Why Aren’t You Listening to Me? The Risk of Taking AI Tax Advice at Face Value
Hoxton Blog • Why Aren’t You Listening to Me? The Risk of Taking AI Tax Advice at Face Value
I use AI a lot in my work and, generally, I am a big advocate for it
It helps me organise thoughts, work through information and, quite often, turn something I have dictated at speed into something considerably more readable.
But anyone who uses AI regularly will probably recognise one of its more frustrating habits. I can dictate something, give what I think is a very clear instruction and then read the answer back only to discover that AI has decided to interpret what I meant rather than simply do what I asked.
Sometimes it changes the emphasis, sometimes it makes an assumption I didn't make, and sometimes it helpfully "improves" something that I specifically didn't want changing. Which quite often leaves me thinking: why aren't you listening to me?
For me, most of the time, that is simply annoying. I know what I wanted to say and, more importantly in the context of my work, I know enough about the underlying tax position to spot when AI has changed the meaning or taken something in a direction I didn't intend. I can correct it, challenge it or go back to the legislation and guidance and check the position.
The bigger risk...
The bigger risk comes when the person asking the question doesn't have that underlying knowledge.
If AI misunderstands part of the question, makes an assumption, or applies its own interpretation to the facts, the answer can still sound completely convincing. It doesn't necessarily come with any obvious warning that something has gone slightly off course, and if you don't already know enough about the subject to spot the problem, why would you question it?
That is where I think the conversation around AI and tax becomes much more interesting. I don't think the issue is simply whether AI can get tax "right" or "wrong". It is whether it understands enough of the context to know which question it actually needs to answer.
Tax has always been full of situations where one seemingly small fact changes the outcome, and these three examples demonstrate that particularly well the danger in relying on it.
Split Year Treatment: Knowing Which Calculation Applies
The UK's Statutory Residence Test is a good example of an area where AI can provide information that appears correct while missing an important part of the analysis.
Someone researching the sufficient ties test may quite reasonably discover the familiar thresholds of 15, 45, 90 and 120 days, depending on their circumstances and the number of UK ties they have. They might then use those figures to determine how much time they can spend in the UK without affecting their residence position.
However, where an individual is relying on certain cases of split year treatment, the position can be more complicated because the day-count thresholds applicable to the overseas part of the year may need to be reduced according to the date on which the relevant split occurs. A January split, for example, can result in thresholds of 12, 37, 75 and 100 days rather than the normal full-year thresholds.
That distinction can materially change the advice being given, particularly for an individual who is planning their UK visits relatively close to one of those limits.
The issue here is not necessarily that AI has performed the calculation incorrectly. It may have accurately explained the sufficient ties test based on the question it was asked, while failing to identify that another provision needs to be considered first. A tax adviser looking at the wider circumstances should instead be considering which split-year case potentially applies, the date from which the overseas part begins, the relevant ties during that period and the resulting permitted day count.
This is one of the fundamental differences between retrieving tax information and applying tax advice: knowing the rule is important, but identifying which rule needs to be applied to the particular facts is often where the real work begins.
When AI Invents The Law
A more obvious risk arises when AI produces information that is not simply incomplete but does not exist at all.
This issue reached the First-tier Tribunal in Harber v HMRC [2023] UKFTT 1007 (TC), where a taxpayer appealing a Capital Gains Tax penalty submitted details of nine supposed tribunal decisions in support of her case.
The Tribunal found that none of those authorities were genuine and concluded that they had been generated by an AI system such as ChatGPT.
Importantly, the Tribunal accepted that the taxpayer did not know the cases were fabricated. She had been provided with case names, dates and summaries that appeared sufficiently credible for her to believe that they represented genuine authorities.
This illustrates one of the more difficult aspects of relying upon AI in a technical environment. An incorrect answer does not necessarily look incorrect; it can be well written, detailed, and presented with the same confidence as an accurate answer.
Unless the person receiving that information has the knowledge, experience or access to appropriate sources to verify it, there may be very little reason for them to question what they have been given.
For a professional adviser, checking legislation, HMRC guidance and case law is a routine part of reaching and supporting a technical conclusion. For someone using AI to answer a tax question independently, there is a danger that the apparent authority of the response becomes the verification process itself.
When the Tax Rules Have Changed
The third difficulty is particularly relevant following significant changes to the taxation of internationally mobile individuals.
For many years, discussions around the UK taxation of foreign income, gains and overseas assets centred heavily on domicile, non-domiciled status and the remittance basis. From 6 April 2025, that landscape changed substantially, with the remittance basis abolished and the introduction of the four-year Foreign Income and Gains regime for qualifying individuals.
Inheritance Tax also moved towards a residence-based framework, with long-term residence becoming an important concept in determining the potential exposure of non-UK assets.
The distinction matters because long-term residence is not simply another name for domicile, nor should advice produced under the previous regime automatically be carried forward into the new one.
An individual might therefore ask AI about their position as a "non-dom", or search for information based upon terminology that would have been entirely appropriate only a few years ago. Depending on the information being accessed and the way in which the question is framed, the resulting answer may blend elements of the old and new regimes or fail to recognise the new regime entirely.
A professional adviser should recognise that immediately. Before answering the tax question, we may need to challenge the terminology being used, establish the relevant tax year, understand the individual's residence history and determine whether the new Foreign Income and Gains or long-term residence rules are relevant.
Again, the difficulty is not necessarily the ability of AI to explain a tax rule; it is whether it identifies that the rule the individual is asking about is no longer the rule that determines their position.
So, should we stop asking AI tax questions?
AI is an incredibly useful tool, and it is only going to become a bigger part of how we work and how clients research their own affairs. If someone comes to a meeting having used AI to understand the basics, that can actually make for a much better conversation.
The important bit is knowing where information ends, and advice begins.
A tax adviser isn't simply there to know a rule that somebody else couldn't find online or ask ChatGPT about. Much of our job is working out which rule applies to the particular person sitting in front of us, asking the questions they might not have realised were relevant, and spotting the small detail that changes the answer.
There is also one final difference which is easy to overlook. When professional tax advice is given, there are professional standards and responsibilities sitting behind it, together with professional indemnity insurance. If that advice is negligently wrong and causes a financial loss, there is a framework for dealing with it.
AI doesn't come with the same protection. There is no professional indemnity policy sitting behind the answer on your screen, and AI isn't taking responsibility for the tax, interest or penalties that could arise if you act on an answer that turns out to be wrong.
None of that means we shouldn't use AI to understand tax. Used well, it can help people research their position, understand unfamiliar terminology, and ask much better questions when they do speak to an adviser.
Perhaps the better way to think about it is that AI gives us access to more information than we have ever had before, but access to information and knowing how to apply it are still two different things.
When the tax consequences matter, there remains real value in having someone who understands the context, knows when the question needs to change and, importantly, stands behind the advice they give.
AI can be part of the conversation. It just shouldn't always have the final word.
How Can We Help You?
If you would like to speak to one of our advisers, please get in touch today.
Contact Hoxton Wealth
We are available to discuss how Hoxton Wealth can help you achieve your financial goals. Together, we can help you build a brighter financial future.
