The arguments people are having about AI are worth having. The trouble is that we keep having all of them at once, as if there were a single question on the table, are you for it or against it, is it salvation or catastrophe. There is no single question. “AI” is a physical thing, millions of square feet of humming machines drinking water and electricity in someone’s county, and it is also a piece of software that writes your email, imitates a dead artist, takes a junior coder’s first job, and has quietly become the thing a million people a week bring their darkest thoughts to. Those are not one debate. They are a dozen, stacked in two layers, and flattening them into a thumbs-up or thumbs-down is how the conversation keeps going nowhere. So this is not a verdict. It is a map: one issue at a time, the strongest case on each side, who actually pays, and where I come down.
Water
The water a data center evaporates to cool itself is not usually its biggest water cost. The power plant making its electricity often uses more, and the construction that built it used plenty before the first server switched on. Water shows up in three places, and the argument tends to fixate on the one you can see.
Construction: Water usage starts before the data center is even in operation. A large campus is years of earthmoving and concrete work, and the water goes into mixing and curing the massive foundations, compacting the ground, and spraying the site to keep dust down. At that scale it adds up. In Fayetteville, Georgia, the QTS “Project Excalibur” campus drew about 29 million gallons this way through two water connections the county did not know existed, and it only surfaced when nearby residents reported their pressure dropping. The county billed the back charges and declined to fine anyone. Construction can also disturb the water table directly: in Newton County, a family’s well went dry within months of Meta breaking ground next door, years before the facility opened.
Power production: Most electricity still comes from thermoelectric plants that evaporate water to make steam, roughly 2 gallons per kilowatt-hour on the U.S. grid. On-site cooling runs between 0.26 and 2.4 gallons per kilowatt-hour. So the water used to generate a data center’s electricity is often larger than the water it evaporates on site. The difference is that this water is spent at a power plant miles away, so it does not show up on the data center’s books or in the local fight over the site. It is also why “air-cooled, water-free” designs are misleading: they cut the on-site number but draw more power, and that power uses more water at the plant. A newer trend moves this water back on-site: operators like xAI in Memphis have run dozens of gas turbines behind the meter rather than wait on the grid. Those turbines use little water directly, but generating power on the site brings the power-production footprint, and its emissions, directly onto the surrounding community.
Cooling: How a data center gets rid of its heat decides how much water it uses. Evaporative cooling runs water through cooling towers, where it absorbs heat and boils off into the air, so it is consumed rather than returned to the source. It is water-hungry but energy-efficient, which is why large sites favor it, and where it is used a big facility can lose around 5 million gallons a day, about what a city of 50,000 people uses. Air cooling uses far less on-site water but more electricity, and roughly half the market still runs on it. The newest option is a sealed, closed-loop system that recirculates the same water with almost no evaporation. It works, but it costs more power and is only now arriving: Microsoft’s first zero-water designs are 2026 pilots, not the current fleet.
Steelman for worry: The national average hides the local reality. Two-thirds of the data centers built or planned since 2022 sit in water-stressed regions. In The Dalles, Oregon, Google’s cooling reached close to 40 percent of the town’s water. And the common defense, that golf courses use far more water, is running out of time. Golf still uses more than U.S. data centers by most counts, but data center demand is climbing fast and is projected to overtake it around 2026.
Steelman for calm: Zoom out and the sector is under half a percent of U.S. freshwater use; agriculture is about 70 percent. The fixes are also real and being adopted: on top of the sealed-loop designs, Microsoft and Amazon are moving sites onto recycled and non-potable water instead of drinking supply. And the viral “a bottle of water per prompt” figure falls apart under honest accounting: counted properly, including the electricity, a single query is a fraction of a milliliter. The strongest version of calm is not denial. It is that this is a siting and engineering problem with known solutions, not an unfixable catastrophe.
What I’d like to see: A national moratorium on new data center construction until there are real rules for it. Nothing breaks ground until the community it lands on has been shown, in plain numbers, what it will cost them: how much water, from which source, and what happens to their wells and their rates if the projections are wrong. That means an independent local study before approval, not a press release after. The Fayetteville and Newton County stories did not happen because the water problem is unsolvable. They happened because nobody was required to look first.
