At the start of July, Turkey announced it was implementing artificial intelligence in its continuing efforts to tackle tax evasion, joining other countries such as the UK, US and Canada in seeking smarter ways to reduce the tax gap.
Within the EU this is not new; 18 member states make regular use of machine learning within their tax administrations, with some models being used as early as 2004.
The EU has developed its own machine learning system to combat carousel fraud.
Machine learning is a great tool for analysing big data, finding commonalities between data sets, clustering information, and highlighting anomalous findings on a large scale.
The use of these tools makes sense; tax administrations can process data in scalable and efficient ways – but the use of such tools is changing the dynamic between authorities and their tax base.
AI concerns
The use of this technology at scale, however, raises several legitimate concerns.
One of these concerns regarding implementing AI as an investigatory tool is the existence of bias within the system.
A typical machine learning model arrives at the desired output after it has been trained using test data and tweaked to provide correct outputs.
The system’s success depends on the quality of the data used and rigorous standards and procedures to ensure correct and consistent outcomes for the developers. Great care must be taken to ensure that test data is absent of social (or other) bias.
This equality is not limited to data; any human interaction manipulating or assessing the output whilst the model is being trained is also crucial to preventing biased outputs.
It may be assumed that the larger the training data, the fairer and more representative the system may be, especially given the large amount of data tax authorities hold.
This was an unfortunate lesson for the tens of thousands of victims of “’toeslagenaffaire’, the Dutch child care benefits scandal.
Great care must be taken to ensure that test data is absent of social (or other) bias.
In 2013, the Dutch tax authority implemented a self-learning algorithm to help uncover the early stages of benefit fraud. Penalties were wrongly levied on tens of thousands of families, typically from lower incomes and ethnic backgrounds.
The impact was harrowing: families were forced into poverty, there were suicides, and more than a thousand children were taken into foster care.
This was not a fault of the system, blame surely lies in the government representatives that allowed the system to go live without the relevant safeguards in place to protect honest taxpayers.
Transparency and explainability are pillars of the OECD AI principles, which state that AI actors should commit to transparency and responsible disclosure regarding AI systems and provide information that enables those adversely affected. No such transparency was offered to those affected in the Netherlands.
The EU AI Act details a risk-based framework requiring private companies operating high-risk AI systems (any system that profiles based on individual data such as processing of personal data, behaviour, location, etc) to be regulated, and for any such system to be designed to implement human oversight.
Why should these rules exclusively apply to the private sector?
Government bodies hold a huge amount of data on individuals, big data that carries significant risk when relying on AI-based determinations.
Interactions with tax authorities are typically opaque. Taxpayers receive notice of determinations or investigations without the authority requiring them to disclose on what basis these matters have been determined.
Taxpayers are more likely to be cynical about an AI-based investigation, potentially eroding the trust between authorities and their customer base.
Similar concerns over privacy, bias, and transparency were raised in the US following the most recent tax filing season.
The IRS is implementing AI in the audit process to tackle sectors where investigations have declined, such as large partnerships and corporations.
This spring, Rishi Sunak’s Conservative government announced the upgrade of the UK’s Snap system, the government’s AI-powered fraud detection tool, by adding three new data sets: UK & US sanctioned entities; World Bank debarments; and UK dormant companies not receiving income.
Taxpayers are more likely to be cynical about an AI-based investigation.
Future projects are considering using AI to tackle ‘phoenixing’, the process of registering and bankrupting successive companies to avoid paying debt.
Machine learning systems operate routinely without an inherent ‘understanding’ of the data. Interestingly, most (if not all) systems used by tax administrations have not been trained using reinforcement learning (the model learns the optimum outcome by undertaking a series of trial and error-based tests).
This may be due to a rewards-based system being ill-fitting for the purposes of the goals of the tax-administration or simply down to costs (extremely resource-intensive to design).
What if a taxpayer or corporation had an abnormal year, with reported figures sitting just on the wrong side of the threshold for such a system to flag an audit?
Audits are costly and time-consuming, and there may be a simple answer to the anomalous data that may mitigate an audit. This may challenge how audits and investigations are conducted, with the need for a more simplified initial approach to deter erroneous suspicions.
We also need to consider taxpayers’ or agents’ use of AI. It is no secret that the Big 4 firms have been developing their own AI models in response to enquiries and investigations, leveraging their big data to predict responses from tax authorities.
Large language models make it easy to process written communication and interpret tone, context and outcome. Historically, all of these elements of communication from a tax authority may have been a summation of the inspector leading the case.
AI can also be used to help build trust between administrations and their tax base and assist with tax collection.
Still, with hundreds and thousands of outcomes available to train their AI models, we are inevitably entering a highly gamified system where larger entities possess greater powers to achieve a desired outcome than other agents.
Shifting the lens away from using AI as a tool to combat fraud, it can also be used to help build trust between administrations and their tax base and assist with tax collection.
In early 2018, the Irish Revenue Commissioners examined the use of a natural language processor to deliver improved customer services in a subset of calls from taxpayers relating to clearance matters.
In 2021, it was reported that 50 per cent of calls were handled from start to finish by the voicebot, 75 per cent of tax clearance holders were able to retrieve access numbers from the bot, and only 10 per cent were transferred to human operators due to the inability of the bot to understand the user request.
In 2017, the Spanish Tax Agency introduced a VAT chatbot to assist taxpayers with their understanding and obligations.
The UK has much to learn from these efforts. With the news of upgrading Snap to combat fraud at a time when HMRC customer service levels are at a historic low, it is evident a shift is needed to use this technology to facilitate willing taxpayer access to HMRC.
Ben Lee is a partner at tax consultancy Anderson
