Riyadh's AI ethics forum brought a toolkit, not a treaty

01

WHAT ACTUALLY CHANGED

UNESCO and Saudi Arabia opened the four-day forum in Riyadh on September 14. The event is hosted with the Saudi Data and AI Authority and the International Center for AI Research and Ethics. Its program combines a scientific briefing, a closed ministerial meeting, plenary discussions, side events, and expert panels focused on readiness, regulation, public administration, finance, inclusion, environment, and international cooperation. This is a policy forum, not a treaty-negotiation conference.

UNESCO announced that three governance products are scheduled for September 16. Readiness Assessment Methodology 2.0 is intended to incorporate more insight from women, people with disabilities, and Indigenous peoples. A cross-country analysis is supposed to extract governance patterns from completed national assessments. An environment and ecosystems toolkit is intended to help policymakers reduce AI's footprint and use AI in climate, biodiversity, and ecosystem work. The announcement supplies the scope and launch date, but not evidence that the products have already changed a national policy or a deployed system.

The existing readiness process is a country-level diagnostic. UNESCO says independent experts and a national multistakeholder team collect qualitative and quantitative information across legal, social, cultural, economic, educational, scientific, and technical dimensions. The work is meant to end in a country report identifying gaps and policy recommendations. UNESCO's September 9 announcement says 58 countries have completed the process and more than 20,000 stakeholders have contributed. Completed assessment does not mean every recommendation was adopted.

UNESCO's system-level Ethical Impact Assessment serves a different purpose. It asks an organization to examine a particular AI system before and during use, identify affected people and possible harms, record mitigation, and revisit the decision as the system changes. A country can be broadly ready on paper while a specific benefits model, school tool, medical triage system, or police system still fails its own impact assessment. The two tools should connect without being confused.

The forum arrives five years after UNESCO member states adopted the Recommendation on the Ethics of Artificial Intelligence. That recommendation calls for human-rights protection, accountability, transparency, environmental assessment, human oversight, and national legal frameworks. It is a normative instrument. It guides governments and creates a shared reference point, but it does not automatically create a court claim, regulator, inspection power, procurement rule, penalty, or appeal process inside each country.

02

WHY THIS MATTERS

International principles can make disagreement legible. A ministry, company, civil-society group, researcher, or community can point to the same commitments and ask whether a proposed system meets them. That shared language is useful in countries whose technical capacity, institutions, and legal systems differ sharply. The danger is that agreement on vocabulary can be mistaken for agreement on action. A declaration that mentions transparency is not the same as a public register of systems, disclosed impact assessments, or a right to see why a benefit was denied.

A readiness assessment works best when it becomes a baseline rather than a graduation certificate. The report should identify gaps, responsible institutions, cost, sequence, and evidence. A later assessment should show what changed and what did not. If the country report remains private, if recommendations have no owner, or if only strengths appear in the public summary, the process risks becoming a reputation exercise instead of a governance tool.

Participation matters most when it changes the decision. UNESCO says the updated method will deepen inclusion of women, disabled people, and Indigenous peoples. Those groups can experience AI systems differently because datasets, interfaces, language coverage, accessibility, surveillance, and access to remedies are uneven. Counting consultations is only the first measure. The public should also be able to see which concerns changed a design, stopped a deployment, altered a budget, or created an appeal route.

Environmental governance needs similar specificity. A toolkit can ask about electricity, carbon, water, land, mining, equipment, network traffic, and discarded hardware across an AI system's lifecycle. The results become useful when agencies define boundaries, use comparable units, disclose assumptions, and attach limits to procurement or permits. A green claim based only on renewable electricity can hide water use, construction, hardware supply, and local ecological effects.

Riyadh's role as host does not certify Saudi Arabia's domestic AI governance, and attendance does not certify any participant's compliance. The same rule applies to every government, company, and institution in the room. Hosting creates an opportunity to publish evidence: the national assessment, named reforms, system registers, impact reviews, enforcement records, and remedies. The credibility test starts after the microphones are packed away.

03

WHERE IT COULD HELP

• Publish every national readiness report with its evidence, limitations, dissenting views, responsible agencies, budget needs, implementation dates, and a dated follow-up scorecard
• Require an ethical impact assessment before a public agency buys or deploys a consequential AI system, then update it when the model, data, purpose, vendor, or affected population changes
• Create a public AI register that identifies system purpose, operator, vendor, data categories, evaluation results, human oversight, incident channel, appeal route, and retirement date
• Give affected communities accessible ways to participate before deployment and show which recommendations changed the decision, including versions in local languages and formats usable by disabled people
• Measure environmental effects across electricity, carbon, water, land, hardware production, network use, and waste, with stated boundaries and local limits rather than one flattering sustainability number