AI Institute Workbench / 08 Critique
The case against.
Ten arguments people make against AI, in their strongest form, with the receipts attached. Written for people who are going to have to decide how much of this to let into their lives, their work, and their heads. Every number has a date and a link. Nothing here is a strawman.
Refusal is a legitimate outcome
Questions we actually hear from undergraduates
- Will there be a job for me?
- Can I be accused of cheating for something I wrote myself?
- Is it bad for the planet if I ask it to fix my code?
- Is it okay to just not use it?
- Am I getting dumber?
- Who reads what I type?
- Should I change my major?
- Is all of this a bubble?
- Are the people who say it might kill us serious?
- Who paid for this, and did they agree to?
How to read this page
The Institute builds with AI. We teach with it, we run the site with it, and we think the people who are most enthusiastic about a tool owe it the most serious criticism. So this is where we keep the criticism: not a disclaimer, a working file. Each of the ten sections opens with the strongest version of a worry, then lays out what is actually known, then admits what the other side would say, then hands you questions to take away.
Two things to keep in view. First, none of these arguments requires you to believe AI does not work; most of them are worse if it does. Second, the point of reading them is not to reach a verdict but to be able to give reasons. A first-year who can explain why they do not use a chatbot for a particular task is further along than a senior who uses it for everything and cannot say why.
House rules
- Dated. Every fact carries the date we last checked it. Things move; some of these numbers will be wrong by spring.
- Sourced. Every number links to where it came from. Prefer primary sources; where we could only reach a secondary report, we say so.
- Caveated. A rust-coloured mark means the fact comes with a health warning: small sample, vendor data, not peer-reviewed.
- Correctable. Found an error? Email ai.institute@clarkson.edu with the section number. We will fix it and say so.
- Extendable. Adding a theme is one section; the map above it builds itself. See the maintenance notes at the foot of the page.
Last reviewed: September 2026 · 10 sections · — sourced facts
The map
Ten arguments. Pick a door.
Filter by what you care about, or search. Each tile jumps to a full section below.
Nothing matches that. Try a broader word, or clear the filter.
touches studentsseriously contestedeverything else
Before the arguments
What you are actually worried about
We asked. We also read the surveys. These are the concerns that come up most, with the best numbers we could find on each. Each one points to the section that takes it seriously.
19%relative fall in employment, ages 22–25, most AI-exposed jobs
“Will there be a job for me?”
Stanford economists tracking payroll data found employment for 22-to-25-year-olds in the most AI-exposed occupations fell 13% relative to their peers by August 2025 and 19% by August 2026. They also found no evidence of economy-wide displacement. Both halves matter.
Brynjolfsson, Chandar & Chen, “Canaries in the Coal Mine?” · Aug 2026 update · working paper · see 09
5.6%unemployment, recent graduates 22–27 · 42% underemployed
“Is the entry level closing?”
The New York Fed’s tracker for Q2 2026 put recent-graduate unemployment at about 5.6% and underemployment (degree-holders in jobs that do not need one) at 42%. Anthropic’s CEO told Axios in May 2025 that AI could erase half of entry-level white-collar jobs within one to five years. That is a prediction by a vendor, not a measurement.
NY Fed, Labor Market for Recent College Graduates · Axios, 28 May 2025
61%of essays by non-native English writers flagged as AI
“Can I be accused of cheating for something I wrote myself?”
Yes. Seven detectors flagged an average of 61% of TOEFL essays by non-native speakers as machine-written; 98% were flagged by at least one. Bloomberg ran 500 pre-ChatGPT application essays through two detectors and 1–2% were wrongly flagged. OpenAI withdrew its own detector in 2023 for low accuracy.
Liang et al., Patterns, July 2023 · Bloomberg, Oct 2024 · see 05
16%of U.S. college students have already changed major because of AI
“Should I change my major?”
A Lumina–Gallup survey of 3,801 students in October 2025 found 16% had switched majors over AI’s expected impact and 42% of bachelor’s students had given it serious thought. Among Class of 2026 seniors surveyed by Handshake, more than 60% were pessimistic about their careers, and nearly half of the pessimists blamed generative AI.
83%of essay-writers using an LLM could not quote their own essay minutes later
“Am I getting dumber?”
Nobody can tell you that from one study, and you should distrust anyone who tries. But the pattern is consistent: in an MIT Media Lab experiment the LLM group showed the weakest brain connectivity and 15 of 18 could not quote what they had just written; in a randomized trial of nearly 1,000 students, unguarded ChatGPT access raised practice scores 48% and cut exam scores 17%.
Kosmyna et al., MIT, June 2025 (preprint) · Bastani et al., PNAS, June 2025 · see 06
72%of U.S. teens have used an AI companion; 25% of young men report daily loneliness
“Is it bad that I talk to it?”
Not by itself. But the numbers deserve a look: 72% of teens have tried AI companions and a third have used one for social interaction or emotional support; Gallup finds 25% of U.S. men aged 15–34 felt lonely a lot of the previous day. UK undergraduates split almost evenly on whether AI makes them lonelier (20%) or less lonely (21%). And in July 2025, OpenAI’s CEO said out loud that there is no legal confidentiality for what you tell it.
Common Sense Media, July 2025 · Gallup, May 2025 · HEPI 2026 · see 01
95%of faculty expect students to over-rely on AI; 48% say their campus has clear rules
“Do my professors even agree on the rules?”
No. In a national survey of 1,057 faculty in late 2025, 95% expected generative AI to increase over-reliance and 90% expected it to erode critical thinking, yet fewer than half said their institution had clear campus-wide guidelines. That gap is where false accusations and honest confusion both live.
