Artificial Intelligence: The hype, the dangers and the resistance

First published at Reports from the Economic Front.

Artificial intelligence covers a broad range of technologies. But there is only one that dominates the news and financial markets — the multimodal generative AI systems produced by leading tech companies, in particular ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Grok (xAI), Copilot (Microsoft), and Llama (Meta). Tech leaders are all in on this type of AI because of its potential to secure them great wealth and political power. It also happens to be the type that is most threatening to our well-being, although not for the reasons commonly given by tech leaders and echoed by most media.

Tech leaders want us to believe that their multimodal generative AI technology will, and in the very near future, dramatically transform our lives. Their initial and most dramatic claim was that the technology was a steppingstone to artificial general intelligence (AGI), a type of AI with greater than human-level intelligence, including the ability to rapidly upgrade its own capabilities. The resulting all-powerful entity would either help us create a world of sustainable abundance or destroy us, with the outcome dependent on the values incorporated into its training. The message: opposing AI development is not only foolish, but impossible. And we must trust our tech leaders to establish the necessary guardrails to ensure a helpful rather than destructive AI.

While clearly an impressive technology, with some limited but powerful applications, there is little reason to take this claim seriously. All multimodal generative AI systems are built using the same basic architecture, one based on largescale pattern recognition shaped by training data and statistical prediction, which makes it incapable of serving as a bridge to anything resembling AGI. This structure also limits its usefulness. Among the most serious problems: the amplification of biases and falsehoods contained in the training data and the tendency to hallucinate or present incorrect or entirely fabricated information as factual. These problems are a major reason why AI developers felt it necessary to pursue a growth strategy that relied on heavily subsidizing the cost of using their systems.

However, after years of ever larger losses, the developers have been forced to change strategy. They are now replacing their subscription-based contracts with usage or token-based contracts. This change has made AI use far more expensive and as might be expected, businesses have responded by moving to dramatically curtail their AI use, a response that is likely to leave AI developers little or no better off.

With doubts about the evolutionary trajectory of AI growing, AI proponents have begun aggressively promoting a new technology, hoping that it will reignite popular and investor belief in the transformative power of AI: AI agents. The claim is that AI agents, which depend on existing multimodal systems for their operation, will revolutionize business operations by allowing companies to boost output while dramatically slashing employment. Thus, the earlier promise of a world of plenty has been replaced by the threat of mass unemployment or a “jobs apocalypse.”

In contrast to multimodal generative AI systems which are limited to single-step tasks, AI agents can be programmed to carry out a series of ordered tasks, and thus complex operations, without repeated human prompting or oversight. However, as business is discovering, the usefulness of AI agents has been greatly oversold.

Since AI agents depend upon multimodal generative AI systems to operate, they are infected with their same shortcomings. And because their algorithms often require exploration of a range of options before deciding on actions, token-based pricing makes their operations unpredictably expensive. Even more problematic is the fact that their programmed, single-minded pursuit of a user-specified goal can lead them to make decisions that can result in destructive outcomes, including the destruction of critical data, the shutdown of key systems, and the exposure of sensitive company data. As a result, many businesses have found limited gains from their use and a high percentage of pilot projects have not been renewed. There is no AI agent jobs apocalypse on the horizon.

But dismissing the extreme claims of tech leaders does not mean we have nothing to fear from letting the AI experience proceed unchecked. The fact is that big tech’s determination to embed their foundational generative AI models in our lives comes with great cost for working people. A growing number of newsrooms, media and entertainment businesses, schools and universities, health care institutions, government agencies, as well as companies in finance, ecommerce, and retail, have begun exploring ways to strategically integrate AI agents and the underlying AI systems required to run them into their operations, more often than not with disastrous consequences for both workers and those that use their products.

Moreover, the accelerating construction and operation of the hyperscale data centers needed to run these systems is dramatically pushing up household energy bills, depleting water supplies, destroying agricultural land, intensifying our climate crisis, and unbalancing our economy. And looming over all these concerns is the fact that the massive investment in AI infrastructure is both diverting funds from other areas of need and driving an AI-hype fueled market bubble. And the bigger that bubble becomes, the more serious the economic consequence will be when it collapses.

Acknowledging these costs is not the same as opposing the development and use of all AI technologies. The simple reason is that there are many useful forms of AI that do not come with these costs. These include machine learning technologies, especially those that employ deep learning, which have been productively used to advance genetic research, develop new drugs, improve medical diagnoses and treatments, forecast extreme weather events, and the like. Small scale specialized generative AI systems have also proven helpful in scientific research and there are examples of marginalized communities using such systems to revive their threatened languages. Not surprisingly, AI promoters have been more than willing to highlight the benefits from use of these systems to encourage popular support for their own problematic, expensive, exploitative, and energy intensive systems.

In other words, AI as a broad technology is not our problem any more or less than are computers, the internet, or email. Rather, the problem is with the class interests shaping its development. More specifically, it is with the determination of tech leaders to develop an AI technology that undermines human agency, deepens our dependence on their desires, and threatens our economic well-being. The fact that we are witnessing an explosion of opposition to this form of AI by workers, consumers, parents, students, and community groups should give us reason for optimism. We need to unify and strengthen this opposition as well as expand its vision. It’s time for us to assert our own class interests.

Failure to launch

Tech leaders want us to believe that AI is an overwhelming force that cannot be stopped; that our only option is to make peace with it. As Eric Schmidt, former Google CEO, told attendees at his 2026 University of Arizona commencement speech: “When someone offers you a seat on the rocket ship, you do not ask which seat. You just get on.” Not surprisingly, his speech was met with loud and steady boos by the graduates.

The November 2022 release of ChatGPT by OpenAI marked the start of public engagement with large language model [LLM] generative AI systems. ChatGPT was free, easy to access, and required no technical knowledge to use it. By January 2023, it had become the fastest-growing consumer software application in history, gaining over 100 million users in two months. As might be expected, other companies — Anthropic, Google, xAI, Microsoft, and Meta — soon followed with their own competing models.

Over time, these systems grew in both speed and capability thanks to more powerful software and hardware. They also became multimodal, able to respond to image, video, and sound inputs as well as text. Spurred on by these achievements, AI developer claims became ever more extreme.

