What the Constitutional Court's President said

The occasion was a symposium at the Constitutional Court marking fourteen years of the individual application mechanism, attended by the President of the Court of Cassation, the Court's vice-presidents and members, rapporteurs and academics.

The argument began from workload. Since the mechanism opened on 23 September 2012 the Court has received 756,726 individual applications and concluded 648,410 of them, finding violations in 86,849 — of which, the President noted, 99.9 per cent have been implemented. On those numbers he made the case that the use of technology is not a choice but a necessity.

Then came the conditions. Artificial intelligence must be used to support the judge and the judicial decision-making process, not to replace the judge. Responsibility for the decision remains with a human. And where AI systems are used, transparency, explainability, personal data protection, prevention of discrimination and the continuity of human oversight must not be compromised.

This is a speech, not a judgment. It creates no obligation and binds no court. But it is the most authoritative statement a Turkish judicial figure has made on algorithmic decision support, and it is a reasonable indication of the constitutional standard the Court would apply if a case reached it.

Why the list is more specific than it sounds

Five conditions were named, and each maps onto an identifiable legal source rather than floating free as principle.

Human responsibility for the decision tracks the constitutional position of the judge. Judicial power is exercised by independent and impartial courts, and a decision not attributable to an identified human judge is difficult to locate within that structure at all. This is the least negotiable item on the list.

Personal data protection is Law No. 6698, and in the judicial context the most relevant provision is Article 11, which gives the data subject the right to object to a result arising exclusively from automated processing where it produces an adverse outcome. A system that produced judicial outcomes without meaningful human involvement would engage that right directly.

Transparency and explainability connect to the constitutional requirement that court decisions be reasoned. A reason that cannot be given is not a reason. If a system contributes materially to an outcome and the contribution cannot be articulated, the resulting decision has a reasoning defect independent of anything to do with technology.

Prevention of discrimination engages the equality provisions of the Constitution and, for the European dimension, Article 14 of the European Convention on Human Rights. Systems trained on historical case outcomes reproduce the distribution of those outcomes, including any unequal treatment embedded in them.

Continuity of human oversight is the operational condition that makes the first one real. Nominal human responsibility with no practical capacity to disagree is a well-documented failure mode — automation bias is strongest exactly where the human reviewer is overloaded, which is the condition the workload figures describe.

That last tension is the interesting one and the speech does not resolve it. The justification for deploying the technology is volume. The safeguard against the technology is meaningful human review. Both cannot scale at once, and the honest question is which gives way when they collide.

The circular the same week

On 19 September, four days earlier, the Official Gazette carried Presidential Circular 2026/11 bringing into force the National Smart Cities Strategy and Action Plan to 2030.

The strategy is reported to contain thirty-two actions, with IoT platform integration, a priority on data governance, and — the part that matters here — AI-supported decision-making mechanisms in municipal service delivery. AI pilot projects are scheduled for 2026–2027, with evaluation of large language model applications for 2027–2029. Implementation runs through the Ministry of Environment, Urbanisation and Climate Change and local administrations.

This is a genuine mandate with a timetable, and it lands on a part of the state that is substantially less equipped than the judiciary to think about what it entails.

Municipal decision-making is administrative action. It grants and refuses licences, allocates social assistance, imposes fines, determines zoning applications, prioritises inspections. These are decisions with direct legal effect on individuals, and they are subject to a body of administrative law that is entirely indifferent to whether a computer was involved.

What administrative law already requires

The obligations that attach to an AI-supported municipal decision exist now, and none of them needed an AI statute.

Reasons. Administrative acts affecting individuals must be reasoned, and the reasoning must be capable of being examined by an administrative court. A municipality that refuses an application because a model scored it poorly has given no reason. The output is not the reason; the reason is why the output is what it is.

Legal basis and competence. Administrative authority is conferred on a specified organ. It cannot be delegated to a system, and a decision taken by a system without the competent organ's genuine involvement has a competence defect.

Objection to automated decisions. Article 11 of Law No. 6698 applies to public controllers. A citizen adversely affected by a decision arising exclusively from automated processing has a right to object, and a municipality must be able to identify which decisions fall into that category — which, in most municipalities, nobody currently can.

Judicial review. Administrative acts are reviewable, and review requires a record capable of supporting it. A system whose inputs and weights at the time of decision were not retained cannot be reviewed in any meaningful sense, and the effect is to insulate the decision from a court — which is a constitutional problem rather than a records-management one.

Procurement. A municipality that buys a decision-support system acquires the legal responsibility for decisions taken with it. The supplier's terms will not transfer that. Contract terms about accuracy, explainability, audit rights and retention of decision records are the point at which most of this is either secured or lost, and they are settled at procurement by people who are not lawyers.

