AI Ethics in Saudi Arabia: What the UNESCO Global Forum in Riyadh Means for Organisations
Riyadh hosted UNESCO’s fourth Global Forum on the Ethics of AI in September 2026. Here is what was verifiably launched, what UNESCO’s and SDAIA’s frameworks ask, and how to turn each principle into a concrete design control for an AI assistant.
From 14 to 17 September 2026, Riyadh hosted the fourth UNESCO Global Forum on the Ethics of Artificial Intelligence, co-hosted by UNESCO and the Kingdom, represented by the Saudi Data and AI Authority (SDAIA) and the Riyadh-based International Center for AI Research and Ethics (ICAIRE). It was the first edition held in the Gulf and the first in an Arab country, and it fell inside the year the Kingdom has designated its Year of AI.
Forums produce speeches more reliably than obligations, so this article keeps the two apart. It sets out what was verifiably said and launched in Riyadh, explains in plain language the two ethics frameworks a Saudi organisation actually meets — UNESCO’s 2021 Recommendation and SDAIA’s AI Ethics Principles — and then does the part most coverage skips: turning those principles into concrete design decisions for an AI assistant. Nothing here is legal advice.
What happened in Riyadh, and what did not
UNESCO ran the forum under the theme Transforming global cooperation for ethical AI governance, and framed it as marking five years since its Member States adopted the Recommendation on the Ethics of AI. Earlier editions had been held in Prague, Kranj and Bangkok. In its post-event summary UNESCO reported more than 6,300 participants and delegations from over 50 Member States, with a closed-door ministerial track attended by 52 delegations.
- Nov 2021Recommendation adoptedUNESCO’s 193 Member States adopt the Recommendation on the Ethics of AI, the first global normative framework on the subject.
- 2022Forum createdUNESCO establishes the Global Forum on the Ethics of AI; the first edition is held in Prague.
- Sep 2023SDAIA principles publishedSDAIA issues its AI Ethics Principles after a public consultation, applying to anyone who builds, deploys or uses AI in the Kingdom.
- Nov 2023ICAIRE recognisedUNESCO’s General Conference grants the Riyadh centre Category 2 status under UNESCO’s auspices.
- Feb 2024Second forum, KranjHeld in Slovenia on 5–6 February 2024.
- 2025Third forum, BangkokThe last edition before the forum moved to the Gulf.
- Sep 2026Fourth forum, RiyadhFour days, co-hosted with SDAIA and ICAIRE, during the Saudi Year of AI.
It is just as useful to be clear about what did not happen. We found no binding declaration, treaty text or new legal instrument coming out of the forum. The written outcome UNESCO published was a Co-Chairs’ summary of the closed ministerial track, in which delegations stressed global cooperation so that AI advances human well-being and sustainable development. If you see a “Riyadh Declaration” cited in a vendor deck, ask for the document.
What was said, and what was launched
The opening statements were consistent with one another. Åsa Regnér, UNESCO’s Deputy Director-General, recalled that Member States had agreed five years ago that AI “must serve humanity and the public good”, and said she hoped Riyadh would be the next step towards implementation at scale. Dr Abdullah bin Sharaf Alghamdi, President of SDAIA, described the forum as an important platform for countries and experts to strengthen international cooperation, and tied it to the Year of AI. Dr Majid Alshehry, Acting Director-General of ICAIRE, put the priority as moving from principles to practical implementation.
The concrete launches were tools rather than rules. UNESCO released RAM 2.0, an updated version of its Readiness Assessment Methodology — the diagnostic it uses to assess a country’s preparedness to govern AI — with an Arabic version developed with ICAIRE’s support. Alongside it came an AI, environment and ecosystems toolkit for policymakers, and a meta-analysis synthesising 55 country assessment reports. UNESCO says 58 countries have now completed the readiness assessment.
On the Saudi side, SDAIA awarded AI ethics badges to more than 20 companies, recognising firms that apply ethical and regulatory principles when developing and delivering AI products and services. It also awarded certified professional badges to 52 people, and, with the ITU and ICAIRE, honoured the winners of an AI readiness hackathon judged partly on adherence to ethical standards.