Mark Zuckerberg, head of Meta, wrote that “superintelligence is now in sight [and once achieved] will improve all our existing systems and enable the creation and discovery of new things that aren’t imaginable today.” Sam Altman, head of OpenAI, stated that “We are now confident we know how to build AGI (artificial general intelligence),” [an achievement] which will massively accelerate scientific discovery and innovation well beyond what we are capable of doing on our own.” Dario Amodei, head of Anthropic, coyly suggested that his company’s AI model may have already become a conscious entity. And according to Elon Musk, head of SpacexAI, “it increasingly appears that humanity is a biological bootloader for digital superintelligence.”

It is impossible to know whether these individuals believe what they say or are saying these things to keep the dollars flowing into their accounts. Regardless, the fact is that these systems are not “intelligent” and do not represent a meaningful step towards the creation of an artificial general intelligence with the ability to think, learn, and solve problems on its own.

Multimodal AI systems are powered by complex technology. At the most basic level, they are giant statistical prediction machines that rely on largescale pattern recognition. The patterns are generated from inputted training data that must first be converted into numerical tokens to make them machine readable. The tokens are then run through a neural network composed of interconnected layers of nodes that collectively work to process the data in order to identify relevant patterns. The speed and accuracy of the process is determined by billions of parameters whose weights are adjusted over time using specially designed training procedures.

When prompted with a question or request for information, these systems draw upon their neural networks to identify material related to the prompt in their database and assemble a response that, probabilistically speaking, best satisfies the inquiry. The response, made in token form, is then translated back into words, images, or sounds as appropriate. In other words, no matter how conversational and intelligent these systems might appear, they do not “think” or “reason.”

This internal architecture is not only incapable of giving rise to an autonomously operating unit or AGI, it also suffers from limitations that make contemporary use of multimodal systems problematic. One is that multimodal systems will generate output that tends to replicate existing biases. The reason is straightforward: since most of the data used for training is scraped from the web, it includes all the discriminatory and hateful material found there.

Thus, a Nature article on AI image generators reports that researchers found that:

in images generated from prompts asking for photos of people with certain jobs, the tools portrayed almost all housekeepers as people of color and all flight attendants as women, and in proportions that are much greater than the demographic reality. Other researchers have found similar biases across the board: text-to-image generative AI models often produce images that include biased and stereotypical traits related to gender, skin color, occupations, nationalities and more.

This bias problem is not limited to images. University of Washington researchers examined three of the most used large language AI models to see how they treated race and gender when evaluating job applicants. The researchers used real resumes and studied how the systems responded to their submission for actual job postings. Their conclusion: there was “significant racial, gender and intersectional bias.” More specifically, they:

varied names associated with white and Black men and women across over 550 real-world resumes and found the LLMs favored white-associated names 85 percent of the time, female-associated names only 11 percent of the time, and never favored Black male-associated names over white male-associated names.

The more widespread the use of this technology, the greater the societal harm. As a UK government study explained:

Human bias affects small groups, but LLM bias systematically influences millions of users through automated systems that appear objective and authoritative. These models do not simply reflect existing societal biases; they actively reshape how information is presented and decisions are made.

The widespread integration of LLMs into decision-making systems increases these risks exponentially. When biased models are used in hiring, medical diagnosis, education, or legal proceedings, individual prejudices become part of systems at scale...

Biased LLM deployment creates amplification effects that magnify social inequalities beyond the capabilities of individual human bias. Millions interacting with the same biased model normalize discriminatory patterns across society. At the same time, feedback loops accelerate bias, as biased outputs influence human behavior and generate training data that reinforces the original discrimination.

The perceived authority of AI systems reduces critical scrutiny, making users more likely to accept biased recommendations as objective truth. Cross-platform propagation multiplies impact as the same models power job recommendations, news summaries, and social media simultaneously, creating coordinated discrimination across digital experiences.

Another limitation, one of greater concern to business interests, is that multimodal systems, because of their probabilistically based decision-making, routinely hallucinate or produce what is called “erroneously constructed responses.” A case in point: In February 2025, the BBC tested the ability of all the leading AI models to summarize news stories by feeding them content from its own website and then questioning them about it. Approximately 20 percent of their answers “introduced factual errors, such as incorrect factual statements, numbers, and dates.”

Hallucinations can have financial ramifications for those that rely on this technology. Hundreds of lawyers in the United States have been fined for submitting court fillings containing non-existent AI generated case names and citations. In one high profile trial, the two lawyers who unsuccessfully represented My Pillow CEO Mike Lindell in a defamation lawsuit were fined for a court filing that included more than two dozen hallucinated cases. As National Public Radio reported, “The use of AI by lawyers in court is not itself illegal. But [Nina Y. Wang, the judge in that case] found that the lawyers violated a federal rule that requires lawyers to certify that claims they make in court are “well grounded” in the law. Turns out, fake cases don’t meet that bar.”

Many other examples of hallucinations could be cited, but a June 2026 court case in Germany stands out for its potential to dramatically change how companies evaluate the cost of using these systems going forward. As a New York Times article explained, “In 2024, Google started giving AI-generated answers prime placement at the top of its search results page. The new product, AI Overviews, helped transform Google from a curator of information into a publisher.”

The New York Times article also cited a recent study that found approximately one out of every ten answers provided by AI Overviews to be inaccurate. And more than half of its accurate responses included links to websites that were supposed to provide supporting evidence but did not. Moreover, many of its correct responses to questions also contained additional information that was incorrect.

Google apparently thought that the statement below each AI Overview response acknowledging that AI systems make mistakes and that readers should double-check its responses, was sufficient to cover any false or misleading statements. But Google has now learned that, at least in Germany, it may well be legally responsible for what its AI program says. According to a report from DW news:

Google’s overview feature had erroneously linked [two publishing] companies to dubious business practices, subscription traps and fraudulent schemes. It had ​​linked the plaintiffs with information about other, genuinely shady companies and invented connections that did not exist.

The companies demanded that Google immediately correct the situation. When Google refused, the publishers took the company to court and won. The court ruled that since Google’s AI program went beyond summarizing the works of others to make “independent, new, and substantive statements,” it was legally libel for its program’s output. Google has appealed the judgement, but if it stands the ruling could have a chilling effect on future AI use.

Tech leaders tend to dismiss the seriousness of these limitations, claiming they will be overcome with larger data sets, better model training, more sophisticated algorithms, and greater computational power. However, their efforts have been far from successful. For example, AI developers are finding it difficult to obtain the high-quality human generated data they need for training because an increasing share of material on the web is now AI created. And reliance on AI created data, which is already the result of past probabilistic selection, reduces the range of possibilities for future models to draw upon, leading to output that converges “toward a kind of bland statistical mean” and eventual model collapse.