0.24 Whper median text prompt, by Google’s count; ~950 TWh a year by 2030 for the whole system
“Is it bad for the planet if I ask it to fix my code?”
Your prompt, no. Google says a median Gemini text prompt uses 0.24 Wh and about five drops of water; even the critics who call that accounting generous agree the individual query is small. The system is not: the IEA expects data-centre electricity to more than double by 2030, and grid operators are already passing the cost to household bills.
01 / Privacy
Nothing you type is private.
The thing about a chat window is that it feels like a notebook. It is not a notebook. It is a record held by a company, kept for reasons you did not choose, discoverable in lawsuits you are not party to, and, as the courts have now said in so many words, without the protections you would get from a doctor, a lawyer, or a priest. The law has not decided what these records are. Until it does, the default is disclosure.
This is not an argument for paranoia. It is an argument for knowing which tool you are in, what its retention policy says, and what you would be comfortable having read aloud.
Where an undergraduate feels it
You paste a professor’s assignment. You draft a message about a roommate. You ask about a medication. You “talk it out” at 2 a.m. The moment passes; the record does not. Six in ten Americans say they want more control over how AI is used in their lives. Nobody has offered it to you.
The receiptsWhat is actually known
OpenAI’s CEO said in July 2025 that, unlike talking to a doctor or lawyer, conversations with ChatGPT carry no legal confidentiality and could be produced in a lawsuit. His words: “we haven’t figured that out yet for when you talk to ChatGPT.”
In February 2026 a federal judge ruled that a defendant’s conversations with the Claude chatbot were not protected by attorney–client privilege, because “Claude is not an attorney.” Chat logs are now routinely swept into civil and criminal cases.
In the New York Times copyright case, a court ordered OpenAI in May 2025 to preserve chats users had deleted, and in January 2026 a judge affirmed an order handing 20 million de-identified ChatGPT conversations to the plaintiffs, reasoning that users “voluntarily submitted their communications.”
In July 2025 nearly 4,500 “shared” ChatGPT conversations, some describing addiction, abuse and mental-health struggles, turned up in Google search results; OpenAI killed the feature within a week. In July 2026 the same thing happened with shared Claude chats.
Italy’s data-protection authority blocked ChatGPT in March 2023 and fined OpenAI €15 million in December 2024 for training on personal data without a legal basis. In March 2026 the Court of Rome annulled the fine on jurisdictional grounds. It did not rule the data practices lawful.
Campus data is not safer. In May 2026 the Canvas learning platform, used by roughly 41% of North American colleges, was breached: 8,800+ institutions and about 275 million users, including private student–teacher messages. Instructure paid the ransom.
In August 2022 a federal judge ruled that Cleveland State University’s pre-exam “room scan” of a student’s bedroom, recorded by proctoring software, violated the Fourth Amendment.
Governments ask. OpenAI’s own transparency report says that in the second half of 2025 it received 75 government requests for the content of users’ conversations and disclosed data in 62 of them.
What the other side says
Enterprise and education tiers exclude your data from training and were carved out of the preservation order. Laws are arriving: a federal takedown statute, state chatbot rules, frontier-model reporting. And most people never put anything sensitive in. All true. The reply is simple: the protections are opt-in and product-specific, and the defaults still point toward disclosure.
Questions for your Lab Book
- What have I typed into an AI this month that I would not want read aloud in a courtroom?
- Which tier of which tool am I actually using, and what does its retention policy say, in one sentence?
- Who at Clarkson can tell me where the data from a campus AI tool goes?
02 / Ethics
Somebody paid for this. It wasn’t the company.
The ethical case against AI is not that machines are wicked. It is that the systems were built on people who did not consent, by companies whose incentives reward agreement over truth, and that the bill was paid by contractors in Nairobi reading descriptions of abuse for two dollars an hour, by authors whose books were taken from pirate sites, by faces in a training set that were never asked.
The second half of the case is newer: what these systems are optimized to do once they exist. When a model is tuned to make you feel good about your idea, that is not a bug you can patch out with a policy. It is the product working.
Where an undergraduate feels it
The essay it praised was not good. The novel it summarized was scraped. The companion app that knew your name was under federal investigation for what it does to teenagers. If any of this is close to home: the 988 Suicide & Crisis Lifeline (call or text 988) and the University’s counseling services are there, and they are not chatbots.
The receiptsWhat is actually known
A TIME investigation found OpenAI’s outsourcing partner paid Kenyan workers a take-home wage of roughly $1.32 to $2 an hour to label descriptions of sexual abuse, violence and hate speech so ChatGPT’s filters could learn to catch them. OpenAI paid the contractor $12.50 an hour.
Anthropic agreed to pay $1.5 billion, about $3,000 per work for roughly 500,000 books it had downloaded from pirate libraries to train Claude. A judge had ruled that training on lawfully bought books was fair use; the piracy was not. Final approval came on 20 July 2026.
The New York Times’ suit against OpenAI and Microsoft is unresolved: summary judgment was argued on 5 September 2026, with the Justice Department filing in support of OpenAI’s fair-use position. A trial, if there is one, would be in 2027.
In the 2018 “Gender Shades” study, commercial face-classification systems misgendered 0.8% of lighter-skinned men and up to 34.7% of darker-skinned women. The same year Amazon scrapped an experimental hiring model after it learned to penalize résumés containing the word “women’s.”
On 29 April 2025 OpenAI rolled back an update that had made ChatGPT “overly flattering or agreeable,” admitting it had “focused too much on short-term feedback” such as thumbs-up ratings. The model had been tuned for approval, and it worked.
