model did not decide that four words were culturally interesting. 2. AIM turned the idea into tests. What exactly preceded the phrase? Was it in the notes? What did Green admit? What evidence would weaken my preferred thesis? 3. AI accelerated discovery. It helped locate original videos, policies, surveys, papers, court records and counterarguments. 4. I opened the sources. The video transcript was checked against the actual recording. Survey figures came from Pew’s tables. The hero image came from Wikimedia Commons with a CC BY 4.0 licence. 5. I hunted contradictions. Young people turned out to be both high-use and unusually pessimistic. The research-assistant defence turned out to have a serious framing problem that Green himself articulated better than his critics. 6. AI helped organise and draft. I treated its prose as material, not authority. 7. I removed machine habits. Generic openings, fake symmetry, repeated conclusions, hedging without purpose and corporate cheerleading were rewritten or deleted. 8. The claims were checked again. Dates, percentages, quotations, links and captions were tested. Unknowns remained labelled unknown. 9. A separate review challenged the package. The article, image rights, embeds, mobile layout and publication state all have to pass before this can go live.

The assistance is material, so the disclosure is material. I used AI for research planning, source discovery, transcription inspection, counterargument search, structure, drafting and editing. I personally opened the cited sources, chose the thesis, revised the language and accept responsibility for every published claim.

That last sentence is not ceremonial. If a cited paper does not say what I claim, that is my error. If a quotation is invented, that is my error. If the argument is shallow, the chatbot cannot be embarrassed on my behalf.

AI should raise the bar

The most depressing definition of productivity is “the same mediocrity, but six times as much”. It is also the definition most likely to fill the internet.

There is a better loop. If AI saves 20 minutes, spend the 20 minutes opening another source. If it can scan a 200-page report, use that ability to inspect evidence you would previously have missed. If it offers ten explanations, reject nine. If it can argue against you, ask it to find the fact that would ruin your headline.

This does not eliminate cognitive risk. Summaries can anchor the reader. Search tools can create a biased evidence set. Fluent drafts can make weak reasoning feel complete. Serious use therefore needs more process, not less: source logs, adversarial questions, uncertainty labels, rights records and a human who can say no.

Historical comparisons help only up to a point. Calculators changed what schools considered legitimate arithmetic. Spellcheck changed proofreading. Photoshop changed photography. Google and Wikipedia changed research. Stack Overflow and autocomplete changed programming. In every case society renegotiated the line between assistance and misrepresentation.

Generative AI is not merely the next calculator. It can produce complete creative works, imitate living styles, manufacture persuasive falsehoods and automate parts of knowledge labour at enormous scale. The new harms are real. The historical lesson is not that critics always lose. It is that “a tool was involved” has never been enough to decide whether a practice is honest.

Material disclosure, not a confession booth

Does an audience need a warning that spellcheck fixed a typo? Probably not. That fact would not change a reasonable reader’s evaluation of the work.

What if AI selected the sources, summarised contested research, proposed the thesis, rewrote half the paragraphs or generated the entire illustration? Now the involvement could change how a reader judges independence, authenticity, copyright, accuracy or craft. Disclose it.

The Associated Press’s July 2026 standards permit early-stage research, document summarisation, transcription, translation and headline suggestions under human review, prohibit generative alteration of news photography, and require disclosure when AI plays a material role. YouTube’s policy similarly requires labels for realistic, meaningfully synthetic or altered content, not for every inconsequential production aid. Medical-journal guidance requires authors to explain AI assistance and states the underlying principle plainly: a chatbot cannot be an author because it cannot take responsibility for accuracy, integrity and originality.

That gives us a better rule than universal labelling:

Disclose AI involvement that would reasonably change how the audience evaluates the work.

The standard is contextual. Translation help may be routine in one article and central in another. AI source discovery matters more in a science explainer than in a personal diary. A synthetic voice matters differently in an animation than in an apparent eyewitness recording. Transparency should communicate, not merely inoculate the publisher with a tiny “AI was used” label.

Who owns the mistake?

In 2023, lawyers in Mata v. Avianca submitted invented cases produced by ChatGPT. The federal judge sanctioned them. The defence was not “the AI made a mistake”. The court dealt with the humans who filed the papers.

The same rule scales.

• A developer owns AI-generated code merged into production.
• A scientist owns a fabricated citation left in a paper.
• A manager owns a decision dressed up by a confident summary.
• A journalist owns a false sentence published under their name.
• A creator owns the gap between what the audience was promised and what it received.

AI can assist. Humans remain responsible.

This is stricter than both popular extremes. It rejects the maximalist fantasy that automation transfers blame to the tool. It also rejects the purity test that any generative assistance invalidates the work. The quality test is more demanding: Is it accurate? Sourced? Original? Appropriate to the tool’s weaknesses? Transparently described when material? Is there a human who owns the result?

And then the question every framework eventually has to face: is it good?

Back to Hank

The simple internet version was excellent content. Hank Green sounded like ChatGPT. He was exposed. His audience revolted. He apologised.

Almost every clause breaks under inspection.

The “AI phrase” was an apparently spontaneous response to an actual disagreement. Green nevertheless had been using ChatGPT heavily in research. Critics were wrong about the smoking gun and right that a creator’s process can change the thing they came to receive. Some collapsed all assistance into cheating. Green accepted a subtler failure: speed and generated notes had diluted the route by which he became himself on camera. His new policy preserves some machine assistance while putting human authorship, sources and disclosure into writing.

Everyone can be partly right. That is why the story matters.

I use AI constantly. I use it to research, code, analyse, challenge ideas and write Nowpinion. I built AIM to make that work more demanding. I am not embarrassed by any of this.

I am also not interested in publishing whatever a chatbot gives me. If it finds ten sources, I can inspect ten sources. If it challenges my argument, good. If it writes a sentence that sounds like committee-approved oatmeal, I delete it. If it gets something wrong and I publish it anyway, that is my fault.

Not the AI’s. Mine.

The machine is not the problem I am defending. Careless people, negligent companies and business models that reward infinite garbage are.

AI is an extraordinary tool. Slop is awful. The missing middle is simply people who use the first because they care too much to make the second.