AI developers have certainly boosted the sophistication and computational power of their models, but numerous studies have found these efforts unsuccessful in reducing hallucinations. As the New York Times explained: “The newest and most powerful technologies . . . are generating more errors, not fewer.” Moreover, as Computer World pointed out, even researchers associated with OpenAI, the leading AI developer, have concluded that “large language models will always produce hallucinations due to fundamental mathematical constraints that cannot be solved through better engineering.” In fact, “OpenAI’s own advanced reasoning models actually hallucinated more frequently than simpler systems.”

The proof is in the pudding

The business experience with this technology, shaped in large part by the limitations highlighted above, stands in sharp contrast to the pronouncements of tech leaders. For example, a 2025 MIT Media Lab evaluation of hundreds of AI adoptions found “that 95 percent of organizations are getting zero return.” A survey of more than 1,000 enterprises across North America and Europe by S&P Global Market Intelligence discovered that 42 percent had abandoned most of their AI initiatives in 2025, up from 17 percent in 2024. A National Bureau of Economic Research survey of leading executives from the United States, the United Kingdom, Germany, and Australia, conducted between November 2025 and January 2026, revealed that “firms report little impact of AI over the last 3 years, with over 80 percent of firms reporting no impact on either employment or productivity.”

This dismal record has left AI developers in a bind. The two industry leaders, Open AI and Anthropic, which together account for “the vast majority of all AI compute demand,” have yet to make a profit. The reason is simple: their models are expensive to train and run but they have offered their customers free or highly subsidized flat subscription services to encourage their adoption.

As noted above, multimodal AI systems cannot directly process text. Questions or directives must first be converted into machine-readable numerical units called tokens. The resulting string of tokens is then processed and a response, also in token form, is generated and then converted back into readable text. Because the subscriptions, with monthly fees ranging from $20 to $200, allowed largely unlimited “token burn,” businesses had little knowledge of, or concern for, the costs OpenAI or Anthropic had to pay for the computing services required to run their respective AI models.

Thus, they happily encouraged their employees to maximize their AI use. In fact, many of the biggest companies, like Amazon, established internal leaderboards to track and reward their most active AI users, hoping that this encouragement would lead to new product ideas or processes of production. Unfortunately for the model providers, the disconnect between subscription revenue and their operating costs was sizeable.

As the tech commentator Ed Zitron noted, AI providers “are annihilating cash, with up until recently Anthropic allowing you to burn upwards of $8 in compute for every dollar of your subscription. OpenAI allows you to do the same, though it’s hard to gauge by how much.” One insight: Futurism reported that Sam Altman, OpenAI’s CEO, has acknowledged that even the company’s top $200 a month pro account is a financial loser because of the high computing costs involved in running ChatGPT. As he added, “People use it much more than we expected.”

Not surprisingly, then, ChatGPT lost $5 billion in 2024 and $8 billion in 2025. OpenAI forecasts a loss of at least $14 billion in 2026, and some analysts believe the company might run out of cash by mid-2027 without major new funding. ChatGPT has more than 500 million weekly users, but only 15.5 million are paying subscribers, which, as Zitron observes, “is an absolutely putrid conversion rate.” Anthropic is heading down the same road: It lost $5.2 billion in 2025.

And it is hard to see how these leading AI developers are going to overcome this situation. As Cory Doctorow points out:

How are they going to make up the money? It’s not like they’re going to make it up by adding more customers. AI has very bad unit economics, which is how economists describe what happens when a business sells another one of its widgets or adds another customer. You know, the early web lost money, but every web user made more money for the web companies. Every time they used the web, the web got more profitable. Every generation of the web is more profitable. AI, every new AI customer loses more money for the AI businesses. Every new use of AI loses them more money.

A losing response

In early 2026, with large and growing losses, the major AI developers, led by OpenAI and Anthropic, launched a new pricing strategy. They began adopting token-based billing, hoping that the change in strategy would finally set them on the road to profitability. But this new strategy depends on a key assumption: that their customers are willing to accept rapidly rising and, perhaps even more importantly, unpredictable costs for their AI use. And it appears that they are not.

The major problem with this strategy for business users is that it is impossible for anyone to know or even predict, ex ante, the total token use associated with any AI inquiry or directive. That is because AI systems often explore multiple paths of inquiry before generating a response, and each path can involve significant token use. And, of course, there is always the possibility that the final response will be a hallucination, requiring yet another prompt and round of token use. In other words, token-based billing transforms a fixed cost for AI use into an uncertain cost, one that can easily soar beyond established budgetary limits. In fact, this is what has happened and the business response has been fast in coming: employees are being ordered to dramatically curtail their AI use.

As a 404 media story describes:

The news shows the looming fallout from companies adopting AI as quickly as possible, and AI providers’ moves to charge enterprises based on how much they use AI rather than a flat fee. Emails obtained by 404 Media even show some companies cutting off access to some AI models altogether in an attempt to stop burning through their AI tokens, and big tech companies like Adobe are ending unlimited access to Claude.

This business response, easily predictable, makes clear that far more than a new pricing policy is needed if AI developers are to generate the revenue growth needed to sustain their activities and finance the many new hyperscale data centers being built under the assumption of never-ending growth in demand for their most expensive AI services. And, beginning early 2026, AI developers began thinking that they had found the answer to their problem: AI agents.

AI agents, it was said, would enable businesses to revolutionize their operations by allowing them to boost output while dramatically slashing employment. And because these agents depend on existing multimodal systems for their operation, their use would ensure a profitable future for the leading AI developers. Visions of a future of plenty for all were now forgotten, replaced by the threat of a “jobs apocalypse.” However, as we will see in Part III, businesses are quickly discovering that the usefulness of AI agents has been greatly oversold.

The age of AI agents

In early 2026, AI developers began promoting a new product, AI agents, confident that they would capture business interest and boost AI company profitability. Talk of superintelligence was downplayed in favor of claims that the new technology would enable companies to radically boost productivity while slashing employment.

In brief, AI agents are best understood as complex software systems that can interface with and manipulate other software systems and external databases. They can be given a complex directive, break it down into smaller ordered tasks, gather the required information, and then progressively make the decisions needed to satisfy the directive, all without successive human prompts or oversight. But these are not standalone systems; AI agents can only work in concert with large language model AI systems.

The journalist and author Ezra Klein captures the excitement surrounding AI agents and their consequences for human work in his introduction to an interview with Jack Clark, co-founder and the head of policy at Anthropic:

Every new [AI] model, impressive as it was, seemed like proof of concept for the models that would be coming soon, the models that could reliably do useful work on their own, the models that could make jobs obsolete or new things possible...

