A new training event for public employees in Castile and León highlights a wider challenge for government: turning familiarity with generative AI into the practical judgement needed to protect information, verify outputs and improve public services
On 28 October 2026, Castile and León brought public employees together in Valladolid for a session on artificial intelligence in administration. The regional government’s account, republished by Spain’s National Institute of Public Administration on 30 October, describes an event with 300 attendees organised by the regional public administration school, ECLAP. It places the session within a wider digital training effort: 184 activities and more than 12,000 participants during 2026, covering cybersecurity, collaborative tools and administrative applications of AI. These are figures for digital training as a whole, not evidence that 12,000 people have completed specialist AI instruction.
The announcement matters because introducing AI into government is partly a question of professional competence. A public employee needs to understand more than how to obtain a polished answer from a chatbot. They must know whether the tool is approved for the material involved, what makes an answer trustworthy and when using it would weaken rather than improve a procedure. The regional announcement presents faster administration and better service as objectives. It does not demonstrate that the training has already produced those outcomes. The useful question is what would have to happen between attending a session and changing the quality of daily administrative work.
Digital confidence is not the same as authority
A generative system can help draft a letter without possessing the authority to decide what that letter should promise. It can summarise a file without knowing which omission would change an applicant’s position. In public administration, that distinction links technical competence to the exercise of institutional power. Consider an illustrative request to summarise a complicated benefits application. A shorter account may save reading time, but a missed exception, date or supporting document could steer the next official towards an incorrect conclusion. Training should make employees capable of recognising that risk and returning to the record, rather than treating the fluency of the summary as a substitute for verification.
This changes what a good learning exercise looks like. A demonstration in which an assistant produces an impressive draft may build interest, but it says little about whether staff can detect a persuasive error. An administration could instead give participants deliberately incomplete records, conflicting instructions or an answer containing an unsupported legal reference. The task would be to identify what cannot safely be accepted and explain how to check it. These are proposed training methods, not reported features of ECLAP’s session. They illustrate why institutional judgement requires practice with uncertainty as well as familiarity with a tool’s capabilities.
Protection begins before the prompt is submitted
The reporting of the Valladolid event emphasises data protection, information security, transparency, equality and human supervision. Europa Press’s account of the same event independently confirms its date and describes the regional emphasis on responsible use. That emphasis needs to become an operational boundary. An employee deciding whether to paste material into an assistant must be able to distinguish an approved environment from a convenient public service, and a harmless example from information that identifies a person or exposes an internal process.
Anonymising a name is not always enough to make a file safe to share: combinations of circumstances may still identify someone. Conversely, an organisation that provides only broad warnings may leave staff unable to use even low-risk applications confidently. Effective guidance would specify permitted tools, suitable categories of information and a route for obtaining advice when the answer is unclear. It should also address retention and access to records created during use. Responsibility cannot rest entirely on an individual remembering a lecture; the working environment must make the approved path understandable and available.
A European framework supports context-sensitive learning
The European Commission’s current AI literacy guidance provides a relevant context. It describes measures to support staff literacy in relation to their knowledge, experience and the circumstances in which systems are used. It also states that no particular certificate is necessary and that organisations can maintain internal records of training and other guidance. The implication for administration is practical: a record of participation can document an effort, but the learning still has to fit the work. General awareness, specialist procurement knowledge and the ability to oversee a consequential system are different needs.
A procurement officer may need to question a supplier about data handling, evaluation and exit arrangements. A communications employee may need to verify facts and prevent an AI-generated explanation from misrepresenting policy. A manager must understand when a proposed efficiency measure changes who reviews a case or who bears responsibility for an error. Giving everyone the same introduction can establish shared vocabulary, but it cannot complete all of these roles. A more durable programme connects learning to specific responsibilities and offers further support as the administration’s systems and procedures change.
Attendance is a starting measure, not a service outcome
The breadth of the reported digital training programme is a sign of organisational activity. Its effects require a different kind of evidence. Departments could examine whether staff select permitted tools correctly, detect unsupported claims and know how to escalate a doubtful result. For a particular workflow, evaluation might compare the time needed to produce a reviewed final output, rather than the time required for the model’s first draft. Corrections, supervisory effort and mistakes discovered later belong in that comparison. Otherwise a faster initial response may conceal work transferred to somebody else.
Public service quality also extends beyond productivity. A tool-assisted letter should remain understandable, accurate and appropriate to the citizen’s circumstances. If an assistant makes correspondence more uniform while overlooking unusual needs, an apparent improvement in consistency may come at the expense of fairness. This is why assessment should include difficult cases and staff feedback, rather than only examples selected to showcase success. The aim would be evidence about the complete administrative process, with explicit limits on where the observed benefits can be expected to apply.
Turning a learning event into institutional memory
The regional account also notes the existence of a practical guide to generative AI for employees. A guide can provide continuity after a training event, but its usefulness depends on how it is maintained and connected to everyday decisions. An administration could collect recurring questions, review incidents and update worked examples as tools change. Staff should be able to report uncertainty without feeling that doing so reveals incompetence. In a new technological environment, knowing when to ask for help is part of professional capability.
Castile and León’s initiative offers a timely entry point into this wider challenge, rather than a completed model of successful adoption. The next test is whether awareness becomes a shared capacity to make sound decisions under ordinary working pressures. Public-sector AI literacy earns its value when an employee can use assistance appropriately, recognise its limits and preserve an accountable route from information to action. That is a more demanding ambition than increasing the number of people who have tried an AI tool, and a more useful one for the citizens whose cases pass through the administration.