Statistics and AI: How Accurate Data Shapes Policy and the Future

Hello everyone, and welcome to "Statistics and Stupidly Smart AI." This article aims to convey the excitement surrounding the intersection of statistics and artificial intelligence, demonstrating their profound impact on our lives and the future.

The Importance of Accurate Information

To illustrate the critical role of accurate information, let's consider a simple game. Imagine being blindfolded and guided through a "dungeon" of chairs by instructions. In one scenario, the instructions are precise, leading to successful navigation. In another, the instructions are intentionally misleading, resulting in collisions. This highlights a fundamental truth: we are only as good as the information we receive.

This principle applies universally, from personal decisions like buying a house or choosing a career to governmental policies. Decisions based on flawed data are akin to building a house on quicksand—they are destined to fail.

My Journey Through Applied Mathematics

My background is in nuclear physics, but the unifying theme throughout my career has been applied mathematics. I've worked in plasma physics, authored a book on nuclear fusion, and contributed to projects at institutions like the Bank of England (including selecting math for the 50-pound note featuring Alan Turing), the Office for National Statistics, and most recently, Google. While I speak for myself, not these institutions, my experiences have consistently shown the immense importance of statistics.

As a William Guy lecturer for the Royal Statistical Society, I've shared with young people why statistics are so exciting and crucial. Currently, statistics face new pressures, which we will explore, but there's also an exciting technological revolution underway: AI.

AI: The Fourth Technological Revolution

AI is poised to profoundly change our lives. It's not just about economic growth; it offers incredible potential for statistics, helping us understand ourselves better and make more informed decisions. My goal is to show how statistics and AI can be brought together for mutual benefit.

Statistics in Action: The Index of Multiple Deprivation

Let's consider a concrete example of how statistics impact society: taxation. Central governments collect taxes and reallocate them to local areas for public services like buses and roads. This allocation is often determined by funding formulas, which are equations that consider various characteristics. A key component of these formulas is the "Index of Multiple Deprivation" (IMD).

The IMD, despite its clunky name, is a sophisticated statistical tool. It aggregates hundreds of statistics about an area, including income, employment, health, and education, to gauge its level of deprivation. Areas with higher IMD scores receive more funding, based on the premise that they need it most. This index is also used for pupil funding in schools.

The real IMD is incredibly complex, incorporating data on homelessness, educational levels, theft, and even broadband speed. If these numbers are wrong, the funding allocated to local areas and schools will be wrong. Similarly, inaccurate migration numbers can lead to misinformed public debates, and incorrect house valuations can result in unfair taxation. These seemingly abstract numbers have a direct and significant impact on our lives, families, and access to services.

Economic Statistics: A Story of Prosperity

My work focuses on economic statistics, which tell the story of our prosperity and highlight areas where things might be going wrong, allowing for timely policy interventions.

One fascinating chart, compiled by colleagues at the Bank of England, illustrates average hourly earnings in the UK over nearly a thousand years (in 2024 prices). For a long time, average wealth remained low, but then, with the Industrial Revolution (starting around 1760), economic growth skyrocketed. This growth enabled leisure time (the concept of a weekend is relatively new, emerging in the 1870s), better healthcare, and improved food.

However, the chart also reveals a stark historical event: a significant spike in wages around the 1340s. This was due to the Black Death, which decimated a third to half of England's population. The drastic reduction in the labor supply led to a surge in wages. This demonstrates how economic statistics can tell powerful stories, but they require careful interrogation to understand the underlying causes.

The Challenges Facing Modern Statistics

Despite their importance, statistics face growing challenges in the modern era. It's becoming harder to gather the information needed to understand the economy and people's well-being.

The Widget vs. Poem Economy

Consider a traditional factory producing identical "widgets." It's easy to measure productivity: how many widgets are produced per unit of labor and raw materials. Now, imagine trying to measure the value of a poem. How do you quantify the inputs? How do you assess its worth?

The modern economy has shifted dramatically. While the UK economy was once dominated by agriculture and manufacturing, today, these sectors account for only about 10% of output. The vast majority is in services—haircuts, holidays, legal advice. These are far more difficult to measure than tangible goods. This makes it challenging to accurately assess productivity in the modern world.

Declining Survey Response Rates

Surveys were once the backbone of official statistics. People would diligently fill out lengthy questionnaires, providing a snapshot of the nation. However, survey response rates have plummeted. Most people now prefer to watch Netflix than complete a 30-page government survey, making it harder to get a comprehensive picture of society.

The Uncounted Economy

Another challenge lies in the boundary between transacted goods and services and uncounted activities. For example, caring for an elderly relative at home is often unmeasured, even though it's significant work. If that relative moves into a care home, it becomes a counted transaction. To get accurate economic statistics, we need to measure both paid and unpaid care, as well as other forms of "home production."

Outdated Institutions

The statistical institutions we have today were largely designed for a "widget world," not the digital economy. It's easier to count cars than poems, highlighting the need for new approaches.

AI to the Rescue?

This is where AI offers a glimmer of hope. AI, at its core, is about pattern recognition. Large language models like ChatGPT recognize patterns in text, vision models recognize patterns in images (like distinguishing a dog from a cat), and AI for economic forecasting recognizes patterns in historical numerical data.

