Blog

Tax Bite – AI and Tax (and Legal) Advice

A recent case of a barrister being referred to the Bar Standards Board after being found to have relied on an AI tool without verifying the outputs, and citing non-existent cases in his pleadings, highlights the dangers practitioners can face when relying on AI.

This will sound obvious to legal practitioners and tax advisers, but what about when members of the public or clients use AI? What happens when a client presents you with advice it has taken from AI and is seeking to rely on it?

We are seeing an increasing number of clients coming to us with what looks like a detailed (and, on the face of the logic presented by the AI tool, seemingly confident and correct) answer from AI and asking us to confirm it. To the untrained eye, the answers look very convincing and often cite HMRC guidance and legislation supporting the conclusions. However, it is clear that AI tools can generate inaccurate material as the recent examples we have had (set out below) show.

Clients should, but perhaps don’t always, appreciate that the terms and conditions of AI tools will make it clear that the answers they give to questions and requests for information cannot be relied upon as legal advice, and the AI provider will have no duty or liability to the user for any incorrect or misleading responses. Having used AI and not sought professional advice is not likely to be accepted as a reasonable defence to tax penalties.

Example – BADR Advice

A recent example saw a seller asking their tax adviser for confirmation on whether Business Asset Disposal Relief was available to a non-employee on the sale of shares after having asked AI for its view. Based on AI’s answer, the seller was convinced the relief was available. The AI result looked very credible and included:

• a very specific piece of legislation (section and subsection);
• a quote from HMRC in a particular piece of guidance; and
• an answer that the relief was available and an assertion that this was confirmed by HMRC’s guidance mentioned above.

On close inspection of the answer however:

• whilst the legislation existed, down to the section, the specific subsection did not exist, and never had done;
• whilst the guidance referred to was a genuine piece of guidance, the quote from HMRC was not contained in it, nor could it be found in any other HMRC guidance; and
• the answer given by AI that the relief was available was incorrect, as the AI tool had correctly identified a way in which the relief could conceivably have applied, but it did not apply on the facts of the case.

This example highlights not only the risk of AI “hallucinations”, about which there has been much publicity, but also - and in defence of AI – the importance of the detail and quality of the inputs. The answer to tax questions such as whether a particular relief is available will of course depend on the detail of the facts and AI is not well placed to draw out all the relevant detail from the user before forming its answer.

Example – Residence

Another recent example related to a client who planned to leave the UK and cease to be UK resident for tax purposes for the following tax year onwards. The client had done his research using AI and again presented this for checking. Although the response included all the relevant aspects of the statutory residence test, it did not accurately cover some of the finer nuances including how the “UK home” limb of the automatic UK resident test works and how long it is necessary to be resident abroad to avoid falling foul of the temporary non-residence rule. Without proper advice this client may have been non resident a year later than he thought, and even if not he may have come back to the UK too soon to avoid UK tax on gains.

This will be an ongoing challenge for advisers. In our experience AI tools have a tendency to mirror the optimism of users, who put questions to the tool hoping to receive the answer they want to hear and, notwithstanding the fact that AI-generated answers clearly cannot be relied upon as legal advice anyway, often won’t give the AI tool all the relevant facts to give it a fighting chance of giving the correct answer.