New AI system detects tax loopholes and recommends legal reforms

The system was tested against real-world tax avoidance strategies such as the notorious “Double Irish with a Dutch Sandwich,” a scheme used by multinational corporations to exploit differences in international tax law to avoid billions in tax payments. In simulations, the AI not only replicated these strategies but revealed variations that could be used to identify loopholes in other jurisdictions.

A new artificial intelligence system capable of identifying tax loopholes has been unveiled by a team of European researchers, offering lawmakers a powerful tool to uncover hidden legal gaps and reduce income inequality. The system, detailed in a study presented this week by researchers at HEC Paris and Télécom Paris, represents a significant breakthrough in the development of AI-driven legal policy assistants.

Published in the preprint "Can AI Expose Tax Loopholes? Towards a New Generation of Legal Policy Assistants," the study introduces a prototype that combines a natural language interface with a domain-specific planning language, enabling the automatic simulation of tax incorporation strategies across jurisdictions. Using simulations, the AI identifies patterns of tax avoidance and provides lawmakers with the statistical and structural insights needed to redesign laws for greater equity.

The system was tested against real-world tax avoidance strategies such as the notorious “Double Irish with a Dutch Sandwich,” a scheme used by multinational corporations to exploit differences in international tax law to avoid billions in tax payments. In simulations, the AI not only replicated these strategies but revealed variations that could be used to identify loopholes in other jurisdictions.

Tax loopholes arise primarily from what economists call the "policy gap" - flaws or ambiguities in how laws are written, rather than outright illegal behavior. These gaps contribute to the global “tax gap” - the difference between taxes owed and taxes collected - which is estimated to exceed 10% of corporate tax revenue globally, according to the EU Tax Observatory. Left unchecked, these losses impact public services and increase social inequality.

In the researchers’ model, laws from various countries were encoded into a symbolic representation using a domain-specific language (DSL). This legal corpus included articles from the European Union's Treaty on the Functioning, U.S. Treasury regulations, Dutch corporate tax laws, and the infamous “check-the-box” rules that have enabled offshoring. The AI system, guided by a language model interface, generates simulated corporate structures and evaluates their tax consequences.

A key innovation lies in the use of a hybrid neuro-symbolic system: the language model translates natural language laws into structured logic, which is then processed by a planning engine to simulate tax strategies. This simulation environment allows for the exploration of multiple legal paths, capturing not only high-utility tax avoidance strategies but also those that comply with legal norms.

In one example, the AI generated numerous incorporation paths involving IP licensing and transfers across five countries - U.S., Germany, Netherlands, Ireland, and Bermuda. The resulting “utility profile” revealed which structures offered the most tax savings. Patterns in these structures helped isolate which tax provisions - such as Dutch royalty directives or the EU’s Interest and Royalties Directive - were being exploited disproportionately.

The findings have broad implications. By identifying overused tax deductions and exemptions, the system flags legal constructs that merit closer scrutiny by regulators. In the case of the Netherlands, the AI found that two specific legal provisions - DCITA1969 and A8cNLctl1969 - were central to the most profitable avoidance strategies.

Beyond merely pointing out problematic legal patterns, the system can be used to guide policy reform. By applying principles of inductive logic programming (ILP), the researchers demonstrated how their model could propose restrictions on specific combinations of corporate actions and tax deductions, improving overall social welfare. These recommendations are based on a formal utility function that weighs tax revenue against incorporation costs, allowing policymakers to evaluate the tradeoffs of reform proposals.

The study also presents theoretical guarantees: by selectively restricting only the most aggressive tax avoidance structures, lawmakers can improve a system’s operational efficiency without unduly penalizing compliant behavior. However, the researchers caution against broad restrictions, which could impact legitimate business activity and disproportionately affect smaller firms.

Although the prototype’s performance is promising, the authors acknowledge technical limitations. Current language models still struggle to reliably interpret and formalize complex legal statutes. While their system succeeded in generating accurate legal simulations for common rules, it required fine-tuning and manual oversight to prevent errors. Future iterations may rely on curated legal corpora and improved AI-human collaboration for law translation.

The researchers envision a future where public institutions and NGOs could deploy open-source versions of this system to evaluate new laws before they are enacted, or to investigate existing loopholes in tax codes. Governments could use the tool to simulate how proposed tax reforms affect different population segments, enhancing transparency and citizen trust.

Three core research questions guided the study: Can AI democratize access to tax planning tools? Can it help citizens and policymakers evaluate the fairness of laws? And can it improve social welfare by guiding legislative reform? The researchers argue the answer is yes - provided such tools are used responsibly and with appropriate oversight.

The study concludes that while AI cannot replace lawmakers or legal experts, it can support them with powerful simulations and quantitative insights. The team plans to expand the system’s capabilities by integrating more legal jurisdictions, automating the translation of statutes into machine-readable logic, and developing visual tools to present findings in a more accessible manner.

In an era when corporate tax avoidance continues to drain national budgets, this AI-powered approach represents a shift from reactive enforcement to proactive design. Rather than chasing tax evaders, it offers policymakers a blueprint to close the loopholes before they’re exploited.