Many of us already interact with AI daily, even if we don't realize it. Recommendation systems on Netflix and YouTube, for instance, use AI to suggest content based on our viewing habits and those of similar users. While AI can seem mystical, understanding it as sophisticated pattern recognition demystifies it. Even a basic large language model can be implemented with surprisingly little code.

AI in Action: Object Recognition

Let's demonstrate AI's capabilities with a simple object recognition task. An AI can easily identify a person or a wine glass, as it has been trained on countless examples of these common objects. However, when presented with an unfamiliar object, like a microphone (which it might misidentify as a tennis racket), it struggles. This illustrates a crucial point: AI is "stupidly smart." It excels in domains it has been trained on but falters when encountering patterns outside its training data.

Superhuman AI in Narrow Domains

Despite its limitations, AI can achieve superhuman feats in specific, narrow domains.

• AlphaFold: This algorithm predicts how proteins fold based on their chemical sequence. This was previously a time-consuming scientific endeavor, but AlphaFold's speed and accuracy are revolutionizing drug discovery and our understanding of diseases.
• Nuclear Fusion: In nuclear fusion reactors, AI is being used to predict "disruptions" (when the plasma inside the reactor becomes unstable and the reaction stops). By predicting these disruptions, AI can adjust reactor settings to maintain the reaction, bringing us closer to a clean energy future.
• Gencast (Weather Forecasting): Developed by Google DeepMind, Gencast uses AI to quickly and cheaply predict the paths of typhoons and other severe weather events. Even a few hours' warning can save lives and livelihoods.
• Project Vesuvius: This project uses AI to "read" carbonized Roman scrolls from Herculaneum, buried by the eruption of Mount Vesuvius in 79 AD. By X-raying the scrolls and using AI to virtually unroll and decipher them, scientists are reading these ancient texts for the first time in 2,000 years, potentially revealing invaluable insights into Roman life.

AI's Potential for Statistics

AI holds immense promise for improving statistics.

Automating Job Classification

Surveys often ask about occupations. Traditionally, statisticians manually classified these free-text descriptions into standardized taxonomies, a time-consuming and often tedious process. AI can automate this. By training an AI on occupational codes and descriptions, it can accurately classify jobs like "zookeeper" (code 6129), saving significant time and resources. This system is already being implemented at the Office for National Statistics.

Statistics from Space: Housing Starts

The US Census Bureau uses AI and satellite imagery to track new housing starts. Instead of relying on error-prone and slow surveys sent to construction firms, they compare satellite images over time to identify new construction sites. This method is faster, cheaper, and more accurate, demonstrating how AI can replace traditional survey methods.

Real-time Movement Data During COVID-19

During the COVID-19 pandemic, real-time data on population movement was crucial for policy decisions. Traditional statistics were too slow. My team at the Office for National Statistics co-opted publicly available CCTV cameras across the country. Using an AI algorithm (similar to the object recognition example), we anonymized faces and license plates, then counted people, cars, vans, cyclists, and lorries passing each location every 10 minutes. This provided an almost immediate read on national movement, informing policy on restrictions. While AI can sometimes struggle with ambiguous situations (e.g., distinguishing a person from a lamp post), it was a significant success.

Measuring the Attention Economy: Time Use Surveys

We live in an "attention economy," where countless things compete for our focus. Understanding how people spend their time is more important than ever, especially with changing work patterns and the need to measure "home production" (like childcare or caring for elderly relatives). Traditional time-use surveys rely on self-reporting, which can be inaccurate due to memory limitations.

A radical idea is to use wearable devices that take pictures throughout the day. An AI on the device (never sending images off-device) could then classify activities into predefined categories (e.g., leisure, work, washing dishes). Even a small AI model, capable of running on an older laptop, can accurately identify activities like inflating a balloon, pouring liquid, or fixing a plug. This approach could revolutionize how we measure time use, providing more detailed and accurate data.

Regional Economic Growth Forecasting

Currently, statistics on regional economic growth in the UK take 14-15 months to be published. While accurate, this delay limits timely policy interventions. My colleagues and I at the Office for National Statistics have developed an experimental AI model that forecasts regional growth much earlier. By recognizing patterns in national and regional data, the AI provides an early indicator of growth, sometimes 14 months in advance of official estimates. While not perfect, this trade-off between timeliness and absolute accuracy is valuable for identifying potential problems sooner and enabling quicker action.

What Does This Mean for You?

1. Statistics are vital: They underpin critical policy decisions, from funding schools and hospitals to broader national strategies. They also influence individual choices, like career paths.
2. The modern economy is complex: Measuring services, poems, and contracts is harder than counting widgets, necessitating new statistical approaches.
3. AI offers solutions: AI can help overcome these challenges, providing faster, cheaper, and more accurate data.
4. Wrong statistics lead to wrong decisions: It's crucial to get the numbers right.

What Can You Do?

• Advocate for statistics: Recognize their importance and demand better statistics. Support those who advocate for evidence-based arguments.
• Participate in surveys: When an official statistics form arrives, fill it in! Your contribution is valuable.
• Get involved with AI: The field of AI is booming, offering exciting opportunities. Competitions like Project Vesuvius offer prize money for hobbyists who can solve AI challenges, demonstrating that you don't need to be an expert to contribute.

We are likely at the beginning of a new industrial revolution driven by AI. Its applications in economic statistics, ancient history, science, and countless other fields are just starting to emerge. While AI is "stupidly smart"—excelling at pattern recognition within its training data but struggling outside it—its potential to transform our world is immense.