Artificial Intelligence (AI) is a threat to democratic norms. Machine learning bots have been used by malign forces to disrupt democratic processes. The literature as a result is replete with warnings! Can the threat be turned on its head – to use AI to advance democratic norms? This is not about the efficiency benefits that AI will produce, often lauded in contradistinction to AI’s problems. There are very important values that can be fostered by AI.

To investigate how to improve democratic participation and render the tax system more legitimate in the UK, I obtained funding from the British Academy and Leverhulme Trust for a project which took place between May and November 2025. The money was used for the purpose of hosting two workshops (one of which was hybrid) were held at King’s College London where the project findings were discussed (and later revised).

A policy paper, available to download here, was the result of these efforts. The paper makes 5 recommendations for the use of AI in the tax system to advance democracy:

Recommendation 1: Feedback on priorities

A Generative AI (GenAI) model could be used to accommodate mass taxpayer input (ca. 1,000 responses) on the priorities of His Majesty’s Revenue and Customs (HMRC), and how success in achieving these priorities is measured. Following a scheme of open-ended, targeted and closed questions over several sittings, such a model would exploit the capabilities of GenAI to summarise and group answers in order to identify key perspectives, areas of agreement, and areas of disagreement. The information generated could then be transmitted directly to HMRC and His Majesty’s (HM) Treasury to be taken into account when determining resource allocation decisions. HMRC and/or HM Treasury will need to explain online the reasoning behind either accepting or rejecting the suggestions. It may be necessary to offer payment to participants at a rate of the minimum living wage to encourage broad representation of taxpayers.

Recommendation 2: Feedback on tax rules

The same process could be used to obtain mass taxpayer feedback on existing procedural tax rules. which have not been significantly reconsidered in light of technological developments.

Recommendation 3: Routine consultation

The same technologies could be used to conduct more bespoke consultations, with a more limited group of participants, to provide a “strong” base of information for government officials.

Recommendation 4: Future automation and reinforcement learning

HMRC and HM Treasury should keep an eye on developments with fully automated surveys and reinforcement learning. Consultations could be automated, with a GenAI model trained to conduct consultations. Such consultations could also be fine-tuned using reinforcement learning to gain an even more incisive understanding of taxpayer concerns and expectations.

Recommendation 5: Accountability for reasonableness

For AI-enhanced consultations on priorities, procedural tax rules and tax policy, HMRC and/or HM Treasury should commit (through a published policy) to give their (fair and convincing) reasons for not making the suggested changes.  

I am indebted to those who generously donated their time to discussing some of the paper’s ideas, namely Victoria Adelmant, Adrian Blau, Iain Campbell, Dominique Chu, Sara Closs-Davies, Francien Dechesne, Sylvie Delacroix, Flynn Devine, Bill Dodwell, Judith Freedman, David Hadwick, Dan Hunter, Vasiliki Koukoulioti, Katarina Lau, Ben Lee, Benita Mathew, Helen Margetts, Kunal Nathwani, Rhodah Nyamongo, Jeffrey Owens, Connal Parsley, Alexandra Pollitt, Siddesh Rao, Niccolò Ridi, Rahmin Sarabi, Joseph Sherlock, Joe Tomlinson and Matthew Vick.

Please read the full report and send me any thoughts (stephen.daly@kcl.ac.uk).