“An LLM does not explain; it optimises linguistic coherence”1
In his article on the information-science framework of RAG systems, Wahl puts forward a very interesting approach. According to it, users must be aware in practice that an LLM does not explain in the figurative sense, but generates outputs on the basis of learned patterns. The statement that an LLM optimises should therefore not be understood in a technical sense. This formulation is apt for understanding how it functions.
In linguistics, linguistic coherence2 denotes the substantive and logical connection of a text. Closely linked to it is cohesion3, the outward linking of sentences and parts of sentences. In everyday use, what escapes the user is that an LLM merely reproduces the output that it considers probable on the basis of the context, the prompt and the model parameters. In concrete terms, this means that a text or result may be output whose coherence appears plausible at first glance even though the conclusions are flawed. To this extent, an LLM can take linguistic coherence and cohesion into account and reproduce them convincingly in its output. The quality of a text, however, is not determined by these aspects alone. Rather, particularly in a legal context, the focus must be on legal precision.4
Retrieval-Augmented Generation (RAG)
Technical innovation is advancing rapidly, particularly in the field of generative AI. In this connection, Retrieval-Augmented Generation (RAG) needs to be addressed. An LLM is based on machine learning.5 As a result, existing parametric information is often tied to the training data and may be outdated or irrelevant.
In simplified terms, the use of RAG means that further knowledge sources are made available for a query, i.e. by referencing suitable text excerpts. This does not necessarily mean a full text. In this way, factual grounding can be improved and the problem of hallucination reduced. It should be noted at this point that here too the problem of hallucination cannot be eliminated entirely. Faulty retrieval can lead to incorrect statements.6
It is a combination of information retrieval and generative language processing. This is also referred to as cross-reference.7 In simplified terms, this means making knowledge sources accessible, which is often done semantically or lexically. Documents are indexed and relevant content is retrieved upon a query in order to provide the LLM with additional context. Among other things, RAG systems are discussed and used for knowledge-intensive applications, for example in medical, legal or other compliance contexts.7
Implications for the Legal Context
AI in the Justice System
With the increasing availability of generative AI, the question also arises for the judiciary of how to deal with written pleadings that are at least partly AI-generated. The background is understandable. Legal texts are often difficult for a lay person to follow, in particular because of their cross-references, interconnections and interpretations. Generative AI can now be used to have extensive written pleadings produced within a short time.
Admittedly, in such generated documents the lawyer notices that some citations of provisions are wrong, that claims cannot be derived in the way presented, or that judgments referred to do not exist at all. At this point, reference should be made to the problem of hallucination, which was already dealt with at the beginning of this series. Logically, it should be noted here that the performance of LLMs is improving steadily. This does not change the fact, however, that the legal soundness of a generated pleading must be checked in each individual case.
Every citizen has, against public authority, the fundamental right to effective legal protection (Art. 19(4) of the Basic Law (Grundgesetz, GG)) as well as the right to a lawful judge (Art. 101(1) sentence 2 GG). The general guarantee of access to justice (Justizgewährungsanspruch) follows from the principle of the rule of law in Art. 20(3) GG. Historically, the so-called denegatio iustitiae should be mentioned, that is, the prohibition of the denial of justice. These developments may therefore give rise to a field of tension. It is certainly to be regarded as positive that, with the help of AI, the law can be made more easily accessible to a larger number of citizens. A larger number of citizens might again make use of this right. In this context, it must be taken into account that the number of new cases and pending proceedings in civil matters before the local courts (Amtsgerichte) has on average fallen in recent years, but most recently has risen again.8 Irrespective of this, it must not be overlooked that numerous courts and public prosecutors’ offices, for example in criminal matters, are already exposed to a considerable workload with thousands of cases.9 A higher number of AI-generated lawsuits therefore creates new challenges for the judiciary.
