R&D tax advisors have been spotting something odd in HMRC compliance letters lately: telltale signs of AI writing. Recently, tribunal ruling has ordered HMRC to come clean about whether they're actually using artificial intelligence to review R&D claims.

Following a Freedom of Information (FoI) request submitted in 2023 that was refused by HMRC, Mr Thomas Elsbury brought an appeal against the Information Commissioner.

The initial request was refused by HMRC on the grounds that it would enable fraudsters to better understand HMRC’s decision making processes when reviewing claims, potentially defrauding the Treasury further.

Mr Elsbury requested an internal review, after which HMRC maintained its position. Mr Elsbury went to the Information Commissioner for a review, at which point HMRC changed its stance to 'neither confirming nor denying' (NCND).

Finally, Mr Elsbury appealed to the First-Tier Tax Tribunal – the ultimate step in his quest for answers.

What Triggered the AI Disclosure Request?

At the hearing in June 2025, it was understood that some R&D tax credit advisors had noticed compliance check letters from HMRC returning with tell-tale signs of AI use, like the “em-dash” and Americanised spellings, as well as some letters being almost non-sensical.

These anecdotal concerns triggered wider worry, as the incidence rate of compliance checks rises. Has HMRC implemented an AI tool in their process that results in some misjudgements and badly written letters? Or have some HMRC agents “gone rogue” and used generative AI tools available to the public (like ChatGPT) to prepare their responses to claimants?

At the hearing, much was made of HMRC’s flip-flopping on their reasoning behind their refusal. The decision to change their refusal from a simple refusal in order to avoid enabling fraudsters to hijack the system to a NCND decision was “beyond uncomfortable and like trying to force the genie back in its bottle.”

HMRC provided no backing for their decision and “over-emphasised the (unsubstantiated and unevidenced) risks of causing an increased incidence of fraud in R&D tax relief claims”, according to the panel.

Ultimately, it was decided that the social need for transparency, particularly in the department of government responsible for the collection of taxes, outweighed the potential but unsubstantiated risks of disclosure.

Therefore, HMRC has been ordered to answer the FoI request within 35 working days or appeal the decision.

Mr Elsbury’s FoI request was as follows:

• Under the Freedom of Information Act 2000, I am writing to request information regarding the use of Large Language Models and generative AI [Artificial Intelligence], such as ChatGPT, by HM Revenue & Customs (HMRC), specifically within the context of the 'R&D Tax Credits Compliance Team'. My request focuses on the following areas:
• Implementation Details: Please provide details on how Large Language Models are currently being used within the R&D Tax Credits Compliance Team. This should include, but is not limited to, the purposes for which these models are employed (e.g., data analysis, decision-making support, enquiry responses, penalty justifications, etc.)
• Model Selection and Development: Information on the criteria used for selecting these models, any custom development or training undertaken, and the providers of these models.
• Data Privacy and Security: Details on the measures in place to ensure the privacy and security of taxpayer data when using these models, including any data protection impact assessments conducted. Given the sensitive nature of R&D claims, detail should be provided if information is being uploaded relating to the claim which is then used to train language models and be available for use by private companies engaged by the Government, who own the overarching tool.
• Outcome and Impact Analysis: Any available information or studies on the impact of using Large Language Models on the efficiency, accuracy, and overall outcomes of the enquiry processes within the R&D Tax Credits Compliance Team.
• Policies and Procedures: Information on the policies and procedures governing the use of these models, including guidelines on human oversight and ethical considerations.
• Future Plans: Details on any planned expansions or changes in the use of Large Language Models within the R&D Tax Credits Compliance Team or other departments within HMRC.

HMRC’s Review Process: Public Interest, Safeguards & Transparency

The primary concern of the appellant, and of many taxpayers interested in this case, surrounds data protection and the safety of confidential information and intellectual property provided to HMRC.

