Correction (19 October 2026): An earlier version of this article described events that we could not verify. It has been rewritten from verifiable sources, which are linked in the text.

On 18 August 2026 the Official Gazette (Resmî Gazete, issue 33344) published Presidential Circular 2026/9, signed the day before, which put the Türkiye Artificial Intelligence Action Plan (2026-2030) into effect. The plan succeeds the National Artificial Intelligence Strategy (2021-2025), which was introduced by Circular 2021/18 in August 2021. It was prepared under the coordination of the Ministry of Industry and Technology, with input from public bodies, the private sector, universities and civil society. The circular asks every public institution to carry out the duties the plan assigns to it.

This review is based on the plan text itself, which the Ministry published and which is also available on the strategic-management portal of the Presidency's Strategy and Budget Office. It covers what the plan contains, how it is governed and what it means legally, and ends with our view of where it could be strengthened.

What the earlier version got wrong

The previous version said the plan rested on "three pillars". It also claimed that the plan required locally trained language models to replace foreign APIs in e-Government, UYAP and e-Nabız, promised "an AI assistant for every student", gave SMEs tax exemptions, and contained a provision on the "algorithmic optimisation of bureaucracy". None of these is in the plan. The plan has four axes and sixteen actions, and its main SME instrument is a voucher scheme. Two of the earlier version's "suggestions", an energy strategy and regulatory sandboxes, are already in the plan. The quotations it attributed to the plan have been removed.

The architecture: four axes, sixteen actions

The plan calls itself a "4×4" framework. Its four axes are presented as stages towards AI maturity:

  • Discover (Fark Et): awareness, skills, the data ecosystem and the rules;
  • Benefit (İstifade Et): building and using infrastructure, and adoption in the public sector and industry;
  • Produce (Üret): production, finance, model development and physical AI;
  • Govern (Yönet): global investment, international diplomacy, regional leadership and model safety.

Each axis has four actions. Running across them is a layered model of the ecosystem: five layers (energy, compute hardware, infrastructure, models and data, applications) and three cross-cutting components (talent, finance, and "regulation and trust"). Delivery is phased. The years 2026-2027 are for governance structures, data-space pilots, the first public-sector applications and settling compute investment. The years 2028-2030 are for scaling pilots, commercialising sectoral models and meeting investment targets.

In his foreword, Minister of Industry and Technology Mehmet Fatih Kacır writes that the plan drew on workshops with nearly 500 professionals and more than 2,200 proposals from a citizen survey. He projects economic value of more than 1 trillion lira, but that figure is not one of the action targets.

Axis 1, Discover: literacy, talent, data and rules

Action 1 sets up a national AI literacy and safe-use programme, open to all and requiring no technical background. It will be delivered through e-Devlet, the education platform EBA, public libraries and community centres. An assessment tool in e-Devlet will let citizens measure their level and steer them to suitable training. The targets are 1.5 million completions in 12 months, 5 million in 24 months, workshops in all 81 provinces and 1,000 trainers in the first six months.

Action 2 aims to train 10,000 advanced AI specialists and 100,000 AI application professionals, and to bring a basic AI course into at least 50 universities by the end of 2027. It also creates an occupational redirection programme for workers most exposed to substitution. A "substitution-impact map" is due within 12 months, and at least one million workers are to receive an occupational transition assessment by 2030.

Action 3 organises access to data in two tiers. Public-interest datasets will go into a National Data Library in machine-readable, licensed and anonymised form. Sensitive data in fields such as health, defence, finance and public security will be available only to authorised actors, through secure data spaces. Library datasets are to contain no personal data or be anonymised, and anonymisation standards will be developed in coordination with the Personal Data Protection Authority (KVKK). Data providers can receive tax and R&D deductions, GPU and data credits, and "Trusted Data Provider" status. The target is at least 2,000 public datasets in the Library, at least 100 of them high-value products available by API.

Action 4: the regulatory environment

This is the action of most interest to lawyers. The plan announces a national legal and ethical framework that takes into account the EU's proportionate-risk approach, with obligations graded by risk. Low-risk uses get simple checklists, and medium-risk uses a short-form Algorithmic Impact Assessment. High-impact systems need an extended Algorithmic Impact Assessment, a public summary model card and a detailed technical file for supervisory authorities. "High-impact" covers health, education, employment, social assistance, credit assessment, biometric recognition, law enforcement and justice, and applications that directly affect children and vulnerable groups.

