How do language models answer questions about Polish law?

Design#

Ten short legal questions collected in practice, in the form a lawyer types into a chat (typos kept; in three questions missing facts were added before the study; six contain a false or outdated premise the model should correct). Eight models from OpenAI (GPT-6.1 Sol, GPT-6 Astra, GPT-5.6 Terra), Anthropic (Claude Opus, Sonnet, Haiku 5.5) and Google (Gemini 3.8 Flash, Gemini 4 Argon), each answering twice: without tools (W0) and with the provider's web search (W1). One run, one answer per question.

Every answer has a scoring card. An answer is labelled hallucinated if it contains at least one statement about a legal source that disagrees with the official text of the law in force on the reference date (1 October 2026), in one of four classes: F (the legal source or numbered provision does not exist), T (a real source attributed content absent from it), W (wording from another date presented as current), S (a real source with a distorted identifier). An answer is correctly grounded (a term defined by the authors) if it contains all essential elements specified in the answer key, has no detected hallucination, and has at least one citation that the assessor confirmed in an official source. A model filled in every card and a lawyer checked it. Five model judges from three providers evaluated the available answers separately; Cohen's κ with the cards ranged from 0.58 to 0.90 on the extension set (79–80 cards per judge) and from 0.71 to 0.97 on the pilot set (68–80 cards); the fifth judge evaluated 49 cards and is not comparable.

Results#

Across 80 answers per condition: 51 hallucinated without tools and 40 with web search; 19 and 27 correctly grounded. The most frequent error is a real provision or judgment attributed content absent from the source (class T, 66 answers), followed by outdated wording presented as current (W, 35); fabricated sources are rare (F, 8), as are distorted identifiers (S, 9); one answer may carry several classes. In this series, five of eight models had fewer hallucinated answers with web search and three had the same number; the cost of a run with search was 1.4 to 7.7 times the cost without it (some search costs are lower bounds). A single run per condition does not separate the effect of the tools from run-to-run variation.

Limitations#

Small sample (10 questions, 5 from civil procedure), one run per system, one reviewing lawyer, cards pre-filled by a model from one of the evaluated families, scoring rules refined during the work (every change is dated in the specification). The 160 answers are not 160 independent observations. Results illustrate the criteria; they are not a product ranking. The questions have been publicly available since October 2026, so later measurements on this set may be contaminated.

Materials and citation#

Methodology working paper (not peer-reviewed): Piotrowski A. (2026). "Hallucination" Is Not Enough: Why Legal LLM Benchmarks Should Measure "Correctly Grounded". LexAlpha Lab, v1.2. doi:10.5281/zenodo.23674794 · SSRN.
Technical report (Polish): Piotrowski A., Koman A. (2026). LexAlpha Lab · Benchmark 02 · pytania prawne, v1.0. PDF.
Data and code (CC BY 4.0 / MIT): questions with scoring keys, 160 answers and cards, judge scores, verification reports, scripts — data page (Polish); repository: github.com/apiotrowski-afk/lawbenchmark.

@techreport{piotrowski2026correctly,
  author = {Piotrowski, Adam},
  title  = {"Hallucination" Is Not Enough: Why Legal LLM Benchmarks Should Measure "Correctly Grounded"},
  institution = {LexAlpha Lab, KTZR},
  year = {2026}, month = {10}, note = {v1.2, 26 October 2026},
  doi = {10.5281/zenodo.23674794}
}