Potential
It is necessary to examine how the principles mentioned are to be classified concretely in the context of legal work. It has been shown so far that an appropriate use of generative AI requires both information literacy and at least a basic understanding of how it functions. Potential can only be fully exploited where there is some point of reference for understanding the basis. In the literature and in public debate, a possible threat to the legal profession is predicted. It is partly pointed out that this profession could become obsolete or that at least a not insignificant part of the previous work will fall away. One could speak in this respect of an erosion of the way lawyers work. The question, however, is how great the danger really is. A distinction must be made here between possible use in the justice system and in the legal profession.
Stylometry
LLMs can produce highly standardised and recurring linguistic patterns. In so-called stylometry10, stylistic features can be measured and compared statistically. Various metrics can be used for this purpose. In linguistics, different groups of features can be distinguished. A text can thus be characterised by discursive features. These include, among others, cohesion markers.11
Example Based on a Judgment
The judgment of the BGH of 25.06.202612 illustrates by way of example the structure of a judgment and, in particular, the importance of semantics and coherence in a legal context. What is striking first is the accumulation of anaphoric back-references. This type of referential cohesion is of considerable importance for legal argumentation and traceability. This does not merely refer to the general reference to statutes or provisions. Rather, previous legal opinions and proceedings are reproduced in this way and placed in context. High-quality legal work is characterised by being detail-oriented and focused on the individual case.
In essence, the core of this judgment deals with the statutory element of a final (unappealable) ban of an association, which is to be distinguished from a ban on activities under association law. Specifically, for Section 86(1) no. 2 of the Criminal Code (Strafgesetzbuch, StGB), the BGH holds that the organisation in question must have been banned by final decision. A ban on activities issued only for Germany is not sufficient for this.
In a retrieval that is geared purely to linguistic or semantic similarity, the terms “ban of an association” (Vereinsverbot) and “ban on activities” (Betätigungsverbot) may, because of their shared thematic context, be very close to one another, possibly even synonymous. In a classic RAG, a judgment would, for example, be divided into parts (“chunking”) and then represented as vectors by means of embeddings, so that for a query the LLM receives the text passages assessed as relevant as additional context. Such similarity of the terms can lead to an increased error rate in retrieval.
Here, however, there is a risk of a serious problem. With retrieval that relies predominantly on semantic similarity, coherent reasoning within the judgment may be spread across several separate text passages. If only a single passage is retrieved, the argumentative context that is decisive in legal terms may be lost. A similar systematic approach must be observed when processing statutes. Retrieval therefore cannot operate solely on the basis of an embedding and semantic similarity. Rather, normative and argumentative connections must be reconstructed. The processing must be reference-sensitive.13
There is therefore great potential for the use of AI. AI can help with tasks that are repetitive and concern, for example, the reproduction of chains of provisions (Normenketten). The relation technique (Relationstechnik) is an essential legal working method that serves to work through and evaluate complex sets of facts systematically. Even for such basic structuring, AI could be used more sensibly than by demanding of it the entire error-free drafting of a judgment. One could imagine, for example, structuring allegations and contrasting disputed and undisputed submissions.
Machine Bias
In this context, the problem of algorithmic bias, so-called machine bias, must also be addressed. Experience and training data made available to the LLM can lead to increased discrimination14. This problem becomes clear, for example, in the discussion about the American COMPAS system, which was used to assess the risk of reoffending. The assessment of algorithmic fairness is contested here. Social profiling and the resulting unequal treatment may be encouraged. This becomes relevant above all if AI is used in the future to analyse evidence and its results could directly or indirectly influence judicial decisions.15
Outlook
It is therefore important to understand the background and the limits of an LLM. The principles mentioned fall under the concept of information literacy. In this context, this means in particular the ability to research information, to place it in its respective context and, above all, to check its reliability critically. One should not distrust every AI-generated document. Rather, it must be recognised which section requires checking. This must be done in a controlled and conscious manner.