UNESCO’s 2021 Recommendation, in plain language
The Recommendation on the Ethics of AI was adopted in November 2021 by all 193 UNESCO Member States. It is addressed to governments: it asks states to build policy and institutions around it, rather than placing direct legal duties on a company. Its structure is easy to remember — four values, ten principles and eleven areas of policy action.
The four values are human rights and human dignity; living in just, peaceful and interconnected societies; diversity and inclusiveness; and the flourishing of the environment and ecosystems. The ten principles are where the practical content sits: proportionality and do no harm; safety and security; privacy and data protection; multi-stakeholder and adaptive governance; responsibility and accountability; transparency and explainability; human oversight and determination; sustainability; awareness and literacy; and fairness and non-discrimination.
Read as a builder rather than a diplomat, three of those carry most of the weight. Proportionality means using no more AI, and no more data, than the purpose needs. Human oversight and determination means a person remains answerable for outcomes and can step in. Transparency means people can tell when they are dealing with AI and can get an explanation of what it did.
SDAIA’s AI Ethics Principles, which apply closer to home
For an organisation operating in the Kingdom, the nearer framework is SDAIA’s own. It was published in September 2023 after a public consultation, and the edition currently on SDAIA’s website is dated May 2025. Its scope is broad: everyone designing, developing, deploying, implementing, using or affected by AI systems in the Kingdom, across public, private and non-profit entities.
It sets out seven principles: fairness; privacy and security; humanity; social and environmental benefits; reliability and safety; transparency and explainability; and accountability and responsibility. Each comes with controls organised around four lifecycle stages — plan and design, prepare input data, build and validate, deploy and monitor — which makes it more operational than most ethics documents.
It also sorts AI systems into four risk levels. Little-or-no-risk systems face no restrictions, though ethical compliance is recommended. Limited-risk systems are subject to the principles. High-risk systems must undergo conformity assessments before and after deployment, alongside relevant statutory requirements. Unacceptable-risk systems — the document cites social profiling, exploitation of children and behaviour distortion — are not allowed. The compliance mechanism it describes is reporting, review and motivational badges rather than penalties; the binding obligations on personal data come from the Personal Data Protection Law. How any of this applies to your specific system is a question for your counsel.
| UNESCO Recommendation (2021) | SDAIA AI Ethics Principles | |
|---|---|---|
| Fairness and non-discrimination | Yes | Yes |
| Privacy and data protection | Yes | Yes |
| Transparency and explainability | Yes | Yes |
| Accountability and responsibility | Yes | Yes |
| Safety and reliability | Yes | Yes |
| Human oversight | Yes | Within accountability and safety |
| Environmental impact | Yes | Yes |
| Awareness and literacy as a principle | Yes | No |
| Tiered risk categories | No | Yes |
| Controls by lifecycle stage | No | Yes |
UNESCO addresses states and sets direction; SDAIA addresses organisations in the Kingdom and adds risk tiers and stage-by-stage controls. A tick means the theme is explicit in the framework, not that one treats it more seriously than the other.
Why the theme was implementation, and why that lands on you
Both frameworks say what good looks like. Neither tells you whether your customer-service assistant should introduce itself as an AI, when it should pass a conversation to a person, or what it should do when asked something its documents do not cover. Those are product decisions, and they are made by whoever deploys the system — usually long before an ethics committee sees it.
SDAIA’s document is explicit on one point that matters for buyers: if an AI system is built by a third party, the owner should make sure AI ethics due diligence is carried out, with documentation accessible and traceable, before procurement or sign-off. In other words, buying a system does not transfer the ethical responsibility for it.
From principle to control: an AI assistant, line by line
The table below is the practical core of this article. It takes the principles that recur in both frameworks and turns each into a design decision for a customer-facing AI assistant, together with a test you can run yourself. It is our reading, not an official mapping by UNESCO or SDAIA.
| Design decision | How to check it | |
|---|---|---|
| Transparency | Say it is an AI at the start of every conversation, in both languages | Open a fresh chat or call and listen for the disclosure |
| Human oversight | A clear route to a person, triggered by request, low confidence or sensitive topics | Type “I want to speak to someone” in Arabic and see what happens |
| Reliability and safety | Answer only from approved material; decline when it is not there | Ask a question next to the source material but not in it |
| Privacy and security | Redact personal data before it is stored or logged; set retention | Send a test ID number, then inspect the stored transcript |
| Fairness | Equal quality across dialects, MSA and mixed Arabic–English | Ask the same question five ways and compare the answers |
| Accountability | Log question, sources used, answer and handovers; name an owner | Pick a past conversation and reconstruct why it answered as it did |
| Proportionality | The smallest model and the least data that do the job | Ask what data the system needs and why each field is required |
Our mapping, for illustration. The left column uses the shared vocabulary of both frameworks; neither prescribes these specific controls.
