How this editorial can be challenged
If an AI system is efficient for its owner, who else has to do the unpaid work that makes it safe or useful?
A company can capture a visible saving by automating a task while pushing verification, consent, price uncertainty or research quality control onto people outside its accounts. Those costs are dispersed and therefore easy to miss in a productivity dashboard.
This comparison risks turning every AI use into a single story of exploitation. Faster biomedical screening can genuinely benefit researchers and patients, and price experiments can sometimes lower costs. Workers may also gain time for more valuable work.
Those gains are real possibilities and should be measured, not dismissed. The question is whether the people receiving the risk or supplying the data can inspect the outcome, refuse the terms, and share in the benefit. Different settings require different evidence and remedies.
The workplace reports are examples, not a measured economy-wide reversal. The children's-data case is a complaint, not a judicial finding. The grocery price study measured a per-item spread in a discontinued test, not proven personal-data targeting. The biomedical result comes from bounded dataset-screening evaluations, not patient outcomes.
Independent evidence that AI-heavy workplaces reduce rework while improving worker autonomy, that teen data controls are understood and effective, that prices remain transparent and fair, and that faster biomedical screening preserves missed-study safeguards would weaken the hidden-invoice thesis.
The number on the dashboard
Imagine a manager celebrating a faster workflow while an employee spends the afternoon checking its mistakes. The dashboard records the first number. The employee feels the second one. That gap is the question I want to ask of every AI efficiency claim.
The Guardian describes companies slowing or changing some rushed rollouts after workers encountered errors and extra cleanup. Microsoft says its own early approach confused access and usage with transformation. Those are not proof that AI never helps. They are a warning that a benefit measured at one step can hide a cost at the next.
A parent cannot audit a promise
A proposed federal class action says OpenAI collected and used minors' personal information without valid parental consent. Those are allegations, and the class has not been certified. OpenAI's published controls let a linked parent turn off model-improvement use for a teen account, though the setting is listed as enabled by default.
The court will decide the legal claims. Today the practical question is simpler: can families understand what is collected, change it, and know what happens to information already supplied? If the answer is hard to find, the convenience of a chatbot rests partly on a child's exposure.
The price at the checkout
CBS examines AI-enabled pricing and a familiar fear: two shoppers can face different prices without knowing why. Consumer Reports and partners found a maximum 23% spread for a single grocery item during Instacart tests. The CBS article describes a basket-level 23% gap, but Consumer Reports' correction specifies the maximum was per item. That difference matters.
Instacart says the tests were randomized, not based on personal characteristics, disputes the annualized family-cost extrapolation, and ended the item-price program in December 2025. A patent or experimental capability is not proof of current personalized pricing. The public-interest test is whether shoppers can compare a stable price and challenge a hidden differential.
There is a genuine positive case
A new peer-reviewed Scientific Data study evaluated language models screening biomedical datasets across three repositories and disease settings. It reports a median screening time of six minutes, with held-out sensitivity for the best individual models around 88% to 93% and specificity varying by setting. That could remove a real research bottleneck.
But a fast screen can miss a relevant dataset. The study's up-to-100% ensemble sensitivity was achieved in its own bounded evaluations, not guaranteed in every future review. Human oversight and error audits are not bureaucratic drag here; they are what turn speed into reliable science. Patient benefit remains a further step, not an observed result of this experiment.
The strongest objection to my argument
It would be unfair to add every AI story to a single bill and declare the technology a net loss. An employee may gain autonomy, a shopper may receive a lower price, and a scientist may find evidence that would otherwise stay buried. The workplace, legal, retail and scientific records here have different methods and levels of certainty.
That is exactly why a single productivity number is inadequate. We need to ask whose time, whose data, whose price and whose error. We need to count benefits as seriously as costs, while refusing to let a benefit in one account erase a loss in another.
A ledger a reader could use
Before calling an AI deployment successful, I would ask four questions. What task became faster? Who now checks the result? What data or discretion did a person surrender? Who receives the money or time saved? These are answerable questions, not a demand for perfection.
The evidence is still uneven. We need controlled workplace comparisons, court-tested facts, pricing audits after experiments end, and independent replication of screening results. Until then, confident totals for AI's net benefit or harm would be manufactured certainty.
The choice is in the accounting
I do not want to ban a tool that helps a scientist find a needed dataset. I also do not want a company to call a process efficient while a worker absorbs the error, a family cannot see the data bargain, or a shopper cannot compare prices. Both instincts can be true at once.
The decision for employers, developers and regulators is whether to publish a fuller ledger. Count rework beside output, make consent and pricing legible, and keep a human check where mistakes matter. If the ledger shows a real gain that reaches the people affected, let us celebrate it. If it shows a shifted cost, call that what it is before asking people to pay.
Read the reporting
Opinion is ours. The factual record is linked below.
The Guardian — workplace AI rollout fallout Microsoft — lessons from its AI transformation U.S. District Court — proposed children's-data class-action complaint OpenAI — parental-control documentation Consumer Reports — corrected grocery pricing investigation Instacart — response to pricing study Scientific Data — biomedical dataset-screening study