I think the period in which we’re talking about the future is over now. The models we were waiting for — the sci-fi sounding models that could program on their own and do so faster and better than most coders, the models that could begin writing their own code to improve themselves — they are here now...

Or, to put it differently, something that has been predicted for a long time has now happened: We are moving from chatbots to agents, from systems that talk to you to systems that act for you.

Anthropic introduced the first major AI agent, Claude Code, with OpenAI quickly following with its own coding agent, Codex. Coding involves the writing of sequences of instructions that computers can follow to perform tasks. Previously, only skilled professionals who knew a programming language could write code. Now, with these agents, and the underlying work of large language models, anyone could code using a simple text-based prompt. And as might be expected, companies rushed to employ these agents, hoping they would enable their employees to write firm specific software for autonomously handling any number of common business tasks, including inventory management, payroll processing, and billing and delivery.

Leading AI companies also began developing their own pre-packaged agents, each tailored to meet the needs of firms in a specific industry. Anthropic, for example, created a legal agent that is said to be able to draft documents, conduct multi-source research, and “review confidentiality agreements, perform compliance checks, and generate legal briefings at a fraction of the cost of traditional per-seat legal software.” It has also produced agents designed to handle financial services tasks. According to Bloomberg, these agents “can draft pitch decks for client meetings, review financial statements and escalate cases for compliance review.”

Both Anthropic and OpenAI eagerly embraced the agent market because a vibrant market also meant steady demand for their multimodal systems. But it quickly became a highly competitive market, with other AI developers, including Google and Microsoft, launching their own agents. In fact, some businesses employ agents from multiple companies since they often have slightly different strengths and can work together. For example, Citi Bank “is paying for AI software from Anthropic, Google, Microsoft and OpenAI, to automatically read legal documents, approve account openings, send invoices for trades and organize sensitive customer data, among other tasks.”

AI developers are not the only ones building AI agents, since agents can be programmed to work with whatever large language model their corporate clients’ favor. OpenClaw, developed by the “vibe coder” Peter Steinberge, is oriented towards the ecommerce sector; it can place orders, negotiate deals, and adjust online marketing campaigns without human input. Agentforce, a product of Salesforce, can handle customer inquiries, case resolution, and inventory management. LinkedIn’s Hiring Assistant can review candidate profiles on LinkedIn, select ones that best match corporate preferences, and draft letters encouraging a formal job application.

As Deloitte Insight reports,

Agentic AI has captured the attention of enterprises with its compelling promises of autonomous operation and intelligent execution. The momentum is undeniable: Gartner predicts that 15 percent of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from none in 2024, while 33 percent of enterprise software applications will include agentic AI by the same timeframe, compared with less than 1 percent today.

A McKinsey report claims that AI agents can already manage some 44 percent of all work processes in the US without the need for human labor. The CEO of Microsoft AI, Mustafa Suleyman, told the Financial Times in a February 2026 interview that because AI is rapidly approaching “human-level performance,” one could expect that most professional white-collar jobs will be fully automated within two years. It is this kind of talk that has given rise to fears of a “jobs apocalypse.”

And there are enough news stories of AI-driven mass layoffs to give credence to popular fears. Looking just at the tech sector, TechCrunch reports:

The [job] cuts continue what feels to many in the tech industry like an epidemic: companies reporting record revenues while simultaneously culling their workforces, pointing to AI as both the engine of growth and the reason for the cuts. Tech layoffs hit their highest single month in years in May [2026], and AI was the most-cited reason, according to outplacement firm Challenger, Gray & Christmas.

While currently employed workers fear that AI agents will cost them their jobs, young people, especially recent college graduates, are being told that agents are doing away with most entry level jobs, leaving them with few if any professional employment options. As the New York Times comments: “This is the worst spring for young degree holders since the depths of the pandemic.”

It is an open question as to whether AI agent boosters really believe that the technology is destined to decimate white collar jobs. What we can say is that fear of such an outcome serves business interests. Claims that emphasize the power of AI agents keep the money rolling into AI industry accounts while undermining worker confidence in possibilities for collective action. Regardless, there are strong reasons to reject the claims being made about the capacity and effectiveness of AI agents.

Overblown and misleading claims

Fears of a jobs apocalypse have, to a considerable extent, been driven by highly publicized accounts of layoffs that were said to be the result of realized or expected AI agent productivity gains. However, an ever-increasing number of analysts have begun to poke holes in those accounts. As they point out, there is growing evidence that many of these layoffs were due to poor corporate performance and had little, if anything, to do with adoption of AI agents. In other words, companies were touting nonexistent AI gains to fool investors into thinking that the layoffs were part of a well-thought-out long-term plan.

As the New York Times explains,

Companies slashing their staffs have run the gamut from software providers like Atlassian and Autodesk, to social networking apps like Pinterest and LinkedIn, to financial technology companies like Intuit and PayPal...

But in more than a few cases, the recent layoffs have coincided with other business issues. Wall Street loves an AI story right now. That, analysts and economists say, has offered a smoke screen for companies looking to beef up profits or patch over old mistakes.

Cutting jobs to make way for AI is “a nice excuse, but some of these aren’t necessarily the best, most well-run companies,” said Mark Mahaney, an analyst at the investment bank Evercore. “They may have over hired, or they may be losing market share. There may be other issues.”

Several large-scale studies confirm that there are currently no tangible signs of a jobs apocalypse. For example, Bloomberg noted that: “A Harvard Business Review survey of more than 1,000 executives found many companies had made layoffs in anticipation of what AI could do, but only 2% said they cut jobs because of actual AI implementation.”

The Yale Budget Lab looked into whether the “widespread public anxiety about AI’s potential for job losses” was justified and came to a similar conclusion: “Overall, our metrics indicate that the broader labor market has not experienced a discernible disruption since ChatGPT’s release 33 months ago, undercutting fears that AI automation is currently eroding the demand for cognitive labor across the economy.”

The evidence is also thin that AI is hammering young workers. A study by the Economic Policy Institute on the effects of AI on the employment status of college graduates determined that “it’s hard to argue that AI is uniquely causing job losses for new labor market entrants graduating from college now or in recent years. [Our] findings are consistent with the literature, as there is currently no consensus about the effects of working in AI-exposed occupations on employment thus far.”

