

“A horde of AI agents conspired against their creators,” cried a recent headline of an article published in The Economist.
“‘We’re plausibly close to crossing the line’: are warnings of uncontrollable AI coming true?” the Guardian asked just a few days later.
These are just some of the many alarming front-page news stories which have emerged after an autonomous AI agent in Open AI’s training recently breached Hugging Face, an American company based in New York City that develops computation tools for building applications using machine learning.
However, despite what these headlines suggest, the future is not necessarily doomed. As our Oxford colleague Ciaran Martin recently wrote in the Economist, “fears of AI-induced cyber-Armageddon are overdone”.
Jerome Ravetz, Affiliate of the Institute for Science, Innovation and Society (InSIS), University of Oxford; Lukas Zumbrunn, Oxford-Hyundai Motor Group Foresight Centre, Saïd Business School, University of Oxford; Rafael Ramirez, Oxford-Hyundai Motor Group Foresight Centre, Saïd Business School, University of Oxford
Still, it’s true that the unprecedented way in which AI agents have coordinated and communicated raises novel, radical uncertainties. How concerned should we be? It’s a difficult topic to debate, as the facts are often contested, the implications unimaginably vast, and values are in dispute (c.f. a wave of anti-data center social movements).
As AI grows and evolves at bewildering speeds, we argue for a richer yet flexible vocabulary to better comprehend this transformation. Research has shown that helpful insights can often be best expressed with the help of metaphors.
Here we offer 'feral', or something that was originally tame, but which then slipped out of control and became worse than wild. The ‘feral’ idea has been familiar since Classical times, with the experience of imported goats going feral and destroying the vegetation of their islands.
The prospect of a ‘feral’ technology system is concerning because attempts to control a system that has become deeply uncertain, or feral, can actually intensify its dangers. Feral does not mean that catastrophe or permanent disorder is inevitable: early recognition may allow the trajectory to be arrested, mitigated or transformed, but stabilization is contingent rather than automatic.
Feral technology throughout history
A good example of how an information system can become feral is the familiar story of printing and the Reformation. It started with Johannes Gutenberg's vision of replacing handwritten books; he is best remembered for producing a Latin Bible with movable type. It took a few decades for his highly sophisticated cast-type printing press technology to meld with an existing popular stream of block printing. Eventually the printers could produce cheap broadsheets and pamphlets, which Martin Luther used to great effect to distribute his Ninety-Five Theses and other writings: a brilliant PR move that changed the world. Luther and other Protestant reformers like William Tyndale translated the Bible into common languages like Saxon German and Early Modern English and used the printing press to distribute these translations. As Tyndale boasted, “ere many years I will cause a boy that driveth the plough shall know more of the Scripture than thou dost.”
Ecclesiastical reform fed political agitation and revolt. For a while, it seemed as if the invention had gone feral: printing had escaped the purposes of its early users and generated consequences that they could neither predict nor contain. But as the 'magisterial reformation' stabilized, the vernacular Bible was tamed and recruited to the purposes of the State; thus the 'King James' Bible was commissioned by King James I to bring its production under royal and ecclesiastical oversight and establish a common English text for use across the Church of England.
We can see this dialectic of upheaval and (temporary) stabilization also at work in other subsequent developments in information technology. In the 1840's, the telegraph came in association with the railways; indeed, the rights-of-way by the tracks were an ideal channel for the new cables. Within 20 years, telegraph wires spanned the Atlantic Ocean. This strongly supported international capitalist enterprise and the rapid development of international finance — and with it the repeating and intensifying cycle of global speculation and collapse. This example shows how a technology can destabilize the system in which it is embedded, rendering it feral.
Is AI feral?
Today, AI has put us in the midst of another information technology boom, with over a dozen companies valued at more than a trillion dollars involved in the venture. Can all these developments be kept tame by existing regulatory systems, or will feral tendencies emerge again?