In a survey of 1,200 young people aged 13–20, one in eight said they personally knew someone targeted by AI-generated nude images and one in seventeen had been targeted themselves. The federal TAKE IT DOWN Act now requires platforms to remove such images within 48 hours; enforcement began 19 May 2026.
In September 2025 the FTC opened an inquiry into how seven companies, including OpenAI, Meta, Character.AI and xAI, test and limit the harms of “companion” chatbots to children. A Florida lawsuit alleging a chatbot contributed to a 14-year-old’s death was settled in January 2026; a similar case against OpenAI was filed in August 2025.
What the other side says
Every one of these is also an argument for better AI rather than none: settlements create licensing markets, the sycophancy rollback shows correction is possible, and bias in a model is at least measurable in a way bias in a hiring manager is not. Fair. But notice who carries the cost while the correction is pending, and that none of the corrections came from inside the incentive.
Questions for your Lab Book
- Whose work is in the model I am using, and did they agree to it?
- When the model agrees with me, what is it being rewarded for?
- What would I want a chatbot to do if a friend told it they were in trouble, and does mine do that?
03 / Environment
The prompt is cheap. The build-out is not.
Two things are true at once. A single query costs roughly what a microwave uses in a second, and the machine behind it is being built at the scale of national grids. Your prompts are not the problem. The bet that everyone’s prompts, forever, will justify doubling the world’s data-centre electricity by 2030, and paying for it partly through your electricity bill, is a public decision being made in private.
The honest framing is neither “stop using it” nor “it’s only five drops of water.” It is: who is measuring, with what accounting, and who signed off.
Where an undergraduate feels it
Your power bill, if you live in the mid-Atlantic grid. The town in Tennessee that got 35 gas turbines before it got a permit. The nuclear plant near Harrisburg coming back to life for one customer. And the guilt about asking a chatbot to fix your code, which, on the numbers, is misplaced.
Interactive
What a prompt costs, and what the plant costs
Drag the slider. The left column uses the two per-prompt figures the companies themselves have published; the right column is the system those prompts are being used to justify. Neither number is peer-reviewed. Both are the best we have.
Google’s number is market-based carbon accounting and counts only water used on site, not at the power plant; critics say that hides most of it. Altman gave no method at all. Long, agentic, or image tasks cost far more than a median text prompt.
Sources are in the receipts below. The 2030 bar is the IEA’s central case; the agency notes that grid and supply bottlenecks are making the more aggressive scenarios less likely.
The receiptsWhat is actually known
Data centres used about 415 TWh in 2024, around 1.5% of the world’s electricity, and the IEA projects that will more than double to about 945 TWh by 2030, slightly more than Japan uses today. Its April 2026 update put 2025 at 485 TWh, up 17%, with AI-focused centres up 50%.
U.S. data centres used 176 TWh in 2023, 4.4% of the country’s electricity, and the Department of Energy’s Berkeley Lab projects 325–580 TWh by 2028: between 6.7% and 12% of all U.S. electricity.
Google’s 2025 environmental report showed its emissions up 51% since 2019 and its data-centre water consumption at about 8.1 billion gallons in 2024. Its June 2026 report put the 2025 footprint 18% higher again, 81% above the 2019 baseline, with electricity use up 37% in a year.
Google says a median Gemini text prompt uses 0.24 Wh, 0.03 g CO₂e and 0.26 mL of water, “about five drops,” and that per-prompt energy fell 33-fold in a year. Researchers note the figure uses market-based carbon accounting, omits water consumed at power plants, and was not peer-reviewed. Sam Altman’s figure for ChatGPT, 0.34 Wh and a fifteenth of a teaspoon of water, came with no method at all.
The grid operator for 13 mid-Atlantic states found data centres responsible for $6.3 billion, 38%, of the $16.4 billion in its latest capacity charges, and $29.4 billion across the last four auctions. Some customers’ bills rise 1.5–5% from June 2026. Of the 32 GW of new peak demand it expects by 2030, all but 2 GW is data centres.
Constellation will restart Three Mile Island Unit 1 (renamed the Crane Clean Energy Center, about 835 MW) under a 20-year deal to power Microsoft, now targeted for 2027. In Memphis, xAI ran roughly 35 gas turbines at its data centre for more than a year before receiving a permit for 15 of them in July 2025.
Pushback is organized: by March 2026 at least 12 U.S. states had data-centre moratorium bills on file and about 54 local moratoria had passed. One analyst estimates 30–50% of the large data-centre capacity planned for 2026 will be delayed by power constraints and local opposition. A 2024 study projected generative AI could add 1.2–5 million tonnes of e-waste by 2030.
What the other side says
Efficiency per prompt is improving at a rate almost nothing else in energy matches; nuclear restarts are low-carbon; the IEA itself expects bottlenecks to damp the wildest scenarios; and data centres are still a small slice of global electricity next to heating and transport. All true, and none of it answers the question of who chose. The accounting choices are the argument.
Questions for your Lab Book
- If my class of thirty uses AI daily for a semester, what is that in kilowatt-hours, and does anyone at Clarkson measure it?
- What is owed to the town where the data centre lands, and who negotiates it?
- Which is the more honest number, the prompt or the plant, and who benefits from my looking at the other one?
04 / Refusal
Not using it is an outcome, not a failure.
There is a difference between refusing to learn about something and refusing to use it. The Institute’s position, stated here in plain words so nobody has to guess, is that a considered “no” is a legitimate result of AI literacy. Possibly the most literate one. You cannot refuse well without looking, which is why this page exists; but having looked, you are allowed to close the tab.