Saying it is an AI, and handing over to a person
SDAIA’s accountability principle states that AI systems “should never lead to people being deceived”, and its transparency principle asks that affected people be informed of how outcomes are produced. For an assistant, the simplest reading is also the strongest: disclose at the first message, do not give the bot a human name and a stock photo, and say so plainly when a person takes over.
Handover is where human oversight stops being abstract. A sound design decides in advance which situations go to a person — an explicit request, repeated failure to answer, a complaint, anything touching health, money or a legal decision — and passes the conversation history along so the customer does not have to start again. SDAIA’s framework goes further for decisions that are irreversible or difficult to reverse, saying these should trigger human oversight and final determination. An assistant that answers questions should not be quietly making those decisions at all.
We built Human Handover as a separate Elbi module for exactly this reason: the route to a person should be designed, tested and logged like any other feature, not bolted on after launch.
Refusing rather than inventing
The most common ethical failure of an AI assistant is not bias or a data leak. It is a confident, fluent, wrong answer — a refund policy that does not exist, a fee that was never charged, a deadline that was made up. Under both frameworks this is a reliability and transparency problem at once, because the user has no way to tell invention from fact.
The control is grounding: the assistant retrieves passages from the organisation’s approved material and answers only from them, and when nothing relevant is found it says so and offers a person or a channel instead. Our retrieval component is built around that rule. Whatever system you use, the test is the same and takes ten seconds: ask something adjacent to the material but not in it. A grounded assistant declines. A weak one improvises.
It also helps to show the user where an answer came from — the policy name, the page, the date of the document. That one detail turns “trust me” into something a customer, an auditor or a supervisor can check.
Privacy and fairness, in Arabic
SDAIA’s privacy controls say the use of personal information should be limited to what the system needs to function. For a chat or voice assistant, that translates into three decisions: strip identifiers such as national ID, phone and account numbers before transcripts are stored or used for analysis; set a retention period and actually delete; and restrict who can read raw conversations. PII redaction is one of the components in Elbi’s Core for that reason, and we have written separately about what the PDPL means for AI chatbots.
Fairness in a Saudi deployment has a specific shape. The risk is rarely that an assistant treats one group worse on purpose; it is that it works well in formal Modern Standard Arabic and noticeably worse for a customer writing in Najdi or Hijazi dialect, in Arabizi, or switching between Arabic and English in one sentence. Those customers get more refusals and worse answers, and nobody notices because the test set was written in MSA.
Bar lengths show suggested relative emphasis in a test set, not measured performance of any system.
The practical control is to measure answer quality and refusal rates separately for each variety, and to treat a gap between them as a defect to be fixed, not as an accepted limitation.
Accountability means you can reconstruct a conversation
Accountability in SDAIA’s framework is concrete: identifiable people responsible for the system, documented decisions, predefined alerts on data and performance assigned to named owners, and a mechanism to log and address complaints. For an assistant, that reduces to a test: pick any past conversation and explain why it answered the way it did.
To pass that test you need the question, the passages retrieved, the version of the knowledge base at that moment, the answer given, and whether and why a person took over — stored with personal data already redacted. You also need a named owner for the knowledge base, a change log for what was added or removed, and a periodic review of refusals and handovers, which are the best early signal of where the system is failing people.
None of this is exotic. It is the same discipline any regulated process already has. The forum’s message, read practically, is that AI systems are now expected to meet it too — the same expectation behind the Kingdom’s wider push in its Year of AI.
The honest limits of this article
Our account of the forum relies on UNESCO’s own pages and on reporting by Arab News and the Saudi Press Agency. Participation figures changed between pre-event announcements and UNESCO’s post-event summary; we have used the post-event figures. We did not attend the closed ministerial sessions and cannot report on them beyond the published Co-Chairs’ summary.
