How this editorial can be challenged
When an AI company calls some harm an acceptable price of progress, who is empowered to accept it on behalf of voters, ratepayers and people recorded in public?
The party choosing speed usually captures the immediate benefit while exposure travels across institutional boundaries. A lab receives adoption, a campaign receives cheap reach and a data-center operator receives capacity; a voter must sort synthetic speech, a household shares grid costs, and a bystander may be recorded. Treating these as one aggregate bargain conceals the transfer and weakens each exposed party's ability to object.
Every widely useful technology has costs, and demanding individual consent for every external effect could freeze innovation. Broad AI access may democratize creation, power expansion can benefit all users, and a narrow ban on AI glasses could also deny accessibility tools to people who need them.
The answer is not a veto over every feature. It is to separate ordinary, reversible inconvenience from concentrated, hard-to-reverse exposure. Publish cost allocation, use independently testable thresholds, protect consent in intimate spaces, and allow narrow exceptions for genuine accessibility uses. Then the claimed public benefit can be compared with the actual parties bearing its cost.
Politico's interview is a reported account of a philosophy, not a quantified risk budget or proof of a particular harm. Wesleyan counted a minimum of 164 AI-enhanced political ads, not their causal effect on votes. Goldman's 66 GW is a May forecast, not measured 2027 demand or an estimate of household bills. Norway has announced a proposal, not enacted a ban. The Bank of Japan describes opposing channels and says their magnitude and timing remain uncertain.
Transparent evidence that campaign AI improves voter understanding, that incremental data-center loads pay their full grid costs, and that wearables can reliably protect bystanders without excluding accessibility users would weaken our concern about unconsented transfer. Documented local bills, persuasion outcomes and privacy incidents would sharpen it.
The part of the bargain nobody signed
In Politico's interview, OpenAI's chief executive makes an argument many builders believe: a world that wants AI's benefits will have to tolerate some bad outcomes, while preserving people's agency. I understand why zero harm is an impossible standard. Still, the sentence leaves me looking for the absent person: the one exposed to the bad outcome who did not choose the product.
A campaign, an electricity customer and a passerby are not interchangeable entries in a social cost-benefit equation. One group decides; another may inherit the consequence. That is the original question of this editorial, not whether a single quote proves a company is reckless.
A campaign can buy reach; a voter does the sorting
Fox News reports that generative tools are reducing the time and cost of political ad production. Wesleyan Media Project had identified at least 164 AI-generated or AI-enhanced political ads by September 4. Its tally is a minimum, not a measure of deception or electoral effect. Fox's examples also show why a viral ad is not automatically a winning one.
The external cost is epistemic: a voter has to determine whether an apparent voice, scene or endorsement is authentic. If the tactic fails electorally, the voter still paid attention to it. That is why an ad's view count cannot stand in for a public benefit or a public harm.
A forecast can become a bill before it becomes a reality
Goldman Sachs projected in May that U.S. data-center power demand would grow from 31 GW in 2025 to 66 GW in 2027. It also expected only roughly 50 to 60 percent of scheduled near-term capacity to arrive on time. The figure describes a forecast with explicit delays and uncertainty, not today's meter reading and not proof that any household's bill will rise by a given amount.
But forecasts can influence real decisions now: interconnection queues, generation plans, finance and rate design. The distributional question is whether the large new load pays for the specific capacity it requires or whether ordinary customers help underwrite speculative infrastructure. That must be answered in tariffs and contracts, not in a slogan about abundant intelligence.
A camera worn by one person can cost another person privacy
Norway's government announced a plan to propose a temporary restriction on AI glasses in selected places, including spaces with children and places where privacy is especially sensitive. It says it does not want a total ban and will consider exceptions for vulnerable groups. The bill has not yet been enacted; even the device scope remains to be worked out.
That openness is the point. An accessibility use may be meaningful. So is the fear of being recorded without knowing it in a clinic, changing room or playground. Neither interest vanishes because someone calls the device innovative. A place-specific rule can test how to preserve the first without forcing the second onto everyone nearby.
The macro picture resists a tidy verdict
In new remarks, the Bank of Japan's deputy governor describes AI as an immediate positive demand shock that can lift prices, a possible source of later productivity growth, a force in stock and bond markets, and a potential restructuring of labor. He does not pretend the relative size or timing of those effects is known. For now, he tentatively sees demand arriving before supply.
That uncertainty is a reason to separate channels and exposed groups. A worker who loses a skill premium, a firm that sells power equipment and a household facing a grid upgrade do not experience one average outcome. We need measurements that can disagree with the optimistic story as well as the fearful one.
The paradox of useful progress
A technology can be broadly beneficial while particular uses transfer a cost to someone who never agreed. The stronger the benefit, the more tempting it is to bury that transfer in an aggregate number. I want the opposite: a public ledger that identifies the decision-maker, beneficiary, exposed group, measurable outcome, remedy and right to refuse in each high-consequence setting.
If that ledger shows that AI-enabled ads inform rather than confuse, that new data centers pay their grid costs, and that wearable safeguards protect bystanders while preserving accessibility, the case for adoption grows stronger. If it shows costs repeatedly landing elsewhere, the words 'acceptable risk' need a different speaker. The paradox is that honest accounting may speed trustworthy adoption precisely by making some deployments slower.
Read the reporting
Opinion is ours. The factual record is linked below.
Politico — interview on AI benefits and risk Fox News — AI in political advertising Wesleyan Media Project — AI-enhanced ad count Goldman Sachs — U.S. data-center power forecast Norwegian government — proposed AI-glasses restriction Bank of Japan — AI, big data and monetary policy