Synthetic intelligence instruments are making waves throughout the authorized sector. ChatGPT particularly is all of the hype, with debates nonetheless ongoing about its potential position in courtrooms: from helping judges in drafting opinions, to legal professionals counting on AI-generated arguments, and even events submitting ChatGPT-produced proof. As a report on Verfassungsblog suggests, the “robotic decide” is already right here, with Colombian judges utilizing ChatGPT to put in writing full verdicts. Within the UK, attraction decide Lord Justice Birss described ChatGPT as “jolly helpful” for offering summaries of an space of legislation. In the meantime, in China, AI is embraced in courtrooms, with the “Little Sage” (小智) system dealing with total small claims proceedings.
Within the Netherlands, ChatGPT made its judicial entrance when a decrease courtroom decide precipitated controversy by counting on info offered by the chatbot in a neighbour dispute over photo voltaic panels. The incident triggered vital dialogue amongst Dutch legal professionals concerning the influence of AI on litigants’ rights, together with the suitable to be heard and celebration autonomy. Nevertheless, a number of different judgments have additionally talked about or mentioned ChatGPT.
On this weblog publish, I’ll have a look at all six revealed Dutch verdicts referencing ChatGPT use—both by the decide or by litigants—and discover whether or not there’s any frequent floor in how AI is approached. I can even sketch the EU-law context surrounding AI-use in courts, and contemplate the implications of this half-dozen rulings for present efforts by the Dutch judiciary to control using AI.
ChatGPT in Courtroom: Guarantees and Pitfalls
Earlier than delving into the particular judgments, it’s useful to know why ChatGPT is drawing a lot consideration in courtroom.
Authorized functions of AI should not new. For many years, the sector of AI and Regulation has researched the probabilities of so-called ‘knowledgeable techniques’, based mostly on logical fashions representing human authorized reasoning, to interchange sure components of authorized decision-making. Presently, such techniques are deployed on a big scale in, for instance, social safety and tax administration. Nevertheless, extra not too long ago, the data-driven strategy to authorized AI has precipitated a revolution. By combining giant datasets (Large Information) with machine studying methods, AI techniques can study from statistical correlations to make predictions. This allows them to foretell the danger of recidivism or use earlier case legislation to forecast outcomes in new instances.
Giant Language Fashions (LLMs) resembling ChatGPT observe related ideas, coaching on huge, internet-scraped textual datasets and deploying machine studying and pure language processing to foretell, in essence, essentially the most possible subsequent phrase in a sentence. ChatGPT can immediately generate responses to complicated questions, draft paperwork, and summarize huge quantities of authorized textual content. As a outstanding byproduct, it could seem to carry out fairly properly in authorized duties resembling analysis help, and it has even confirmed able to passing first-year American legislation exams.
But, these potentialities include dangers. ChatGPT can produce inaccurate solutions (so-called “hallucinations”) and lacks real-time entry to personal authorized databases. Analysis by Dahl et al. demonstrated that ChatGPT-4 generated inaccurate authorized info or sources in 43% of its responses. In a now-infamous incident, a New York lawyer was reprimanded after ChatGPT cited non-existent case legislation. Moreover, the know-how is akin to a black field: as a result of complicated nature of neural networks and the huge scale of coaching information, it’s usually tough—if not unattainable—to hint how particular outputs are generated. Lastly, bias can come up from incomplete or selective coaching information, resulting in stereotypical or prejudiced output. Over- or underrepresentation within the enter information impacts the system’s outcomes (rubbish in, rubbish out).
Regardless of these vital caveats, the next Dutch judgments present how AI is more and more making its look in courtrooms, probably shaping judicial discourse and apply. First, using ChatGPT by a decide is mentioned, adopted by the instances by which litigants used the chatbot.