The gap between the two documents

Place the two side by side. The Constitutional Court's President states principles for judicial AI, in a forum of judges, about systems that mostly do not exist yet. A circular directs municipalities to deploy AI-supported decision-making, on a timetable beginning immediately, in a domain where such systems very much will exist.

Between the principle and the deployment there is no instrument. No binding rule on public-sector AI, no procurement standard, no impact assessment requirement, no register of automated decisions, no guidance to municipalities on what Article 11 means for a licensing workflow. The AI Action Plan for 2026–2030 promises a proportionate risk-based framework and sectoral sandboxes; those are commitments, not rules, and they will not arrive before the pilots do.

This is the ordinary sequencing problem of technology governance, and Türkiye is not unusual in having it. What is unusual is the clarity with which one week displayed it.

The comparison worth drawing is with the European approach, where AI systems used by public authorities to determine access to essential public services and benefits, and systems intended to assist judicial authorities in researching and interpreting facts and law, are classified as high risk under the AI Act and carry obligations of risk management, data governance, logging, human oversight and registration. One can argue about whether that regime is well calibrated. It is not seriously arguable that the category is wrong: these are the decisions where the state acts on the individual, and they are the decisions Türkiye is about to automate without an equivalent framework.

What a municipality should do before the pilot

The circular's timetable means this is a 2026–2027 question, not a future one.

Classify the decision before selecting the tool. Is the output advisory or determinative? Does it affect a right, a benefit or an obligation? Is a human genuinely able to depart from it? These answers determine the legal regime and they should precede procurement, not follow it.

Retain the decision record. Inputs, model version, output, the identity of the official who decided, and the reasons given. Without this an administrative court cannot review the act, and the municipality cannot defend it.

Write the reasons from the file, not from the score. The reasoning must stand on the underlying facts. A reason that reduces to "the system indicated" will not survive review and should not.

Make departure real. If officials never disagree with the system, oversight is nominal. The rate at which recommendations are departed from is the single most useful metric a municipality can track, and a rate of zero is a finding, not a success.

Fix audit rights at procurement. Access to model documentation, the right to have outputs examined, retention obligations, and an obligation to notify material changes to the model. These cannot be added afterwards.

What the courts are already using

It is worth being clear that judicial technology in Türkiye is not hypothetical, because the debate is often conducted as though it were.

UYAP, the national judicial informatics system, has been in operation for two decades and handles case management, document production, electronic filing and integration with other state databases. It is one of the more comprehensive judicial information systems anywhere, and it means the Turkish judiciary already runs on software that touches every case.

The distinction that matters is between systems that manage process and systems that contribute to outcome. Document assembly, hearing scheduling, deadline calculation and file routing are process. They raise ordinary questions about reliability and records and no constitutional ones. Systems that classify a case, predict an outcome, rank applications by likely merit or draft reasoning are a different category, because they influence what is decided rather than how it is administered.

The Constitutional Court's own workload provides the obvious application. With more than a hundred thousand applications still pending, triage — identifying which applications are plainly inadmissible, which raise a repeated issue already resolved, and which require full examination — is the function where automated assistance would deliver most. It is also the function where the risk sits, because an application wrongly triaged as repetitive is an application that never receives the examination the mechanism exists to provide.

That is the concrete version of the President's principles, and it is why the speech reads as more than ceremony. The conditions he named are the conditions that would have to be built into a triage system: a human who can escalate, a record that shows why a case was classified, and a way to detect whether the classification is systematically wrong for a particular category of applicant.

What to watch next

Whether the Council of Judges and Prosecutors or the Ministry of Justice issues anything operational on judicial AI, which would convert the principles into practice. Whether the smart cities pilots are accompanied by guidance to municipalities on reasons, records and Article 11 — and whether KVKK, which has been the most active Turkish regulator in this area, takes up public-sector automated decision-making as it has taken up agentic AI and legal practice. Whether the promised risk-based framework in the AI Action Plan arrives before the 2027–2029 large language model evaluation phase. Whether the draft Turkish AI framework proposal is introduced in the parliamentary session beginning in October. And whether any administrative court is asked to review a decision in which a system materially contributed, which would settle more than any amount of strategy.

Frequently asked questions

Did the Constitutional Court rule on artificial intelligence?

No. This was a speech by the Court's President at a symposium. It has no binding force. It indicates the standard the Court would likely apply, which is useful precisely because no Turkish judgment yet exists.

Does the circular require municipalities to use AI?

It brings into force a strategy and action plan that includes AI-supported decision-making in municipal services, with pilots scheduled for 2026–2027 and evaluation of large language model applications for 2027–2029. It is a policy instrument with a timetable rather than a rule imposing an obligation on any particular authority.

