Artificial Intelligence: the hype, the dangers, and the resistance—Part II

AI

Artificial Intelligence: the hype, the dangers, and the resistance—Part II

By Martin Hart-Landsberg (Posted Aug 10, 2026)

Originally published: Reports from the Economic Front  on August 6, 2026 (more by Reports from the Economic Front)

InternetGlobalNewswireAI (Artificial Intelligence), ChatGPT (OpenAI), Claude (Anthropic), Copilot (Microsoft), Gemini (Google), Grok (xAI), Llama (Meta)

This is the second post in a multipart series aimed at cutting through the fog of AI hype in order to help us understand some of the real dangers we face from AI use and highlight hopeful avenues of effective resistance.  Part I provides an overview of my argument.  This post debunks the hype surrounding the large language model (LLM) generative AI systems (later to become multimodal AI systems) produced by leading tech companies, in particular ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Grok (xAI), Copilot (Microsoft), and Llama (Meta).

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.

Monthly Review does not necessarily adhere to all of the views conveyed in articles republished at MR Online. Our goal is to share a variety of left perspectives that we think our readers will find interesting or useful. —Eds.

About Martin Hart-Landsberg

Martin Hart-Landsberg is Professor Emeritus of Economics at Lewis and Clark College, Portland, Oregon; and Adjunct Researcher at the Institute for Social Sciences, Gyeongsang National University, South Korea. His areas of teaching and research include political economy, economic development, international economics, and the political economy of East Asia. He is also a member of the Workers’ Rights Board (Portland, Oregon) and maintains a blog Reports from the Economic Front where this article first appeared.

 AI (Artificial Intelligence)ChatGPT (OpenAI)Claude (Anthropic)Copilot (Microsoft)Gemini (Google)Grok (xAI)Llama (Meta)

A tribute to Johanna Fernandez

India supplying arms to a genocidal regime

Also by Martin Hart-Landsberg

Monthly Review Essays

Lost & Found

Leave a comment