To the left,
If our politics of artificial intelligence ends with refusal, the decisions do not stop. Employers still decide how the systems enter workplaces. Technology companies still decide what gets built and whose data pays for it. Police and immigration agencies still decide how automated analysis expands surveillance. Utilities and local governments still decide who pays for new data centers.
Refusal can be a moral boundary. It is not, by itself, a governing program.
This letter makes a narrower claim than the original version did. The left is not a single bloc, and many labor organizers, privacy advocates, environmental groups, artists, and researchers are already doing serious work on AI. The strategic mistake is not criticism. The mistake is allowing criticism to substitute for institutions, rules, technical competence, and public alternatives.
Start with what the evidence says
Public concern is substantial. In a February 2026 survey of 5,119 U.S. adults, Pew Research Center found that 40% expected AI to have a negative effect on society over the next twenty years, compared with 16% who expected a positive effect. Sixty-three percent said AI was advancing too quickly, and 71% expected it to make their personal information less secure. Those are not fringe concerns.
Some AI marketing is demonstrably deceptive. The Federal Trade Commission’s Operation AI Comply brought cases involving an alleged fake-review generator, a service marketed as “the world’s first robot lawyer,” and business-opportunity schemes that used AI claims to promise consumers large returns. That record supports scrutiny of particular companies and claims. It does not justify saying every AI product is fraudulent or useless.
The labor evidence is serious, but it is not simple. Challenger, Gray & Christmas reported that U.S. employers cited AI in 54,836 announced job cuts during 2025. That is a count of employer announcements, not an independent causal audit. At the same time, the Yale Budget Lab’s January 2026 update found no clear relationship between measures of AI exposure or use and changes in economy-wide employment or unemployment through December 2025. The International Labour Organization estimates that one in four jobs worldwide has some exposure to generative AI, while emphasizing that transformation of tasks is more likely than full replacement for most occupations.
These findings can coexist. Companies are already attributing some planned cuts to AI. Some occupations and workers face more pressure than others. The aggregate labor market has not yet shown the economy-wide disruption predicted by the most alarming forecasts. Exposure is not displacement, an announced cut is not proof of cause, and an absence of aggregate effects so far is not a guarantee about the future. A credible politics should be able to hold all four statements at once.
The resource cost is also real. Lawrence Berkeley National Laboratory estimated that data centers used about 4.4% of U.S. electricity in 2023 and could use between 6.7% and 12% by 2028. The International Energy Agency reported that global data-center electricity demand rose 17% in 2025, with faster growth at AI-focused facilities, and projected total data-center consumption to double by 2030. Not all data-center use is AI use, and the Government Accountability Office warns that separating the energy and water used by generative AI from other computing can be difficult. That uncertainty is an argument for disclosure and measurement, not for pretending the burden is unknowable or assigning every server to AI.
The surveillance risk does not begin with generative AI. It begins with institutions that already collect data and exercise coercive power. Documents obtained through public-records requests showed that Palantir software was used in operations by ICE’s Enforcement and Removal Operations division, despite the company’s earlier effort to distance its work from deportations. Adding faster classification, link analysis, or prediction to an existing surveillance system can expand its reach. The central questions are who has authority, what data can be used, whether a person can challenge a decision, and what uses are prohibited.
What the evidence does not say
Documented harms do not establish that every use is harmful. Documented benefits do not establish that the technology is good in general.
A field study of 5,179 customer-support agents found that access to a generative AI assistant increased issues resolved per hour by 14% on average, with larger gains among novice and lower-skilled workers in that setting. A randomized trial in one Harvard physics course found that students assigned a carefully designed AI tutor had greater learning gains than students assigned an active-learning class lesson.
Both studies are meaningful. Both are narrow. Neither proves that a generic chatbot improves every job or classroom. They do show why “the tools do nothing” is not a defensible position. The serious question is which systems improve which tasks, under what conditions, with what error rate, at whose expense, and under whose control.
This distinction matters because AI is not one object. A model that helps a worker retrieve a document is not the same use as a model that scores the worker for dismissal. A tutor that gives a student practice is not the same use as software that falsely accuses the student of cheating. A tool used to locate a paper is not the same use as generated prose published under a human byline.
The Hank Green dispute was a standards problem
The reaction to Hank Green’s AI use is useful because it exposed how poorly developed our public standards remain.
Green said that he had been relying too heavily on AI as a research aid to locate papers and other resources. He also said the claims in his videos did not originate from accepting a chatbot’s assertions, that the words remained his, and that the tool had begun to distort his research process. He apologized and said his main channel might pause. The immediate accusation that a phrase in a video had been generated by AI was, according to Green, wrong: he said the phrase was an unscripted remark. Axios later described the episode as part of a broader tendency to treat very different forms of AI use as equivalent.
