On November 12, 2026, the ELTE CSS Institute for Legal Studies and the MTA/ELTE Algorithmic Constitutionalism Lendület/Momentum Research Group hosted a doctoral conference as the seventh event of the series Sustainable Use of Artificial Intelligence in Legal Work, now in its second season. The aim of the conference was to give the youngest generation of legal researchers a voice in the debate on how artificial intelligence (AI) is reshaping legal work, and to move that debate beyond ad hoc fixes towards an understanding of the deeper layers of the problem.
Following an open call that attracted applications from a wide range of universities, five doctoral students representing five Hungarian institutions presented their research: Lavinia Regina Simon (University of Szeged, Faculty of Law and Political Sciences), Mónika Morvay (University of Miskolc, Ferenc Deák Doctoral School of Law), Katalin Podonyi (Ludovika University of Public Service, Doctoral School of Public Administration Sciences), Renáta Papp (Károli Gáspár University of the Reformed Church, Doctoral School of Law and Political Sciences) and János Szabadkai (Pázmány Péter Catholic University, Doctoral School of Law and Political Sciences).
The conference was opened by Márton Varju, director of the Institute for Legal Studies. He recalled that the Institute became involved in AI regulation research through the former Artificial Intelligence National Laboratory, at a time, around 2020–2022, when the regulation of AI still seemed utopian. The adoption of the EU AI Act has simplified the picture for European lawyers, but the field remains in motion: in his words, the goalposts shift roughly every twelve months, so the target a doctoral thesis aims at today may have moved by the time it is completed. Varju also raised a methodological challenge that concerns doctoral candidates and research institutes alike: how to make research "AI-proof", that is, how to convince universities and reviewers that a dissertation reflects substantial human contribution at a time when generative AI can draft text, summarize data and build tables. Boldizsár Szentgáli-Tóth, principal investigator of the research group, added that supporting young researchers is a core aim of the Lendület programme. The conference was designed both as a platform for doctoral students from different universities and as an opportunity for the group to meet young researchers working on topics relevant to its research. He announced that the conference papers will be published in a Hungarian-language edited volume in 2027.
The first presentation, by Lavinia Regina Simon, asked under what conditions the use of AI in legal work can remain compatible with the requirements of the rule of law in the long term, and not only advantageous in technological or economic terms. Simon began with the physical side of sustainability. Citing the International Energy Agency's 2025 Energy and AI report, she noted that data centres consumed around 415 TWh of electricity in 2024, roughly 1.5% of global consumption, a figure projected to more than double to about 945 TWh by 2030. She illustrated the tension between efficiency gains and rising aggregate demand with four examples: an underwater data centre in Shanghai's Lingang area cooled by seawater and powered by offshore wind; research published in Applied Energy showing that 72% of China's computing capacity was located in severely water-stressed regions in 2022; Microsoft's closed-loop, chip-level liquid cooling design; and the IEA's forecasts on overall data centre demand. Her conclusion was that technological progress alone does not solve the problem of sustainability, and that AI in legal work likewise cannot be judged solely by speed or by the degree of automation.
Turning to legal and institutional sustainability, Simon argued that the measure of sustainable legal AI is whether, despite technological involvement, a legally relevant outcome remains verifiable, attributable to a human or institutional actor, contestable and correctable. Since summarizing a contract, searching for a statutory provision and constructing the substantive argument of a court brief carry very different risks, she rejected uniform oversight requirements for all AI uses. She distinguished technical explainability from legal justifiability: lawyers need not reconstruct the mathematics of a model, but they must be able to check the legal claims they adopt, which she called functional verifiability. She illustrated the point with Lnu v. Blanche (9th Cir., June 3, 2026), in which two attorneys were each fined USD 2,500 and suspended from practice before the court for six months, not for using AI as such, but for filing unverified briefs containing non-existent cases and misattributed quotations and for failing to be candid about the errors afterwards. After reviewing the relevant provisions of the AI Act (Article 4 on AI literacy, Article 14 on human oversight, and point 8 of Annex III on the administration of justice), she proposed a three-pillar model of sustainable legal AI. In this model, legal certainty is operationalized as verifiability, accountability as attributability and adaptivity as correctability, with contestability serving as a guarantee common to all three.
Mónika Morvay compared the EU's top-down, risk-based model with the bottom-up American approach, which relies on professional responsibility, bar self-regulation and judicial sanctions. She stressed that the two models are complementary rather than opposed. At the centre of the American model is the duty of technological competence, laid down in Comment 8 to Rule 1.1 of the ABA Model Rules of Professional Conduct since 2012 and applied to generative AI by ABA Formal Opinion 512 (2024). Morvay walked through the opinion's implications. Lawyers need not understand the algorithm in detail, but they must understand its basic functioning, its risks and its benefits. Clients must be informed of AI use when they ask or when it significantly shapes the litigation strategy. Time saved through AI may not be billed by the hour. The confidentiality risks of self-learning, open systems call for closed systems and internal information barriers. Supervising lawyers are responsible for the AI use of their staff. Among the US states, she contrasted Florida's soft, ethics-based approach (Ethics Opinion 24-1, which treats AI by analogy with non-lawyer assistants), California's statutory route (SB 574, signed in September 2026, which prohibits delegating the practice of law to generative AI), and New York's pragmatic court rules, under which the signature on a filing certifies that it is free of hallucinated authorities.