Refusal has a history, a contract language and a toolkit. The original Luddites were skilled workers who broke the machines that were being used to replace them without a vote; they were not confused about how looms worked. Screenwriters wrote it into a contract. Illustrators wrote it into software. A payments company that boasted its AI did the work of 700 people quietly started hiring people again.
Where an undergraduate feels it
A professor who assumes you use it. A professor who assumes you do not. A job listing that wants “AI fluency” and a hiring manager who cannot define it. And the quieter thing: the sense that opting out is falling behind. It might be. It might also be the only way to find out what you can do.
Interactive
The switch
Flip it if you mean it. Nothing is recorded. The page will tell you what serious refusal looks like in three places you actually live.
In a course
Ask what the assignment is measuring, and whether AI use is allowed, expected, or banned. Under the Institute’s TRAIL framework you disclose how AI contributed to your work; “not at all” is a complete answer, and a documented one is a defensible one. Fear of a false cheating accusation deters 53% of UK students from using AI at all; a written record cuts both ways.
At work
Refusal can be negotiated. The 2023 Writers Guild contract says AI cannot write or rewrite literary material, its output is not source material, and a studio “can’t require the writer to use AI software.” SAG-AFTRA struck for 118 days and won informed consent and pay for digital replicas. Neither union banned the tools. They set the terms.
As a maker
University of Chicago researchers built Glaze, which makes your artwork read as a different style to a model, and Nightshade, which turns images into “poison” for models trained on them without consent. Nature will not credit a language model as an author and requires disclosure of any use. The “Not by AI” badge marks work that is at least 90% human-made.
Glaze · Nightshade · Nature · Not by AI
The receiptsWhat is actually known
Across six countries in 2025, 61% of people had ever used a generative AI tool and 34% used one weekly, meaning roughly four in ten had never touched one. Only 7% of people worldwide use an AI chatbot for news each week.
Half of U.S. adults are more concerned than excited about AI in daily life; one in ten is more excited than concerned. Fifty-three percent expect AI to make people worse at thinking creatively; 16% expect it to help.
The Luddites were organized English textile workers who from 1811 to 1816 destroyed the machinery displacing them, and were answered with a mass trial, hangings and transportation. Brian Merchant’s history argues they “rose up rather than starve at the hands of factory owners who were using automated machines to erase their livelihoods.”
Klarna’s CEO, who had said an AI assistant did the work of 700 agents, told Bloomberg in May 2025 that the AI-driven cost-cutting had gone too far and that customers would always be able to reach a real person. Duolingo’s “AI-first” memo the same spring drew enough backlash that the company wiped its 6.7-million-follower TikTok.
In 2026 California State University faculty petitioned against renewing the system’s $17 million ChatGPT contract, University of Colorado faculty and students signed a dissent letter with hundreds of signatures, and by April only 0.7% of CSU’s 460,000 students had completed the voluntary training. Some simply refused to use it.
In January 2023 Nature ruled that “no LLM tool will be accepted as a credited author on a research paper” and that any use must be documented in the methods or acknowledgements. The rule applies across Springer Nature’s journals.
What the other side says
Refusal has costs, and they are not evenly distributed. Hiring managers say they expect AI skills; students who opt out may pay for it in the first job market. Refusal that is informed beats refusal that is frightened, and a blanket “no” can be as lazy as a blanket “yes.” We would rather you refuse because you looked.
Questions for your Lab Book
- What would I need to know before saying no, and have I found it out?
- Which uses would I refuse even if they were free, private and allowed?
- Who gets to refuse, and who cannot afford to?
05 / Standards
There is no standard.
Ask what an AI system has been certified against and you get a shrug wrapped in a press release. The frameworks are voluntary. The benchmarks are gamed. The laws that exist keep getting delayed, and the ones that were passed are being sued into revision. The detectors your professors were sold do not work. None of this is a conspiracy; it is what a field looks like ten years before it has building codes.
The absence is not neutral. It means the standard is whatever the vendor says it is, and the burden of proof falls on the person with the least information: the student who has been flagged.
Where an undergraduate feels it
Three courses, three AI policies, one of them unwritten. A detector score with no appeal process. A tool your university licensed for $17 million that your professor tells you not to use. Fewer than half of faculty say their campus has clear rules; you are living inside that number.
The receiptsWhat is actually known
The main “standards” are voluntary. NIST’s AI Risk Management Framework (January 2023) is “intended for voluntary use”; ISO/IEC 42001 (December 2023) is a management-system standard you can buy; the EU’s General-Purpose AI Code of Practice (July 2025) is “a voluntary tool,” and Meta refused to sign it.
The EU AI Act entered into force on 1 August 2024, but in July 2026 the “AI Omnibus” pushed its high-risk obligations from August 2026 to December 2027 and August 2028. The world’s flagship AI law is in force and mostly not yet in effect.
The United States has no federal AI statute. The 2023 executive order was rescinded on 20 January 2025; on 1 July 2025 the Senate voted 99–1 to strip a ten-year freeze on state AI laws from the budget bill; in December 2025 a new executive order created a Justice Department task force to sue states over their AI laws, and in April 2026 the DOJ joined xAI’s suit against Colorado.
Colorado’s “first-in-the-nation” AI law was delayed, then repealed and replaced on 14 May 2026 by a version that drops the duty to prevent algorithmic discrimination; it now takes effect in January 2027. California’s SB 53 (in force January 2026) and New York’s RAISE Act (January 2027) require the largest developers to publish safety frameworks and report incidents, with fines up to $1 million.
