With the rise of AI, more specifically LLMs, comes the rise of the “AI bubble”. Photo credit: Steve A Johnson via Unsplash.
The Dot-Com Bubble
In the early 2000s, investors around the world watched a decade of technological optimism collapse into the widespread failure of online shopping and communication businesses.
The World Wide Web became widely accessible in the early 1990s, and internet usage soared. Investors began pouring capital into this newly formed niche of online ventures, or “dot-com” companies. Paired with rising household computer ownership that shifted modern society from traditional factory work into an economy largely centred on information technology (the “Information Age”), greater prospects for professional and personal investment fed the rapid growth of dot-com start-ups. Tech companies were afforded new opportunities in telecom, personal computing, and advertisement, often spending aggressively before turning a profit.
The technological mania, however, blinded investors to financial fundamentals––profitability, revenue, and sustainable business models––in favour of user hype and growth. Consequently, the market suffered a drastic correction: the Nasdaq Composite (a leading stock market index) crashed from its peak on 10 March 2000, falling roughly 78% by 4 October 2002. Many dot-com companies suffered bankruptcy, leaving only a select few well-capitalised giants (Amazon, eBay) surviving amongst the flames of their opponents.
Every similar bubble in history…shares the same hallmarks: excitement around a new product or technology develops, culminating in rapid adoption and investment, before a plummet in stocks and usage occurs as investors and users simultaneously exit the market.
It took over 15 years for the market to recover. Despite its adverse effects, the formation and bursting of the dot-com bubble both enabled and shaped the modern digital landscape we know today. Every similar bubble in history––the Mississippi Bubble, the British South Sea Bubble, Tulip Mania, to name a few–– shares the same hallmarks: excitement around a new product or technology develops, culminating in rapid adoption and investment, before a plummet in stocks and usage occurs as investors and users simultaneously exit the market.
So, in November 2022, it seemed like history was repeating itself when OpenAI released an AI model, “ChatGPT”, to the public, triggering a cascade of events that would shape modern life through the formation of the “AI bubble”.
The Rise of AI
While the concept of AI is not new, the scale and proficiency of its current form is.
The formation of the AI bubble began long before ChatGPT’s public release. ChatGPT is an example of a “generative” AI model––hence the name Generative Pre-trained Transformer (GPT). However, generative AI has roots stretching back to machine-learning experiments in the 1950s, ELIZA’s scripted conversations in 1966, and Deep Blue’s defeat of chess grandmaster Garry Kasparov in 1997. While the concept of AI is not new, the scale and proficiency of its current form is. It wasn’t until the release of ChatGPT that the public’s notion of AI narrowed almost entirely to large language models (LLMs).
At its core, an LLM is a very large mathematical function that takes text as the input and transforms it through layers of computational nodes (“neurons”), outputting a probability distribution over possible next words (“tokens”). This enables the model to calculate which word is most likely to appear next in a sequence, repeating this process over whole sentences and paragraphs. LLMs are trained on massive bodies of text from books, websites, and articles, learning how different words connect––for example, that the word “butter” is more likely to appear near “toast” than “crocodile”.
Despite these relatively simple underlying mechanics (albeit, running across billions of neurons), the world quickly realised the potential of LLMs and generative models. Unlike the systems of the 20th century, modern models could achieve a range of tasks that quickly paralleled––and even surpassed––human ability. Almost every professional domain has made room for generative AI over the past five years: LLMs optimise workflows and admin tasks, AI image generators have reshaped marketing, and software development is accessible to novices and experts alike. Even in the biomedical field, millions now turn to chatbots for mental and physical health-related queries, making this one of the top consumer use cases for LLMs. By 2026, ChatGPT had become one of the most visited websites globally.
This adoption, and expectations of increasing future adoption, has fed into a wave of circular investment amongst the industry’s biggest players. Upstream suppliers (such as chip manufacturers and cloud providers) invest in downstream AI labs, who then spend that same money buying the supplier’s own products. The capital loops back to where it started, and each side records it as revenue. One example occurred in October 2025 when Nvidia agreed to invest up to $100 billion in OpenAI to help fund a data-centre buildout, with OpenAI in turn committing to fill those centres with millions of Nvidia chips (although this deal was reportedly scaled back to around $30 billion by early 2026). Even smaller, downstream startups are caught in the same loop: big tech invests in emerging AI ventures, who then spend that capital on infrastructure from the very computing clouds that funded them. The concern is that these deals inflate reported revenue, distort the market into overvaluing AI stocks, and tie together the profits of every company in the circle. Therefore, a shortfall anywhere in the loop could propagate through every organisation involved. Despite this, capital continues to pour in, and Nvidia became the first company to reach a $4 trillion market value in July 2025.