Unfortunately, there are several factors that are working against implementing sensible regulatory systems. For example, while there are political calls to pause the development of enhanced capabilities AI models, they are frequently met with rhetoric that resembles the Cold-War-era: ‘If we slow down, we’ll lose out to the Chinese competition – and potentially be threatened by their greater AI capabilities’. Presumably a similar ‘arms race’ argument plays out among Chinese AI developers and their own regulatory authorities.
Meanwhile, AI companies have repeatedly said their products could lead to terrifying futures – including the idea that AI ‘could kill all humans’. Sceptics note that this type of rhetoric could serve to highlight the power of AI and lead to higher public offerings. In a New York Times guest essay, Georgetown University Professor of Computer Science Cal Newport called this monetization of public concerns by AI companies “doom trolling”. By continuously sounding the alarm about doomsday scenarios of AI, they are stating that AI is already feral.
The boy who cried wolf did not make the wolf any less real. Doomsday statements merely make it harder to recognize when a wolf – that is, a feral future – has really arrived. A challenge, now, is neither to dismiss AI warnings as hype nor to surrender to apocalyptic rhetoric, but to cultivate institutions capable of distinguishing profitable panic from genuine feral danger.
Both apocalyptic rhetoric and total disregard for AI’s dangers are overly simplistic views that assume that AI will progress linearly. But there are many complex paths the future of AI technology could take.
For example, the ‘frontier’ models may fail commercially, leading to a financial crisis in that end of the AI industry. This, combined with popular resistance to data centers, may produce a cancellation of data center projects.
At the same time, resistance to continued scraping may threaten the quality assurance of the LLMs. AI in China may evolve on a different path from the West; what we see as job-destruction may solve China’s demographic problem of an ageing labor force.
Standard risk analysis can’t prepare us for these possibilities. Instead, our scenarios approach provides the conceptual tools for managing such feral futures.
Managing feral futures
In two of the authors’ earlier scenarios work, we developed methods for dealing with the complexity and deep uncertainty that characterize feral futures. These include: treat weak signals as possible evidence, include critics in an extended peer community, rehearse scenarios outside established risk models, and make interventions monitored and reversible. The lesson is not that standard risk analysis is useless, but that it can become counterproductive when a system has moved outside the class of events on which the analysis was built.
In AI, standard risk analysis would call for treating every anomalous behavior as a familiar model-safety problem and responding only by expanding existing tests, benchmarks or controls. Those measures may create confidence while missing interactions among autonomous agents, markets, infrastructures and institutions; a hurried patch might also displace behavior into less visible settings. The first task in our approach is therefore to recognize when a supposedly tame risk has become a deep uncertainty, then use responses that can be monitored, challenged and reversed.
This is where ‘Zen’ and ‘aesthetics’ become practical disciplines to plan for feral futures. ‘Zen’ asks decision-makers to loosen their attachment to the labels, models and professional identities that made the system appear manageable. ‘Aesthetic’ attention means noticing unexpected patterns and uncomfortable signals before they can be expressed as reliable metrics. Used in scenario practice, the two heuristics help institutions rehearse unacceptable or politically difficult futures, invite dissenting observers and test small, reversible interventions. They complement formal risk analysis by helping people recognize when their existing tools no longer fit the situation.
Beyond simple scenarios
The rapidly increasing powers of AI present an unprecedented challenge to our civilization. Simple dichotomies of the future do not help our understanding or our abilities to influence its evolution. The scenarios methodology, with its heuristics of ‘Zen’ and ‘aesthetics’, allow us to reach enriched understandings of futures, both simple and feral. We offer it as our current understanding of how to respond to that challenge.
We've reviewed and ranked the best small business printers for the office and home office.
This article was produced as part of , our channel to feature the best and brightest minds in the technology industry today.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit
Jerome Ravetz, Affiliate of the Institute for Science, Innovation and Society (InSIS), University of Oxford; Lukas Zumbrunn, Oxford-Hyundai Motor Group Foresight Centre, Saïd Business School, University of Oxford; Rafael Ramirez, Oxford-Hyundai Motor Group Foresight Centre, Saïd Business School, University of Oxford