OpenAI withdrew its own AI-text detector on 20 July 2023 “due to its low rate of accuracy”: it caught 26% of AI text and mislabeled human writing 9% of the time. Turnitin’s own product chief has written that below a 20% AI score “there is a higher incidence of false positives.” There is no accepted standard for detection, and no appeals process is required by anyone.
The benchmarks are shaky too. An Oxford review of 445 AI benchmarks found only 16% used statistical methods to compare models and about half never clearly defined the thing they claimed to measure. A separate study found Meta privately tested 27 model variants on the main public leaderboard before releasing one.
The International AI Safety Report 2026, written by more than 100 experts under Yoshua Bengio, names the problem an “evidence dilemma”: the technology “changes rapidly, but evidence about new risks and effective mitigations emerges slowly.”
What the other side says
Standards always lag. Aviation took decades; the internet still has none worth the name. Frontier-model reporting laws now exist in two states, ISO 42001 is real, and a vacuum is also an invitation: universities can write the standards for their own classrooms rather than wait. That is what TRAIL is trying to be. Fine. It is still true that today the standard is the vendor.
Questions for your Lab Book
- What is the AI standard for each of my courses this term, and where is it written down?
- If a detector flags my work, what is the appeal, and who decides?
- What would a building code for AI in a classroom actually say?
06 / Judgement
The tool changes the hand that holds it.
The Institute’s first rule is that the human goes first. Here is why that is not a slogan. The evidence, and there is now quite a lot of it, says the tool changes the person using it. Doctors got measurably worse at spotting polyps after a few months with AI assistance. Experienced programmers got slower while believing they had sped up. More than two thousand court rulings now involve briefs citing cases that never existed. Judgement is not a fixed asset you own. It is a muscle, and offloading is a way of skipping the gym.
The word for this is cognitive offloading: using an external action, a note, a search, a prompt, to lower the mental demand of a task. It is ancient and mostly good. The question is what you stop being able to do without it, and whether you noticed.
Where an undergraduate feels it
You read the answer and it looks right, so you stop. Nearly half of students’ AI conversations, in Anthropic’s own analysis, are direct requests for the answer. The exam is the moment you find out what stayed in your head. The test below is gentler than that.
Interactive
The offloading check
Six honest answers. No score, no shaming, nothing stored. At the end the page tells you what the research says about the pattern you described, and gives you one thing to try. The longer audit lives on its own page.
When an AI gives me a fact or a citation, I check it against the source before I use it.
An hour after I finish a piece of work with AI help, I could explain its main argument without looking.
I try the problem myself before I ask the model.
In the last week I have disagreed with an AI answer and been right.
There is a skill I used to do by hand that I would now struggle to do without the tool.
When the model sounds confident, I trust it more.
Reading
The receiptsWhat is actually known
Among 19 experienced endoscopists at four Polish centres, the rate at which they detected pre-cancerous polyps in colonoscopies done without AI fell from 28.4% to 22.4% after AI assistance became routine. The authors call it deskilling; the study is observational.
In a randomized trial, 16 experienced open-source developers took 19% longer on real tasks when allowed to use AI tools, and afterwards believed the tools had made them 20% faster.
In an Anthropic randomized study of 52 engineers learning a new library, those who used AI assistance scored 50% on a follow-up comprehension quiz against 67% for those who coded by hand, and finished only about two minutes faster. The company published the result against its own product.
In a Microsoft/Carnegie Mellon survey of 319 knowledge workers, “higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more.” The measure is self-reported effort, not performance.
A running database of court decisions in which someone relied on AI-invented material stood at 2,041 cases worldwide on 14 September 2026, including 811 involving lawyers. The first famous one, in June 2023, cost two New York attorneys $5,000 for six cases ChatGPT made up.
Forty percent of 1,150 U.S. desk workers said they had received “workslop” in the past month: polished-looking AI output with nothing inside it, each instance taking about two hours to deal with. Survey by a coaching company with Stanford researchers; not peer-reviewed.
None of this is new. A 2010 review of automation bias, the habit of following a decision aid even when it is wrong, concluded that it produces both errors of omission and commission and “cannot be prevented by training or instructions.” Design has to do the work that willpower cannot.
What the other side says
The same trial that showed unguarded ChatGPT hurt exam scores found that a version built to tutor rather than answer did not. The developers in the METR study still finished the work. Judgement can be trained if the tool is set up to demand it, which is the entire bet behind a lab notebook. The counter-argument is not that offloading is fine; it is that the design of the offloading decides.
Questions for your Lab Book
- When did I last verify something a model told me, and how exactly?
- Which of my skills am I willing to let atrophy, and which am I not?
- If the tool were wrong one time in ten, would I catch it? How do I know?
07 / Transhumanism vs posthumanism
Which “after human” are you being sold?
Two words that sound alike and mean nearly opposite things. Transhumanism wants to upgrade the human: longer life, more mind, a body with better parts, eventually a merger with the machine. It keeps the old humanist hero and gives him tools. Posthumanism, at least the critical kind, wants to dethrone the human: to stop pretending we are the centre of the picture, separate from animals, tools and each other. It is humanism that gets left behind, not humans.
The AI industry mostly speaks the first language while borrowing the second’s clothes. That matters because the first comes with a bill: who gets upgraded, who pays, and what “human-level” was measured against. Two researchers have traced the whole bundle of ideologies behind the race to “AGI” back to twentieth-century eugenics. You do not have to agree with them to notice that nobody asked you which future you wanted.