There were legitimate questions here. How should a creator disclose AI-assisted research? Can model-generated recommendations narrow what a researcher sees? What verification is required when a system is known to invent citations? Did a tool change the work even if it did not write the final script? Green himself raised versions of those concerns.
The overreach came when some commentary treated research assistance, ghostwriting, plagiarism, and labor replacement as the same act. They are not. Defending that distinction does not require defending every choice Green made. It requires standards proportionate to the conduct.
A workable standard asks:
- What function did the system perform?
- Did generated material reach the audience, and was that use disclosed?
- Were factual claims checked against the underlying sources?
- Did the use replace paid labor or change working conditions?
- What private or sensitive data entered the system?
- Who remains accountable when the output is wrong?
Those questions produce better judgments than either “AI touched it, so it is tainted” or “a human checked it, so it is fine.”
Power is already organizing
While the public argues about personal tool use, companies and their allies are spending to shape policy. Wired reported that Leading the Future, a pro-AI super PAC supported by technology executives and investors, said it had received $140 million in contributions and commitments, with $51 million available as of April 2026. A related nonprofit paid influencers to promote U.S. AI and frame Chinese AI as a threat. Its recruiters sought both left-leaning and right-leaning creators.
That is the structural imbalance. Corporate interests are not abstaining. They are financing infrastructure, lobbying, political advertising, research, and public narratives. A movement that limits itself to consumer refusal leaves those decisions to the institutions with the most money and the least democratic accountability.
Engagement does not mean adopting the industry’s claim that deployment is inevitable or that regulation is surrender. It means contesting ownership, design, procurement, workplace use, and enforcement before private defaults harden into public infrastructure.
What a serious left program could demand
Worker power before deployment
Workers and unions should receive notice before AI systems are introduced, access to impact assessments, and the right to bargain over monitoring, workload, staffing, evaluation, and job redesign. A worker should be able to challenge an automated decision and reach a human with authority to reverse it. If a system produces measurable gains, workers should share them through higher pay, shorter hours, stronger benefits, or ownership, not only through layoffs and higher output targets.
Public capacity, not permanent dependence
Public institutions need the ability to evaluate and operate technical systems without relying entirely on the vendors they regulate. The National Science Foundation’s National Artificial Intelligence Research Resource is one existing model for widening access to compute, data, models, and expertise for researchers and educators. It is a public-private program, not a complete substitute for publicly owned infrastructure, but it shows that access to advanced computing does not have to be limited to a handful of firms.
Ownership also belongs in the debate. Senator Bernie Sanders’s American AI Sovereign Wealth Fund Act proposes a one-time 50% tax paid in stock by the largest AI companies, with the shares placed in a public fund overseen by an independent commission. The sponsor’s own estimate puts the fund at $7 trillion, with possible annual payments above $1,000 per person. Those figures are projections from the bill’s sponsor, not guaranteed outcomes. The proposal is sweeping and open to serious legal, economic, and governance objections. It is also a concrete attempt to answer who should own the gains.
Rights that technology cannot waive
Some uses should be prohibited, not merely audited. A serious program should define clear boundaries for biometric surveillance, predictive policing, immigration enforcement, benefits decisions, hiring, firing, housing, credit, and health care. Where automated systems are permitted, due process, data minimization, independent testing, public records, and an enforceable right of appeal should be baseline requirements.
Environmental accounting with consequences
Data-center operators should disclose projected and actual electricity and water use, the source of that power and water, backup-generation emissions, and the public infrastructure costs attached to a site. Local communities need a meaningful role in siting decisions. Utilities should not shift grid expansion and reliability costs from large operators onto households without transparent public review.
Evidence and disclosure standards
AI-assisted research should lead readers back to primary material, not replace it. Publishers, schools, governments, and creators need rules that distinguish brainstorming, retrieval, translation, editing, generation, and automated decision-making. Disclosure should describe the material use that affected the work, not rely on a vague badge that says “AI assisted.” Accountability remains with the person or institution that publishes or acts on the result.
Criticism is the beginning
The choice is not acceleration or abstention. It is private control or democratic contest, weak labor rights or strong ones, hidden resource use or measurable obligations, surveillance by default or enforceable limits, concentrated gains or shared gains.
AI systems are already entering workplaces and public institutions. Their long-term effects remain uncertain, and anyone claiming certainty is selling something. That uncertainty does not excuse passivity. It increases the value of institutions that can measure outcomes, stop harmful uses, revise rules, and distribute gains.
The left’s strongest tradition is not technological optimism. It is the insistence that power be named, organized, and made answerable to the people who bear its costs.
Apply that tradition here.
Suspicion is justified. Build the program that makes it useful.
Sources and scope
This essay relies on public polling, government enforcement records, labor research, energy analysis, public statements, and reporting available as of August 14, 2026. It distinguishes measured outcomes from exposure estimates, employer announcements, sponsor projections, and forecasts. Links are provided at the claim level so readers can inspect the underlying evidence.