ChatGPT in Motion: From the Bench to the Bar
A. Judicial Use Circumstances
1. Gelderland Courtroom (ECLI:NL:RBGEL:2024:3636)
On this neighbour dispute over rooftop building and diminished output from photo voltaic panels, the courtroom in Gelderland used the chatbot’s estimates to approximate damages. It held:
“The district courtroom, with the help of ChatGPT, estimates the common lifespan of photo voltaic panels put in in 2009 at 25 to 30 years; it subsequently places that lifespan right here at 27.5 years … Why it doesn’t undertake the quantity of € 13.963,20 proposed by the claimant in the principle motion has been sufficiently defined above and in footnotes 4–7.” (at 5.7)
The decide, once more counting on ChatGPT, additionally held that insulation materials thrown off the roof was now not usable, thus awarding damages and cleanup prices. (at 6.8)
B. Litigant Use Circumstances
2. The Hague Courtroom of Attraction (ECLI:NL:GHDHA:2024:711)
On this appellate tax case in regards to the official valuation of a lodge, the Courtroom of Attraction in The Hague addressed arguments derived from ChatGPT. The appellant had launched AI-generated textual content to contest the assessed worth, however the courtroom discovered the reference unpersuasive. It held:
“The arguments put ahead by the celebration that had been derived from ChatGPT don’t alter [the] conclusion, significantly as a result of it’s unclear which immediate was entered into ChatGPT.” (at 5.5)
3. The Hague Courtroom of Attraction (ECLI:NL:GHDHA:2024:1771)
In a tax dispute relating to the registration tax (BPM) on an imported Ferrari, the Courtroom of Attraction in The Hague rejected the taxpayer’s reliance on ChatGPT to determine appropriate comparable autos. The claimant had requested ChatGPT to record autos that shared an analogous financial context and aggressive place with the Ferrari in query, resulting in a number of ten luxurious automobiles. The courtroom explicitly dismissed this strategy, contemplating that whereas AI would possibly group autos based mostly on common financial context, this methodology doesn’t replicate what a median client (a human) would contemplate genuinely comparable. (at 5.1.4)
4. District Courtroom of The Hague (ECLI:NL:RBDHA:2024:18167)
On this asylum dispute, the claimant (an alleged Moroccan Hirak activist) argued that returning to Morocco would expose him to persecution as a result of authorities there routinely monitor protestors overseas. As proof of this state surveillance, his legal professional cited a ChatGPT response. The courtroom dismissed the argument as unfounded:
“That the claimant’s consultant refers on the listening to to a response from ChatGPT as proof is deemed inadequate by the courtroom. First, as a result of the claimant has not submitted the query posed nor ChatGPT’s reply. Furthermore, the legal professional admitted on the listening to that ChatGPT additionally didn’t present any supply references for the reply to the query posed.” (at 10.1)
5. Amsterdam District Courtroom (ECLI:NL:RBAMS:2025:326)
On this European Arrest Warrant case, the defence submitted a Polish-language report about jail circumstances in Tarnów, hoping to show systemic human rights violations. Nevertheless, on the listening to, counsel acknowledged utilizing ChatGPT to translate the report into Dutch, and the courtroom discovered that inadequate. Missing an official translation or a model from the issuing group itself, the courtroom dominated it couldn’t confirm the authenticity and reliability of the AI-generated translation. Because of this, the ChatGPT-based proof was dismissed and no additional questions had been posed to the Polish authorities in regards to the Tarnów jail. (at 5)
6. Council of State (ECLI:NL:RVS:2025:335)
In a dispute over a compensation declare for lowered property worth, the claimant (Aleto) tried to indicate that the knowledgeable’s use of sure comparable properties was flawed and submitted late-filed materials derived from ChatGPT. The Council of State dominated:
“In the course of the listening to, Aleto defined that it obtained this info by ChatGPT. Aleto didn’t submit the immediate. Moreover, the ChatGPT-generated info didn’t present any references for the reply given. The data additionally states that, for an correct evaluation, it might be advisable to seek the advice of a valuer with particular data of commercial websites within the northern Limburg area.” (at 9.2)
Double Requirements: AI-Use by the Courtroom vs. the Litigants
Easy methods to make sense of those rulings? Let’s begin with the ruling by which the decide used ChatGPT to estimate damages, which generated by far essentially the most debate. Critics of the ChatGPT use on this case argue that the decide primarily launched proof that the events had not mentioned, thereby operating counter to basic ideas of adversarial proceedings, resembling the suitable to be heard (audi et alteram partem), and to the requirement underneath Dutch civil legislation that judges base choices solely on information launched by the events or on so-called “information of common data” (Article 149 of the Dutch Code of Civil Process).
