Representing yourself at the First Tier Tribunal is one thing, but relying on artificial intelligence (AI) as a basis for appealing an HMRC decision, without human verification, is quite another.

Discovery assessment

A recently heard case concerned a discovery assessment for just over £2,500 in tax from a high-income child benefit charge for 2018/19. The delay was largely caused by uncertainty over whether HMRC could rely on discovery assessments. The government subsequently legislated in HMRC’s favour, meaning the use of a discovery assessment cannot be used as a basis for defence even though the legislation is retrospective in its application.

Artificial, not entirely intelligent

The taxpayer’s entire defence was suspect due to over-reliance on AI:

• One aspect of the defence put forward was that HMRC should have notified the taxpayer of the charge, despite the primary responsibility for declaring tax liabilities resting with the taxpayer.
• Although none of the cases pulled up and cited by the AI used were entirely fictitious, the cases that were put forward in the taxpayer’s defence were irrelevant to the tax charge in question.

This is, of course, the problem with AI if not used correctly – the AI may not fully understand what is being asked or may miss relevant information. The result is an answer that seems plausible, but is not relevant or accurate. Indeed, the judge for the case said that there is no reason why AI should not be used to research a defence, but the results need to be checked carefully.

HMRC’s use of AI

HMRC itself, of course, makes extensive use of AI, for example, to analyse large amounts of data, such as when a taxpayer’s declared income is insufficient to support their lifestyle. Landlords should be aware that AI allows HMRC to search a range of databases, including the Land Registry, listing websites and various tenants’ deposit schemes to establish whether property income is being correctly declared.

Although aimed at public bodies, the government’s artificial intelligence playbook clearly explains the limitations of AI. The playbook can be found here.