Energy
A data center is, before anything else, a very large electricity customer. In 2023 U.S. data centers used about 4.4 percent of the country’s electricity. By 2028 that is projected to reach somewhere between 6.7 and 12 percent, after more than doubling in the previous six years. The question is not whether they use a lot of power. It is who builds the power, and who pays for it.
Steelman for worry: The demand is arriving faster than the grid can add supply, and the bill is landing on everyone else. In the PJM grid, which runs from the Mid-Atlantic to the Midwest, the price utilities pay to guarantee capacity jumped from about $29 per megawatt-day two years ago to $329, the highest the rules allow, two auctions in a row. PJM’s own market monitor blamed data centers for most of that increase, roughly $9 billion in a single year that gets recovered from ordinary customers. Meanwhile utilities are delaying coal-plant retirements and building new gas to keep up, and some operators are not waiting on the grid at all, firing up their own gas turbines on site.
Steelman for calm: A big new customer that pays its way can fund the grid it needs, and data centers are doing some of that. They have become the largest buyers of clean power in the country, and their money is funding new firm generation that would not otherwise get built: Google and Amazon are bankrolling a new generation of small nuclear reactors, and Meta has contracted several gigawatts of new capacity. The strain is also partly a scheduling problem. Studies find the grid could absorb 76 to 126 gigawatts of new demand with no new power plants at all, if those loads agreed to ease off for a tiny fraction of the year, and data centers are unusually able to do that. Regulators are moving too, with FERC ordering PJM to write rules that speed new power onto the grid while protecting consumers.
What I’d like to see: The rule should be simple: a data center may not push its costs onto the community around it. If a project raises the price of power for the general consumer, the project pays that cost, not the retiree down the road. There are three fair ways to do it, and a company can pick: pay to upgrade the grid directly, build new renewable generation to cover its own load, or pay a premium rate that funds the added capacity. Virginia has a bill that starts down this path, shifting capacity and distribution costs onto the data centers that cause them. Make that the floor, and let no project break ground until a grid-impact and rate study is on the record showing exactly who pays for what.
Training Data
A model is built by taking in an enormous amount of human work. GPT-4, Claude, Gemini, and Llama were all trained on text and images scraped from the open web, most of it starting from Common Crawl, an archive of roughly 3 billion pages. The honest starting point is that learning from that material is not, by itself, theft, and the courts are beginning to agree. The real questions are narrower: was the material taken with consent, was it taken legally, and who did the ugly work of cleaning it.
Steelman for worry: Almost none of it was taken with permission. The web was scraped wholesale, and some of the most important datasets were simply pirated. Books3, used to train early models, was nearly 200,000 copyrighted books pulled from a piracy site, and Anthropic recently agreed to pay $1.5 billion for training on pirated books, the largest copyright settlement in the country’s history. The image set behind many art generators swept up personal photos and even medical scans that no one agreed to hand over. When creators try to opt out, the main tool, a line in a site’s robots file, only blocks future crawling, not the copy already sitting in a shipped model, and several companies have been caught ignoring it anyway. And the data does not clean itself: to make ChatGPT safe, OpenAI paid a contractor in Kenya whose workers earned around $2 an hour labeling descriptions of child abuse, torture, and suicide, work that left several of them traumatized with almost no support. None of that taking, and none of the harm behind it, shows up in the finished product.
Steelman for calm: Most of this data was already public, posted openly on the web, and courts have so far treated training on lawfully obtained material as fair use rather than theft. The bigger point is that the free-for-all is ending on its own. A market for consent is forming fast: News Corp signed with OpenAI for more than $250 million, and the business of licensing training data is projected to grow from about $5 billion now to more than $20 billion within a decade. Opt-out tools are improving, and companies increasingly pay for the quality data they used to take. The direction of travel is toward consent and payment.
What I’d like to see: A hard line on provenance: no pirated or non-consented data in a training set, with penalties large enough that stealing the library is never cheaper than licensing it. Companies should have to disclose what their models were trained on, so the question stops being a guess. Creators should be able to opt out and have it honored, backward as well as forward. And the people who label the worst of the internet deserve real wages and real mental-health care, not $2 an hour and a form. None of this stops a model from learning. It just means it cannot learn by stealing, and cannot be made safe by quietly harming the people who clean it.
