There is little doubt that the release of ChatGPT in November 2022, and the tsunami of artificial intelligence (AI) investment that followed had a seismic impact on the global economy. In Q2 2025 $40 billion, or 45% of all venture capital (VC) funding, went to AI-related companies, according to Crunchbase. On the day that I wrote this blog, NVIDIA, which makes the chips that power the AI wave, was worth $4.4 trillion, 13 times more than it was worth in October 2022, just before ChatGPT was released by OpenAI. NVIDIA is now the most valuable company in the world, and represents over 7% of the Standard and Poor 500 stock market index.
As of now, the AI picture is looking a little frothy. A Goldman Sachs report in August 2025 found that only 9% of US companies have actually used generative AI in production in recent weeks. Still, it projects that around 6% of US jobs may be displaced, but that new jobs will also be created, with a projected 15% improvement in labour productivity. The largest impacts so far on employment have been in industries where AI has an obvious impact, such as graphic design, marketing consulting, call centres and technology. A Microsoft study identified 40 jobs most exposed to AI, with translators, sales reps, writers and travel agents and concierges among those most affected. Unsurprisingly, jobs involving manual labour, such as hospital orderlies, tree logging and rail track laying, were the safest areas. No chatbot is going to be cutting your hair, painting your house or putting up a new roof any time soon.
One interesting economic impact of AI has been on the world of online advertising. Google made 78% of its revenue from advertising in 2024. However, people are increasingly using AI chatbots to do their online searches rather than just relying on search engines like Google. The advertising revenue of Google in Q2 2025 was actually down 1% year-on-year, which is unprecedented. A study by Pew Research Center in 2025 found that Google users were about half as likely to click on a Google AI summary answer as one without such a summary. In March 2025 about a fifth of Google searches produced such an AI summary. These summaries, incidentally, cost around ten times as much as a regular search to process, so Google’s response of fighting AI with AI does not come cheap. The combined effects of AI summaries by search engines and users bypassing search engines altogether are hurting the publishers that advertise online. If someone sees a Chatbot-generated summary in response to a query, then the end user may not visit the source websites, and so will not see any adverts. Publishers are complaining, and with reason. In one 2023 survey, 5% of users claimed that ChatGPT was their primary search engine, up from 1% a year before. By 2025, that pace has accelerated greatly. A February 2025 Bain survey found that 80% of consumers now rely on AI-written results for at least 40% of their searches, reducing their organic web traffic by up to a quarter. Even basic questions are now being answered by AI chatbots, with 48% of consumers using AI to answer questions about the weather and 42% for shopping recommendations. Advertisers need to respond to these changes in the market, for example, by using more video or interactive formats, and by making their adverts more visible to generative AI search.
Are we in an AI bubble? Marz Zuckerberg has been hiring top AI talent for unheard-of compensation packages, some in excess of $100 million per year. Does that sound a little bubble-like to you? In August 2025 Sam Altman, CEO of OpenAI, said in an interview with Verge magazine: “Are we in a phase where investors as a whole are overexcited about AI? My opinion is yes,” Even back in the heady days of 2023, some economists, such as the World Economic Forum, were warning that AI may have less impact than some enthusiasts were projecting at that time. The August 2025 release of the much-awaited ChatGPT5 by OpenAI was widely regarded as a flop. This was a major blow to industry confidence, as OpenAI had long been hinting that they were on the path to “artificial general intelligence“ (AGI). This is a stage of AI where the AI becomes able to understand, learn and apply knowledge across a wide range of tasks, like humans. Computers have long been able to outperform humans at specific tasks such as multiplying large numbers together, analysing large datasets or playing chess, but an AGI could do all this and more. Yet ChatGPT 5 was little more than a remarketing and aggregation of existing OpenAI models. Even its own demo at the launch showed ChatGPT planning a baseball park tour that involved missing out most of the East Coast of the US, yet included a non-existent, and presumably rather damp, baseball park in the middle of the Gulf of Mexico.
It is telling that large language models, which underlie ChatGPT and all its generative rivals like Claude, Perplexity and Grok, are absolutely terrible at playing chess, yet a specialist computer beat the chess world champion in a match as far back as 1997. In an August 2025 LLM chess championship hosted by Kaggle, the leading LLMs played one another. The results were hysterical. The eight LLMs were allowed five attempts at each turn just to play a legal move, and within a few moves of the games, many of them failed to leap over only that pathetically low bar and were disqualified. So, in mid-2025, many of the latest LLMs are unable to play chess as well as a small child that has just been taught the basic rules of the game. Even in the final, one LLM (Grok) played dismally, blundering its queen in multiple games. AGI, this was not. On a more serious note, the persistent issue of hallucinations by LLMs, where they produce fabricated or meaningless answers, is seriously hindering the take-up of AI by business, where people are used to computers that give consistent, correct answers. With hallucination rates around 15% and getting worse, this is a problem.
The sluggish adoption of AI by the corporate world, partly due to the thorny issue of hallucinations, spells trouble for the AI industry and its investors. Executives report a 90% failure rate for their AI projects in an IDC survey, and over 80% failure rate in a RAND report. Consultants, who make money from such projects whether or not they succeed, urge executives to believe harder. The wave of AI hype crashing on the rocks of reality is starting to hit AI companies, which have been buoyed up by wild valuations and have been spending heavily on NVIDIA chips and data centres. One such company, chatbot vendor character.ai, announced in August 2025 that its aspirations for AGI had been abandoned. Meta’s chief scientist Yann LeCun has stated that LLMs will be obsolete within five years, and that alternative AI research paths and architectures are the way forward. He is not the only one. Many researchers now think that LLMs are hitting a point of diminishing returns.
Investor Paul Kedrovsky told the Wall Street Journal that there were parallels between the current AI boom and the boom in infrastructure spending in the internet era of the late 1990s We all know what happened to the technology economy from March 2020 to 2023: a 78% drop in the NASDAQ. It remains to be seen whether the parallels with the 1999 tech boom and the current valuations of AI companies are real. However, if we really are in an AI bubble, and if it bursts, then the results will not be pretty.