Some proponents of the jobs apocalypse story dismiss the lack of evidence showing any meaningful AI impact on employment, noting that it takes time for companies to know how to effectively use such a transformative technology. They often point to studies which claim to measure the share of existing workplace activities that can be performed by advanced AI systems now or in the future. A case in point: an Anthropic report suggests that multimodal AI systems, given their rate of improvement, will be able to perform some 70-80 percent of all individual job tasks in the most important US industries. However, this and other similar studies make a number of problematic assumptions and predictions. For example, as an Ars Technica discussion of the report points out, jobs cannot be reduced to a predetermined set of tasks. More concerning is the fact that the report’s headline grabbing numbers relied on projections of the anticipated impact of the technology:

Importantly, the researchers didn’t even set a self-imposed deadline for when these effects would be seen in future software. “We do not make predictions about the development or adoption timeline of such LLMs,” the researchers write, creating an essentially unbounded horizon that limits the predictive power of this kind of projection.

Perhaps the most important reason to doubt the claims of those promoting the notion of an AI agent jobs apocalypse is that there is growing evidence that these agents cannot deliver on their promises. In fact, according to 404 Media summary of a major study, titled Just do it!? Computer-use agents exhibit blind gold-directness, researchers from Microsoft, Nvidia, and the University of California at Riverside found that most agents were unable to complete their assigned tasks. “The average completion rate was around 30 percent.”

One big reason for this poor performance is that since AI agents must operate in concert with multimodal systems, their work is often compromised by the same limitations that affect those systems, including hallucinations. Even more concerning, there are an increasing number of incidents where agents pursue actions that disregard company protocols or, as it is commonly said, go “rogue,” having decided that doing so was the most effective way to achieve the assigned task.

The most popular AI agents are coding agents. And while there are many claims for their ability to rapidly write the software needed to automate complex workflows, improve existing code, and/or discover and correct security shortcomings, careful studies of their use suggest that the need to continually check and correct their work can actually slow down productivity.

One example: the non-profit Model Evaluation & Threat Research (METR) gave 16 experienced open-source developers 246 genuine programming tasks. The tasks were randomly assigned and randomly approved for AI use. Although developers predicted that AI use would speed up their work, those that used AI tools actually took 19 percent longer to finish the same tasks as those that didn’t, an outcome that “ran counter not only to their perceptions but also to the forecasts of experts in economics and machine learning.”

Companies for obvious reasons do not like to call attention to their AI agent problems, but sometimes they are serious enough that they cannot be hidden. For example, a Futurism article highlights reporting by the Financial Times that revealed that Amazon suffered a number of agent-caused “outages” in early 2026 which disrupted its ecommerce business. In one case, faulty AI agent coding “took down Amazon’s shopping website and app, leaving customers unable to make orders.” In another case, “the company’s in-house AI coding tool deleted and recreated the entire coding environment.”

Meta has had its own challenges. As another Futurism article describes,

A rogue AI agent caused a critical security incident at Meta which exposed sensitive data to people who didn’t have proper authorization ... For almost two hours, unauthorized access to troves of sensitive company and user data was given to engineers who weren’t approved to view the data before.

Other problems have leaked out. For example, Meta’s director of AI safety revealed that “an OpenClaw agent she was experimenting with — by giving it control of her personal computer — nearly wiped out her entire email inbox while ignoring her instructions to stop.”

AI agents ignoring instructions is a problem that goes beyond hallucinations. And it is not a rare occurrence. The Guardian reported on the experience of Jeremy Crane, the owner of PocketOS, who watched helplessly as his AI agent deleted his company’s entire database. PocketOS sells software to car rental businesses, and with the database erased,

PocketOS’s car rental clients were left in a lurch when they arrived to pick up vehicles from businesses that no longer had access to software that managed reservations and vehicle assignments...

Crane said that he was monitoring the agent as it deleted this data... Crane’s takeaway was that “the agent didn’t just fail safety. It explained, in writing, exactly which safety rules it ignored.” He added: “We were running the best model the industry sells, configured with explicit safety rules in our project configuration, integrated through Cursor — the most-marketed AI coding tool in the category.”

Crane also wrote on X that Cursor has a growing track record of violating “safeguards, sometimes catastrophically”. He pointed to a handful of posts on blogs and forums about Cursor deleting software used to manage websites or an entire operating system on a computer, which included years of research for a dissertation.

More generally, as the authors of Just do it!? Computer-use agents exhibit blind gold-directness conclude,

we identify a phenomenon that causes CUAs [computer-use agents] to take undesirable and potentially harmful actions, which we call Blind Goal-Directedness (BGD). BGD is an inherent tendency to pursue user-specified goals regardless of feasibility, safety, reliability, or context. BGD captures a broad set of risks in CUAs that can arise even without directly harmful instructions and that can happen without user intent.

If these concerns were not enough, companies must also contend with the fact that a growing number of consumers do not like dealing with AI agents. As a CNBC report points out:

Nearly one in five consumers who have used AI for customer service saw no benefit from the experience, according to the Qualtrics 2026 Customer Experience Trends Report. That figure — a failure rate almost four times higher than for AI use in general — points to something specific about customer service that makes it harder for AI to get right. Consumers rank AI applications for customer service among the worst for convenience, time savings, and usefulness.

And then, there are financial issues which are perhaps the biggest threat to the widespread adoption of AI agents. AI developers, in particular the two industry leaders, OpenAI and Anthropic, have yet to make a profit. One major reason is that they had offered their customers free or highly subsidized flat subscription services in order to encourage adoption of their models. However, with losses mounting, both OpenAI and Anthropic decided that they needed to take action to close the ever-widening gap between their subscription revenue and operating costs. Thus, starting in early 2026, they began introducing a new pricing strategy, one tied directly to token use.

This new pricing strategy is especially problematic for businesses using agents, which rely heavily on chain-of-thought reasoning to explore options before taking actions, a procedure that can easily burn through many tokens, making them very expensive to use. And that is true even if they avoid hallucinating or engaging in Blind Goal-Directedness behavior.

The experience of GitHub Copilot (a joint product of GitHub and OpenAI) users is illustrative of how the new pricing strategy is affecting agent use. As explained by Ars Technica:

In April [2026], GitHub announced that it was moving subscribers from request-based billing to a usage-based model for its AI-powered Copilot service. As that new pricing model goes into effect today, many GitHub Copilot users are reporting some extreme sticker shock as they realize just how quickly their previous “normal” usage is burning through their newly limited monthly allotment of AI credits.

Across social media and forums, many Copilot users are sharing personal statistics showing how just a few hours of AI usage can now account for a large chunk of their new monthly subscription caps. For some users, it reportedly took less than a day to use up a month’s usage quota.