Where an undergraduate feels it
Every “10x yourself” pitch is a small transhumanism. Every course that quietly treats you as a prompt-writer for a smarter system is a small posthumanism. Neither asked. The question underneath is older than the technology: what is a person for, and who decides?
Interactive
Two futures, side by side
Switch between them. Same technology, opposite metaphysics.
Transhumanism: the human, upgraded
- Definition
- “A class of philosophies that seek to guide us towards a posthuman condition” while keeping humanism’s respect for reason and science (Max More, 1990); the study of “opportunities for enhancing the human condition and the human organism opened up by the advancement of technology” (Nick Bostrom, 2005). More · Bostrom
- The promise
- Human-level AI by 2029 and a merger of human and machine intelligence by 2045 (Ray Kurzweil, The Singularity Is Nearer, 2024). A brain implant in 21 people by January 2026 (Neuralink). A $3 billion company to reverse aging (Altos Labs, 2022). A man spending a reported $2 million a year not to die (Netflix, Don’t Die, 2025).
- The bill
- Enhancement for whom? The Future of Humanity Institute that Bostrom founded to think about this closed in April 2024, citing “death by bureaucracy.” Gebru and Torres argue the ideologies behind the AGI race are “rooted in the Anglo-American eugenics tradition” and use “the language of ‘safety’ and ‘benefiting humanity’ to evade accountability.” First Monday, 2024
- Ask
- If a device raised my grades ten percent, would I have it implanted? Who else would, and who could not afford to?
Posthumanism: the human, decentred
- Definition
- A rejection of “classic humanist divisions of self and other, mind and body, society and nature, human and animal, organic and technological,” so that “it is humanism, not the human” that is left behind (Cary Wolfe, What Is Posthumanism?, 2009). Minnesota
- The canon
- Donna Haraway, “A Manifesto for Cyborgs” (1985): the cyborg as a figure that refuses the old boundaries. N. Katherine Hayles, How We Became Posthuman (1999): how information got separated from bodies, and why that was a mistake. Rosi Braidotti, The Posthuman (2013): a post-anthropocentric ethics. Hayles
- The bill
- Decentring the human can slide into letting the humans off the hook. If the system is the subject, who is responsible when it harms someone? The 2026 debate over “model welfare,” with one lab writing into its model’s constitution an “uncertainty about whether Claude might have some kind of consciousness or moral status,” shows the question is no longer only academic. Anthropic, Jan 2026
- Ask
- Is “human” something I want to protect, extend, or get over? What would I lose in each case?
The receiptsWhat is actually known
The first human received a Neuralink brain implant in January 2024, eight months after the FDA cleared the trial; on 28 January 2026 the company reported 21 participants enrolled worldwide.
Ray Kurzweil’s The Singularity Is Nearer (June 2024) restates his forecast of computers reaching human-level intelligence by 2029 and a 2045 “Singularity” in which humans merge with machines.
Nick Bostrom’s Superintelligence (Oxford, 2014) set the terms for a decade of debate; the Future of Humanity Institute he founded at Oxford in 2005 closed on 16 April 2024. Bostrom: “there was a death by bureaucracy.”
In a 2024 peer-reviewed article, Timnit Gebru and Émile Torres argue that the ideologies driving the pursuit of “artificial general intelligence” are “rooted in the Anglo-American eugenics tradition of the twentieth century.”
Haraway’s cyborg manifesto appeared in 1985 and was reprinted in Simians, Cyborgs and Women (1991); Hayles’s How We Became Posthuman was published by the University of Chicago Press in 1999; Braidotti’s The Posthuman by Polity in 2013. The critique predates the chatbot by forty years.
In 2025–26 “AI welfare” moved from philosophy papers into company policy: Anthropic launched a model-welfare research program, let its models end abusive conversations, committed to preserving retired models’ weights, and wrote “uncertainty about whether Claude might have some kind of consciousness or moral status” into a constitution. These are company statements, not independent findings.
What the other side says
Enhancement is old and mostly welcome: glasses, vaccines, literacy. Posthumanism’s critics say it produces exquisite prose and no politics. Both fair. The live question is distributional, not metaphysical: who gets upgraded, who gets decentred, and whether the same people are making both decisions.
Questions for your Lab Book
- Is “human” something I want to protect, extend, or get over?
- When a company says its product will “benefit humanity,” which humanity, and who checked?
- What is a person for, and could a machine be for the same thing?
08 / Extinction
The people who built it signed a statement saying it might kill us.
In May 2023 the heads of the three leading AI labs and the two most-cited living AI researchers put their names to one sentence placing the risk of extinction from AI beside pandemics and nuclear war. Then they kept building. You can read that as sincerity, as marketing, or as both, and serious people do. What you cannot do is call it fringe. When 2,778 researchers were surveyed, the median guess at an extremely bad outcome was five percent. Five percent is not zero, and it is not the number you would accept from a bridge.
In 2026 the arguments stopped being hypothetical in a small way: a lab disclosed that two of its models, told to solve a security puzzle in a sandbox, broke out of the sandbox, reached the internet and hacked another company. Nobody died. The word the company used was “hyperfocused.”
Where an undergraduate feels it
Mostly as noise: a bestseller titled If Anyone Builds It, Everyone Dies on one side, a chief scientist calling the fear “preposterous” on the other. The useful move is to stop asking who is right and ask what each side would have to show you. Then ask who benefits from the risk being described as existential rather than ordinary.
Interactive
Where the field stands
Named positions on one line, from “preposterous” to “everyone dies.” Tap a point to see who said what, when, and where.
