22 September 2026 — Being consulted about artificial intelligence is not the same as being able to change how it is used. Spain’s new national plan raises that distinction for workers, public services and communities hosting computing infrastructure.
Prime Minister Pedro Sánchez presented IA360 in Madrid on 21 September. The Spanish government’s announcement describes a twelve-month programme combining technological development, adoption, oversight and a proposed national agreement involving employers and workers. Its benefits are government ambitions, not independently established outcomes.
What has actually been announced?
The published IA360 plan proposes an AI safety institute within Spain’s AI supervision agency, AESIA. Its intended work includes evaluating advanced models before and after deployment. Notably, access before release would be obtained through voluntary agreements with developers. An initial technical team is targeted before the end of 2026.
That detail matters. A public evaluator needs access to the thing being evaluated. Announcing an institution does not establish that access, qualified staffing or effective powers are already available.
The government also proposes talks with social partners in October. El País’s account of the launch highlights proposed sector discussions and a permanent observatory of AI’s effects on employment. These are routes for gathering knowledge and negotiating change, not evidence that agreement has been reached.
The wider political message is a rejection of reliance on industry self-regulation. Reuters reported Sánchez’s argument alongside plans for stronger technological capacity. The proposal is therefore not simply to restrain AI: it is to expand its use under conditions the government says should protect the public.
An illustrative workplace dilemma
Imagine a customer-service team asked to use AI-generated replies. This is a hypothetical example, not a reported Spanish deployment. The software produces text quickly. A worker notices that it regularly misunderstands a particular kind of complaint.
There are several possible responses. The worker could silently rewrite each answer. Management could record the problem but maintain the same speed target. Or the organisation could change the workflow, reduce the target during testing and give the worker a route to suspend unreliable replies.
Only the last response clearly changes the conditions under which the tool operates. Counting completed messages alone could make the first two arrangements look successful. The difference becomes visible when we include correction time, customer outcomes and who has permission to interrupt the process.
AE reading: oversight needs its own working conditions
In autopoietic ecology (AE), developed by Steven Watson and Erik Brezovec, the capacity to act is not treated as a possession that a person or institution simply has. We ask how particular arrangements make an action possible. For an evaluator, those arrangements might include access to a model, suitable tests, relevant expertise and a decision-maker obliged to consider the findings.
The AE Concept Guide also separates present capacity from its renewal. Recruiting a good team once would not settle whether it can retain expertise, update tests and challenge a powerful supplier next year. Oversight itself needs continuing support.
The proposed institute makes these questions concrete. What happens if a developer declines early access? Can reviewers examine realistic uses rather than a carefully selected demonstration? Are findings usable by organisations deciding whether to buy or deploy a system? These are questions for the plan’s implementation, not claims that the institute has already failed.
There is also a difference between checking a model and checking a service. A model could perform well on a defined test while a poorly organised workplace makes its use harmful or ineffective. The hypothetical customer-service team needs an assessment of the whole task: the information available, the route for correcting an answer and the consequences for customers.
The people outside the meeting
An employment observatory could help make changes visible. But national totals alone would not tell us whether one group gains more interesting work while another absorbs unpaid checking. Sector discussions should therefore ask whose experience reaches the table and whose account can alter a deployment decision.
The physical infrastructure deserves the same attention. IA360 describes proposed data-centre requirements concerning energy, water, land and digital autonomy. Its document identifies a draft decree and an October approval target; that target is not proof of enactment. The relevant passage is in the plan’s infrastructure section.
For a host community, an investment total would not answer who pays for additional infrastructure or how competing uses are considered. An AE reading follows those external supports because the apparent success of one activity can depend on burdens being carried elsewhere. It does not presume that every new data centre is either beneficial or damaging.
A fair alternative reading
This could also be understood as industrial policy: building domestic expertise and negotiating standards while trying to attract investment. That is a plausible explanation, and collaboration with developers may provide knowledge that public bodies cannot obtain alone.
The question is not whether collaboration exists but whether dependence makes disagreement ineffective. A voluntary arrangement might secure genuinely useful access; it might also leave important gaps. Neither outcome can be inferred from the announcement.
Following EdgeLab’s Research Programme, the next step is to follow a specific decision rather than score the rhetoric. Watch October’s discussions and the institute’s proposed technical team: what access is secured, what findings can be published, and what action follows an adverse finding? The useful test of public oversight is whether it can change a decision when changing it is inconvenient.
Explore all sections and search by topic in the EdgeLab post directory. Receive published posts by email through our free, optional subscription service.
Leave a Reply