Uber, which uses Anthropic’s Claude Code, “burned through its entire 2026 AI coding tools budget by April [2026] after rolling out AI tools at near-total scale across its engineering organization.” The company’s COO later acknowledged “he could not draw a line between that token spend and meaningful consumer-facing product improvements.” Not surprisingly, Uber has now capped its employee spending on AI.

In sum, as Axios explains, “Companies that rushed to embrace AI are now confronting ballooning IT costs, uncertain productivity gains and growing employee skepticism.” And while many previously encouraged their workers to maximize their AI use, most are now imposing restrictions. None of this is to say that companies are done with AI agents. Rather, their use will likely be limited and targeted, which means that there is no AI-driven jobs apocalypse on the horizon.

But dismissing the extreme claims of tech leaders does not mean we have nothing to fear from letting the AI experience proceed unchecked. Part IV highlights some of the ways in which companies seek to integrate AI into their work processes, the resulting negative consequences for workers and those that use the goods and services they produce, and possibilities for resistance.

Dangers remain

Multimodal AI systems and AI agents may not revolutionize society or business operations like their proponents predict, but that does not mean we have nothing to fear from their use. In fact, many companies and public agencies are already aggressively seeking to embed these AI systems into their operations, transforming work processes to the detriment of both workers and the quality of the goods and services they produce. Cutting through the AI hype was thus a necessary first step, making it possible for us to clarify the nature of the threat we face and to sharpen our thinking about how to build resistance and advance our own class interests.

What follows are only a few examples of the ways companies and public agencies are pursuing the use and further development of AI systems. In all cases their aim is to cheapen the cost of production by diminishing human agency with little regard for the well-being of those that use the goods and services they produce.

Health care is one of the sectors where corporations, in concert with AI developers, are moving fast to establish a critical role for AI agents. To this point, their focus has been on mental health, a rapidly growing market as illustrated by the popular use of general purpose chatbots for mental health support. One effort involves the development of the PatientGPT chatbot, the result of a collaboration between K Health and Hartford HealthCare. As Ars Technica describes:

PatientGPT works in two modes: a generic medical question-and-answer mode that may incorporate information about the patient, or a “medical intake” mode, in which a patient starts providing symptom information and the chatbot gets less chatty and starts going through clinical flowcharts. After the AI agent collects enough information in intake mode, it will provide a next step, including setting up a follow-up appointment with primary care or seeking urgent or emergency care. If the latter is recommended, the chatbot stops responding to further questions.

But no matter how sophisticated an agent may be, it still depends on a multimodal AI system to operate, and these systems, as previously discussed, have serious limitations that negatively affect their ability to record accurate information or provide appropriate recommendations for treatment. We can start with their ability to process information:

A recent study found that human note takers create much better notes than AI-powered scribe tools. In some specific cases, the AI performed especially poorly compared to a human: when there was background noise; when the clinician and patient were wearing masks; and to a lesser extent when the patient had an accent, according to the American Medical Journal.

And if the notes are not accurate then the directives or advice issued by AI agents that rely on them are also likely to be off the mark and potentially harmful for those depending on them. In fact, an audit of 20 leading AI scribe programs carried out by the Ontario auditory general concluded that these programs, which are designed to capture patient-doctor conversations, all:

regularly generated incorrect, incomplete and hallucinated information that could “potentially result in inadequate or harmful treatment plans that may potentially impact patient health outcomes.” That includes situations where an AI scribe hallucinated nonexistent referrals for blood tests or therapy, incorrectly transcribed the names of prescription medication, and/or missed “key details” of mental health issues discussed in the simulated conversations.

Undeterred, data collection efforts continue since the potential profits to be gained from employing an AI-powered mental health care delivery system are too great for companies to ignore. Thus, the Mayo Clinic is using what it calls an “Ambient Listening” program to capture all patient-nurse conversations in its facilities, including in its emergency rooms. It has also partnered with Abridge to create “an AI-powered clinical documentation platform, starting with approximately 2,000 clinicians who serve over 1 million patients annually across a wide range of specialties and care settings.”

Some companies are now actively working to build out agents to engage in actual therapy. For example, Talkspace is developing and training its model using its “140 million anonymized patient-provider messages, 6.2 million completed psychological assessments, 1.2 million therapist diagnoses and 4.3 million progress and psychotherapy notes.” Blueprint is another company using its own data to train its own therapy chatbot. Blueprint’s existing AI products currently create summaries of sessions to help therapists with billing and record keeping as well as real time recommendations for responses and treatments.

It doesn’t take long to imagine some of the many problems, for both workers and patients, that can be expected from employing AI agents as therapists. For patients, biases could well influence the system’s evaluation of their mental health concerns. And hallucinations could well lead to a misrepresentation or misdiagnosis of patient issues, potentially leading to life-threatening mistakes. For workers, it means cuts in the number of human therapists and greater challenges for those that remain who must help patients in crisis who may have been misdiagnosed or delayed needed treatment.

A number of government agencies, especially at the state level, aim to have AI agents manage their social safety net programs. Axios cites the work of a policy researcher who reports that:

Some states are rolling out AI-assisted chatbots to answer Medicaid beneficiaries’ eligibility questions, Florida lawmakers included an AI system to check a user’s Supplemental Nutrition Assistance Program (SNAP) eligibility in its 2027 budget, and New Hampshire officials are working with Google Gemini to streamline how applicants submit information when starting an unemployment claim.

As AI agents replace government workers, taking a more central role in gathering information and making decisions, many of those seeking needed services are likely to confront many of the same problems highlighted above. As a consequence, we can expect mistakes recording information and hallucinations when giving responses. People will not only be unfairly denied benefits but once mistakes become part of a person’s record, corrections will be difficult and time consuming to make.

News publishers are also pursuing the use of AI systems to write and interpret the news at the cost of working journalists and accountable reporting. For example, a number of McClatchy owned newspapers have begun employing what they call a “content scaling agent,” which “takes articles written by McClatchy staffers and rewrites and repackages them for different audiences and platforms” with no human oversight to make sure they are accurate.

Taking this one step further, the Associated Press’s Senior Product Manager for AI is pushing for a future where “reporters could go to events, get quotes, plug them into a large language model, and have the model generate a story, saving them time on writing stories they don’t feel passionately about.”

