Abu Dhabi is training 20,000 public workers for AI-native government

THE SIGNAL IN ONE SENTENCE

There are two ways to announce an artificial-intelligence strategy. One is to buy the tools. The other is to teach people what to do with them, when not to trust them and how to tell whether they helped. Abu Dhabi is trying the second route at unusually large scale. The emirate's Department of Government Enablement has launched AiNative, a training program built with OpenAI, Microsoft and Mohamed bin Zayed University of Artificial Intelligence. The program aims to train 20,000 government employees by the end of 2026. It also includes a public learning journey that is supposed to open after an initial government pilot. That is a meaningful change in emphasis. Government AI programs are often described through models, data centers and productivity promises. AiNative starts with the people expected to use the systems. Foundation courses will cover practical tools including ChatGPT, Codex and Microsoft 365 Copilot. Department leaders get a separate executive program. Later phases are expected to add function-specific certifications in areas such as coding and data. Citizens and residents are part of the design too. The public journey is planned for Tomouh, Abu Dhabi's AI-powered learning app. People will log in with UAE Pass and use the free version of ChatGPT to work through examples involving personal finance, small businesses, job searches, studying and responsible content creation. That could make AI training more useful than the familiar hour-long seminar followed by a certificate and a brave return to the inbox. It could also become a large enrollment counter with very little public evidence behind it. The number 20,000 is a target, not a result. The official announcement does not yet say how many employees have started, how many will finish, how practical skill will be assessed, what accessibility support will be available, which languages the material will cover or how the government will measure changes in service quality. The public course is not generally available yet. A pilot is under way inside government. Those gaps are normal at launch. They should become the reporting plan. Abu Dhabi already has a substantial installed base for the effort. The government says Microsoft 365 Copilot was rolled out to 35,000 employees in July through a separate Frontier Employee Program. AiNative is therefore not teaching into an empty room. Many workers are already surrounded by AI features in tools they use. That makes the training question more immediate. When a civil servant asks an assistant to summarize a case, draft a public notice or organize a spreadsheet, the output can look polished while carrying a wrong date, a missing exception or a private detail that should never have entered the prompt. A general warning to "check the answer" is not enough. Training should be tied to the real decisions people make. A procurement officer needs examples about confidential bids, traceable research and conflicts of interest. A call-center worker needs to know when a generated answer must be checked against the official rulebook and when a person should take over. A policy analyst needs to distinguish brainstorming from evidence. A software team needs secure coding practices, testing and approval gates. A manager needs to know how productivity targets can quietly pressure employees into using a tool where it does not belong. The useful unit is not a completed video. It is a safer workflow. That is why each learning journey should end with a job-relevant task, a documented review and a record of what the learner caught. Can the employee identify a fabricated source? Can they remove personal information before using a system? Can they explain why a result should be rejected? Can they compare the AI-assisted process with the old one on time, error rate and user satisfaction? Can they stop the tool and finish the job manually when the network, model or policy fails? Those questions turn confidence into competence. The public course needs equally practical guardrails. Helping someone draft a business plan or study outline is relatively low risk. Personal finance and job hunting can become consequential quickly. A model may confidently invent an eligibility rule, produce unsuitable investment guidance, expose sensitive details in a resume or reproduce bias in a hiring suggestion. A responsible lesson should show both the useful prompt and the exit ramp. For personal finance, that could mean organizing expenses while clearly separating general information from regulated advice. For a small business, it could mean drafting product descriptions without uploading customer records. For job search, it could mean tailoring a resume while checking every claim and preserving the applicant's voice. For studying, it could mean using AI to quiz understanding rather than outsourcing the assignment. The program should also account for people who do not begin at the same digital starting line. Tomouh access through UAE Pass may make identity and progress tracking straightforward for many residents. It can also create friction for people who share devices, have limited connectivity, use assistive technology, read more comfortably in a language not prioritized by the course or need in-person support. Public access is not the same as a public link. Abu Dhabi should publish the languages, accessibility standards, device requirements, support channels and community delivery partners before claiming broad reach. Libraries, universities, community centers and employers can matter as much as the app. Curriculum ownership is one of the more promising details in the announcement. The Department of Government Enablement says it owns the curriculum and designed AiNative to be technology-agnostic, leaving room to add partners over time. That could protect the program from turning into permanent product training for whichever vendor arrived first. The test will be in the lessons and assessments. People should learn transferable skills: how to define a task, inspect a source, protect data, test a result, record a decision and escalate a risky case. Product-specific instructions are useful, but they should sit beneath those durable habits. If the interface changes or a contract ends, the public investment in judgment should survive. Technology neutrality also needs procurement evidence. Can a department move a workflow to another provider without rebuilding the course? Are data-retention and model-training settings explained? Are instructors able to compare tools honestly? Can independent researchers examine outcomes? Does the government publish which partner influenced which part of the curriculum? The announcement says government, industry and academia shaped the program together. That can combine practical deployment experience with research and public responsibility. It can also blur the line between education and market development if the roles remain vague. Name the roles. Publish the rubrics. Keep the receipts. The best public dashboard would not rank departments by how many employees clicked through modules fastest. It would show enrollment and completion, assessment performance, retraining, participation by role and language, accessibility use, reported incidents, human overrides, employee confidence before and after training, and changes in service measures connected to specific workflows. Those service measures need restraint. If permit processing becomes faster, check whether error and appeal rates changed. If call handling becomes shorter, check whether people had to contact the government again. If document drafting speeds up, measure corrections and review time. A productivity gain that moves work onto residents or reviewers is not a gain. It is an accounting trick wearing comfortable shoes. Employees also need a safe way to report where AI made work worse. Training programs often reward adoption and quietly discourage dissent. A worker who identifies that a tool is unreliable for Arabic dialects, inaccessible to a colleague or prone to missing a legal exception is providing valuable operational evidence. That report should improve the system, not damage the employee's performance review. Independent evaluation would make the whole effort more credible. MBZUAI is already a curriculum partner. Abu Dhabi could also invite outside researchers, worker representatives, accessibility experts and service users to test the program, publish aggregate findings and examine whether measured improvements persist after the course ends. The initial pilot is the moment to build that evidence loop. Start with a few workflows. Establish a baseline. Train people. Test the exact task. Observe what happens in live work. Ask service users what changed. Record harms and near misses. Revise the curriculum. Then expand. That method is slower than treating 20,000 people as one giant attendance list. It is also how a government learns. AiNative's most important promise is not that every employee will become an AI expert. They will not, and they do not need to. The promise is that people across government and the wider public can gain enough practical judgment to use these systems without confusing fluency for truth. Abu Dhabi has announced the scale, the partners and the route into the classroom. The next signal is whether it publishes what people learned, what services improved and where the tools were wisely left alone.