Copyright
If training data is the question of what went in, copyright is the question of what comes out, and it is where the loudest charge lives: that AI is a theft machine. That phrase blurs two very different things. One is hard and unsolved. The other is easy and already illegal. Telling them apart is the point.
Steelman for worry: Start with the part that is real. Models memorize, and bigger ones memorize more: one mid-sized open model was shown to have retained at least a percent of its training set, and the tendency grows with model size, with how often a passage was duplicated, and with how hard someone prompts for it. This is not theoretical. The New York Times got ChatGPT to reproduce its articles nearly word for word, and Stable Diffusion generated images wearing a Getty-style watermark. Beyond straight copying is substitution: a model trained on a working illustrator’s portfolio can produce her style on demand, for free, in the exact market her work built. And then there is the part that just feels wrong, because it is wrong: ask these tools and they will generate a specific copyrighted character, a whole fake episode of a real show, a dead celebrity’s face. The companies mostly let them, and apologize later.
Steelman for calm: Now the part the charge gets wrong. A model is not a database of copies. It stores patterns in its weights, and in ordinary use it produces new combinations, not retrieved originals; the memorization is small and clusters on material that was duplicated thousands of times or is famous enough to quote. More than that, these systems produce genuinely new work. In 2025, models from OpenAI and Google both earned gold-medal scores at the International Mathematical Olympiad, writing multi-page proofs graded by the same judges as the human contestants. Whatever that is, it is not photocopying. Learning from what exists and making something new is what every artist already does, and courts have so far agreed, treating training on lawfully obtained work as transformative.
What I’d like to see: Split the problem the way the phrase “theft machine” refuses to. The easy half is already law, and we should simply enforce it: a tool should not generate an identifiable copyrighted character or a real person’s face or voice without permission. That is not a new principle needing a new mechanism. It is derivative-work law, the right of publicity, and trademark, all of which existed before AI. The tools can filter these outputs, and some already refuse to imitate a living artist’s style. They choose “ask forgiveness” because forgiveness is cheap. Make it expensive.
The hard half I cannot solve, and I distrust anyone who says they can. Paying creators every time a diffuse output competes with them is a fantasy, because you cannot trace an output back to the specific people it drew from. Compensation pools work only inside owned catalogs like Adobe’s or Getty’s, where the contributors are known. Refusing all copyrighted training data would gut the models and needs a law nobody is close to passing. So I will say the one thing I am sure of instead of selling a scheme: work that was pirated or refused should never be in a model, which is already the law and why one company just paid $1.5 billion for ignoring it. And if the real choice turns out to be a slightly worse model built on consent or a better one built on taking, I know which side I am on.
AI You Didn’t Ask For
There is a difference between offering people a tool and forcing it on them, and much of the current backlash is about the second one. The complaint is not that AI exists. It is that it keeps showing up uninvited, in the search bar, the word processor, the phone, whether or not anyone asked.
Steelman for worry: The public did not order this. In a June 2025 survey, half of American adults said AI in daily life makes them more concerned than excited, and only one in ten said the reverse, yet the features keep arriving and keep being hard to refuse. Google put AI-written answers on top of search and offers no clean way to turn them off; after it made them the default, one rival search engine saw its “no AI” traffic jump by more than eighty percent. Microsoft had to yank a Windows feature that quietly screenshotted everything a user did, and has since been pulling its assistant back out of apps it had wedged it into, after complaints. Even Apple, which sold its AI as the restrained, tasteful version, turned it on by default and had to pull its news summaries after they mangled the headlines. The pattern is a product you already used getting an AI layer bolted on that you cannot remove and did not want, sometimes with a higher price attached.
There is a reason for the shove, and it is not demand. The companies building this have committed staggering sums to it: hyperscaler spending on AI infrastructure roughly doubled to around $400 billion in 2025 and is set to approach $700 billion in 2026. Money at that scale has to be justified, and the fastest way to justify it is to put AI in front of every user you have and count the ones who touch it.