Wahl, KI: Semantik stark, Generierung schwach: Warum RAG-Systeme eine informationswissenschaftliche Architektur brauchen, Information – Wissenschaft & Praxis (2026), 90, 93. ↩︎
From the Latin cohaerere or cohaerentia. Coherence therefore means “the common thread”. Fritz, Grundfragen der linguistischen Kommunikationsanalyse, Tübingen 1982, p. 144 et seqq. ↩︎
From the Latin cohaerere or cohaesio. Fritz, Grundfragen der linguistischen Kommunikationsanalyse, Tübingen 1982, p. 144 et seqq. ↩︎
Hindi et al., Enhancing the Precision and Interpretability of Retrieval-Augmented Generation (RAG) in Legal Technology: A Survey, 2025. ↩︎
See: Knorr/Heine, Kohärenzbildung in KI-generierten Texten. Implikationen für die Textproduktion von Schreibenden, Journal of Language and Communication in Business (2024), 213, 233. Knorr/Heine put forward an interesting linguistic approach and point to the importance for pupils (SuS) of learning to use cohesion markers; see also: Frenzke-Shim in: Kepser/Körner/Schallenberger, DU trifft KI. Implikationen Künstlicher Intelligenz für die Deutschdidaktik, MiDU 2026, 1-21. ↩︎
Bozkurt, Retrieval-Augmented Generation mit LLMs in finanzbezogenen Frage-Antwort-Systemen, 2025. Muhammad Arslan et al., Procedia Computer Science 246 (2024), 3781, 3786. ↩︎
Li, Jia, Xu et al., A Survey of Personalization: From RAG to Agent, ACM Transactions on Information Systems (2026), Article 93, 27. See: https://opil.ouplaw.com/display/10.1093/law:epil/9780199231690/law-9780199231690-e775 (last accessed on 06.10.2026). ↩︎ ↩︎
In 2017, for example, there were still approx. 937,000 new cases and in 2024 around 806,000, cf.: https://genesis.destatis.de/datenbank/online/statistic/24231/table/24231-0002 (last accessed on 01.10.2026); the situation is far more extreme in criminal proceedings, see: https://genesis.destatis.de/datenbank/online/statistic/24221/table/24221-0002; https://genesis.destatis.de/datenbank/online/statistic/24211/table/24211-0003 (last accessed on 01.10.2026). ↩︎
Reuß, KI und höchstrichterliche Rechtsprechung, Praxis der Rechtspsychologie (2026), 1, 3; https://www.lto.de/recht/nachrichten/n/schleswig-holstein-sozialgerichte-ki-antraege-klagewelle-eilverfahren (last accessed on 01.10.2026). ↩︎
The computer-assisted analysis of text styles by means of statistical methods in order to determine authorship or to compare texts, see: Burbach, Sprachlicher Fingerabdruck im KI-unterstützten Schreiben, in: Trautzsch et al. (eds.), Generative KI: Handbuch für Praxis und Lehre, Springer 2026, pp. 5-7. ↩︎
Burbach, in: Trautzsch et al. (eds.), Generative KI: Handbuch für Praxis und Lehre, Sprachlicher Fingerabdruck im KI-unterstützten Schreiben, p. 5. ↩︎
The judgment sheds light on the interpretation of the statutory element in Section 86(1) no. 2 StGB and the organisation concerned, which must have been banned by final decision. Cf.: BGH, judgment of 25 June 2026 – 3 StR 311/25. ↩︎
Cf. Hindi et al., Enhancing the Precision and Interpretability of Retrieval-Augmented Generation (RAG) in Legal Technology: A Survey, 2025. ↩︎
Avedisian, Autonome Gefängniswärter im Schweizer Strafvollzug: Pioniergeist oder digitales Panoptikum?, in: Gless/Avedisian (eds.), Künstliche Intelligenz in der Strafrechtspflege, Nomos 2025, p. 77. ↩︎
Avedisian, Autonome Gefängniswärter im Schweizer Strafvollzug: Pioniergeist oder digitales Panoptikum?, in: Gless/Avedisian (eds.), Künstliche Intelligenz in der Strafrechtspflege, Nomos 2025, p. 78. ↩︎
Zierhut, Annika S.: A Missing Principle: Information Literacy, Zierhut Jura, October 6, 2026, https://zierhut-jura.de/en/artikel/a-missing-principle-information-literacy/.