A comparability has been drawn to prior case legislation involving judges who independently searched the web (the so-called “Googling decide”). As a vital criterion, the Dutch Supreme Courtroom has dominated that events should, in precept, be given the chance to precise their views on internet-sourced proof. Nevertheless, others are much less vital of the decide’s ChatGPT-use, mentioning that judges have appreciable discretion in estimating damages underneath Article 6:97 of the Dutch Civil Code.
Wanting extra intently on the means the decide used ChatGPT on this case, it stays unclear whether or not the events had been afforded the chance to contest the AI-generated outcomes. Nor does the ruling specify what prompts had been entered or the exact solutions ChatGPT produced. That stands in stark distinction to the Colombian decide’s use of ChatGPT, described in Juan David Gutiérrez’s Verfassungsblog publish, the place the courtroom totally transcribed the chatbot’s responses, comprising 29% of the judgment’s textual content.
It additionally contrasts sarcastically to the judicial perspective towards litigants’ personal ChatGPT submissions. In truth, three causes have been cited by Dutch judges for rejecting ChatGPT-generated proof:
It’s unclear which immediate was entered into ChatGPT;
The claimant has not offered ChatGPT’s solutions;
The ChatGPT-generated textual content didn’t cite sources.
In different instances, ChatGPT use was dismissed as a result of ChatGPT’s views had been seen as incomparable to these of the common client, or as a result of the interpretation was deemed unreliable.
Returning to the Dutch decide’s use of ChatGPT, plainly it suffers from the exact same shortcomings courts have recognized as grounds for rejecting ChatGPT-based proof when launched by the events. Such double requirements, in my opinion, level to an pressing have to develop constant tips.
Rising Tips for Judicial AI-Use?
Though new steering paperwork on AI use by judges and legal professionals are rising in jurisdictions such because the EU, the UK, New Zealand, and the Flemish Area, they hardly ever spell out express necessities for introducing AI-generated content material in authorized proceedings, and as an alternative emphasize common ideas resembling acknowledging limitations of AI and sustaining attorney-client privilege. Against this, the Dutch case legislation to this point suggests not less than three components which may form greatest practices: (1) making certain the immediate is disclosed, (2) requiring that ChatGPT’s full solutions are shared, and (3) demanding correct references.
Whereas such necessities align with widely known ideas of transparency and reliability, they alone could not suffice. The Dutch response to the decide’s use of AI displays deeper issues about honest trial ensures and the significance of human oversight, which carry the matter into the area of constitutional significance. Consequently, when judges make use of LLMs to help in decision-making, they need to be vigilant as to the influence on events’ rights to be heard, to submit proof, and to problem proof.
Moreover, it is very important contemplate how the necessities of the EU AI Act come into play when LLMs resembling ChatGPT are utilized by judges and litigants. Annex III of the AI Act qualifies sure judicial functions of AI as ‘excessive threat’, triggering the AI Act’s most stringent obligations, such because the requirement for deployers to conduct a Basic Rights Influence Evaluation (Article 27), in addition to inter alia obligations relating to data-governance (Article 10), human oversight (Article 14), and transparency (Article 13). These obligations come into play with regard to:
“AI techniques supposed for use by a judicial authority or on their behalf to help a judicial authority in researching and decoding information and the legislation and in making use of the legislation to a concrete set of information, or for use in an analogous means in various dispute decision” (Annex III(8)(a)).
Because the corresponding recital notes, this high-risk qualification serves to “handle the dangers of potential biases, errors and opacity” (Rec. 61).
