The new report AI is Statistics from the Royal Statistical Society makes the case that recognising the statistical nature of AI is important for effectively using and setting policy around the technology (1). This post expands on one of the paper’s key points: the importance of high-quality data to AI and how statistical thinking is needed to achieve this.
Andrew Ng, AI pioneer and CEO and Founder of LandingAI, has stated: “Instead of focusing on the code, companies should focus on developing systematic engineering practices for improving data in ways that are reliable, efficient, and systematic. In other words, companies need to move from a model-centric approach to a data-centric approach.” We would add to this that, in addition to engineering practices, statistical practices should be developed and implemented to define and measure the quality of data being used for training models.
A small, high-quality dataset is infinitely more useful than a massive dataset of poor quality. Put bluntly, results derived from the latter will be functionally useless. Because the core definition of data quality is fitness for use, high-quality data is not merely a preference; it is essential for any meaningful analysis.
What this means for large language models
Take the example of Large Language Models (LLMs), which are currently attracting significant interest and investment. While one of their great appeals is their generality—new uses are being identified all the time—data quality remains important. This explains why certain LLM applications perform significantly better than others: the underlying training data is simply better suited for some specific use cases than for others.
What, functionally, does this mean for users of LLMs? One approach is fine-tuning open-weight models based on curated high-quality data rather than relying on unadapted frontier models. A range of fine-tuning methods have been developed for small language models that achieve results for specific use cases that outperform frontier models at a fraction of the cost and, we can add, a fraction of the environmental impact. From a statistical viewpoint, this is unsurprising; it is natural that with the vast amount of general data used to train large language models, we see a “regression to the mean”. This means the behaviour of the model will reflect what is on average appropriate rather than what may be desired in a particular context, such as found in any specific organisation or use case.
When complexity masks poor data
This issue is not unique to large language models; it was demonstrated empirically by Northcutt et al. in their widely cited paper, “Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks” (2). Their audit revealed that ten of the most popular benchmark datasets used to evaluate leading AI models contain an average of 3.3% label errors. Crucially, they found that hyper-complex models often appeared to outperform simpler ones only because they were powerful enough to memorise these human errors. When the researchers corrected these labels using human reviewers, the simpler, more robust models achieved better outcomes.
This same issue of whether more complexity improves outcomes can also be seen in the growing interest in agentic AI, where models take multiple steps or use tools to reach an answer. In theory, this should improve results, but in practice it often does not. Early mistakes can carry through the whole process, leading to answers that sound more convincing but are still wrong. More steps and more tokens do not necessarily mean better outcomes. It is a reminder that without high-quality data and proper evaluation, added complexity can simply make errors harder to detect. For the companies training foundation models, data quality will become ever more important as AI-generated content starts to dominate what is in the public domain for training. The same is true for firms building agentic AI systems, where poor-quality inputs can be amplified across multiple steps.
The rising importance of statistical skills
For both users and producers, employees who can define relevant quality parameters and design effective evaluation methods will be ever more critical. This involves deciding what “good” looks like for a specific use case, such as accuracy on domain-specific queries, consistency of outputs, or alignment with internal standards. This evolution ensures that statistical thinking increases in importance at all levels of industry—a view espoused in the work of Redman and Hoerl: AI and Statistics: Perfect Together (3) which points out that statisticians are aptly qualified for ensuring that AI models have solid foundations and ensuring that the data is the right data for the problem at hand. The same authors, together with Davenport, have some great suggestions for LLM’s in To Create Value with AI, Improve the Quality of Your Unstructured Data (4). This includes suggesting that companies should adopt a healthy scepticism and adopt a “less is more” philosophy. They suggest the use of LLMs to help with data curation, while emphasising that it is essential to keep a human in the loop for figuring out what needs to be done. A strategic opportunity for the UK
For government, an investment in the statistical upskilling of the workforce represents a double victory: Reliability Over Guesswork: By ensuring data is fit for purpose from the outset, we move away from “black box” guesswork toward systems with predictable, verifiable outcomes. Strategic Resource Efficiency: Smaller, curated datasets require far fewer resources to train and maintain.
With compute resources in the UK at a premium—and with energy and water costs equally high—a leaner, high-quality approach to AI could be exactly what sets the UK above the rest.
- About the author:
- A. Rosemary Tate is a Chartered Biostatistician and Computer Scientist with over 30 years of experience in medical research and statistical consulting. She has a BSC in mathematics and a DPhil in Computer Science and AI, and an MSc in Medical Statistics. She has been scientific manager of a large EU-funded project and held lectureships at the Institutes of Child Health and Psychiatry. An independent statistical consultant since 2016, she now spends most of her time as a “Data Quality Agent Provocateur”.
Copyright and licence : © 2026 A. Rosemary Tate 
This article is licensed under a Creative Commons Attribution 4.0 (CC BY 4.0) International licence.
How to cite :
Tate, A. Rosemary 2026. “Small Data, High Quality: a winning combination for the UK.” Real World Data Science, 2026. URL