A multi-stakeholder National AI Ethics Board will publish guidance on bias, explainability, data provenance, robustness, human oversight and uses involving children. It will also receive complaints from citizens and refer them onward. The plan is explicit that the Board gives guidance only: binding supervision and sanctions stay with the sectoral regulators, and AI that touches the judicial function is left to the judicial institutions. TÜRKPATENT is to run an "AI Fast Track" for patent and other industrial property applications. KVKK-GDPR alignment is to be strengthened on explainability in automated decisions, personal data in model training and cross-border transfers. Within 12 months the plan expects a proportionate-risk guide with sectoral annexes, the Ethics Board, and impact assessments and model cards for every newly deployed high-impact public AI system.

Axis 2, Benefit: compute, energy and public adoption

Action 5: a low-carbon compute backbone

A National AI Compute Strategy and investment plan are due within six months. The 2030 targets are at least 1 GW of installed AI and data-centre capacity, at least 10 TWh a year of low-carbon power purchase agreements, 1 exabyte of storage and 100 Tb/s of backbone capacity. At least USD 10 billion of mainly private investment is to be mobilised, using the data-centre and AI calls of the HIT-30 high-technology programme as incentives. Contracts are to require at least 60% low-carbon (renewable and nuclear) electricity, and data-centre permits are to go through a "one-stop office" at the Investment and Finance Office. The plan also treats Türkiye's participation in EuroHPC and the EU's "AI Factories" approach as complementary capacity.

Action 6: GPU for All

"GPU for All" will pool accredited domestic data centres and TÜBİTAK ULAKBİM's TRUBA infrastructure, and offer GPU-hour credits to start-ups, SMEs and researchers. The first call is due within six months. Initially, at least five accredited data centres are to provide at least 2 million GPU-hours a year, 40% of it for technology start-ups. This is to rise to 20 million GPU-hours a year by the end of 2028.

Action 7: the state as first customer

The plan targets at least 2% of the total public investment programme budget for AI projects. Every ministry is to scan its AI use cases, run quick pilots and scale successful ones within 6 to 12 months. Each public body is to appoint a senior official for AI and produce a three-year AI roadmap. Start-ups bidding in public AI procurement are expected to hold a technology start-up certificate. Solutions proven in the public sector will get a "digital badge", and rules are to be made so that they can be listed in the State Supplies Office's Tekno Katalog. e-Devlet comes first: personalised service routing, application pre-checks and document support will be piloted with regard to data protection, explainability and human oversight. Basic AI training becomes mandatory for public personnel. The lead body is the Cyber Security Directorate, not the Ministry.

Action 8: sectoral pilots and SME vouchers

Sectoral pilots start in health-service efficiency, grid optimisation, smart manufacturing and geographic information systems, and run as staged competitions in 12- to 18-month cycles. Safety and ethics checks and incident reporting are mandatory. For SMEs the instrument is an AI Voucher and Ready-Solution Package programme, covering software subscriptions, cloud compute, data services and conformity testing. The 12-month target is at least 1,000 vouchers. There is no general tax exemption for SMEs.

Axis 3, Produce: zones, funds, models and robots

Action 9 creates AI Growth Zones: special-status sites where energy, land and telecoms are ready and permits are fast-tracked. They will host multi-tenant "AI factories", alongside regional Centres of Excellence. At least one zone is to be operating by the end of 2027. Action 10 sets up two funds: a National AI Research Fund of at least TRY 10 billion and an AI Growth Fund of TRY 15 billion. The Growth Fund is to back at least 20 start-ups at Series A/B and mobilise two units of private capital for each public unit. Action 11 opens at least two calls a year for sectoral foundation models in areas such as health, drug discovery, disaster management, energy and agriculture. It also continues work on Turkish language models and introduces an "export-ready package" of safety, performance and compliance documents. Action 12 covers physical AI and robotics, including autonomous-driving test environments with insurance and liability rules still to be defined.

Axis 4, Govern: investment, diplomacy and safety

Action 13 offers international investors a single window, with a pre-eligibility review and investment roadmap within 30 business days. Under Action 14, Türkiye is to work in the OECD, G20, UN, ISO/IEC, ITU and UNESCO, and to sign bilateral AI frameworks with at least five countries by 2027. Action 15 plans a common Turkic-languages large language model with at least two Organization of Turkic States members, with a first version due by the end of 2027.