The principle-to-control table is our interpretation. Neither UNESCO nor SDAIA prescribes those specific controls, and the legal effect of SDAIA’s framework on your organisation depends on your sector, your system and other regulation that applies to you. This is not legal advice. What we can describe precisely is our own product: it answers from your approved material, on hardware you control inside the Kingdom, with no external model call in the answering path, and hands over to a person when it should.
Sources
- Global cooperation for ethical AI governance reinforced at UNESCO’s 4th Global Forum on the Ethics of AI in Saudi Arabia — UNESCO, Sep 18, 2026
- UNESCO launches new tools to reinforce ethical AI governance at global forum in Riyadh, Saudi Arabia — UNESCO, Sep 9, 2026
- Global Forum on the Ethics of AI — UNESCO, Sep 14, 2026
- UNESCO’s Ethics of AI forum opens in Riyadh — Arab News, Sep 14, 2026
- Companies, tech specialists, and global hackathon winners at Riyadh UNESCO forum — Arab News / Saudi Press Agency, Sep 18, 2026
- Recommendation on the Ethics of Artificial Intelligence — UNESCO, Nov 1, 2021
- AI Ethics Principles (document SDAIA-P114E) — Saudi Data & AI Authority (SDAIA), May 1, 2025
- Saudi Arabia – AI Ethics Principles (register entry) — Regulations.ai, Sep 1, 2023
- UNESCO grants Saudi AI center Category 2 status — Arab News / Saudi Press Agency, Nov 17, 2023
Common questions
It was the fourth edition of UNESCO’s Global Forum on the Ethics of Artificial Intelligence, held in Riyadh from 14 to 17 September 2026 and co-hosted with Saudi Arabia, represented by SDAIA and the International Center for AI Research and Ethics (ICAIRE). It was the first edition in the Gulf and in an Arab country. UNESCO reported more than 6,300 participants and delegations from over 50 Member States.
We found no binding declaration or new legal instrument. The published written outcome was a Co-Chairs’ summary of the closed ministerial track. The forum’s concrete launches were tools: UNESCO’s updated Readiness Assessment Methodology (RAM 2.0, with an Arabic version), an AI and environment toolkit, and a meta-analysis of 55 country reports. SDAIA also awarded AI ethics badges to more than 20 companies.
SDAIA’s framework sets out seven principles: fairness; privacy and security; humanity; social and environmental benefits; reliability and safety; transparency and explainability; and accountability and responsibility. It attaches controls to four lifecycle stages and sorts AI systems into four risk levels, from little or no risk to unacceptable risk, which is not allowed.
The framework applies to anyone building, deploying or using AI in the Kingdom, but the compliance mechanism it describes is reporting, review and motivational badges rather than penalties. Binding obligations on personal data come from the Personal Data Protection Law, and sector regulators may add their own. How it applies to your system is a question for your legal counsel; this is not legal advice.
SDAIA’s principles say AI systems should never lead to people being deceived and that affected people should be informed about how outcomes are produced. The practical reading is to disclose at the start of every conversation, avoid presenting the bot as a person, and say clearly when a human takes over. Sector rules may set further requirements.
Keep reading
SDAIA AI Training Programmes: How Saudi Arabia Is Teaching a Million People AI
1.56 million beneficiaries, 1.2 million SAMAI completions, 14,495 specialists: what SDAIA’s training numbers really measure, which programme suits whom, how to apply, and what a trained workforce changes for companies deploying AI.
Read Article
Sovereign AI vs Cloud AI in Saudi Arabia: Two Gates, Seven Scores, and Why One Policy for the Whole Company Is Wrong
The same company can correctly run one workload on a cloud API and another on-premises. A seven-dimension scoring sheet, which form of hybrid survives staff turnover, and a worked example for a Saudi insurer.
Read Article
SDAIA’s National AI Risk Management Framework: A Practical Guide for Saudi Organisations
SDAIA launched its National AI Risk Management Framework in July 2026. Here is what it contains, how it relates to PDPL and earlier guidance, and how to turn it into a risk register and controls for an AI assistant.
Read Article