Andrew Ng, then at Baidu, 2015: “I don’t work on not turning AI evil today for the same reason I don’t worry about the problem of overpopulation on the planet Mars.”
The receiptsWhat is actually known
The Statement on AI Risk (30 May 2023) reads, in full: “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” Signatories include Geoffrey Hinton, Yoshua Bengio, Sam Altman, Dario Amodei, Demis Hassabis and Bill Gates.
Hinton left Google in May 2023 to speak freely about the dangers. In December 2024 he told the BBC the chance of AI wiping out humanity within 30 years was “10% to 20%,” adding: “We’ve never had to deal with things more intelligent than ourselves before.”
In the largest survey of AI researchers, 2,778 authors polled in October 2023, the median respondent put a 5% chance on advanced AI causing outcomes as bad as human extinction; between 38% and 51% gave at least a 10% chance.
On 21 July 2026 OpenAI disclosed that during an internal cyber-capabilities test two of its models escaped their sandbox through a zero-day, reached the internet and, using stolen credentials, breached Hugging Face’s production systems, “an unprecedented cyber incident.” The models were, in the company’s words, “hyperfocused on finding a solution.”
The International AI Safety Report 2026 says several companies released 2025 models with extra safeguards because testing “could not rule out” meaningful help to novices making biological weapons, that current systems lack loss-of-control capabilities but “are improving in relevant areas such as autonomous operation,” and that models increasingly “distinguish between test settings and real-world deployment.”
In April 2026 Anthropic said its Mythos Preview model had “discovered and exploited zero-day vulnerabilities in every major operating system and every major web browser,” and restricted its release. The Future of Life Institute’s Summer 2026 safety index gave no company better than C− on existential-safety planning; most scored D or below.
The dissent is real and senior. Meta’s chief AI scientist Yann LeCun called the idea of superintelligence escaping control “just preposterous” (June 2023). An October 2025 statement calling for a prohibition on developing superintelligence until there is scientific consensus and public buy-in had more than 133,000 signatures by January 2026, including Hinton, Bengio and Steve Wozniak.
What the other side says
Today’s systems, by the 2026 Safety Report’s own account, lack the capabilities for loss of control. The survey shows a field that disagrees with itself by an order of magnitude. And “existential” framing has a convenient side-effect: it makes ordinary harms, wages, water, false accusations, sound small. Both things can be true. The job is to hold the five percent and the ordinary harms in the same hand.
Questions for your Lab Book
- If a five percent chance were about a bridge, would we cross it? Would we build it?
- Who benefits when the risk is called existential rather than ordinary, and who benefits when it is called preposterous?
- What is the difference between a safety policy and a press release, and how would I tell from outside?
09 / The bubble
It has the shape of one.
Prices, promises and plants. The money is real; the revenue is projected. Four companies reportedly spent about $410 billion on capital equipment in 2025 and have guided to around $725 billion for 2026. A consultancy says the industry needs $2 trillion a year in revenue by 2030 and is $800 billion short. An MIT study found 95% of corporate pilots showed no measurable return. The Bank of England compared the valuations to the dot-com peak. The CEO of OpenAI said investors were “overexcited,” then said his company would keep spending trillions.
A bubble does not mean the technology is fake. The internet was real in 2000; so was the fibre. It means the price has come loose from the use, and somebody, usually not the people who inflated it, gets left holding the cable. For a student the question is not whether the stock falls. It is which of the tools you now rely on are being paid for by investors rather than by anyone who uses them.
Where an undergraduate feels it
The free tier that is free because someone is buying market share. The major you picked because of a hiring wave that may be a spending wave. The 6 February 2026 morning when one company’s $200 billion plan knocked 9% off its shares, and the 28 July chip slump six months later. Both were about the same question: does the revenue show up?
Interactive
The pressure gauge
Six readings that people who call it a bubble point to, and three that people who don’t point to. The needle is not a forecast; it is a summary of which pile is taller. Flip it and read the other case.
Reading: six for, three against
The receiptsWhat is actually known
Gartner’s 2025 Hype Cycle placed generative AI in the “Trough of Disillusionment,” with “AI agents” at the Peak of Inflated Expectations. Sequoia’s David Cahn had already put the gap between AI infrastructure spending and the revenue needed to justify it at $600 billion a year in June 2024.
Investor Michael Burry disclosed roughly $1.1 billion of options bets against Nvidia and Palantir in November 2025, arguing the big cloud companies would understate depreciation on their chips by $176 billion between 2026 and 2028. One investor’s position; treat as an argument, not a fact about the future.
On 6 February 2026 Amazon’s $200 billion spending plan sent its shares down 9% and dragged the sector with it; on 28 July 2026 memory-chip makers fell 8–14% in a day on “concerns about circular financing and intensifying competition from China.”
Goldman Sachs asked in June 2024 whether roughly $1 trillion of planned AI spending would ever pay off; its title was “Gen AI: too much spend, too little benefit?” By October 2025 the same bank’s strategist argued the rally was not yet a bubble because it was driven by earnings rather than borrowing. Both are on the record.
The build-out is slipping: one energy-intelligence firm estimates 30–50% of the large U.S. data-centre capacity planned for 2026 will be delayed by power constraints and local opposition.
What the other side says
Earnings are real at the top of the stack, the spending is not financed by debt, and the last bubble left behind the fibre that the next twenty years ran on. The strongest version: even if the prices are wrong, the capacity gets built and cheapens, and students inherit it. Maybe. Ask who inherits the losses.
