Human vs. AI: Who is responsible for AI mistakes? Accountability, law & best practices

Responsibility for AI mistakes does not rest with a single party. Developers, organisations that deploy AI, and the people who use it all share responsibility. Under UK law, AI itself cannot be held liable. Instead, accountability sits with the individuals and organisations under existing legal frameworks, including negligence, contract, data protection and equality law.

At OneAdvanced, we believe AI should enhance human expertise, not replace it. Consider Google Maps: it suggests the best route, but the driver stays in control. AI can recommend a course of action, but people remain responsible for the decisions they make. It can generate insights, draft content and automate routine tasks, but every meaningful decision should be reviewed, validated and owned by a human.

This guide explains how human vs AI responsibility plays out in practice, what AI accountability looks like under UK law in 2026, and what organisations can do to reduce their exposure.

What counts as an AI mistake?

Not every AI error looks the same. In reality, most fall into these categories: output errors (a system produces a factually wrong result), bias (a system systematically disadvantages a group), hallucinations (a generative model invents information with confidence), and misclassification (a system wrongly categories a person, image, or transaction).

These errors have real consequences too. In February 2024, an individual was misidentified by the Metropolitan Police's live facial recognition system in London, detained, and questioned by officers before the error was uncovered. He later received compensation, and the case remains before the courts on appeal following an April 2026 High Court ruling. Elsewhere, AI hiring tools have been shown to score candidates less favourably based on proxies for age, disability, or background, and clinicians have raised concerns about over-reliance on AI-generated diagnostic suggestions without adequate human sign-off.

The scale of the problem is growing quickly. According to the Stanford AI Index 2026, the AI Incident Database recorded 362 documented AI incidents in 2025, up from 233 in 2024 – a rise of more than 50% in single year. OneAdvanced's own Annual Trends Report 2026 found a closely related pattern: AI adoption is now organisations' number one priority, yet only 49% of organisations report AI powering less than a quarter of their work, which means ambition is running ahead of the oversight needed to catch mistakes before they cause harm.

Download the full report: Annual Trends Report

The human role in AI mistakes

Humans remain at the centre of every AI system, and human decisions are frequently where mistakes originate.

Developers and data scientists

Developers and data scientists determine how models are designed, what data they’re trained on and how they’re tested. If training data is incomplete or biased, the AI is likely to produce biased outcomes. For example, Amazon's Rekognition facial recognition system was found to have significantly higher error rates for women and people of colour, demonstrating how flaws in training data can translate into unfair results.

Organisations deploying AI

Responsibility doesn’t end once an AI system is purchased or built. Organisations must ensure AI is deployed responsibly, monitored continuously and governed throughout its lifecycle. This includes establishing clear policies, maintaining human oversight and ensuring compliance with legal and ethical requirements.

End users

The people using AI also have an important role to play. Relying on AI outputs without questioning their accuracy or applying professional expertise can lead to poor decisions, particularly in high-risk areas such as healthcare, legal services and financial advice. End users should validate recommendations, understand the technology's limitations and then, make informed decisions based on both AI insights and human expertise.

The AI system's own role

Although humans design, deploy and oversee AI systems, they aren’t the only source of AI mistakes. Unlike traditional software, AI learns from data, identifies patterns and generates outputs without fixed rules. As a result, it can behave unpredictably, especially in unfamiliar situations.

• AI can produce unexpected results:Machine learning systems develop their own way of recognising patterns, which can sometimes lead to inaccurate or biased outcomes. For instance, Google's image recognition system once incorrectly labelled photos of Black people as gorillas – a serious error that highlighted the consequences of biased training data and weaknesses in AI models.
• AI can spread mistakes quickly and at scale: AI operates and affects decisions on a scale and at a speed unmatchable by humans. For example, Meta's news feed algorithm has faced criticism for amplifying and speeding up the fake news and polarising content, influencing public opinion and even election outcomes with its automated content curation.
• AI can struggle with unfamiliar situations: AI performs best in the scenarios it has been trained to recognise. When faced with unexpected or complex real-world situations, it can make poor decisions. Investigations into Tesla's Autopilot system following several accidents highlighted the challenges AI faces when interpreting unpredictable driving conditions.
• AI can amplify flaws in its design: AI systems are only as reliable as the models, data and rules they are built on. If those foundations contain weaknesses, AI can magnify them. The 2010 Flash Crash, in which algorithmic trading contributed to a sudden and dramatic stock market plunge, showed how automated systems can accelerate errors when safeguards are insufficient.