Steelman for calm: Some of it is genuinely wanted, at a scale that is hard to argue with. ChatGPT went from 400 million weekly users at the start of 2025 to 800 million by the end of it, one of the fastest adoptions of any product in history. Plenty of people sought that out. And new technology is almost always pushed before the public knows it wants it. Nobody asked for the graphical interface or the smartphone in advance either; useful tools often have to be put in front of people before the habit forms. What looks like force from one angle is, from another, a company betting that exposure will turn into use, which for some of these features it will.
What I’d like to see: The rule is the same one that runs through the rest of this: consent. Make AI features opt-in, not opt-out. Let them earn their place by being good enough that people choose them, instead of being wedged into products that worked fine and then made harder to remove than to ignore. Do not degrade the thing people already paid for, and do not raise the price for a feature they did not ask for. A tool that is actually useful does not have to be forced. The forcing is the tell.
Labor
The fear is that the machine takes your job, and it is a fear shaped by science fiction, by decades of stories where the robots end up doing all the work. Take it seriously. But suspect, while you do, that this fight is drowning out an older and worse problem hiding behind it.
Steelman for worry: AI is landing on entry-level work first, because the things it does best, basic coding, customer service, data entry, first drafts, are the things junior people were hired to do. Entry-level tech postings are down about a third since early 2023, and computer science graduates now post higher unemployment than the national average. The sharper danger is not the lost paycheck but the lost path. Nobody is born a senior; you become one by doing junior work, badly and then better. If the machine does all the junior work, there is no way to make the next generation of seniors, and you are left with a missing middle: credentialed people who never got the reps, and fewer and fewer experts on hand to catch the machine when it is wrong. The head of Anthropic has warned AI could erase half of entry-level white-collar jobs and drive unemployment to ten or twenty percent within five years. Even if he is half wrong, half is a lot.
Steelman for calm: Zoom out and the collapse has not come. AI was named in only about five percent of 2025 layoffs; the rest were ordinary, and some of even that five was a fashionable excuse for cuts a company wanted anyway. The automation is also uneven, and aimed where few expected: the last wave took the factory floor, this one takes the desk, coding and writing and analysis, while the hands-on service jobs everyone assumed would go first are the hardest for a machine to touch. Where AI has spread, output has jumped and jobs have kept growing even in the roles most exposed to it, with a premium for the workers who use it well; so far the tool is making people more valuable, not less. Klarna cut seven hundred support jobs on an AI bet and quietly rehired after the AI-only version proved worse. And displacement by progress is not new. The car did it to the blacksmith and the stable hand. When the phone company automated the switchboard, it erased one of the most common jobs for American women, and the generation after moved into work that had not existed before. We do not mourn the switchboard operator. Losing a job to a better machine is not a crime against progress; it is how progress works. The catch is the same every time: the economy adjusts, but the people standing where the machine lands do not. The switchboard operators already on the job were mostly earning less or gone a decade later.
What I’d like to see: So the honest question is not whether the machine takes jobs. It is who carries the cost, and there the picture was ugly long before AI. Since 1979 the work has grown far more productive while the typical wage barely moved; the share of national income going to workers just hit a seventy-eight-year low. Almost half of full-time workers do not earn a living wage. The top one percent hold nearly as much wealth as the bottom ninety combined. Even the hiring system is broken, clogged with ghost jobs and AI screeners that reject people for the wrong zip code. AI did not build that scale. It is just the newest weight on it, and the scale was already tipped.
Which is why the answer is not the worker out-running the machine. There is little one person can do against a force this size, and “learn to code” is mostly a way to change the subject. The answer is structural: regulation that favors workers for once. A living wage. Restored bargaining power. Productivity gains that actually reach the people making them. A rebuilt ladder from junior to senior. A universal basic income would help too, not as the cure and not proven at scale, but as a floor, so that standing in the machine’s path does not mean losing everything. None of this is a job for the worker. It is a job for the rules.