Since the introduction of the Additional Information Form, companies have been required to provide project details for their R&D tax relief claim to be accepted. Sometimes companies will provide sensitive project specifics within their AIF, on the understanding that it will be read by a caseworker and then go no further. Finding out this information is being used to train an LLM, or being held by unknown third parties, would be a significant breach of trust.

Without knowing if HMRC uses AI and, if so, in what ways, the public can have little confidence in their data being properly handled with guardrails in place. Ethics around AI are constantly being reworked as the technology progresses; avoiding hallucinations and bias is critical for the technology to progress rather than creating additional problems. Security around the data shared is equally important for users to have confidence in the responses received.

It's crucial that HMRC divulge if AI is being used and, if so, how; what safeguards are there around transparency? What information is training the models? Has HMRC implemented their own tool or modified an existing one?

This becomes particularly significant given the current heightened focus on AI across all sectors, where artificial intelligence initiatives are fast-tracked to the top of government priorities. The government has made it clear that adoption of AI in the public sector is an urgent matter, releasing an AI Opportunities Action Plan in January 2025 and allocating nearly £2 billion over the next four years to implementing this plan.

However, the Public Accounts Committee, which examines the value for money of Government projects and programmes, is calling on the government to address public concerns over the sharing of sensitive data in AI’s use.

Promoting transparency and accountability should be at the forefront of HMRC’s policies in this matter, to demystify the R&D tax credit and incentivise companies to invest in R&D. Knowing how your claim is treated and what goalpost you’re actually aiming for may serve to encourage eligible companies to make their claim without fear of lengthy compliance checks.

The Potential Risks to the Treasury

The primary reason for HMRC’s refusal was that a disclosure could “prejudice the assessment or collection of tax or duty”; in other words, understanding HMRC’s review policies could enable some actors to make ineligible claims and submit them with greater confidence they would evade HMRC’s controls.

Error and fraud is a hot topic in the R&D tax sphere, as HMRC’s investigation released in 2024 estimated that the level of error and fraud in the 2020-21 period was £1.13 billion.

Understandably, HMRC is trying to decrease that number (and has purportedly succeeded, according to the last publications on this topic, with the comparatively lower figure of £475 million being mistakenly claimed in the 2023-24 period).

In its written arguments (HMRC was unrepresented at the hearing), HMRC claimed that the level of information in the public domain satisfies the requirement for transparency.

They also claimed that disclosing the requested information may undermine compliance activity, by undermining public confidence in the tax system.

Insights from Myriad

It feels like every piece of content we read has the dog whistles of AI; em-dashes, a simplified trio of bullet points and Americanised spelling permeate LinkedIn, blogs, newsletters and emails from your boss.

The one place I don’t expect to see ChatGPT responses is in official letters from HMRC, especially in letters that reject a R&D tax relief claim for thousands of pounds. All this can do is discourage companies from making claims for fear of having to defend their claim against a computer; the antithesis of HMRC’s stated objective of the R&D tax credit scheme – to incentivise innovation and reward companies performing R&D.

It’s also a somewhat disrespectful practice. Improperly reviewing a claim is not considerate of all the time and effort put in by a claimant, both on the claim and the actual R&D project. Receiving a rejection of your R&D tax credit claim, often worth thousands of pounds and offering a lifeboat in terms of cashflow, is already a crushing moment for a business; reading on to find that the rejection was written by a bunch of ones and zeroes instead of an actual person is even worse.

The rise of AI being inevitable, we need to set a precedent of good behaviour now. Insisting on assurances from HMRC that decisions on tax matters are not being mined from easily manipulated AI tools would form a good faith foundation for AI to be responsibly implemented in the R&D tax relief claim process.

Clear transparency from HMRC will go some way to quell fears of rejection quotas, biased algorithmic decision-making and inadequate data protection safeguards.

Concerned about how AI might affect your R&D tax credit claim? Our experienced advisors can help you navigate the changing compliance landscape and ensure your claim is properly protected. Contact us today for a free consultation.