The enthusiasm for artificial intelligence among investors, CEOs and journalists has been at a fever pitch since the release of ChatGPT four years ago. Elected officials, executives, investors and journalists were quick to hail it as the fourth industrial revolution, despite statements from OpenAI and Anthropic that their products would lead to massive layoffs and possibly the extermination of humanity. Four years after the launch of ChatGPT, it is time to acknowledge that the reality of AI has fallen so far short of the promise that we should cut our losses and do what we can to protect ourselves.
Few doubt the enormous potential of AI to transform our economy and society, but it now appears likely that the transformation will be for the worse. We can see now that the industry talked about existential risk in order to create the impression that AI was inevitable and that the people who created it should not be held accountable for the negative consequences, of which there are many.
The recent hacks of Hugging Face, agencies of the government of Australia, and many other organisations are violations of anti-hacking laws, which could be a felony. At a minimum, the AI companies have been negligent, and it is possible they are also incompetent. The role of ChatGPT in an elementary school shooting in British Columbia and of several chatbots in assisting suicides suggests that the current products also do not meet consumer safety standards. There are studies that say the use of chatbots leads to cognitive impairment, something that is particularly threatening in the young. And, finally, the entire AI project has depended on the theft of copyrighted materials and private data. The latter is particularly troubling in the context of national security, where government data and secrets can become training data.
Roger McNamee. Photo: Horacio Villalobos - Corbis/Getty Images
Anthropic’s IPO is an opportunity for existing investors to cash out, rather than for new investors to get rich
The people who argue that AI is transforming the world for the better should be forced to explain the apparent absence of enterprise customers who pay more for AI services than it costs the vendor to provide them. The promised benefits of scaling training sets and data centres have yet to materialise. Worse yet, surveys of enterprises indicate that the productivity gains promised from AI have been slow to materialise. How is it that the industry continues to attract unprecedented levels of investment for a technology whose harms are evident, but whose benefits are difficult to measure?
Into this frenzy, Anthropic has filed paperwork for an initial public offering (IPO). The desired valuation could exceed $2 trillion. No matter the price, we should recognise that the IPO is an opportunity for existing investors to cash out, rather than for new investors to get rich.
I do not doubt the long-term opportunity for AI. I believe it may one day be transformative. However, the current generation of technology falls far short of that goal. Worse yet, past choices made by industry players have made the hurdle for financial success impossibly high, irrespective of the quality of any company’s products. Because of the scale of AI, the industry’s financial problems may undermine the entire economy. Here is why:
The industry may be riding the wrong horse. Large language models (LLMs), the technological approach used by the industry today, may be a dead end. Yes, LLMs can do really cool things. Yes, chatbots are fun and often helpful. At the same time, the limitations of LLMs are increasingly evident. The systems are inherently probabilistic, which results in varying levels of errors and non-predictive results. The consequences of this unfixable error rate may be fatal to a court case, business deal, or human life. The industry’s initial hypothesis – that the shortcomings of LLMs could be overcome by scaling computing power and training data – has hit a speed bump, necessitating new approaches, such as neurosymbolic technology.
The decision to go to market with one-size-fits-all products may backfire. For 30 years, Silicon Valley has known that the best way to create a market is with narrowly focused applications that deliver an immediate return on investment. This was the basis for Google Search (“Are you feeling lucky?), Amazon (“What book would you like to buy?), eBay (“What would you like to sell/buy today?”) and nearly all the other tech giants. The problem with general purpose technologies is that customers must invest huge amounts of time and money to integrate them into existing systems, and few have the capacity and funding to do that.
