To cultivate robust AI, fertilize the soil it grows in

The race to develop more powerful artificial intelligence is truly global. So, increasingly, are the risks. An AI system developed in one country may run on infrastructure in another, draw on data stored in a third and affect people everywhere.

No country can govern that chain on its own. Yet meaningful international coordination will remain elusive unless we understand what exactly needs to be governed.

For years, the dominant metaphor was that "data is the new oil". It is memorable, but somewhat misleading. Oil is scarce, costly to extract and depletes through use. In contrast, data is potentially infinite, often surrendered for free and capable of being copied and reused indefinitely.

A better metaphor may be soil.

Just as crops grow from soil, AI systems grow from data. The quality, composition and provenance of that soil matter. Contamination can spread. Soil that nourishes one crop may be unsuitable for another. Above all, to remain productive, the soil requires continuous care and cultivation.

This updates the old warning: "garbage in, garbage out". With generative AI, bad soil does not necessarily produce a visibly bad crop. A system trained on flawed or contaminated data may generate a response that appears polished, persuasive and entirely plausible.

That output may then be published, relied upon and eventually ingested as training data for another model. The danger is not just that contamination is consumed, but that it is replanted.

Early debates about responsible AI focused mainly on training data: whether it was biased, legally and ethically obtained, and of sufficient quality. Those questions continue to be important, but the focus has now shifted from training to operation.

Many newer AI systems retrieve live information, connect to databases, use software tools and remember previous interactions. AI agents may read and write documents, communicate with customers, make purchases or modify records. The data environment is no longer fixed when the model is trained. It changes continuously as the system operates.

Therefore, it is impossible to ensure responsible AI simply by inspecting a dataset once and declaring it clean. We need governance of data flows, not merely datasets. Organizations must know what information a system can access, what it can infer, where it can transmit that information and what actions it is authorized to take.

Search engines learned what information we wanted. Social media platforms learned whom we knew and what kept us engaged. AI companions and agents may infer something even more sensitive: what we feel, fear and trust.

Over time, systems can infer preferences, vulnerabilities and personal relationships from trivial interactions.

No individual exchange may appear sensitive, yet the accumulated data can create a profoundly intimate profile.

Data protection has traditionally relied heavily on user consent. But few people can anticipate every future use or inference an AI system may make. Before entering a continuing relationship with such a system, users may need the digital equivalent of a prenuptial agreement: clarity about what it will learn, retain, share and delete when the relationship ends.

Consent remains important, but it cannot ensure comprehensive protection. We need accountability and contestability. Who approved access to the data? Can decisions be traced? Can affected people challenge or reverse them? Who is responsible when something goes wrong?

Some see governance as a stop sign in the path of innovation. A better comparison is a seat belt. It does not prevent movement; rather, it allows us to move faster with greater confidence. Poor governance may permit a rapid experiment, but it rarely supports deployment at scale. Trust, documentation and clear responsibility are part of the infrastructure of innovation.

Controls should depend on capability and consequences. The greater a system's autonomy, the wider its access to data and tools, and the more irreversible its actions, the stronger those safeguards should be.

The aim cannot be to eliminate all risks. The only completely safe AI system would be one switched off and locked away. Instead, organizations should seek to bound risk by restricting access, requiring approval for consequential actions, preserving audit trails and expanding autonomy only when evidence justifies it.

The world is unlikely to agree on a single global AI regulator or universal rulebook. Countries differ in legal traditions, economic priorities and social values. But agreement should be possible on minimum expectations concerning data provenance, security, testing, incident reporting, auditability and responsibility among developers, deployers and users.

This is not an argument for regulatory uniformity, still less for leadership by any single country. It is an argument for interoperability. Without it, companies face incompatible obligations, while individuals could fall through the gaps between jurisdictions.

Major powers will continue to compete over AI. That competition makes coordination more urgent, not less. Aviation, telecommunications and finance combine national regulation with international standards because cross-border systems require common foundations. AI will be no different.

The essential principles are simple: know your data, know your AI agent, and know who is responsible for its actions.

AI may flourish in many kinds of soil. But neglected ground does not remain neutral. Weeds take root, contamination spreads and healthier growth is eventually crowded out. Building responsible AI therefore requires more than celebrating the harvest of technological breakthroughs. It requires sustained cultivation — and cooperation wherever the roots and seeds cross borders.

The author is David Marshall professor of law at the National University of Singapore, where he is also AI Governance and Policy lead at the NUS AI Institute and dean of NUS College.

The views don't necessarily reflect those of China Daily.

If you have a specific expertise, or would like to share your thought about our stories, then send us your writings at opinion@chinadaily.com.cn, and comment@chinadaily.com.cn.