This article examines specific AI deployments in the legal systems of five jurisdictions – not concepts from white papers, but operational systems with names, dates and documented outcomes. For lawyers and regulators who have been observing this trend from a distance: the time for distance has just run out.
Estonia: the most widely repeated legal-technology myth of the decade
In March 2019, WIRED magazine published an article in which Ott Velsberg – Chief Data Officer of Estonia – mentioned the concept of an AI system to resolve minor civil disputes up to €7,000. Media outlets worldwide reported this as an accomplished fact: "Estonia deploys robot judge." The narrative was too good to fact-check – a digital nation, the firm Nortal, automation of the justice system. The story fitted perfectly with the technological enthusiasm of the era.
The problem: it never happened. The project never progressed beyond the conceptual stage. In February 2022, the Estonian Ministry of Justice officially denied the reports: "There hasn't been that kind of project or even an ambition." No Estonian procedural law was amended, no regulations were modified, and no adjudicating system was deployed. Nortal collaborated with the government on other e-governance projects, but not on building an "AI judge".
Lesson for the industry: The story of the Estonian "AI judge" is a cautionary tale about media hype in legal tech. A single article in a prestigious magazine, unverified by subsequent editorial teams, became a "fact" cited in academic papers, conference presentations and regulatory analyses worldwide. Before citing a foreign AI deployment in the judiciary – check the primary source.
China: Internet Courts and an AI judge conducting hearings in real time
China's Internet Courts (Hangzhou 2017, Beijing 2018, Guangzhou 2019) are considerably more advanced than most Western legal tech projects. In 2022, Chinese media promoted the concept of a "virtual judge" with an avatar interface – demonstrations and pilots of such AI interfaces exist (including virtual assistants guiding parties through procedure), but the operational reality is an "AI-assisted judging" model: the system supports the proceedings, while a human judge formally approves the final ruling. It is unclear, however, to what extent judges actually scrutinise the algorithm's suggestions.
The jurisdiction of these courts covers e-commerce disputes, online copyright infringement, disputes with streaming platforms and online loans. In 2021, the Beijing Internet Court handled over 300,000 cases. The entire process – from filing a claim via WeChat to service of the judgment – takes an average of 38 days instead of the standard 6 months in ordinary courts [TO BE VERIFIED – data from China Internet Court Annual Report 2021].
- The Smart Court (智慧法院) system now covers more than 3,500 ordinary courts across China
- AI assists with identity verification, hearing transcription and precedent retrieval
- Since 2023, the Supreme People's Court requires that judgments produced with AI involvement be clearly indicated in the reasoning
- Governing legislation: Regulations on Internet Courts 2018 (最高人民法院关于互联网法院审理案件若干问题的规定) – Articles 1–16
The scale is incomparable with any Western project. But scale also masks a problem: the Chinese system does not have an independent judiciary in the European sense. AI in Chinese courts operates within a system in which the Party exercises institutional oversight. Importing this model into democratic legal orders requires more than importing software.
USA: from hallucinations in a brief to a USD 5,000 fine – and lessons the industry should not ignore
Mata v. Avianca (S.D.N.Y., case no. 22-cv-01461, judgment of Judge P. Kevin Castel of 22 June 2023) is the most high-profile AI disaster in the history of the American bar. The firm Levidow, Levidow & Oberman filed a pleading containing citations from six judgments – all generated by ChatGPT, none of them existing in reality. The court imposed a fine of USD 5,000 on the firm and individually on two attorneys. More significantly, Judge Castel ordered that a copy of the judgment be sent to every court in which those lawyers had active cases.
This was not an isolated incident. In 2023, similar cases were recorded in courts in Texas, Georgia and British Columbia (Canada). The American Bar Association issued Formal Opinion 512 on 29 July 2024 – the first nationwide binding ethical guidance on the use of generative AI by lawyers (previously, in 2023, opinions had been issued only by individual states: California, Florida, Texas). The key requirement: every AI output must be independently verified before submission to a court. Using AI without verification may constitute a breach of ABA Model Rule 3.3 (candor toward the tribunal).
Expert Tip: Several federal district courts (including S.D. Tex. and N.D. Tex.) have already introduced mandatory AI disclosure statements in pleadings – so-called AI disclosure orders. A lawyer swears under oath whether they used AI tools and whether they verified the output. This is a standard that is spreading rapidly.
At the other end of the spectrum: tools such as Harvey (used by Allen & Overy, and since 2024 by PwC Legal), Casetext CoCounsel (acquired by Thomson Reuters for USD 650 million in 2023) and Lexis+ AI operate on closed, verified case law databases – without the risk of hallucinations about non-existent cases. The difference between a general-purpose LLM and a legal tool built on verified datasets is fundamental and is frequently overlooked in discussions about AI in law.
United Kingdom: HMCTS, court automation and the AI project in the Crown Prosecution Service
The UK's HM Courts & Tribunals Service (HMCTS) Reform Programme, launched in 2016 with a budget of £1.3 billion, is the largest justice digitisation project in Europe. By 2023, proceedings had been digitised for small claims (Money Claims Online), uncontested divorces (following the entry into force of the Divorce, Dissolution and Separation Act 2020) and immigration tribunal appeals.