Together, these examples show that AI mistakes, whether caused by humans or machines, are rarely down to a single cause, which is exactly why shared responsibility is the only realistic model.

Who is legally liable for AI mistakes?

As of 2026, the UK has no AI-specific liability statute. Instead, responsibility is worked out through existing legal frameworks:

• Equality Act 2010: Applies where an AI system produces discriminatory outcomes, including in recruitment, lending, or service provision.
• UK GDPR and the Data (Use and Access) Act 2025: Section 80 of the DUAA came into force on 5 February 2026, replacing the old Article 22 near-prohibition on solely automated decision-making with a permission-with-safeguards regime. Organisations relying on automated or AI-assisted decisions must now provide meaningful information about the logic used, offer genuine human intervention from a reviewer with real authority to change the outcome, and give individuals a route to contest decisions.
• Consumer Rights Act and tort law: Standard negligence and contract principles continue to apply, alongside the Consumer Protection Act 1987 where AI is embodied in a physical product.

The direction of UK law is clear. In its draft Legal Statement on Liability for AI Harms, published in January 2026, the UK Jurisdiction Taskforce (UKJT) reaffirmed that AI has no legal personality under English law and cannot be held liable for harm. Instead, responsibility rests with the individuals and organisations that design, deploy and use AI. In practice, courts are likely to treat AI as a tool, with liability determined under existing principles of negligence, duty of care and other applicable laws.

In practice, this points to a three-party model of liability, summarised below.

Party	Typical responsibility	Where liability often attaches
Developer	Model design, training data quality, pre-release testing, post-deployment monitoring	Negligence in design, misrepresentation of capabilities, product liability where AI is embedded in a physical product
Deploying organisation	Procurement, configuration, oversight, monitoring in production	Negligence in deployment; Equality Act claims; UK GDPR/DUAA breaches for inadequate ADM safeguards
End user	Appropriate use, professional judgement, escalation of concerns	Professional negligence where AI outputs are relied on without scrutiny; contractual breach of use terms

The shared responsibility model

Shared accountability is the only realistic framework for managing AI risks. This means embedding ethical considerations at every stage – from development to deployment, continuous learning, oversight, and responsible day-to-day use – rather than treating accountability as a single checkbox.

The UK Government's AI White Paper set out five cross-sector principles that continue to anchor regulatory expectations in 2026: safety, transparency, fairness, accountability, and contestability.

The Information Commissioner's Office (ICO) plays an increasingly central role in translating those principles into practice. Its updated guidance on automated decision-making, published in March 2026 alongside the DUAA reforms, sets out concrete expectations: organisations must be transparent about when Automated decision-making (ADM) is used, provide decision-specific explanations, and treat AI impact assessments as standard practice for higher-risk use cases.

Board-level oversight of AI is mandatory. Regulators and industry bodies increasingly look for a named accountability owner, a documented governance structure, and evidence that risk is on the board agenda.

What should UK organisations do? A practical accountability checklist

According to OneAdvanced’s Annual Trends Report 2026, skills gaps rank as UK organisations' second-biggest operational challenge – yet talent development sits at 10^th position on the investment priority list. The same report found that organisations are investing roughly 5x more in AI tools than in preparing the people who use them. That gap is exactly where accountability breaks down.

Use the checklist below to assess whether your organisation has the right foundations for AI accountability.

Area	Ask yourself
Ownership	Have you appointed a named individual or committee with overall responsibility for AI governance and accountability?
Risk assessment	Do you carry out AI impact assessments before deploying systems that could affect people's rights, finances, or wellbeing?
Human oversight	Are high-risk AI decisions reviewed by people with the authority to challenge or override AI outputs?
AI inventory	Do you maintain an up-to-date inventory of AI systems, including their purpose, risk level and ownership?
Training	Have employees been trained on AI capabilities, limitations and responsible use?
Monitoring	Do you regularly monitor, test and audit AI systems to identify issues before they become business risks?
Governance framework	Is the AI governance of aligned with recognised frameworks such as ISO/IEC 42001 or the NIST AI Risk Management Framework?

What does your result mean?

If you answered "No" to two or more questions, your organisation may have gaps in its AI accountability framework. Addressing these areas can help reduce risk, strengthen governance and build greater trust in how AI is developed, deployed and used.