Action 16 brings safety and experimentation together. Safety-evaluation capacity is to be built up at the TÜBİTAK BİLGEM AI Institute, and high-impact systems are to be evaluated before going live and at each major update. Regulatory sandboxes are to start in finance, health, energy, mobility and telecommunications. They will offer a single application point, time-limited trials and a "graduation gate" into permanent rules or public procurement. Cloud and cross-border transfer rules in regulated sectors are to be reviewed, with an action plan due within six months. For the long term, the key point is that the plan says the AI legal framework will rest on "a general law" setting out principles, mandates and responsibilities, with details left to secondary regulation.

Who runs it

The National AI Council, chaired by the President, is to be the highest decision-making and coordination body. It will approve annual implementation plans, resolve conflicts of competence and publish an annual "National AI Policy Letter". Its secretariat is the Ministry's Directorate General for National Technology and AI, which will also house a Programme Office. A public Progress and Transparency Portal is to show, for each action, the responsible body, timetable, budget against spending and a green, yellow or red status. The first public progress report is due within 12 months, and later annual reports are to include an independent evaluation.

What the plan means legally

A presidential circular is addressed to the administration. Neither the circular nor the plan imposes enforceable duties on private companies, and the plan's glossary says that concepts needing legal force will be defined in legislation. In our view, however, the plan will reach the private sector in four ways:

  1. Public procurement. The 2% target, the digital badge, the start-up certificate requirement and the documentation expected of high-impact public systems will shape what suppliers to the state must show.
  2. Documentation norms. Impact assessments, model cards and technical files apply first to public systems, but the plan presents them as standards. They are likely to find their way into sectoral regulation.
  3. Data. The KVKK anonymisation standards and closer GDPR alignment on training data will matter to anyone building models on Turkish data.
  4. A framework law. The plan commits to a general AI law but sets no timetable.

The plan's high-impact list overlaps substantially with the high-risk areas in Annex III of the EU AI Act: education, employment, essential services and benefits (including credit scoring), biometrics, law enforcement and justice. On our reading, it omits critical infrastructure, migration and democratic processes, and adds health in general and children. Its tools also differ: impact assessments and model cards where the EU uses conformity assessment. Companies active in both markets should not assume that complying with one framework satisfies the other.

Where the plan could be strengthened

The points below are our analysis, not part of the plan.

  • Oversight. The plan creates no independent AI authority. It splits the work between a Council chaired by the President, an Ethics Board with no binding powers and the existing sectoral regulators. That avoids a new bureaucracy, but it leaves open who supervises AI uses that fall between regulators and how individuals can obtain a binding remedy. The promised general law is the natural place to answer this.
  • Consistency. The plan's summary of the actions describes GPU for All as offering 10 million GPU-hours a year, and its closing "where are we going" page also speaks of 10 million GPU-hours of annual capacity. Action 6 itself sets at least 2 million a year in the first phase, rising to at least 20 million by the end of 2028. The progress portal should state which figure is the benchmark.
  • Money. Apart from the two funds, budgets are left to the annual implementation plans. The 2% share is worded as a target, not an appropriation.
  • Safeguards. Requiring impact assessments for every new high-impact public system is a strong commitment. Publishing the assessments themselves, and not just summary model cards, would make it stronger.
  • Sandboxes. These need a legal basis that lets regulators grant time-limited exemptions. Without one, the "graduation gate" will be hard to operate.

Burhan Doğuş Ayparlar's View

Read on its own terms, the 2026-2030 plan is more concrete than its predecessor. Almost every action has a lead body, a deadline and a measurable target, and the plan commits to reporting its own progress in traffic-light colours. That self-imposed transparency may be its most valuable feature, because it gives the public and the market a yardstick.

Legally, the plan's centre of gravity lies in Actions 4 and 16. Together they point to a Turkish regime that borrows the EU's risk-based logic but relies on impact assessments, model cards and sectoral supervisors rather than a single regulator. Until the general law is drafted, that regime will exist mainly in procurement conditions and guidance. Organisations that supply AI to public bodies, or work in regulated sectors, may find it sensible to keep an inventory of their AI systems, record the data sources and legal bases used for training, and prepare documentation of the kind the plan describes for high-impact systems.

Several targets fall due within the plan's first six months, which on our reckoning runs to about February 2027. They include the compute strategy, the first GPU for All call and the review of cloud rules. The first progress report is due within a year of publication. Those milestones, more than the plan's ambitions, will show how much of it becomes practice.

Sources

This article is general information and does not constitute legal advice. It reflects the position as of 19 October 2026.