The AI Making Decisions
for Billions of People

How We Got Here

For decades, artificial intelligence existed in research laboratories as a promising but distant idea — too expensive, too fragile and too far from the messy reality of human systems to matter much outside academia. The pivot point arrived in 2017, with the publication of "Attention Is All You Need", the paper that introduced the Transformer architecture underlying every major language model in use today.

In less than a decade, the technology moved from university servers to the pocket of every smartphone user on Earth. What changed wasn't just model quality — it was scale of adoption. ChatGPT reached 100 million users in under two months, a feat that took Facebook four years. From that moment, we stopped experimenting with AI and started depending on it.

The critical leap happened quietly. Governments, hospitals, courts and financial institutions adopted AI-powered tools not because they fully understood them, but because the performance gains were too compelling to ignore. A diagnostic algorithm that outperforms junior radiologists. A credit-scoring model that processes thousands of variables in milliseconds. A recidivism predictor that claims to forecast criminal behaviour before it happens.

Governments: AI at the Heart of Power

In 2026, 132 countries use artificial intelligence systems in at least one area of public policy, according to data from the Oxford Internet Institute. But what does that mean in practice?

It means that in the United Kingdom, an algorithm partially determines whether you are investigated for social welfare fraud. In the United States, predictive policing tools inform where law enforcement resources are deployed. In China, social credit systems process behavioural data to assign citizens opportunity scores that affect travel, employment and education access. In the European Union, border control systems use AI to flag travellers for secondary screening.

These are not experimental programmes. They are operational infrastructure — running daily, affecting millions of people, with minimal public visibility into how they work or how to challenge them.

The Transparency Problem

When a government official makes a decision that harms you, there are legal mechanisms for challenge: appeals, judicial review, freedom of information requests. When an algorithm makes that decision, the accountability chain fractures. The developer may argue the model is proprietary. The deploying agency may argue it merely surfaces recommendations. The legislator may argue regulation hasn't caught up.

Healthcare: When the Algorithm Prescribes

The application of AI in healthcare represents both its most promising potential and its most serious risk. Diagnostic algorithms trained on millions of medical images can identify certain cancers earlier than human radiologists in controlled conditions — this is a genuine and important capability.

The problem emerges in deployment. Major hospital systems in the United States, United Kingdom and Australia have integrated AI tools into clinical workflows in ways that make it difficult to distinguish AI recommendation from physician judgement. A 2024 study published in The Lancet found that in 60% of cases where AI flagged a patient as low-risk, clinicians did not order additional tests — even when their own assessment suggested further investigation might be warranted.

The AI wasn't wrong in those cases. But the effect — clinicians deferring to algorithmic confidence — represents a shift in how medical decisions are made that happened faster than the evidence base supporting it.

Bias in Medical Data

Medical AI systems are predominantly trained on datasets that over-represent white male patients from high-income countries. The consequences are documented: pulse oximeters calibrated on lighter skin tones systematically underestimate oxygen desaturation in patients with darker skin. Dermatology AI performs significantly worse on darker skin types. Cardiac risk algorithms show different accuracy profiles across demographic groups.

Financial Markets: 70% Algorithmic

More than 70% of daily trading volume on major global stock exchanges is now executed by algorithms. These systems react to news, economic indicators, earnings reports and pattern signals in milliseconds — at speeds that make human monitoring in real time practically impossible.

The 2010 Flash Crash — when the Dow Jones Industrial Average plunged nearly 1,000 points in minutes before recovering — was an early demonstration of what happens when multiple algorithmic systems interact in unexpected ways under stress. The 2024 "AI Contagion" event, where correlated AI-driven selloffs across multiple asset classes triggered circuit breakers simultaneously in seven markets, was a more recent reminder that the risk has not been resolved.

Credit decisions are similarly algorithmic. Mortgage applications, small business loans, credit card limits — these decisions, which determine access to economic opportunity for hundreds of millions of people, are increasingly made by models whose internal logic is inaccessible to the individuals they affect.

The Accountability Gap

The defining challenge of AI in critical systems is not malice — it's accountability. When a human decision-maker causes harm, there are legal, institutional and social mechanisms for redress. When an algorithmic system causes harm, responsibility disperses across a chain of actors — developers, deployers, regulators, legislators — in ways that existing legal frameworks are poorly equipped to address.

The "black box" problem makes this worse. Modern deep learning models — the foundation of most high-performing AI systems — produce outputs through processes that their creators cannot fully explain. An AI system can correctly identify a pattern in data without being able to articulate, in terms a human can evaluate, why that pattern is predictive.

This matters because explanation is the foundation of accountability. Without it, individuals cannot challenge adverse decisions, regulators cannot meaningfully audit systems and courts cannot evaluate whether an AI-assisted verdict was fairly reached.

The Regulatory Response

The most substantive regulatory response to date is the European Union's AI Act, which entered enforcement in phases from 2025. It classifies AI systems by risk level and imposes binding requirements for transparency, accuracy, human oversight and documentation in high-stakes applications — healthcare, law enforcement, credit, employment and critical infrastructure.

The AI Act represents the most comprehensive attempt by any jurisdiction to govern AI deployment in consequential domains. Its requirements for high-risk systems include: mandatory conformity assessments, human oversight obligations, data governance standards and the right of individuals to receive explanations for AI-assisted decisions that affect them.

Whether enforcement will match ambition remains to be seen. The UK has adopted a principles-based approach, relying on existing sectoral regulators rather than a dedicated AI authority. The United States has no comprehensive federal AI legislation, operating instead through executive orders and sector-specific guidance. China has implemented AI regulations focused primarily on content moderation and recommendation systems.

What Needs to Change

The researchers, policymakers and civil society organisations working on AI accountability broadly agree on the direction of necessary change, even when they disagree on pace and mechanism.

• Explainability requirements: AI systems used in consequential decisions should be required to produce explanations that affected individuals can understand and challenge — not as a technical nicety, but as a legal right.
• Independent auditing: High-risk AI systems should be subject to mandatory third-party auditing against defined performance and fairness benchmarks, with public reporting of results.
• Procurement transparency: Governments should be required to disclose when AI systems are used in public decision-making, what data they are trained on and what performance assessments have been conducted.
• Redress mechanisms: Individuals adversely affected by AI-assisted decisions should have accessible, meaningful pathways to challenge those decisions — including access to human review.
• International coordination: AI systems operate across borders; effective governance requires coordination between jurisdictions to prevent regulatory arbitrage.