Conrad Young gave the Chartered Institute of Taxation (CIOT)'s 2024 CTA address last week to a packed audience on The impact of AI on tax. What do tech expert Chartered Tax Advisers (CTAs) think is going to be the impact of AI on tax?
Conrad Young, Chair of the Advisory Board and A visiting Policy Fellow of the Oxford Internet Institute, Oxford University and lead speaker for the 2024 CTA Address took the audience through the evolution of AI from the 1950s up until Chat GPT passed the US bar exams in 2022.
He considered where we are now.
The early focus, and a key focus in AI remains in Rules-Based Learning. This is one way to pass a US Bar exam, however when something is rules-based this method results in some extremely long and complex decision trees.
Greater research into the workings of an actual human brain have led to the development of neural networks and on to the creation of Discriminative AI.
Discriminative AI is great for spotting differences and therefore useful for sorting data, spotting patterns and trends. For example, taking photographs and deciding whether it is looking at a cat or dog. To achieve its ends, it has to be trained on data and this done in a supervised or unsupervised form with reinforcement of adding more positive training data.
Generative AI is a newer development: this is evidenced in the large language models that we all note - e.g. Chatbot GPT. This is great for creating models and content. Generative AI does suffers from ‘hallucinations’, we cannot accurately track the way that it creates its content and therefore is not a reliable method of generating advice.
None of these AI actually think or reason in the same way as humans, all require vast amounts of data, human intervention and ‘prompts’ to steer them.
The OECD’s Tax Administration 3.0: The Digital Transformation of Tax Administration paper explores a digital transformation of tax administration which could potentially lead to embedding taxation in processes into the natural systems that taxpayers use in their daily lives and businesses. This will means looking at far broader concepts than simply digitalising existing tax processes, (although, this is not to say that tax legislation could not be drafted in a more AI friendly way).
In terms of how we process tax legislation, Conrad noted that tax legislation is set to achieve different ends, depending on the policy objectives of government and may be designed to create fairness and influence taxpayer behaviours as well as raise revenue.
Responding to the address were:
Bivek Sharma, Chief Technology Officer for PwC UK
Bivek has described his firms development of AI tools. He gave the example of a data room of some 55k documents. AI is now able to scrutinise those documents spotting words and clauses that have a potential tax consequence. This task can be done in minutes, and it saves several weeks of human time. The results are about 80% accurate, human intervention is required. AI is useful as it can also be trained on data sets which aids comparison of companies. This saves humans hours and hours of sifting through company data.
Shan Sun, Tax Technology Lead at Deliveroo
Shan noted that AI whilst extremely useful for things like transfer pricing as who like spending hours on that kind of analysis. She noted that the current AI models are just not very good! The lack of transparency in AI modeling and creation of bias in algorithms is an ongoing issue. She cited the book "Weapon of Maths Destruction" as recommended reading in this area: no one actually knows how generative AI creates its outputs.
Her thoughts are that modelling must allow for AI to make errors to validate testing outputs. So much depends on answering the right questions. She concluded that we have a considerable way to go before AI does what we want it to do.
All panellists agreed that 'garbage in, garbage out' is very much a thing, AI as it is now requires a vast amount of human intervention and prompts. There is the problem of a lack of audit trail in decision making, particularly when it comes to the use of Generative AI. Although AI can be used to test the output of AI, that is one highly useful feature. Those relying on AI must necessarily consider safeguards before they can place reliance on its output.
Will there be changes? Definitely, AI is expanding rapidly, its impossible to predict where we will be in two years , let along ten.
When asked should we abandon a rules based tax legislation, Conrad said 'No'. Clearly he believes there is too much politics at stake for the UK to take on the OECD concepts just now.
External links
OECD: Tax Administration 3.0: The Digital Transformation of Tax Administration paper
Wikipedia: Weapons of Maths Destruction
CIOT Diploma in tax technology
Learn more about tax and the impact of AI
