TL;DR

AI skepticism is not the reflex to declare every new model useless, fraudulent, or dangerous. It is the obligation to demand evidence that a system improves an outcome without hiding costs, shifting risk, or making people less capable. The strongest AI products are already useful; the weakest habit is treating impressive demos as proof of dependable value. Healthy skepticism gives us a better way to identify which products deserve scale—and which deserve scrutiny.

The setup

AI has moved from research labs into hiring software, customer-service portals, coding environments, classrooms, and hospitals. ChatGPT reportedly reached 100 million monthly active users within two months of its release in November 2022, and the subsequent competition among OpenAI, Google, Anthropic, Meta, and others turned model quality into a consumer product and an industrial policy.

The conventional argument is straightforward: resistance to AI is like resistance to electricity or the spreadsheet. Early systems fail, institutions adapt, and productivity eventually follows. Under this view, early adopters who learn to use the tools will gain an advantage while skeptics accumulate obsolete skills.

That argument contains a truth and a dangerous omission. AI systems can produce large gains on specific tasks. The omission is that a task improvement is not the same thing as a better organization, a more valuable worker, or a more just economy.

The International Monetary Fund estimated in 2024 that generative AI could affect about 40% of jobs globally and 60% in advanced economies. But “expose” does not mean “replace,” and even the IMF’s own analysis said the effect would vary sharply by sector and country. The distinction matters. Exposure means questions are being asked about workflow, wages, managerial control, and the distribution of gains.

The important question is not whether AI is powerful. It is whether deploying it makes people and institutions better off after accounting for errors, displaced work, degraded skills, legal exposure, and concentration of power.

That is where skepticism becomes necessary.

The argument

1. A convincing demo is not a reliable system

AI companies are exceptionally good at showing what a model can do. They are less forthcoming about how often it fails, which failures matter, and what happens when a user does not know when to distrust the output.

The distinction became legally concrete in the 2024 Moffatt v. Air Canada decision. Air Canada’s website gave a customer incorrect information about bereavement-fare rules. The airline argued that it should not be responsible because the answer came from a separate chatbot. British Columbia’s Civil Resolution Tribunal rejected that argument and held the airline liable. The decision did not establish a broad chatbot doctrine, but it exposed a basic problem: organizations cannot outsource accountability to an interface while retaining the profits generated by automation.

Companies have learned to say “human in the loop.” That phrase is useful only when the human has enough context, time, authority, and incentive to intervene. A reviewer who sees dozens of generated documents per hour is not meaningfully supervising the system. They are rubber-stamping a queue.

Healthy skepticism therefore asks for more than a feature list. What is the error rate on ordinary inputs? What happens on unfamiliar inputs? Can the system recognize its own uncertainty? Who is liable when it is wrong? Does escalation to a human work, or merely appear in a product screenshot?

A model that can answer 90% of routine questions may be valuable if the remaining 10% is rare, detectable, and inexpensive to review. It may be reckless if the 10% contains medical, legal, financial, or safety-critical errors that users cannot identify.

The demo says what the system can do. Skepticism asks what it does when nobody is watching.

2. Productivity claims are real—and routinely overstated

Skeptics often dismiss AI productivity gains because they are bad at reading experimental evidence. That is its own failure.

In a well-known 2023 study published in Science, Harvard researchers Shakked Noy and Whitney Zhang assigned 453 college-educated professionals to write tasks such as product descriptions, press releases, and short reports. Generative AI reduced completion time by about 40% and improved evaluated quality by 18%. The average writing task fell from roughly 11 minutes to 6 minutes.

That is a substantial gain. Pretending otherwise is not skepticism; it is ideological denial.

But the result also demonstrates why “AI will raise productivity” is an incomplete claim. The experiment covered a bounded category of writing performed by experienced professionals. It did not measure a company’s total output, long-term skill development, coordination costs, customer satisfaction, or the time required to detect confidently wrong answers. It did not tell us whether a faster draft was easier to revise, whether workers became better at the underlying skill, or whether managers captured the savings through increased targets rather than higher wages.

A separate study of 5,179 customer-support agents found that AI assistance raised average productivity by about 14%, with gains of roughly 34% among the least-experienced workers. Again, the result is useful. Again, it does not prove that every support organization should replace people. Faster ticket resolution can mean better service—or fewer difficult cases receiving the attention they require.

Klarna’s announcement that its AI assistant could perform the work of 700 full-time customer-service agents became a symbol of AI-driven labor reduction in 2024. The company later moved to hire human agents amid complaints about service quality. The most defensible conclusion is not that Klarna’s automation failed. It is that a headcount-equivalence claim is not the same as evidence of customer satisfaction, lower costs, or equal service.

Healthy skepticism preserves the measured task-level gains while refusing to inflate them into claims about whole occupations.

3. The labor question is about power as much as unemployment

The public argument is usually framed as “AI replaces jobs.” The more plausible near-term mechanism is subtler: AI changes the tasks inside jobs, reduces the number of workers needed for existing output, and shifts bargaining power toward the organizations that own the models.

Anthropic’s Economic Index, released in 2025 after analyzing millions of Claude conversations, found that AI use was divided roughly 57% toward augmentation and 43% toward automation. It also found that about 36% of occupations used AI for at least a quarter of their tasks, while only 4% used it for at least three-quarters. That is evidence of broad penetration without mass displacement.

The index has limitations. Conversations with Claude are not a representative sample of every workplace, and usage is not the same as causal impact. But it offers a more useful picture than anecdotes. AI is entering many jobs without immediately eliminating them.

That can still be bad for workers. Consider customer support, legal research, software development, or graphic design. If AI compresses the output expected from each junior employee, companies may use fewer junior workers while continuing to demand senior-level output. The first people harmed are not necessarily the most productive; they are the workers who historically learned by doing repetitive work under supervision.