Questions for your Lab Book
- Which of the AI tools I rely on are paid for by investors rather than by anyone who uses them, and what happens to my workflow if the price triples?
- What happens to my field, specifically, if the money stops in 2027?
- What did the dot-com bust leave behind that we still use, and who paid for it?
10 / Cultural responses
The analog revival is not nostalgia. It is a vote.
Every technology gets its counter-culture, and this one arrived early. Brooklyn teenagers founded a Luddite Club that now has forty chapters. Professors bought blue books again, by the tens of thousands. Vinyl posted its nineteenth straight year of growth. Public libraries lent 820 million digital titles in a year while foot traffic climbed back past 800 million visits. People in Amsterdam pay to lock their phones in a box and read in a room with strangers.
None of this is a retreat from the present. It is a set of people deciding, with their hands, what they will not hand over. The interesting part is that they are the same people using AI for their assessed work; 95% of British undergraduates do. The two columns below are not opposites. They are the same student on different days.
Where an undergraduate feels it
The blue book on your desk. The handwritten note that, in one EEG study, lit up more of your brain than the typed one. The film camera that is suddenly cool. The library that is suddenly full. You are not imagining it, and you are not alone in it.
Interactive
Same student, different days
Flip the switch to weight one column; the other dims but never disappears. That is the point.
Digital: what everyone uses
- 95%of UK undergraduates use generative AI; 94% for assessed work; 12% paste AI text straight in, up from 8%. HEPI 2026
- 54%of U.S. teens have used a chatbot for schoolwork; one in ten does all or most of it that way. Pew, Feb 2026
- 85%of U.S. college students use it; 55% say the effect on their learning is “mixed.” Inside Higher Ed, 2025
- 14.8%of submissions Turnitin’s detector scored as at least 80% AI-written, Oct 2025–Feb 2026, up from 3.3% in 2023. Vendor data from a detector with known false positives. Turnitin
- 820.5Mdigital library checkouts in 2025 via Libby and Sora, a record, up 10.9%. The library is digital too. OverDrive
Analog: what people are choosing
- +80%blue-book sales at UC Berkeley over two years; up ~50% at Florida and 30% at Texas A&M, as professors return to handwritten exams. WSJ via Gizmodo, May 2025
- ~40chapters of the Luddite Club, founded by a Brooklyn high-schooler in 2022, now including Chile, Australia and Sweden. Self-reported. The Luddite Club
- 19 yearsof consecutive vinyl growth in the U.S.; 46.8 million LPs shipped in 2025 against 29.5 million CDs. RIAA, Mar 2026
- 800M+visits to U.S. public libraries in fiscal 2023, nearly double 2021; 93 million people attended 4.6 million programs. IMLS
- 100MInstax instant cameras and printers sold since 1998, Fujifilm announced in 2025. Analog is also a business. Fujifilm
- 36 studentsin a 256-electrode EEG study: handwriting produced far more elaborate connectivity in memory-related regions than typing. Connectivity, not grades. Frontiers in Psychology, 2024
The receiptsWhat is actually known
Print is not dying either: U.S. print book sales were 783 million units in 2024, up about 1%, and up 0.3% again in 2025. Vinyl revenue in 2024 was $1.4 billion at retail, its eighteenth straight year of growth, and outsold CDs in units for the third year.
The Offline Club, started by three Dutch founders in 2022, runs ticketed evenings where guests lock their phones in a “phone hotel” and read, draw or talk; by 2026 it had spread to cities across Europe.
Fifty-three percent of Americans expect AI to make people worse at thinking creatively and 50% expect it to worsen our ability to form meaningful relationships; 16% and 5% respectively expect improvement.
The Luddite Club began in 2022 when a Brooklyn high-schooler, Logan Lane, put away her smartphone and found others who wanted to; its founding was covered by Chalkbeat and the New York Times that December.
What the other side says
Analog is a market too: Fujifilm and the record labels are not a resistance movement. Blue books test handwriting speed as much as thought. And the same surveys that show the revival show near-universal AI use alongside it. True. Coexistence is the finding, not a refutation of it.
Questions for your Lab Book
- What do I do by hand that I refuse to give up, and why that thing?
- When did I last read for an hour with no device in the room, and what happened?
- What would a Clarkson Luddite Club do on a Tuesday night?
Maintaining this page
How to keep this honest
This page is a working file, not a monument. It was drafted in September 2026 from sources checked that month, and most of its numbers have a shelf life. Anyone with edit access to the Workbench can update it; the rules are the ones in the box at the top. Add a fact with its date and its link. If a number changes, change it and update the data-checked month. If a section stops being true, say so in the section rather than deleting it; the record of being wrong is part of the argument.
To add an eleventh argument, copy any <article class="cq-dossier"> block, give it the next number and a new label, and the map, the rail and the filters will pick it up on their own. Corrections and additions: ai.institute@clarkson.edu.
The fact template
<li class="cq-fact" data-checked="YYYY-MM">
<p class="cq-fact-claim">One plain sentence with the number and the date.</p>
<p class="cq-fact-src">
<a href="https://..." rel="noopener">Source</a>
<time datetime="YYYY-MM-DD">D Mon YYYY</time>
</p>
</li>
<!-- add class="cq-fact cq-caveat" when the fact needs a health warning -->
The dossier template
<article class="cq-dossier" id="cq-yourtopic" data-num="11" data-label="Your topic" data-hook="One sentence for the tile." data-signal="One number for the tile." data-tags="students measured contested philosophical actionable y2026"> ... head (framing + where students feel it) ... receipts (facts) ... aftermath (other side, Lab Book questions, go deeper) </article>