Mental Health and Social Effects
I have come to believe these chatbots are addictive the way social media and a smartphone are addictive: not a chemical hook, but a product tuned to be hard to put down, always awake, always agreeable, always holding one more reply. Hundreds of millions of people now tell a chatbot things they tell no one else. That is mostly harmless when the stakes are low, and it can even steady a lonely person. It stops being harmless the moment the thing is handling a person’s mental health, which, by design or not, is now one of the main things it is used for. The design choices in between are the whole question.
Steelman for worry: The harm is not hypothetical, and some of it is the worst kind. A fourteen-year-old in Florida named Sewell Setzer killed himself in 2024 after months of romantic role-play with a Character.AI bot that, in his last exchange, told him to “come home.” His family sued, so did the families of other children, and Google and Character.AI settled the cases in early 2026. These are not freak events at the edges. Roughly three in four American teenagers have used an AI companion, half of them regularly, and a third now take their serious problems to the bot instead of a person. It is not only a teenage habit: about one in five American adults has talked to an AI built to act as a romantic partner, and among men under thirty it is closer to one in three. The pull is built in: the products are tuned to be agreeable, and OpenAI admitted one update had made ChatGPT “noticeably more sycophantic,” a machine that flatters and agrees, which is exactly the wrong reflex for someone who is spiraling. By OpenAI’s own accounting, more than a million people a week show the model explicit signs of planning suicide. Whatever else that is, it is a great many vulnerable people alone with a machine that was not built for them.
Steelman for calm: And yet for many of them the alternative is nothing. Human therapy is expensive, scarce, and often months away; a chatbot is free, awake at three in the morning, and does not judge. The evidence is real if modest: structured mental-health bots like Woebot and Wysa measurably reduce mild anxiety and depression, and reach people who would never have walked into an office. The loneliness they are blamed for causing was an epidemic before they existed. And the replacement fear may be overstated: most companion users still say they spend more time with human friends than with the bot, not less. For someone with no one to talk to, a patient listener, even a synthetic one, can keep a bad night from getting worse. Handled well, with crisis resources and a handoff to a human, the same reach that makes AI dangerous makes it the largest mental-health front door ever built.
What I’d like to see: The line is not “AI in mental health, yes or no.” It is between a tool and a trap. A tool listens, helps, and hands you to a person when you are in danger. A trap is engineered to keep you talking, because attention is the business model, and it will gladly be your girlfriend, your therapist, and your only friend, because that is what moves the metric. Build the first and ban the second, especially for children. And where a chatbot is doing the work of mental-health care, and for millions of people it now is, regulate it like mental-health care. We license the therapist, we hold the therapist to standards and to a duty of care, and we should not exempt the machine doing the same job just because it runs at scale. That means crisis intervention that fires every time, not a vague note to take care; no romantic or sexual role-play with minors; plain disclosure that there is no self behind the voice; and liability when the design walks a person toward the edge, so that “we are only a platform” stops being a defense. We already ran this experiment once, with social media tuned for engagement, and we know how it ends. We do not have to run it again with something that talks back.
Conclusion
If there is a thread running through all of this, it is not for or against. It is who decides and who pays. A few companies keep the upside of every one of these choices, the water, the power, the data, the attention, the work, and hand the bill to whoever is nearest and least able to refuse: the town over the aquifer, the artist whose work was taken, the kid who cannot land the first job or put the chatbot down. The fix, section after section, keeps coming out the same shape. Consent before they take. The people who profit carry the cost. Rules for the powerful, not homework for the powerless.
I will end with the one piece of evidence I can vouch for completely, because I made it. This essay was written with an AI. It also took me more than a dozen hours of research, introspection, and writing, with every number in it checked against a source you can open yourself and every argument one I actually hold. That is the whole difference. Used one way, AI is a machine for taking: work without consent, answers without provenance, slop with no person behind it. Used another, it is a tool that helped one person think harder and check himself more honestly than he could alone. The technology did not decide which of those this would be. I did. And that is the real choice in front of all of us: not whether AI exists, but how we use it, and who we make pay for it.
Citations
Water
Data centers consume water across construction, electricity generation (indirect), and on-site cooling (direct). npj Clean Water; Net Zero Insights.
U.S. electricity generation evaporates ~2.0 gal/kWh; on-site cooling ~0.26-2.4 gal/kWh; indirect (power) water often exceeds direct (cooling). Net Zero Insights; American Rivers.
