AI is no longer a new concept in the world of justice. Many law firms already have internal AI tools, and large legal databases now offer their own AI tools. State courts in Germany are also pushing ahead with the digitisation of civil proceedings, as demonstrated by a resolution of the Conference of Ministers of Justice in June 2025 and the adoption of a “Joint Declaration on the Use of Artificial Intelligence in the Justice System”.

AI applications can contribute to the effective and rapid processing of legal cases in the context of legal practice. For example, AI tools can be used as analysis and document management systems. Modern AI systems can analyse large quantities of documents in a very short time and extract relevant information.

The use of AI in document analysis appears to be unproblematic in principle. However, all AI tools have their respective risks and limitations in the context of legal practice. Not all AI tools are suitable for each individual case, and hallucinations are a problem. The following case example is based on the use of one of the currently available AI tools.

AI to support evidence presentation – Case example

A current case concerns a claim relating to several thousand damaged items. The client has documented the damages individually in damage reports and documented it with photos. Accordingly, the damage reports are correspondingly extensive.

How best to review thousands of documents?

In this context, the question arises as to how such a large number of damaged items should be dealt with in view of legal proceedings and the presentation of evidence. In terms of the burden of proof according to German law, the claimant would in principle have to prove precisely every instance of damage to each individual item. In return, however, the court would also have to examine the relevant evidence in detail in order to be able to pass a judicial judgement. In view of the sheer volume of damaged items, this proves to be extremely time-consuming and inefficient in practice, making the use of AI an obvious choice.

The damage reports can be uploaded to the AI platform internally, i.e. without any data protection concerns. Questions can be asked of the AI with regard to the uploaded documents, i.e. the damage reports. In principle, AI is able to scan all files and present the information in a collated form.

Not all the information contained in the documents can always be filtered out. For example, each document lists the damaged items with ID numbers. However, if AI is asked, for example, for the total sum of all items listed in the damage reports with the ID number, the AI can only filter out this information for a small number of items, even though the information is actually contained in every document. Incidentally, in addition to the list of damaged items with their ID numbers, each document also contains the total number of items recorded in the damage report.

Certain AI tools reach their limits when it comes to image analysis. The damage reports contain all images that visually document the damage to the individual items. However, according to AI, the images are only available as graphics embedded in PDF files. As a text extractor, AI can only recognize text modules and not image content in such a “mixed” PDF file consisting of text and images. AI can only analyze the damage or categorize it based on a separate table which lists all damaged items again with an additional description of the type of damage.

Review by a lawyer is still essential

In view of the increasing relevance of AI for legal practice, the German Federal Bar Association (“BRAK”) published a corresponding AI guide with information on the use of artificial intelligence. In these guidelines, the BRAK emphasizes the need for lawyers themselves to carefully review the results generated by AI in order to avoid errors and the resulting liability consequences.

The German Federal Lawyers’ Act (“BRAO”) standardizes the obligation of a German Rechtsanwalt to exercise their profession conscientiously. Particularly relevant here is the principle of highly personal service provision, which states that a lawyer must perform its work independently and, in case of doubt, personally. Consequently, the use of AI systems may not replace the work of a lawyer, but may only support it. It is undisputed that independent review and final control of the AI results by the lawyer is necessary.

Standards of care and confidentiality

The specific duties of care increase in their requirements when dealing with AI with the degree of automation and the intended use. For example, a higher standard of care applies if AI tools are used not only to support internal workflows but also in relation to clients (e.g. in automated communication with clients, auto-responders, use of chatbots for client intake, etc.).

Implications for the courts

With the increasing use of AI in legal practice, the question arises for the courts as to the extent to which judges can trust an AI mechanism in their evaluations.

In recent years, various automated programs have already been tested and used in numerous German courts, primarily to provide support and relief in mass proceedings. Examples include the OberLandesGerichts-Assistent (“OLGA”), the Massenverfahrens-Assistenz durch Künstliche Intelligenz (“MAKI”) and Codefy.

In view of the difficulties that arise when assessing the admissibility of AI tools, it seems sensible from the perspective of legal certainty for the legislator to include specific provisions on permissible areas of application in the Code of Civil Procedure, taking into account fundamental procedural rights.

It seems convincing to apply the principles of expert evidence (“Sachverständigengutachten”). The judge must be able to understand how the AI mechanism works (in order to rely on its results without reviewing the documents themselves). The main problem is that both the functioning and the results of the relevant systems are not comprehensible to users. The computer works like a “black box” that is fed with data and ultimately produces a result without it being possible to see what happens in between. If its process cannot be comprehensively understood from the outside (“black box AI”), but its use is nevertheless necessary, the expert’s duties are limited to the design, control and comprehensible presentation of the process and its prerequisites and consequences, so that the court can fulfil its task of evaluating and classifying the expert opinion and, in this respect, assess the evidence. Conversely, the use of AI below this threshold is generally permissible, but may then require disclosure and transparency.

It should also be in line with general quality standards to check the results obtained with the help of AI systems for their accuracy or at least plausibility. Here, too, no universally applicable standards have yet been developed. But insofar as it is part of good scientific practice to explain how results are obtained and thus make them comprehensible, this should also apply to the use of AI – or even more so, given the risk of “hallucination” in large language models.

Conclusion

Despite the existing challenges, AI offers considerable potential for processing tasks more quickly and effectively in the context of legal practice. Nevertheless, further developments are still needed, and professionals are always obliged to review and critically question the results produced by AI. In view of recent developments, it is to be expected that the initial problems with AI mentioned above will be resolved in the future and that risks will be further minimized as developments in the field of AI progress.