An AI Action Plan Is Not Yet an Accountable Justice System

A justice system is not judged only by how quickly it processes work. It is judged by whether people can understand what happened to them, whether errors can be challenged, and whether someone remains responsible when a decision has serious consequences.

That is why the Ministry of Justice’s new update on its AI Action Plan for Justice deserves close attention. Published by GOV.UK on September 24, the first-year update says the original plan’s three priorities remain: strengthening foundations, embedding AI across the justice system, and investing in people and partners. It adds a fourth strategic priority: identifying and responding to emerging AI risks.

The update names risks that deserve public attention. It discusses the misuse of AI to support criminal activity, the growing prevalence of synthetic and manipulated content, and the changing demands AI tools may place on justice services. It also links to an Action Plan Explorer intended to provide more detail about individual actions.

Those are important developments. They show that the Ministry is treating AI not only as a tool for efficiency, but also as a source of new risks for a justice system already responsible for difficult, high-stakes work.

But a plan is not yet an accountable justice system.

What the update establishes—and what it does not

The reported facts are clear. The Ministry has published a one-year progress update. It retains the plan’s original three priorities and adds a fourth focused on emerging risks. The source says that risks can include AI-enabled criminality and synthetic or manipulated content entering or affecting the justice system.

The public document does not, by itself, prove that every accountability safeguard needed for justice AI is already in place. It does not establish that every system has a named decision owner, that every affected person can obtain meaningful human review, or that fairness has been tested in every real-world setting. The update discusses future-oriented work, including possible tools, guidance, safeguards, and broader strategies as appropriate. It should not be read as proof that specific safeguards have already been implemented across the system.

That distinction is not a criticism of planning. It is a way to keep public discussion honest. Justice institutions should be able to explore useful technology. They should also be held to a higher standard of explanation when technology helps shape evidence, triage, services, recommendations, or other consequential decisions.

Unautomated analysis: accountability has to survive contact with reality

The Ten Commitments offer a practical lens for that standard. They do not tell the Ministry what its policy should be. They ask what responsible leadership must still be able to answer when an AI-supported process affects a person’s rights, safety, opportunity, or trust in the justice system.

Commitment Four: Every consequential decision made with the assistance of artificial intelligence remains the responsibility of a specific, identifiable person or institution, and no system may be offered as the final answer to why something happened.

In justice, responsibility cannot disappear into a workflow. If a tool helps prioritize a case, flag a risk, summarize material, identify a pattern, or guide a service interaction, officials still need to know who can question its use, pause it, and explain the result. A system may inform a decision. It cannot become the final answer to why a person was treated a certain way.

Commitment Five: Every person affected by a consequential automated decision retains the right to a human review, a clear explanation, and a real path to appeal, correct, or reverse that decision.

Not every justice process can disclose every detail. Privacy, safety, investigations, and legal protections can properly limit what is shared. But limits on disclosure do not erase the need for meaningful review. A person should not be trapped in a process where an automated conclusion is treated as unchallengeable because no human can or will take responsibility for looking again.

Commitment Six: Fairness must be tested, not assumed, and any system that performs well on average must still be examined for how it fails the people the average conceals.

This is especially important when synthetic or manipulated content can enter a case or service channel. A tool may appear accurate in a controlled test and still fail in conditions that are uneven, pressured, or difficult to audit. Testing must ask who is represented in the evidence, whose circumstances are missing, what errors recur, and whether those errors fall more heavily on people with less power to correct the record.

Commitment Nine: Progress must be measured by whether human beings are living fuller, more capable lives, not merely by whether a system has made a task faster.

Faster administration may be valuable. It can reduce delay, help staff find relevant information, and make services easier to use. Yet speed is not the whole public purpose of justice. The meaningful measure is whether a person receives a fairer hearing, clearer information, a more reliable service, and a real opportunity to be treated as a human being rather than a data point moving through a queue.

Four practical accountability questions

The Ministry’s fourth priority is a reason for every justice organization, public body, and technology partner to ask practical questions before a concern becomes a failure. These four questions are a useful starting point:

1. Who is accountable for this AI-supported decision? Name the person or institution with authority to approve, pause, explain, and correct its use.
2. Can an affected person obtain meaningful human review? Make clear when a person can ask for a reconsideration, who performs it, and what authority that reviewer has.
3. How has fairness been tested in the actual context? Look beyond average performance to the people, communities, evidence types, and edge cases most likely to expose a harmful failure.
4. What human outcome would show that this use is worthwhile? Measure more than throughput: consider dignity, access, understanding, accuracy, trust, and the ability to challenge mistakes.

These questions do not require an organization to predict every future risk. They require it to build a habit of responsibility before technology becomes too embedded to question. They can also make the Ministry’s new priority more concrete. Horizon scanning is useful, but it must lead to clear decisions about ownership, review, evidence, and remedies.

A constructive standard for the year ahead

The update is right to recognize that AI risks change. AI-enabled criminality and manipulated content are not abstract concerns for a justice system. They can affect evidence, public confidence, casework, and the ability of people to distinguish reliable information from material designed to mislead.

The response should not be panic or a promise to automate nothing. It should be disciplined, public-minded work: identify where AI is used, set boundaries before serious harm occurs, involve legal and operational expertise, listen to people who may bear the cost of an error, and keep a human authority capable of intervening.

An action plan can be a useful beginning. It can bring attention, resources, and structure to a difficult task. But in justice, accountability cannot live only in a plan, a dashboard, or a future commitment. It has to be visible in the ordinary moments when a person asks: What happened? Who decided? Can anyone review it? What will be done if it was wrong?

Technology will keep evolving. The standard for justice must remain human judgment, human responsibility, and a real path to correction when the system gets it wrong.