Nevertheless, the classification turns into problematic for AI-systems that serve greater than strictly authorized functions, like LLMs. LLMs like ChatGPT fall underneath the umbrella of general-purpose AI (GPAI) techniques, which means they’re extremely succesful and highly effective fashions that may be tailored into quite a lot of duties. In apply, the exact same LLM is likely to be employed to extend effectivity in organizational processes (in precept a low-risk utility), but even be employed to help judicial authorities in “researching and decoding information and the legislation”. Whether or not a given use case falls underneath ‘high-risk’ can subsequently be a high-quality line, although it issues tremendously for the relevant obligations.
Furthermore, even when AI use in courtroom wouldn’t meet the standards for high-risk classification, sure common provisions nonetheless apply, resembling Article 4 on AI literacy. The AI Act’s inclusion of guidelines for AI techniques able to inter alia textual content (Article 50(2) AI Act) additional complicates issues by imposing transparency necessities in a machine-readable format. It stays undefined how this transparency obligation would possibly function in, for instance, advert hoc use of ChatGPT by judges and litigants. In that regard, the transparency obligation is directed primarily at suppliers, who should implement technical options, quite than the customers themselves.
Lastly, with regard to privateness and information safety, the EU AI Act merely refers back to the common framework laid down by the Normal Information Safety Regulation ((EU) 2016/679), the ePrivacy Directive (2002/58/EC), and the Regulation Enforcement Directive ((EU) 2016/680). Nevertheless, the data-driven strategy inherent to machine-learning algorithms like ChatGPT—involving huge processing of (private) information—opens up a Pandora’s field of novel privateness dangers.
As Sartor et al. emphasize of their examine for the European Parliament, this strategy “could lead […] to an enormous assortment of private information about people, to the detriment of privateness.” But, from a privateness perspective, the AI Act doesn’t itself lay down a framework to mitigate these dangers, and it stays questionable whether or not current rules suffice. Particularly, exactly how the GDPR applies to AI-driven functions stays topic to ongoing steering by information safety authorities. It goes with out saying that, in delicate contexts resembling judicial settings, making certain compliance with the GDPR’s necessities is important for preserving public belief.
Prospects for a Dutch Judicial AI-Technique
On a concluding observe, the jurisprudence mentioned above carries direct implications for the Dutch judiciary’s forthcoming AI technique. The Council for the Judiciary’s 2024 Annual Plan mentions the event of such a technique by mid-2024, but no doc had been revealed by the top of that 12 months. In a letter to the Dutch Senate, the State Secretary for Digitalization and Kingdom Relations acknowledged that the judiciary intends to current its AI technique in early 2025; nevertheless, no such technique has been made public thus far.
What is obvious from the rulings thus far is that AI and LLMs are more and more discovering their means into courtrooms. But, the present state of affairs is way from supreme. It displays a fragmented patchwork of double requirements and authorized uncertainty regarding a know-how that intersects instantly with constitutional ensures resembling the suitable to a good trial, together with the suitable to be heard and equality of arms.
In mild of this, it appears very important that any overarching AI coverage be accompanied by clear and sensible tips for judicial AI use. Based mostly on the developments reviewed on this publish, three components seem particularly pressing:
1. Establishing applicable preconditions for AI use by each judges and litigants (thus avoiding double requirements whereas safeguarding events’ basic rights and due course of);
2. Introducing a transparent threat categorization consistent with the EU AI Act (benefiting from lower-risk functions whereas exercising warning with high-risk ones); and
3. Guaranteeing sturdy information privateness and safety.
Though it could be tempting to undertake AI instruments just because they’re “jolly helpful,” disregarding these ideas jeopardizes the belief and legitimacy wanted for a accountable integration of AI inside the judicial system. If AI is to develop into a part of the way forward for the courts, then the time to put down the principles is now.
D.G. (Daan) Kingma holds an LLB in Worldwide and European Regulation from the College of Groningen and a grasp’s in Asian Research from Leiden College (each research cum laude). He at the moment research authorized analysis (LLM) at Groningen College, specializing in the intersection between know-how, privateness, and EU digital regulation.