Sector-specific responsibility considerations

The shared responsibility model above is the starting point, but what it demands in practice looks different in every sector. Here's what that means in practice.

Healthcare

The Medicines and Healthcare products Regulatory Agency (MHRA) regulates most clinical AI tools as Software as a Medical Device (SaMD), but regulatory approval doesn’t remove clinical responsibility. Under Good Medical Practice, clinicians remain accountable for the decisions they make, even when supported by AI. NHS organisations must also comply with DCB0160 clinical risk management standard, which requires a named Clinical Safety Officer to maintain a hazard log for the tool throughout its lifecycle. See how AI in healthcare is being applied responsibly across the sector.

Legal services

AI can help legal professionals research cases, draft documents and improve efficiency, but it does not replace professional judgement. The Solicitors Regulation Authority (SRA) expects solicitors to meet the same professional and ethical standards whether work is completed by a person or assisted by AI. In short, AI can support legal work, but accountability always remains with the solicitor and their firm. Explore AI in legal services built around that principle.

Finance

The UK Financial Conduct Authority (FCA) does not have separate rules for AI. Instead, firms must apply existing regulations to AI-driven decisions in areas such as lending, fraud detection and claims processing. This means organisations remain responsible for ensuring AI systems are fair, transparent and properly monitored, with clear accountability, regular model reviews and human oversight where needed. Learn more about OneAdvanced Financials.

HR and recruitment

Under the Equality Act 2010, employers are responsible for ensuring AI hiring tools don’t discriminate against candidates. Even if AI makes the decision, organisations remain accountable. To reduce this risk, employers should regularly test AI systems for bias and ensure important hiring decisions are reviewed by a person with the authority to challenge or overturn AI recommendations. See People Management solutions.

Public sector

Public bodies carry an added layer of accountability to citizens, including Freedom of Information obligations and heightened transparency expectations. When AI influences decisions about individuals or is used in public services, organisations are expected to be transparent about how it works and provide meaningful explanations where automated decisions affect people. This helps build public trust, supports accountability and ensures AI is used fairly and responsibly. Explore AI in the public sector.

How OneAdvanced approaches responsible AI

At OneAdvanced, we recognise that responsibility for AI mistakes cannot be solely assigned to humans or machines. It requires a shared commitment to the ethical and transparent use of AI. This thinking sits at the heart of OneAdvanced IQ, our intelligent system of work, which is built on three principles that map directly onto responsible AI use:

• Connected: people, data and AI working together, with context carried across every decision, not isolated in a black box
• Trusted: a secure, sovereign platform with compliance and human sign-off built in, not bolted on afterwards)
• Intelligent: AI-driven insight and automation embedded into the flow of work, so people get real-time input without losing control of the outcome.

Ready to strengthen your AI accountability? Explore OneAdvanced IQ or read our Responsible AI Framework guide to get started.

Frequently Asked Questions

Who is responsible when AI makes a mistake – the developer, the company, or the user?

All three can share responsibility, depending on where the failure occurred. Developers are accountable for design and testing, deploying organisations for oversight and monitoring, and users for appropriate, professionally judged use.

Can an AI system itself be held legally liable for errors?

No. Under English law, AI has no legal personality. The UK Jurisdiction Taskforce's 2026 draft Legal Statement confirms that liability must attach to the people and organisations behind the system, not the system itself.

Is there a law governing AI liability in the UK?

Not a dedicated one, as of 2026. Liability is currently established through existing law — the Equality Act 2010, UK GDPR and the Data (Use and Access) Act 2025, tort and contract law — while the UK Jurisdiction Taskforce's draft Legal Statement works to clarify how the courts should apply these principles to AI-specific harms.

What should organisations do to reduce AI risk and liability?

Establish a named AI accountability owner, run impact assessments before deployment, keep a human meaningfully in the loop for high-risk decisions, maintain a model inventory, train staff on AI's limitations, and audit outputs regularly. See the full checklist above.

What is a 'human in the loop' and why is it important for AI governance?

It refers to a human reviewer who genuinely understands an AI system's outputs and has real authority to challenge or overturn them. Regulators increasingly treat token human involvement as insufficient to avoid automated-decision-making obligations.

How does OneAdvanced ensure its AI tools are used responsibly?

Through ethical AI principles embedded across the product lifecycle, transparent decision-support design, human oversight built into managed services, and ongoing governance support for customers. Read more in our Responsible AI Framework guide.