Can a municipality make a decision with AI today?

Yes, subject to law that already applies: reasons, competence, reviewability, and Article 11 of Law No. 6698 on decisions arising exclusively from automated processing. The absence of an AI statute is not an absence of constraint.

What is the biggest practical risk?

Records. A decision whose basis was not retained cannot be reviewed by an administrative court, and unreviewable administrative action is a constitutional problem. Most of the exposure is created at procurement, by contracts that do not require retention.

How does this compare with the EU?

Under the EU AI Act, public-benefit eligibility systems and systems assisting judicial authorities are high risk, with obligations including risk management, logging, human oversight and registration. Türkiye has no equivalent framework in force; the Action Plan promises one.

Does Article 11 apply to public authorities?

Yes. Law No. 6698 applies to public controllers, and the right to object to a result arising exclusively from automated processing is not limited to the private sector.

What counts as "exclusively" automated?

The qualifier does a great deal of work and has not been tested in Turkish practice. A human who signs whatever the system produces is unlikely to take the decision outside the provision; the involvement has to be capable of changing the outcome. This is the question on which the first Turkish case in this area will probably turn.

Is UYAP an AI system?

UYAP is a judicial informatics system handling case management, documents and filings. Process automation of that kind raises reliability and records questions rather than constitutional ones. The distinction that matters is whether a system contributes to what is decided, not whether software is involved.

Burhan Doğuş Ayparlar's View

This section sets out my personal assessment as the founder of this site and an AI ethics & compliance counsel.

I have read a great many statements of AI principles this year and most of them are interchangeable. The Constitutional Court President's is not, for one reason: he tied the argument for the technology to a specific number, and the number is the problem.

Seven hundred and fifty-six thousand applications. On that volume, the case that technology is a necessity rather than a choice is unanswerable, and I would not want to argue against it. But look at what the same figure does to the safeguard. The condition that makes AI acceptable in adjudication is meaningful human oversight — a human who can and sometimes does disagree. Meaningful review takes time. The reason the technology is being deployed is that there is not enough time. The safeguard is therefore under pressure from precisely the thing that justifies the deployment, and no list of principles resolves that. It has to be resolved by design: by limiting what the system is allowed to touch, by measuring how often reviewers depart from it, and by treating a zero departure rate as evidence that oversight has failed rather than that the system is good.

The juxtaposition with the smart cities circular is what I find genuinely concerning, and I want to be precise about why. It is not that municipal AI is inherently worse than judicial AI. It is that the judiciary has just had this conversation at the highest level, in public, with the right conditions named — and the municipalities have not had it at all. A district that automates part of its social assistance triage in 2027 will do so having received a policy mandate to deploy and no guidance whatsoever on what Article 11 requires, what the reasons obligation means when a score is involved, or what needs to be retained for an administrative court to review the decision. They will buy a product, the supplier will explain that it is decision support, and nobody in the room will ask whether the official who signs the refusal can actually explain it.

The thing I would fix first is not a statute. Statutes take years and the pilots start next year. It is a record-keeping rule, which could be issued as guidance tomorrow: any administrative decision in which an AI system materially contributed must retain the inputs, the model version, the output, the deciding official, and reasons drawn from the underlying file rather than from the output. That single requirement does most of the work. It preserves judicial review, which is what protects the citizen. It creates the evidence base a court needs. It forces the reasoning discipline that prevents the score from becoming the decision. And it costs almost nothing if it is specified before procurement rather than discovered afterwards.

My wider view, and I hold it more strongly after this week: Türkiye's AI governance problem is not the absence of a law. It is the absence of connective tissue. We have an action plan, a draft framework, a signature on an international declaration, a head of state calling for a UN convention, a data protection authority producing genuinely good sectoral guidance, and now a constitutional court president articulating sound judicial principles. What we do not have is anything that translates any of that into an instruction a municipal official can follow on a Tuesday. The judiciary will probably be fine — courts are institutionally careful and this one has now said the right things in public. It is the six hundred-odd municipalities working to a 2026–2027 pilot timetable, with no guidance, buying from vendors who will tell them it is only decision support, where the first Turkish case about an automated administrative decision is going to come from. We could write the guidance now, or we could read about it in a Danıştay judgment in 2029.

This article is for information only and does not constitute legal advice. The facts are based on press reporting of remarks made on 23 September 2026 and on the circular published in the Official Gazette of 19 September 2026; the remarks were made in a speech and carry no binding force. Statements about Turkish law are general in nature; specific cases require individual assessment. The analysis and assessments are the author's own.