There is another cost: loss of institutional memory. An organization that lets a model produce answers without maintaining human expertise can become unable to audit the system. The worker still exists, but only as a reviewer of output generated from data the organization no longer fully understands.

This is why “AI as augmentation” is not automatically a labor-friendly answer. Who receives the extra output? Who learns the necessary skills? Does the employee get time to question the tool, or is the tool always right because refusal is treated as poor performance?

4. Skepticism matters most where truth has become a product

Education provides a clear example of what happens when society adopts AI faster than its ability to evaluate it. Universities moved from debating whether to allow ChatGPT to building detection systems, honor codes, and assessment redesigns.

The detection systems were not reliable. In a Stanford study of 61 TOEFL essays written by non-native English speakers, every essay was labeled as AI-generated by at least one of seven tested detectors. Such a high false-positive rate makes “AI detection” inappropriate for consequential punishment.

This does not mean students should be allowed to submit generated work as their own. It means institutions need assessment methods based on process, oral defense, in-class work, and transparent standards—not a statistical guess about stylistic patterns.

The same pattern appears in search, customer support, and professional advice: AI can create a plausible answer before anyone checks whether it exists. A fabricated case can be syntactically perfect. A plausible citation can point to a nonexistent paper. A confident summary can erase the exception that determines the legal or medical outcome.

Skepticism is therefore not merely an economic position. It is a defense of the human ability to distinguish reliable knowledge from fluent text.

What the conventional view gets wrong

The conventional view says skepticism has already been incorporated into “responsible AI” reviews, model cards, safety teams, and staged rollouts. Sometimes it has. More often, those practices become institutional decoration.

A model card describing training data does not establish whether the product is accurate in a particular workflow. A pilot with 50 users does not reveal behavior at 50,000. A disclaimer saying “AI can make mistakes” does not prevent a customer from acting on a confident mistake. Governance fails when organizations substitute documentation for measurement.

The opposite convention is also wrong: the assumption that falling model prices will automatically deliver shared prosperity. The Stanford AI Index reported that the inference cost for performance equivalent to GPT-3.5 fell more than 280-fold between November 2022 and October 2024. Lower costs expand access and make bad deployments affordable too. Cheap errors at scale remain errors.

There is a legitimate counterargument to skeptical positions: skepticism can delay beneficial tools and protect inefficient incumbents. AI-assisted coding, accessibility tools, scientific search, and protein analysis already create value. DeepMind’s AlphaFold database contains more than 200 million predicted protein structures and has been used by millions of researchers. A philosophy that treats every AI claim as suspect would miss these advances.

But “some applications work” does not justify “all deployments deserve trust.” The answer to real benefits is better evaluation, not surrender.

The implications

For builders

Builders should stop using benchmark scores as the main evidence of product value and start publishing task-specific evidence: accuracy, severe-error frequency, human-review time, user correction rates, and performance under real distribution shifts.

They should also make escalation operational. If a model is uncertain, the interface should identify the uncertainty, preserve the source material, and route the case to someone with authority. Logging, rollback, and incident review should be default capabilities rather than premium trust features.

The defensible moat will not simply be access to a model. It will be trustworthy integration into a valuable workflow.

For users

Users should treat AI as a fast junior colleague, not an oracle. That means using it to outline alternatives, generate drafts, and expose missing questions—but verifying consequential claims against primary sources.

We should also protect skills that make verification possible. A person who can no longer write, calculate, program, or evaluate an argument may use AI efficiently for a few months while becoming dependent on it for years.

For regulators

Regulation should focus on outcomes and evidence rather than declaring all AI good or bad. Consumer protection, employment, education, and healthcare need rules suited to their risks: incident reporting, documentation, audit access, and transparency about consequential automated decisions.

The EU AI Act is a useful test because it introduces risk tiers and enforcement dates rather than a blanket prohibition. Its general-purpose AI obligations began applying in 2025, with broader transparency requirements scheduled for August 2026. The real test is whether those rules produce measurable reductions in harm rather than expensive compliance documents for systems that never mattered.

For the industry

AI companies cannot rely on market enthusiasm indefinitely. Enterprise buyers will eventually ask whether lower labor costs or faster output compensate for review expense, liability, customer churn, and regulatory risk. The strongest vendors will be the ones capable of saying, with evidence, where the system should not be used.

What I’m watching next

First, I’m watching whether independent evaluations become tied to actual work rather than public leaderboards. Leaderboard saturation, benchmark contamination, and cherry-picked prompts make scores increasingly poor procurement tools. A bank, hospital, or software team needs an evaluation on its own tasks, including the ugly edge cases.

Second, I’m watching labor data for changes in task mix and entry-level hiring. Aggregate employment can remain stable while junior roles, apprenticeship paths, and wage ladders deteriorate. The signal will come from occupation-level hiring, hours, pay, and the share of work delegated to AI—not from how many companies announce an “AI strategy.”

Third, I’m watching the economics of agents. As models become cheaper and more capable, they will attempt actions rather than merely return text. The next major incidents may involve software agents making reversible mistakes, or irreversible ones. Systems that can send email, move money, modify code, or update medical records raise the standard from “does the answer sound right?” to “what damage can one wrong action cause?”

The bottom line

AI skepticism is not a prediction that the technology will fail. It is a refusal to confuse capability with value. Some systems already save time and improve real work. Others create legal exposure, weaken expertise, and conceal their true operating costs behind “augmentation.” The answer is not unconditional adoption or reflexive rejection. It is evidence: measured error rates, meaningful human oversight, transparent costs, and consequences for the people whose labor is being reorganized. We should be enthusiastic about useful tools and ruthless about unsupported claims. That is not anti-innovation. It is how innovation becomes something more than a fundraising story.