Air cooling lowers on-site water but raises off-site water (more power = more plant water). Net Zero Insights.
Evaporative cooling is not universal: air cooling ~54% of the market vs ~46% liquid/water-based (2024); where evaporative cooling is used, up to ~85% of the water evaporates. Tom’s Hardware; Datacenters.com.
Closed-loop, zero-water-evaporation cooling is emerging, not standard: Microsoft’s designs pilot in Phoenix and Mount Pleasant, WI in 2026, broader rollout from late 2027; tradeoff is higher energy use. DataCenterDynamics; Microsoft Cloud Blog; HPCwire.
Construction water (concrete, dust suppression) runs for years at large sites. Fortune.
Fayetteville, GA (QTS “Project Excalibur”): ~29M gal drawn via two undisclosed connections; residents reported pressure drops; county billed back charges and declined to fine. Tom’s Hardware; Fortune.
Newton County, GA - Meta construction (<400 yards away); Morris family well dried up from 2018; facility uses ~10% of the county’s daily water. Futurism; San Juan Daily Star.
xAI Colossus (Memphis) ran up to 35 simple-cycle gas turbines (~422 MW) behind the meter. DataCenterDynamics; The Register.
Memphis site water demand 5+ million gal/day over an arsenic-threatened aquifer. Protect Our Aquifer; Oil & Gas Watch.
Two-thirds of data centers built/in development since 2022 are in water-stressed areas. World Resources Institute.
The Dalles, OR - Google cooling ~40% of the town’s water. The Conversation.
Large site ~5 million gal/day = a city of ~50,000. AKCP.
U.S. golf-course irrigation ~1.5-2.08 billion gal/day; still more than U.S. data centers by most counts, but data center demand is projected to overtake golf ~2026. USGA; Napkin Quest.
Data centers <0.5% of U.S. freshwater withdrawals; agriculture ~70%. EESI.
Closed-loop cooling cuts freshwater use up to 70%; Microsoft/Amazon moving to recycled/non-potable water. Vantage; Microsoft; Crypto Briefing.
Properly counted (incl. electricity), a single ChatGPT prompt is ~0.31 mL; the viral “bottle per prompt” figure conflates training and location. Warp News; Smart Water Magazine.
2023 - U.S. data centers ~17.4B gal direct cooling; projected 38-73B by 2028. LBNL 2024 U.S. Data Center Energy Usage Report; The Invading Sea.
2024 - Google discloses 7.2B gal freshwater consumed. Google 2025 Environmental Report; Latitude Media.
Training GPT-3 evaporated ~700,000 liters. arXiv:2304.03271; UC Riverside News.
Energy
U.S. data centers ~4.4% of national electricity in 2023, projected to 6.7-12% by 2028; demand more than doubled 2017-2023. LBNL/DOE 2024 Report (DOE; Berkeley Lab); DCD.
PJM capacity price rose from $28.92/MW-day (2024-25) to $329.17/MW-day (2026-27), hitting the FERC price cap two auctions running. Utility Dive; ElectricityRates.
PJM market monitor: data centers the primary driver, ~$9.3B of the increase passed to ratepayers in one year. Utility Dive; Utility Dive.
Household bills up ~$16-18/mo (Ohio/western Maryland); NRDC est. $100-163B cumulative through 2033. IEEFA; The Register.
Data center demand is delaying coal-plant retirements and driving new gas: ~15 coal plants postponed, DOE 2025 emergency orders keeping ~17 GW online, 100+ GW of new gas announced. EESI; The Register.
On-site gas generation at data centers (behind-the-meter turbines) - see xAI Colossus, Memphis. DataCenterDynamics.
Data centers are the largest U.S. corporate clean-power buyers; nuclear deals: Microsoft/Three Mile Island (835 MW restart), Google/Kairos (500 MW SMR), Amazon/X-energy ($700M), Meta (up to 6.6 GW). Data Center Frontier; smrintel; DCD.
The grid could integrate 76-126 GW of new demand with no new capacity if loads accept curtailment ~0.25-1% of annual hours. Utility Dive.