Scale and timing of capital investment. In the interest of speed above all, the industry did not wait for innovation to drive down the price of infrastructure before investing $1.4tn, and committing an additional $3tn or more in future obligations. That was a huge unforced error that threatens the bond market and potentially the entire economy. In the internet era, the cost of infrastructure had declined by nearly three orders of magnitude by the time Google built its search infrastructure. The scientific years of the internet required Silicon Graphics supercomputers. The commercialisation wave, led by Yahoo, was enabled by Sun Microsystems servers, at roughly 10% of the cost of the supercomputers. Google took a giant leap, building its network around bare Intel motherboards, at a few per cent of the cost of the Sun servers they replaced, and approximately 0.1% of the early supercomputers. Meanwhile, Big AI has invested at least $1.4tn during an era when the most expensive components, Nvidia processors, were priced like supercomputers. In some contexts, unit costs of AI services are still rising. Alternative technical approaches to that of Anthropic and OpenAI, known as open-weight and open-source LLMs, are gaining market share because these systems cost far less, can substitute for frontier models in many use cases, and allow customers to retain control of their own data. Control of your own data is a strategic necessity for governments and enterprises.
Competitive Dynamics. There are 10 giant companies competing in AI with the same tech and the same training data, pursuing the same use cases for the same customers. Even if AI works, experience suggests that only two companies are likely to be profitable. Instead of the traditional model seen in the early stage of new sectors – where there are often lots of seed-stage companies in healthy competition – AI was backed by the largest tech companies with a level of capital that discouraged competition. The inevitable result will be hundreds of billions in write-offs, if not more. What remains unknown is the extent of the collateral damage.
The scale of investment in AI creates systemic risk. Much of the infrastructure for AI is being paid for with debt, at a scale that competes with the refinancing of the US national debt. In combination with the inflationary pressures of the war in Iran and tariffs on imported goods, the borrowing for AI infrastructure has exhausted the appetite of investors. Infrastructure players both large (Oracle, SoftBank) and small (CoreWeave) are in financial distress, while profitable big-tech players like Meta and Amazon find themselves with balance-sheet issues that would have been unimaginable two or three years ago.
Young people have already concluded that AI is a mistake. Now it is time for the rest of us to see the truth
The most important thing each of us can do today is to understand what is really going on in AI. When OpenAI and Anthropic use anthropomorphic language, saying that their products go “rogue”, that their agents “collaborate,” and that they “scheme” to avoid detection, we need to recognise that they are trying to evade accountability. The truth is that AI is software and software does exactly what its programming dictates. The AIs that hacked Hugging Face and other enterprises were programmed to hack and to lie about it. Hacking is a felony, which suggests that the executives who are responsible may face the risk of imprisonment.
Young people have already concluded that AI is a mistake. Now it is time for the rest of us to see the truth about this industry. Be sceptical about everything you hear from AI companies. Existential risk is not real. Today’s AIs fail at interacting with the real world, so they are not going to kill us. However, there is a high probability that AIs will make our lives miserable.
The real threats from AI products today include cybersecurity, fraud and reduced productivity. For reasons that defy explanation, the AI industry prioritised hacking and fraud in their product designs, which is why you see so many stories about that. In addition, the firehose of AI slop is ruining the internet. AI is deskilling students and workers. AIs continue to give bad advice, especially as relates to health and wellness, finance, and the law. They increase social isolation. When used in warfare, AIs contribute to terrible mistakes that kill people or potentially start wars. Perhaps worst of all, the widespread deployment of AI will concentrate power in the hands of a very small number of people whose authoritarian tendencies are well known.
The AI industry may have great promise for the future, but according to the International Monetary Fund, its flaws are threatening the global economy. The markets are increasingly focused on default risk on AI debt. Banks are working overtime to repackage the risky debt so that it can be sold off to pension plans and insurance companies, which ensures that the harm from a financial crisis would be absorbed by the entire population instead of by the people who created the problem. The scale of current harms are increasingly out of proportion to the future opportunities.
None of these risk factors and issues contradicts the notion that AI can be valuable to people or important to the economy. They speak to the risk to investors, customers, and society as a whole from an industry whose business model does not appear to be viable. The question is whether past choices made by the people running the industry limit the opportunity for economic and investment returns over the next few years. I believe they do.
Roger McNamee began investing in technology companies in 1982. He was an early investor in Facebook and Google, is the author of Zucked: Waking Up to the Facebook Catastrophe, and is an executive producer of the film The Social Reckoning