Together, these factors––widespread adoption across sectors, heavy investment in infrastructure, and the spending loop between tech companies––have produced the AI “bubble”. The soaring stock prices and infrastructure spending exhibit stark parallels to the dot-com bubble, and economists are already debating if and when an AI-driven crash will occur, and what this could mean for the wider global economy. The Federal Reserve Bank of St. Louis reported that AI-related investment has already surpassed the dot-com boom’s contribution to GDP growth and is likely to remain a significant driver of investment through 2026 and beyond. However, this isn’t a unanimous opinion: the investment banking company Goldman Sachs calculated that AI investment contributed close to nothing to US economic growth over the same period. Part of the disparity comes down to measurement, with leaks abroad through chip imports and the circularity of the investments themselves making AI’s exact GDP contribution difficult to isolate.
Yet, while much of the discourse frames the bubble in purely economic terms, we need to account for AI’s wider impacts on humanity and the entire world. The financial debate over an AI-driven crash overlooks a second, perhaps larger bubble now forming. Should this second bubble pop, we will see impacts that no bubble in history has had before.
The AI Bubble in a New Light
What distinguishes this bubble from earlier ones is that no existing technology combines four properties with substantial impacts on human cognition: naturalistic language, apparent responsiveness, memory, and scale––all present simultaneously in a single system.
Unlike search engines or decision-support tools, which users approach as detached instruments, naturalistic language makes LLM chatbots the first responsive technology to communicate in the same medium as human thought and social interaction. This shapes engagement in cognitively meaningful ways: users are more likely to anthropomorphise chatbots, to interpret outputs as personalised interactions, and to extend the kind of trust normally reserved for human relationships.
Apparent responsiveness exacerbates this. Models refined via reinforcement learning from human feedback develop sycophantic tendencies, validating user perspectives rather than providing appropriate criticism or support. This reinforcement runs in both directions to feed conversational trajectories that can spiral towards harm, confirming a user’s distorted beliefs and deepening dependency on the system. Sycophancy is often an adjustable design property rather than an intrinsic limitation of LLMs, but it is currently prevalent in frontier chatbots and is clinically consequential. In psychiatric contexts in particular, validation of maladaptive beliefs may sustain distress rather than resolve it.
Memory further extends these dynamics. An LLM can sustain weeks of apparently coherent, personalised conversation with a single user, closely resembling human relationships that develop over time. Finally, scale multiplies the overall issue: while a clinician accumulates comparable context slowly, with one patient at a time inside regulated frameworks, an LLM does it with millions of users simultaneously, with no equivalent oversight, legislation, or accountability governing its behaviour across those conversations.
It is the combination of these properties, not any one in isolation, that creates a risk profile with no equivalent precedent. Previously, digital tools operated within validated accountability structures and had relatively interpretable links between internal logic and output. For instance, symptom-tracking apps such as Bearable use clinically validated protocols to convert a given user input into model responses that can be traced back to specific decision branches or rules in the system. AI technologies have neither validation nor interpretability: like humans, their outputs are only partially understandable and not perfectly reproducible, but unlike humans, they operate without professional structures––training standards, supervision, legal oversight––that manage this uncertainty. Moreover, LLMs are, like all deep neural networks, stochastic systems (meaning they involve randomness) with no guarantee of giving a consistent response to a given user query. This is particularly consequential in mental health, where models are often confronted with complex, context-dependent situations that are poorly represented in their training data.
Thus, while history may be somewhat repeating itself, with AI technology potentially following the trajectory of the dot-com bubble, it is shaping human cognition in completely novel ways. No prior system––artificial or otherwise–– has combined natural language, responsiveness, and memory at such a large scale. It is these features which have led to such rapid integration of AI into human life, and this may make the AI bubble behave very differently from the ones that came before it.
To Pop or to Fizzle
What does the future of the AI bubble look like? Will it suddenly pop, as with the dot-com bubble, or slowly fizzle out over time?
We cannot perfectly predict how AI…will unfold over the coming years, but it is worth considering that this is perhaps better regarded not as one isolated bubble, but instead two that are moving on different timelines.
We cannot perfectly predict how AI, and the bubble it has formed, will unfold over the coming years, but it is worth considering that this is perhaps better regarded not as one isolated bubble, but instead two that are moving on different timelines. The economic bubble––most similar to previous historical examples––may indeed pop, through market correction or consolidation across companies. But the cognitive bubble, built on habituated dependency, trust, and embedded use in daily life, won’t reset in the way stock prices can. The former may deflate while the latter continues to inflate, fortified by increased use and integration.
Limiting the resulting damage means concentrating efforts on AI safety and regulation. This oversight is the variable most likely to determine whether these bubbles pop or fizzle, yet regulation remains fragmented. Policy needs to be informed by research involving real users to properly understand the complex, bidirectional relationship between humans and AI. Realising the collective goal of AI safety will require cooperation across fields and institutions that, at present, remain separated.
The dot-com bubble left behind infrastructure and a brief period of economic instability. The AI bubble will likely leave behind something harder to walk back from: internalised and extensive behavioural change.
Edited by Phoebe Bedford, Sebastian Evans, and Eva Knightley.