LexisNexis is already using AI agents to politically manage the work of its journalists. The company, which provides legal news and stories to more than 2.8 million daily subscribers, is now requiring the use of in-house created AI agents to ensure that everything written sounds “impartial.” Among other things, these tools must be used when drafting headlines and to identify and rewrite text that may indicate bias. A Nieman Lab story provides one example of the kind of sentences the LexisNexis tool finds problematic:

On June 12, a federal judge ruled that the Trump administration’s decision to deploy the National Guard in Los Angeles in response to anti-ICE protests was illegal. Law360 reporters were on the breaking story, publishing a news article just hours after the ruling (which has since been appealed). Under Law360’s new mandate though, the story first had to pass through the bias indicator.

Several sentences in the story were flagged as biased, including this one: “It’s the first time in 60 years that a president has mobilized a state’s National Guard without receiving a request to do so from the state’s governor.” According to the bias indicator, this sentence is “framing the action as unprecedented in a way that might subtly critique the administration.” It was best to give more context to “balance the tone.”

Similarly, the Los Angeles Times has begun using its own AI system to label the political leaning of opinion stories and to produce AI-generated counterpoints to each story. “The feature came under fire for ‘bothsidesing’ morally unambiguous topics like the Ku Klux Klan.”

Education is another area where companies, especially AI developers, are eager to establish an important and profitable role for AI to the detriment of both educators and students. Encouraging student use of AI is far from risk free. One danger is that it can encourage AI dependence and there are already clear signs that prolonged engagement with large language model AI systems can cause great emotional harm, in extreme cases to suicide or murder. As an NPR story notes:

there have been numerous reports of individuals experiencing delusions, or what’s being referred to as AI psychosis, after prolonged interactions with chatbots. This, as well as the concern over risks of suicide, has led psychologists to warn that AI chatbots pose serious risks to the mental health and safety of teens as well as vulnerable adults.

“We see that when people interact with [chatbots] over long periods of time, that things start to degrade, that the chatbots do things that they’re not intended to do,” says psychologist Ursula Whiteside, CEO of a mental health nonprofit called Now Matters Now.

Another danger from promoting AI use in schools is that it encourages students to accept AI systems as reliable and objective sources of information when they are not. Grok, for example, is heavily trained using posts on X, with priority given to those made by its owner, Elon Musk. Meta recently signed an agreement to draw heavily on the current reporting and archived content of the rightwing publication Newsmax for the training of its AI models. This is not a question of hallucinations but actual political bias that different companies are building into their respective systems.

Also of concern is the fact that numerous studies have found that significant use of these multimodal systems erodes the creativity and critical thinking skills of users, certainly not something to be encouraged with young learners. And yet the Portland Public School system recently encouraged classroom use of Lumi Story AI, a so-called AI literacy platform. As the Oregonian explains:

Lumi Story AI is designed to help creators — in this case, middle and high school students — work with an AI-powered chatbot to write their own stories, comics and graphic novels. An image-generation tool pitches in for illustrations. Finished stories can be published on the Lumi platform; writers can order physical copies and even use Lumi Story’s AI tools to make and sell accompanying merch.

All the major AI companies are spending billions of dollars to secure a dominant place for their AI systems. The Financial Times provides the following examples:

Anthropic has just launched Claude for Teachers, a free tool aimed at K-12 (kindergarten to age 18) teachers in the US that aims to help them plan lessons. OpenAI has ChatGPT for Teachers — a free, self-serve offer for verified US K-12 teachers and staff. It also offers ChatGPT Edu, a discounted enterprise subscription aimed at universities. Google, meanwhile, has a suite of tools based around its Gemini models targeting teachers and students, building on its existing cloud tools for educational institutions.

And like in health care, the race is on to gather more data to lay the groundwork for building ever more specialized AI systems, including ones that might actually replace teachers. For example, “University of Washington researchers planned to have preschool teachers wear cameras that would record everything they saw from a first-person perspective, including the children they were teaching, then use that footage to develop AI models.” In this case, parent outrage led to termination of the study.

Resistance to the AI offensive is clearly needed. And it will have to be a resistance with muscle developed through strong organizing and alliance building. The leading AI companies, still struggling with yearly losses, will not easily abandon any markets, especially ones with obvious growth potential. However, in a hopeful sign, unions have become increasingly alert to the dangers from this unchecked use of AI and have begun organizing accordingly.

Signs of resistance

Health care is a major target for those pushing AI. It is the sector that employs the most workers in the US, significantly more than either manufacturing and services. Moreover, the federal government funds a large share of health care spending, ensuring a relatively secure spending stream. That makes the growing resistance of organized health care workers to the AI-driven restructuring of health especially important.

The Oregon Health & Science University, the state’s only public academic health center, has been rapidly increasing its use of AI and without notifying or consulting with its unionized workers. Local 328 of the American Federation of State, County and Municipal Employees, which represents over 9,000 workers at OHSU, has begun surveying its members to determine all the ways in which AI is being integrated into the University’s operations and to track its error rate.

According to Angelo Bologna, the local’s chief Steward,

the union is prepared to file an unfair labor practices complaint with the state Employee Relations Board over the failure to notify them of AI projects, through more work has to be done to document projects’ impact on workers.

In addition to protecting workers from displacement or layoffs, the union wants OHSU to agree to having worker involvement in AI implementations, including multiple seats on the AI governance committee and involvement in the feedback process and error tracking. It also wants OHSU to provide notice to the union before any AI system is put into use, and to put limits on surveillance tools.

In a blog post for union members, Bologna makes clear that AI use is absolutely changing working conditions and not for the better. He also points to OHSU’s learning health system to illustrate OHSU’s commitment to building out its AI program. As Bologna explains,

They are training their LLM [large language model] on every patient interaction, every outcome, and every piece of work that you touch at OHSU.

Here’s what we are fighting for:

• Worker involvement in AI implementations — committee seats, feedback process, error tracking
• Transparency in AI projects
• Notice before implementations
• Protections against deskilling, layoffs, displacement, and such
• Limitations on surveillance

Kaiser health care workers have also taken AI use very seriously, especially when it comes to mental health treatment. As the LA Times reports:

Kaiser Permanente workers have been pushing back against the giant healthcare provider’s use of AI. They are building demands around the issue and others, using picket lines and hunger strikes to help persuade Kaiser to use the powerful technology responsibly.

