How accurate is AI on Turkish law? A citation and current-legislation test (November 2026)

Short answer: In November 2026 we asked four AI models 28 questions on Turkish law and looked up every citation they gave in its source. Without tools, Gemini 4 gave 40 Yargıtay (Court of Cassation) citations and wrote a decision number that does not exist in 32 of them; only 7 were verified. Of the 28 citations GPT-6.1 Sol gave, 15 were verified; of GPT-6 Astra’s 24, 16 were verified; 10 and 7 of them, respectively, were fabricated. Without tools, Claude Opus 5.5 mostly declined to give citations. The same Claude, connected to Yargı PRO, got all 40 of its 40 citations right. On 8 questions about provisions amended by the 12th Judicial Package, none of the models without tools got more than one right; Claude connected to Yargı PRO answered all eight correctly.

The results at a glance#

A real search: the citation test#

On 20 different topics (such as interest on severance pay, rent determination, the family-home annotation on the land register and aggravated fraud) we asked the models for two Yargıtay decisions each, with chamber, docket (esas) number, decision (karar) number and date. We searched for every citation in Yargı PRO’s decision search by docket and decision number; where nothing came up, we also searched for the docket number alone within that chamber.

Gemini 4: In 32 of the 40 citations, the docket number really exists in that chamber, but the file was decided under a different number and date; the decision the model wrote does not exist. For example, for interest on severance pay the model wrote “9th Civil Chamber, E. 2016/14258, K. 2020/7542, 11.06.2020”. In reality this docket file was decided on 12.05.2016 under K. 2016/11930; there is no decision four years later. In one citation the date was a day out and the decision was not on the topic asked. Of the 7 verified citations, 4 matched the record exactly; 3 are not in the decision archive but are cited with the same details in other Yargıtay decisions.

This also shows why a fabricated citation is so convincing: because the docket number is real, at first glance the record seems to exist. You need to check the decision number and the date as well.

GPT-6.1 Sol and GPT-6 Astra: The two models gave no citation in 6 and 8 of the 20 questions respectively, returning an empty list in the requested format. More than half of the citations they did give were verified: 15 of GPT-6.1 Sol’s 28 and 16 of GPT-6 Astra’s 24. Of the 31 verified citations, 21 were decisions of the Civil or Criminal General Assembly or decisions unifying conflicting case law; 17 of the 21 wrong citations were chamber decisions. For example, for the calculation of notice pay GPT-6.1 Sol wrote “9th Civil Chamber, E. 2009/48920, K. 2012/9400, 21.03.2012”. In reality this file was decided on 08.03.2012 under K. 2012/7809, and the decision is not about notice pay. For interest on severance pay, GPT-6 Astra gave two citations whose docket numbers do not exist in the 9th Civil Chamber at all. One GPT-6.1 Sol citation matched the record exactly, but the decision was about amending a claim in a rent determination case, not about the topic asked (partition of co-owned property).

Claude Opus 5.5, no tools: In 19 of the 20 questions it declined to give a citation; it said numbers written from memory could turn out to be fabricated, and explained the general rule on the topic instead. The one citation it did give (a decision unifying conflicting case law) was verified. This is honest behaviour, but since it gives the lawyer no citations, it does not move the research forward either.

Claude Opus 5.5, connected to Yargı PRO: All 40 citations matched the record exactly, with chamber, docket number, decision number and date. Every decision dealt with the topic asked; in five of them the link to the topic was loose (for example, for a question on reinstatement, decisions on the wage claim after reinstatement).

The current-legislation test: they do not know the 12th Judicial Package#

Law No. 7589 was published on 31 July 2026; the models’ knowledge ends before that. We asked about eight provisions amended by this law “as of November 2026” and scored the answers against the current text of the law.

• When we asked whether an action for an unquantified claim (belirsiz alacak davası) can be brought today, Claude and Gemini without tools said “Yes, it can.” The GPT models also set out the conditions of HMK (Code of Civil Procedure) Art. 107 and added that they could not verify whether it is still in force. Art. 107 was repealed by this law.
• Claude and Gemini without tools gave the objection period of the Chief Public Prosecutor of the Court of Cassation as “thirty days”, the two GPT models as “one month”. The new period is three months from the date the file is delivered to that office.
• The only answer the two GPT models got right was the single-judge monetary limit in the administrative courts (486,000 TL); they based it on the yearly revaluation of the limit, not on the amendment made by Law No. 7589.
• Claude connected to Yargı PRO cited the amending provision (for example “Law No. 7589, Art. 19”) in every answer.

All the changes, and the results question by question: What changed with Turkey’s 12th Judicial Package (Law No. 7589)?

Method#

• Date: Claude and Gemini were tested on 3 November 2026, the GPT models on 6–7 November 2026.
• Models: Claude Opus 5.5 (Claude Code command line), Gemini 4 (Antigravity command line, plan mode), GPT-6.1 Sol and GPT-6 Astra (Codex command line, high reasoning effort). We ran Claude without loading user settings or memory, with web search turned off and, in the no-tools arm, from a neutral folder whose name does not mention Yargı; in the Yargı PRO arm the only tool enabled was Yargı PRO. We told Gemini in the prompt not to search the web. We ran Codex with an empty home folder and a separate settings folder holding only the sign-in data; web search, running commands, connectors and plugins were turned off.
• Questions: 20 citation questions and 8 current-legislation questions; the same prompt for each arm, each question in a separate session.
• Verification: Citations were searched in Yargı PRO’s decision search by docket and decision number, and where needed by docket number and chamber; every classification rests on a search result. A citation that is not in the decision archive was counted as verified only if other Yargıtay decisions cite it with the same details. Legislation answers were scored against the current text on mevzuat.gov.tr.
• Limits: The sample is small; each question was asked once, and models do not give the same answer to the same question every time. Four models were tested; only Claude was tested connected to Yargı PRO. In our first attempt we ran Claude without tools from a folder named “yargi-mcp-pro”; in some answers the model saw this name and mentioned Yargı MCP. So we re-ran that arm from a neutral folder and report the results of that run. In Codex’s first run, the user skills on the computer were also loaded and the model tried to search the web; we discarded those answers and re-ran the test in the isolated setup described above. The test was prepared by Yargı PRO; we publish all the prompts, answers and scores.

Raw data: yapay-zeka-dogruluk-testi-2026-11.json (the questions, the models’ answers, and the verification result for each citation).

What do these results mean for a lawyer?#

• Do not trust a citation that comes from memory. Even if the docket number looks real, the decision number and date may be fabricated. Check all four together (chamber, docket number, decision number, date). A model that gets half its citations right writes the other half with the same confidence; only a search in the source tells you which is which.
• Have current legislation read from its source. A model does not know about changes after its knowledge cut-off date and often presents the old rule as if it were current.
• Work connected to a source, and still check the result. When connected to Yargı PRO, the model retrieves the citation from the source; the risk that a real decision is summarised wrongly remains.

Cases that have reached the courts in Turkey: AI hallucination cases in Turkey. Verifying citations step by step: AI hallucinations: how to verify fake Turkish court citations. To connect your AI: How to connect Claude, ChatGPT and Gemini to Turkish law sources