On September 28, 2026, the UNESCO AI justice panel in Riyadh put a hard number on a soft spot: 44% of judicial operators are already using some form of AI, but only 9% have received formal training. That finding, shared by UNESCO from a global survey, framed a blunt question for courts going digital: who gets left out when AI moves in? UNESCO reported the figures alongside first‑hand concerns from judges, ministers, and civil society.
What the UNESCO AI justice panel heard: access before acceleration
Alpha Sesay, Attorney General and Minister of Justice of Sierra Leone, told the forum that less than 30% of the country has mobile internet access. Many people meet their justice needs outside the formal system. For any court digitization effort, he argued, the first design question is who could be excluded—like a domestic violence victim in a rural area or a farmer who cannot afford to travel. His remarks, cited by UNESCO, turn the usual AI pitch on its head: unless tools reduce distance, cost, and fear, they entrench the gap.
Judge Abdulaziz bin Mohammed Al‑Samaani of Saudi Arabia’s Board of Grievances pressed a similar test. A judicial AI strategy should be judged by whether it improves access to knowledge and services without eroding core safeguards, he said, according to UNESCO. More models and more dashboards do not equal more justice if litigants can’t use them safely.
Smita Mutt of India’s DAKSH warned about adoption running ahead of governance. With AI already shaping outcomes but little training in place, small operational choices carry outsized effects. Case lists released late at night, for example, can make it harder for caregivers—often women—to prepare, tilt hearing attendance, and nudge outcomes, UNESCO reported. These are not headline‑grabbing systems. They are calendars, queues, and defaults—where many fairness failures begin.
Why this matters for court technologists and clerks
Access to justice is already thin. The World Justice Project estimates that billions of people face unmet justice needs. Its last comprehensive analysis put the number at 5.1 billion people living with justice problems they cannot resolve, have unresolved serious civil or administrative issues, or are excluded from the legal system. That gap predates any AI rollout—but AI can widen it at scale if routine court functions replicate offline barriers. See the World Justice Project’s data for context: Measuring the Justice Gap.
The risk is practical, not abstract. A scheduling algorithm optimized for docket throughput can extend travel times for rural litigants. A chatbot that assumes constant connectivity can freeze out the very people who most need help. A biometric login can exclude those whose documentation is patchy. Each fix looks small; together they shape whether an e‑court serves or filters.
There is also a regulatory undertow. Under the European Union’s AI Act, systems used in the administration of justice fall into high‑risk territory, triggering requirements for risk management, data governance, transparency, human oversight, and post‑market monitoring. Court projects that ignore inclusion metrics will not only miss their mission, they will miss compliance. The Commission’s overview is a useful primer: European approach to AI.
Five near‑term fixes the UNESCO AI justice panel points toward
The discussion in Riyadh did not arrive with a shopping list, but it implied one. Here are concrete steps court leaders can take now—especially in places where digital adoption is uneven.
- Make training the first milestone. If 44% are using AI and only 9% trained, reverse that ratio project by project. Track training coverage by role and court level, then gate new features on minimum training thresholds. That aligns with the UNESCO Recommendation on the Ethics of AI, which calls for human oversight and capacity building.
- Design for low bandwidth and no bandwidth. Every AI‑assisted service should have an SMS, USSD, or staffed hotline path for filings, notices, and status checks. Publish calendar changes by SMS as a default, not a side channel.
- Put calendars on an inclusion SLA. Lock in cut‑off times for releasing case lists—e.g., by 3 p.m. the prior day—so caregivers and long‑distance litigants can plan. Audit actual release times monthly and publish the metric.
- Procure for auditability. Require model cards, versioned prompts, and immutable decision logs in contracts. If a tool influences scheduling or recommendations, mandate exportable logs that link model outputs to case IDs and timestamps.
- Evaluate with real users before scale. Run structured pilots with rural litigants, pro se users, and disability advocates. Measure success as completion rates and time saved for those groups, not just average throughput.
None of these steps requires a new model. They do require intent. The UNESCO AI justice panel made clear that fairness in courts is often operational: who sees a notice, when a list is posted, whether a digital door has a handle for someone on a feature phone.
Governance that matches the risk in AI in courts
Courts do not need bespoke theory to start. They need governance that maps AI to existing safeguards: explainability for any recommendation that could affect a right, appeal channels for automated triage, and red‑team exercises aimed at common failure modes like name, address, and language mismatches.
Three practices deserve early adoption. First, pre‑deployment risk assessments that explicitly test exclusion—simulate outages, throttle bandwidth, and attempt filings with basic phones. Second, post‑deployment monitoring that checks error and complaint rates across gender, location, language, and representation status. Third, a standing change control board that includes clerks and legal aid groups, not just vendors and IT.
For high‑risk deployments under the EU AI Act—like tools that assist fact research or legal interpretation for judges—documentation and human oversight are mandatory. That should be treated as a floor, not a ceiling. Courts can go further by publishing public‑facing model summaries and by setting up quarterly open sessions where practitioners can raise issues with AI‑supported tools in use.
What progress should look like by 2027
The right scoreboard is simple and human. By 2027, a court modernizing with AI in justice should be able to show four lines moving in the right direction:
- Training coverage above 80% for every role that touches an AI‑assisted tool, with refreshed certification each year.
- Calendar reliability: 95% of case lists released before a fixed daily cut‑off, verified by automated timestamp audits.
- Access gains for offline users: the share of filings and notifications completed via low‑bandwidth channels rising quarter over quarter.
- Equity checks: complaint and no‑show rates narrowing across gender and rural‑urban lines, with published methodology.
If those numbers move, the technology is serving its purpose. If they stall, the process—not just the product—needs to change. That is the thread running through the UNESCO AI justice panel: measure the things that matter to the people the courts exist to serve.
The promise is real. AI can reduce travel, shorten queues, and help judges and clerks find the right precedent faster. The warning is just as clear. Without guardrails and training, it will sort people before it serves them. The work now is to act on the panel’s test—access first, safeguards intact—and to publish the evidence that courts are getting it right. For more on this, see bloomberg.com and nytimes.com.