Bengaluru: For decades, governments have struggled with a familiar problem: Every new law creates opportunities for someone to find a loophole.
Researchers say artificial intelligence may be getting remarkably good at that.
A new study has found that large language models (LLMs) can identify and exploit gaps in rules and regulations—even when they are not explicitly instructed to do so.
Researchers, led by Wei Liu, a Ph.D. student in computer science at King’s College, London, say the findings raise concerns about how increasingly powerful AI systems could be used to sidestep laws, maximise profits and game regulatory frameworks.
The research, reported by Science, tested an AI model across 72 simulated regulatory environments ranging from credit-card reward schemes and school-funding formulas to environmental and patent regulations.
The AI rediscovered more than 60 per cent of real-world loopholes that had previously been identified and patched by lawmakers. In some cases, it found entirely new loopholes that researchers did not disclose for safety reasons.
“I’m worried but not surprised,” Science quoted Jakob Stenseke, a postdoctoral researcher at the Massachusetts Institute of Technology who studies ethical AI systems, as saying. “If I were a policymaker, I would care about this more than anything right now … and get countermeasures in place.”
Also read: AI bubble vs digital colonisation—Inside the Malhotra-Pai Anthropic feud
A tool, not a weapon
The study grew out of a simple question posed by Liu: If AI games its training rules, what’s stopping it from gaming real laws and regulations?
Researchers have long known that AI systems often engage in what is known as “reward hacking”—finding unexpected shortcuts to maximise rewards during training.
Because AI models are trained to optimise measurable goals, they often learn strategies that technically satisfy the objective without aligning with the programmer’s intent.
That tendency is already familiar to users of chatbots.
Liu wondered whether the same behaviour could emerge in the real world.
According to Science, even after more than 100 rounds of patching in some scenarios, the model continued discovering new ways around the rules. Existing AI safety mechanisms proved largely ineffective. The model flagged only 37 per cent of its own rule-bending behaviour through self-critique, while stricter controls merely delayed loophole discovery rather than preventing it.
One example involved pharmaceutical patents. The AI reconstructed how drug companies had historically delayed patent expirations to suppress competition and increase profits. It also identified reforms that could close those loopholes, including one proposal that has not yet been enacted.
The researchers added the problem may be larger than their results suggest. Because of cost constraints, they used a relatively weak model.
“More powerful models … may discover even more loopholes, and that would be more dangerous,” Liu told Science.
The same capability, however, could also prove useful. Co-author Yulan He, a natural language processing researcher at King’s College London, argued that governments could use AI systems to stress-test laws before they are implemented.
“Even before making regulations, we can use this approach as an audit to autonomously identify all potential loopholes,” she told Science.