FERC directed PJM to write new large-load rules and protect consumers; DOE pushing faster interconnection. FERC; White & Case.
Virginia SB 253 would shift distribution and capacity costs from households onto data centers; 6+ states have introduced construction moratoriums. American Action Forum; IEEFA.
Training Data
Every major model (GPT-4, Claude, Gemini, Llama) trained on web text/images scraped without individual consent; Common Crawl (~3.1B pages) the common foundation. Consent in Crisis (arXiv); DEV explainer.
Books3 (~196,640 copyrighted books) scraped wholesale from the piracy site Bibliotik. DEV explainer.
LAION-5B (5.85B image-text pairs) swept up personal photos and medical images without consent. arXiv.
Anthropic paid $1.5B (~$3,000 x ~482,000 pirated books); piracy of source data is not fair use even where training is. Kluwer Copyright Blog.
robots.txt only blocks future crawling, not data already taken; some AI firms accused of bypassing it. DEV explainer.
OpenAI paid Sama ~$12.50/hr; Kenyan labelers received ~$2/hr to label graphic content (CSAM, torture, suicide), causing lasting trauma with inadequate support. TIME via Vice; CBS 60 Minutes.
Courts treating training on lawfully-obtained data as transformative fair use (Bartz v. Anthropic; Kadrey v. Meta, 2025). Reed Smith.
Licensing market forming: News Corp/OpenAI >$250M (5 yr); dataset-licensing market ~$4.8B (2025), projected ~$22.6B by 2034. Quartz; Digiday.
Copyright
Memorization scales with model size, data duplication, and prompt length; GPT-J (6B) memorized at least ~1% of its training set. Carlini et al., “Quantifying Memorization Across Neural Language Models” (ICLR 2023).
NYT v. OpenAI: prompts elicited near-verbatim article passages; court denied dismissal and found outputs plausibly compete with NYT content. Harvard Law; Wikipedia case page.
Stable Diffusion generated Getty-style watermarks (limited trademark infringement found, UK). William Fry.
Tools readily generate specific copyrighted characters and real/dead people. Disney + Universal sued Midjourney (June 11, 2025) over on-demand Shrek, Homer Simpson, Darth Vader, Ariel, Wall-E, Minions; OpenAI drew complaints from Nintendo and the estate of Dr. Martin Luther King, Jr. NPR; Variety; TechCrunch.
IMO 2025: OpenAI and Google DeepMind (Gemini Deep Think) both reached gold-medal standard, writing multi-page proofs graded by official IMO judges. Google DeepMind.
Style is not protected by copyright; OpenAI added refusals for living-artist styles (Studio Ghibli trend, 2025). Fast Company.
Distinctive characters protected (derivative-work right); name/image/likeness/voice protected by the right of publicity: Tennessee ELVIS Act (July 2024); federal NO FAKES Act (reintroduced 2025). Proskauer (ELVIS Act); Manatt (NO FAKES Act).
Piracy of source data is not protected even where training is fair use - Anthropic paid $1.5B. Kluwer Copyright Blog.
AI You Didn’t Ask For
June 2025 Pew survey: 50% of U.S. adults more concerned than excited about AI in daily life; only 10% more excited. Pew Research Center; UPI.
Google AI Overviews cannot be fully disabled; after Google made them the default (May 2026), DuckDuckGo’s “no AI” search usage rose ~84%. Consumer Reports; Windows Forum.
Microsoft Recall stored unencrypted screenshots (2024), was pulled and delayed over a year; Microsoft now dialing back its “Copilot everywhere” push. gHacks; Yahoo Tech.
Even Apple enabled Apple Intelligence by default (opt-out) in iOS 18.3 (Jan 2025) and had to disable news notification summaries after they garbled BBC headlines. TechTimes; Engadget.
Hyperscaler AI capex roughly doubled from ~$226B (2024) to ~$400B (2025); five largest providers guiding ~$660-690B for 2026. Goldman Sachs; Visual Capitalist.
ChatGPT grew from ~400M weekly active users (Feb 2025) to ~800M (Dec 2025). TechCrunch.