AI appears here specifically in two areas. First: the Crown Prosecution Service has been testing an AI system since 2022 for the preliminary analysis of evidence in cybercrime cases – processing data from electronic devices, a task that previously took analysts weeks. Second: the Judicial Office published in December 2023 its Guidance for Judicial Office Holders on Artificial Intelligence – the first official guidance for judges in the UK on the use of AI. The document explicitly prohibits the use of public chatbots to analyse case facts or draft rulings.
- Online Court for claims up to £25,000: full digitisation without mandatory attendance
- AI in the CPS: preliminary analysis of digital materials – reducing analysis time from 3 weeks to 48 hours [TO BE VERIFIED]
- The Law Society published its Technology and the Law Policy Commission Report in 2023 with regulatory recommendations for AI in law firms
- The Solicitors Regulation Authority (SRA) has not yet issued binding standards for AI – as of Q1 2025
India: SUPACE and 50 million pending cases – AI as a lifeline for an overburdened justice system
India has approximately 50 million pending court cases – one of the largest judicial backlogs in the world. The Supreme Court of India launched SUPACE (Supreme Court Portal for Assistance in Court Efficiency) in 2021 – an AI system that does not issue rulings, but processes procedural documents and prepares case summaries, extracts from case law and factual analyses for judges.
This was a deliberate design decision: the creators of SUPACE defined the role of AI from the outset as a judge's assistant, not a decision-maker. Chief Justice N.V. Ramana, inaugurating the system, stated plainly: 'AI provides the facts; the judge provides the justice.' The system operates in English and Hindi; work is under way to extend it to India's 22 constitutional languages – which is itself an engineering challenge without precedent.
Alongside SUPACE, SUVAS (Supreme Court Vidhik Anuvaad Software) operates as a tool for translating judgments from English into regional languages, directly implementing Article 348 of the Constitution of India with regard to access to justice. This is an example of AI solving a specific constitutional problem rather than searching for a problem to solve.
Expert Tip: The Indian model – AI as a tool supporting access to justice rather than replacing the judge – is probably the most transferable template for developing countries and for Polish discussions about the digitisation of the judiciary. It requires no constitutional amendments and operates within the existing constitutional order.
graph LR
A[ESTONIA\nMedia myth\nconcept 2019] -->|never deployed| A2[Ministry denial 2022]
B[CHINA\nAI assists\nproceedings] -->|human approves| B2[Human judge]
C[USA\nAI assists\nlawyers] -->|verification mandatory| C2[Lawyer accountable]
D[UK\nAI processes\nevidence / case summaries] -->|decision rests with| D2[Prosecutor / Judge]
E[INDIA\nAI translates\nand summarises files] -->|judge decides| E2[Human judge]
style A fill:#ef4444,color:#fff
style B fill:#FF9800,color:#fff
style C fill:#2196F3,color:#fff
style D fill:#9C27B0,color:#fff
style E fill:#009688,color:#fff
What this means for Polish lawyers and regulators: five operational conclusions
Poland has not yet deployed AI in adjudication at any of the levels described above. The Ministry of Justice is pursuing digitisation (e-Delivery, the Information Portal), but this is procedural infrastructure, not a decision-making system. Meanwhile, Polish law firms – particularly those serving foreign clients – are already using Harvey, Casetext or Lexis AI for analyses of foreign law. The absence of domestic regulation does not mean the absence of risk.
- 1. ABA Opinion 512 and the UK Judicial Office guidance are not Polish law, but they set a standard of care that Polish courts may take into account when assessing malpractice
- 2. The AI Act (EU Regulation 2024/1689, applicable from 2 August 2026) classifies AI systems assisting the administration of justice as high-risk (Annex III, point 8) – entailing obligations of conformity assessment, registration and human oversight
- 3. AI systems used by Polish law firms for legal analysis are already subject to GDPR (processing of client data) and the rules on attorney-client privilege and legal professional secrecy
- 4. AI hallucinations in pleadings filed by Polish lawyers may constitute a breach of Article 8 of the Code of Ethics for Legal Advisers (integrity) – there is no disciplinary case law yet, but the risk is real
- 5. Public procurement for AI systems for courts must comply with the requirements of the AI Act and Directive 2014/24/EU on public procurement – IT departments of ministries should be factoring this into their plans now
The AI Act is being phased in: prohibitions on unacceptable-risk systems have applied since 2 February 2025. Provisions concerning high-risk systems, including those used by the justice system, will be enforceable from 2 August 2026. This is not a distant future – it is the planning horizon for every institution currently considering a pilot.
timeline
title AI Act Timeline for Legal Systems
2024-08-01 : AI Act enters into force
2025-02-02 : Prohibition of unacceptable-risk systems
2025-08-02 : Provisions on GPAI models
2026-08-02 : Full application - high-risk systems in the administration of justice