Silicon Valley’s AI Race Risks Becoming a Strategic Deadlock, Oxford Researcher Warns

“We’ve got a small number of very wealthy companies pursuing AI while simultaneously warning that it could go badly wrong.”

Oxford computer scientist and artificial intelligence researcher Michael Wooldridge says the rapid expansion of artificial intelligence is being shaped less by scientific inevitability than by competitive pressures among a small group of technology companies racing to avoid falling behind rivals.

In an interview discussing his latest book, Life Lessons from Game Theory: The Art of Thinking Strategically in a Complex World, Wooldridge argued that many of the current tensions surrounding artificial intelligence can be understood through the framework of game theory, particularly scenarios in which competitors continue escalating despite recognizing collective risks.

Wooldridge, a professor at the University of Oxford and one of Britain’s most prominent public communicators on artificial intelligence, said the industry increasingly resembles a strategic trap in which companies continue investing heavily in advanced systems because they believe competitors would gain advantage if they slowed development.

“We’ve got a small number of very wealthy companies that are busy pursuing AI, while at the same time saying that they are afraid that something’s going to go horribly wrong with it,” Wooldridge said. “So why are they busy pursuing it? Because they think if we back down and we don’t pursue it, somebody else will.

”The comments come amid intensifying global competition over artificial intelligence infrastructure, computing capacity and access to data. Major technology firms including OpenAI and Google DeepMind have expanded investments in large-scale machine learning systems, while governments in the United States, Europe and China are increasingly treating AI as a strategic industry tied to economic growth and national security.

Wooldridge said many of the core technologies underpinning today’s AI systems are not recent discoveries. He noted that key neural network techniques central to modern machine learning were developed by the mid-1980s, but computing power and data limitations prevented their wider deployment at the time.

“The only obstacle standing in the way of the AI revolution in the 1980s, really, was that computers weren’t powerful enough and we didn’t have enough data,” he said.He described the emergence of GPT-3 in 2020 as a turning point driven largely by scale rather than a fundamentally new scientific breakthrough.

According to Wooldridge, many researchers initially doubted whether simply expanding computational power and training data would substantially improve performance. He said the success of that approach surprised a significant portion of the research community.

OpenAI’s development strategy demonstrated that scaling existing methods could generate major commercial results, he said, although he cautioned against interpreting those advances as evidence that artificial general intelligence, or AGI, is imminent.Executives including Sam Altman and Demis Hassabis have publicly discussed the possibility of achieving human-level general intelligence within years. Wooldridge said those forecasts remain overly optimistic.

He argued that current systems still struggle with tasks requiring physical reasoning and adaptation in unfamiliar environments. While advanced chat systems can process complex linguistic queries, he said they remain unable to reliably perform many basic real-world activities that humans execute routinely.

“You can talk to ChatGPT about quantum mechanics in Latin,” Wooldridge said, “but at the same time, we don’t have AI that could come into your house, that it had never seen before, locate the kitchen and clear the dinner table.”Wooldridge said data availability may become one of the industry’s most significant constraints.

He noted that large language models already consume enormous quantities of text and digital material, creating pressure to secure new sources of information for future training cycles.“The whole of Wikipedia made up just 3% of GPT-3’s training data,” he said. “Where do you get 10 times more data from next time around?”That search for data, he argued, could reshape relationships between governments, corporations and individuals.

Wooldridge pointed to healthcare systems, wearable devices and online content creators as examples of potentially valuable data sources for future AI development.“The NHS is sitting on a huge amount of data about human beings,” he said. “That’s the most valuable kind of data imaginable.”He warned that commercial pressure to obtain increasingly detailed behavioral information could create incentives for broader surveillance and monitoring.

Wooldridge suggested future generations of online influencers may routinely agree to extensive data collection arrangements in exchange for visibility and commercial opportunity.The professor’s latest work focuses primarily on game theory, which he defines as the study of interactions between self-interested actors.

He said many geopolitical disputes, commercial rivalries and social conflicts can be interpreted through a relatively small number of strategic models.One recurring example in his analysis is the “game of chicken,” in which opposing sides continue escalating until one party backs down or both suffer severe consequences.

Wooldridge compared the framework to current tensions involving the United States and Iran, describing unpredictability as a recognized strategic tactic within game theory.“You’ve got two sides with ever-escalating threats against each other,” he said. “Somebody’s got to back down at some point.

”Wooldridge added that highly unpredictable behavior can complicate strategic decision-making because opponents struggle to assess likely responses and risks. Under such conditions, he said, game theory often encourages actors to prepare for worst-case outcomes.He also criticized what he described as a growing “zero-sum” political mindset in parts of modern public discourse.

In game theory, he said, zero-sum situations are not merely competitions where one side wins and another loses, but systems where actors are incentivized to maximize damage to opponents.“This zero-sum mentality is very damaging,” Wooldridge said.

“One of the important lessons from game theory is that, actually, the majority of interactions that we’re in are not zero-sum.”He linked that framework to populist political narratives that portray economic or social gains by one group as direct losses for another. As an alternative, Wooldridge highlighted the “Veil of Ignorance,” a philosophical model developed by political philosopher John Rawls in 1971.

The thought experiment asks individuals to design a society without knowing which position they themselves would ultimately occupy within it.Wooldridge said the model creates incentives for fairer social systems because participants must account for the possibility of ending up disadvantaged. He noted that former U.S. presidents Bill Clinton and Barack Obama had both expressed interest in Rawls’ ideas.

Despite concerns surrounding AI development, Wooldridge said he remains optimistic about technology and scientific inquiry. Growing up in rural Herefordshire, he taught himself programming after repeatedly visiting a local electronics shop that displayed a TRS-80 computer in its storefront during the early 1980s.

He later completed a doctorate in artificial intelligence and went on to publish more than 500 scientific papers and multiple books, while also presenting public lectures on the social implications of AI.

Asked whether students should avoid fields vulnerable to automation, Wooldridge rejected the idea that education should be driven solely by labor market forecasts.

“I didn’t get into computing because I thought it was going to give me a good job,” he said. “I got into it because I was just really interested in it.”