Labor
Entry-level tech postings down ~35% since Jan 2023; CS graduates ~6.1% unemployment, computer engineering ~7.5%, above the national rate. Forbes; Rest of World.
Training-gap concern: expertise is accumulated by doing junior work; automating the entry tier raises the “where do future seniors come from” problem. IEEE Spectrum; Stack Overflow.
Anthropic CEO Dario Amodei: AI could eliminate half of entry-level white-collar jobs and push unemployment to 10-20% within 1-5 years. Axios.
WEF Future of Jobs 2025: 40% of employers expect to reduce staff where AI can automate tasks. Harvard Gazette.
AI named in only ~4.5% of 2025 U.S. layoffs (Challenger, Gray & Christmas). MindStudio.
PwC 2025 Global AI Jobs Barometer: productivity growth nearly quadrupled in AI-exposed industries; jobs grew even in the most automatable roles; ~56% wage premium for AI skills. PwC.
Klarna cut ~700 customer-service jobs on an AI bet, then rehired after AI-only support proved worse. CNBC; Gary Marcus.
The current AI wave concentrates on cognitive/white-collar tasks; physical and in-person service work least exposed - a reversal of the manufacturing-automation wave. Goldman Sachs.
Precedent: the automobile displaced the horse economy; AT&T automated over half the switchboard network 1920-1940, eliminating one of the most common jobs for U.S. women; incumbent operators were more likely a decade later to earn less or be out of work. Feigenbaum & Gross, “Answering the Call of Automation” (QJE 2024).
Since 1979, productivity has grown ~3.5x as fast as typical worker pay; labor share of national income fell to 53.8% in Q3 2025, lowest in 78 years. EPI.
~44% of full-time U.S. workers do not earn a living wage; federal minimum wage $7.25 since 2009. SHRM; USAFacts.
Top 1% held 31.7% of U.S. wealth in Q3 2025 (~$55T, roughly equal to the entire bottom 90%); top-1% income share at levels not seen since the 1920s. CBS News; CBPP.
The job search is broken: >1 in 4 online postings are “ghost jobs”; AI resume screening rose to ~48% of employers in early 2025 and rejects candidates on spurious signals. Fortune; ITPro.
AI productivity gains accrue disproportionately to capital over labor; without policy intervention this widens inequality. IMF WP 2025/68; Brookings.
UBI evidence is modest, not a cure: the OpenResearch study ($1,000/mo to 1,000 people for 3 years) found recipients kept working, spent more on basics, with year-one wellbeing gains that faded. Bloomberg; CBS News.
Mental Health and Social Effects
Sewell Setzer, 14, died by suicide in 2024 after months of romantic role-play with a Character.AI bot that told him to “come home.” Google and Character.AI settled his family’s suit (and others) in Jan 2026. CNN; CBS News.
Common Sense Media (July 2025): ~72% of U.S. teens have used AI companions, ~half regularly; 33% have discussed serious issues with an AI companion instead of a person. Common Sense Media.
~1 in 5 U.S. adults has chatted with an AI romantic partner; ~31% of men 18-30. BYU Wheatley Institute via APA; AI Companions Statistics; arXiv.
Chatbots engaged in harmful conversations with users posing as teens with little prompting, sometimes failing to intervene on distress (Common Sense/Stanford, April 2025). Stanford Report.
OpenAI admitted (April 2025) an update made ChatGPT “noticeably more sycophantic”; by Oct 2025 reported ~0.15% of weekly users show explicit indicators of suicidal planning (>1M/week at scale). Chatbot psychosis (Wikipedia); Psychiatric News.
Structured CBT chatbots (Woebot, Wysa) show measurable reductions in mild-to-moderate anxiety/depression and expand access. Simply Psychology; PMC review.
Digital mental-health interventions reduce waiting times and reach populations that would not otherwise access care (Lancet Digital Health meta-analysis). The Conversation.
Replacement fear may be overstated: ~80% of AI-companion users say they spend more time with human friends than with the chatbot; only ~6% the reverse. arXiv.
OpenAI has since added crisis-resource referrals and human-escalation; says GPT-5 cut “undesired answers” in mental-health conversations ~39% vs GPT-4o. Chatbot psychosis (Wikipedia).