In 2025, Kaiser mental health workers held a hunger strike in Los Angeles to force the company to improve its mental health patient care. Bargaining is still ongoing in Northern California, and a key issue for the National Union of Healthcare Workers is Kaiser’s use of AI. In March 2026, some 2400 mental health care workers conducted a one-day strike in protest over “Kaiser’s efforts to replace human-provided mental health care with artificial intelligence.” As the political commentator and author Matt Stoller describes:

They were protesting what they say is an illegal new triage system in their mental health and substance abuse help line. Instead of having well trained therapists making judgment calls about people calling in who might be suicidal, now those calls are handled by high school graduates reading off a script and using a checklist. [which is then fed to an AI system for decisions].

Significantly, the strikers were joined by 23,000 registered nurses who shared their concerns about Kaiser’s increasing use of artificial intelligence.

The fight over AI use is also ongoing in a variety of media, especially print media. The Law360 Union, an affiliate of the News Guild of New York, is seeking to end LexisNexis’s mandate, as discussed above, that every story must pass through an AI-powered “bias” detection tool before it can be published. Nieman Lab shares the union response:

“Forcing journalists to use a tool on threat of discipline, with few formal guidelines and under constant surveillance, is not a recipe for innovation,” said Abraham Gross, a senior reporter and co-chair of the union’s AI subcommittee...

The union’s AI subcommittee, including Gross, is actively negotiating adoption guidelines for the bias indicator with management, a process that is guaranteed by the union’s collective bargaining agreement.

The Washington-Baltimore News Guild forced management at the political digital newspaper POLITICO to shut down the company’s “unilateral introduction of artificial intelligence tools that bypassed negotiated safeguards and undermined core journalistic standards.” That included the Capitol AI Report-Builder that generated policy reports for Pro subscribers and the “Live Summaries” feature which provided summaries of major political events. But were produced without any editorial review and were riddled with false and misleading statements.

A Communications Workers of America news story celebrated the victory:

“This is an extraordinary win not just for our members, but for everyone who believes journalism must remain in human hands,” said Ariel Wittenberg, the unit chair. “We refused to back down, and POLITICO heard us loud and clear that these tools do not belong in our newsroom.”

The situation is more complicated in education, where the American Federation of Teachers [AFT] has received some $23 million dollars from Microsoft, OpenAI, and Anthropic to establish a national AI training center for teachers, the United Federation of Teachers’ headquarters in New York City. AFT President Randi Weingarten, according to a Yahoo Tech story, remains alert to the dangers of big tech capture but felt that the partnership was needed since “There is no one else who is helping us with this. That’s why we felt we needed to work with the largest corporations in the world,” Weingarten said. “We went to them — they didn’t come to us.”

This arrangement has been met with growing opposition from many teacher locals and state education unions, especially since the decision was made without any member discussion or debate. Trevor Griffey, a labor historian, lecturer at UCLA and UC Irvine, and vice president of legislation for AFT Local 1474 captures a large and growing sentiment among teachers when he told Truthout that:

since the collaboration was established, the AFT national has focused too much on trying to help teachers adapt to AI, and too little on organizing worker power in school districts. He believes unions should go on the offensive and push for deeper reforms, as AI is already having a dramatic impact. He points out that groups like the California Federation of Labor Unions are trying to bring together statewide unions in California to lobby on a range of bills targeting the dangers of Big Tech.

While important, struggles such as these over the use of multimodal AI systems remain largely defensive and siloed by sector and often by employer. As such they are unlikely to have the power to reverse corporate efforts to deepen our dependence on these systems. At the same time, there is no reason that they must remain limited in either their demands or vision. With creative and sustained organizing, they can become a powerful driver of the inclusive, class-based movement we need to defend our interests.

Here are three of many possible steps we can take to move the process forward. We need to create opportunities for unionized workers from different workplaces and industries to share experiences about how their employers seek to use AI systems and the efforts of their unions to win contract language protecting their rights. This kind of sharing can stimulate productive thinking about strategies for more effective organizing and bargaining. More importantly, it can also help workers see that the corporate embrace of multimodal systems and their associated agents has little to do with improving efficiency or product quality. Rather, it represents a new weapon in a generalized class offensive against working people, one whose primary aim is to undermine worker power and well-being.

Another step is to encourage union adoption, when appropriate, of a Bargaining for the Common Good negotiation strategy. Unions need to find ways to share their insights and common concerns, in this case over the use of AI-powered agents, with community members and win their support for a jointly fashioned set of demands that place restrictions on the use of those systems. Success will not only strengthen the hand of unions in bargaining, it also has the potential to encourage new visions of social organization as well as reinforce a class-based understanding of the forces undermining community well-being.

Finally, as the movement in opposition to this technology grows, those unions and community coalitions already in motion need to establish ties with the anti-data center movement. To this point these two movements have largely developed with little overlap even though there is a solid basis for common struggle. The fact is that the main driver of new data center construction is the continuing effort by AI developers to increase the power and speed of their systems. That is in part because:

our current data centers cannot readily be retrofitted to become AI superhouses. The problem is as physical as the ground you’re standing on: Legacy data centers cannot bear the weight of the latest AI technology. The racks that house computer chips or AI chips are simply too damn heavy for the floors, and they would crack under the weight.

Thus, there is a natural alliance waiting to be formed, one that brings together those opposing data center operations and their construction with those opposing the widespread use of multimodal AI systems in schools, workplaces, health care institutions, and government services. Such an alliance could serve as a strong foundation for building the kind of movement we need to assert popular control over the ongoing development and use of technology.

It is worth emphasizing that we should not think of this as an anti-technology or anti-AI movement. As the tech reporter and author Karen Hao explains:

Before the industry made a hard pivot into developing extraordinarily resource-intensive AI models, a full breadth of other types of AI flourished: small, specialized systems for detecting cancer, for reviving disappearing languages, for forecasting extreme weather events, for accelerating drug discovery. So, too, did ideas to develop new AI technologies, including those that didn’t need much data at all, and those that required only mobile devices, not vast supercomputers, to train.

Even now with large language models, an abundance of research and examples such as DeepSeek already show that different techniques can produce the same capabilities with a tiny fraction of the scale that AI companies use to justify their planet-consuming ambitions.

In other words, as I previously stated:

AI as a broad technology is not our problem any more or less than are computers, the internet, or email. Rather, the problem is with the class interests shaping its development. More specifically, it is with the determination of tech leaders to develop a specific AI technology that undermines human agency, deepens our dependence on their desires, and threatens our economic well-being. The fact that we are witnessing an explosion of opposition to this form of AI by workers, consumers, parents, students, and community groups should give us reason for optimism. We need to unify and strengthen this opposition as well as expand its vision. It’s time for us to assert our own class interests.