Morvay traced the evolution of sanctions through four cases: Mata v. Avianca (2023), Park v. Kim (2024), in which an appellate court held that no special rule is needed to remind lawyers not to make false statements, Kruse v. Karlen (2024), where a self-represented litigant relied on an online consultant whose brief contained 22 fictitious citations out of 24, and a 2026 case in which counsel, besides citing hallucinated authorities, misrepresented the record and refused to cooperate with the court. Building on the work of Richard Susskind and Andrew Perlman, she also highlighted the "five-year problem": if AI takes over the routine tasks through which junior lawyers have traditionally gained experience, the profession must find new ways to train practitioners capable of checking and overriding machine output. Perlman's view is that avoiding AI altogether may itself amount to a competence failure. Morvay concluded that the Hungarian Bar Association's rules should consider recognizing technological competence explicitly, adopt a New York-style certification for filings, and offer lawyers a self-assessment checklist on the use of open and closed systems, prompting, verification of citations and the review of AI-suggested strategies.
Katalin Podonyi examined chatbots in administrative information provision, one of the most dynamically developing areas of e-government. She described a two-stage development, from a simple change of channel (paper to electronic, with human case handling in the background) to automation and AI-driven solutions, and she proposed a typology of chatbots by decision logic. Rule-based systems, such as the former chatbot on magyarorszag.hu, are fully predictable but rigid. Generative, LLM-based systems are linguistically fluent but rely on statistical probability without genuine legal understanding. Agent-based hybrid systems, as used in Estonia, let the AI interpret natural-language queries while verified state back-end systems and authorized officials handle the substantive data processing. On the legal nature of chatbot answers, Podonyi drew on the general clause of the Code of Administrative Court Procedure, the uniformity practice of the Curia and the concept of authority in the General Administrative Procedure Act (Ákr.). She concluded that chatbot answers are not automatically communications attributable to the authority, but technical tools of official communication, distinct both from automated decision-making under the Ákr. and from formal service of a decision.
Podonyi identified four main procedural risks: erroneous or inconsistent answers, which can lead clients to miss deadlines or procedural steps; black-box opacity, which calls for multi-level transparency; algorithmic bias against non-native speakers and clients with lower digital or legal literacy, which raises equality concerns under the Fundamental Law; and the diffusion of responsibility among the many actors involved. The risk of technology, she argued, must never be shifted onto the client. She closed with two de lege ferenda proposals: amending the Ákr. rules on applications for excusing a missed deadline so that misinformation from an official chatbot is expressly recognized as a possible ground (without automatic acceptance), and requiring authentic, retrievable logs of chatbot interactions that affect clients' rights and obligations, in compliance with data protection and data minimization requirements.
Renáta Papp presented research conducted within the Ministry of the Interior's doctoral internship programme at the National Police Headquarters on the role of AI in crime prevention. She outlined three technological directions (facial recognition, intelligent video surveillance and predictive policing) and pointed to the algorithmic video surveillance deployed at the 2024 Paris Olympics, which expressly excluded biometric identification, as an example of bounded use. She warned that predictive algorithms trained on biased historical data may work as self-fulfilling prophecies. After summarizing the AI Act's prohibitions, its restrictions on real-time remote biometric identification and the high-risk status of most law enforcement applications, she noted that, under the July 2026 amendment, the detailed rules for Annex III high-risk systems will in principle apply from December 2, 2027. Her comparative overview covered the United Kingdom, France, the United States, and China.
The empirical part of Papp's research rested on 25 questionnaires completed by police leaders, investigators, IT specialists and data protection experts. Respondents reported that no autonomous AI-based decision-making without human intervention currently operates in the Hungarian police. They regarded AI almost unanimously as a decision-support tool and emphasized explainability, auditability and human control. Papp paid special attention to petty offence procedures, where facial image analysis is becoming an everyday tool: its use was extended in 2024 to offences punishable by confinement and in 2025 to all petty offences. According to the data she cited from the Hungarian Institute for Forensic Sciences, more than 1,700 facial image analysis requests were made in petty offence cases in the second half of 2024, more than 900 of which led to successful identification. In the first half of 2025, requests rose to more than 2,900, with more than 1,650 successful identifications. She stressed that in the Hungarian system the algorithm only suggests possible matches, which several independent experts then assess, and that this multi-level human control is the system's key safeguard.
The final presentation was given by János Szabadkai, who manages the Hungarian manufacturing subsidiary of an American multinational producing equipment for ultrapure water plants serving microchip production. He connected his doctoral research on green considerations in public finances with corporate practice. He explained why legal certainty (predictability, transparency and the protection of acquired rights) is a precondition for investment decisions, and how science-based environmental law and the logic of the circular economy put pressure on that stability. He then surveyed the corporate sustainability toolkit: the ISO 9001, 14001 and 45001 management standards, the voluntary EU Eco-Management and Audit Scheme (EMAS, Regulation (EC) No 1221/2009), the Corporate Sustainability Reporting Directive (CSRD), the Corporate Sustainability Due Diligence Directive (CSDDD) and the Hungarian ESG Act (Act CVIII of 2023). He illustrated the risks of greenwashing with the sanctions imposed on Deutsche Bank's asset manager DWS in the United States and Germany, the Dutch court ruling against KLM's "fly responsibly" campaign, and the climate litigation brought by Milieudefensie against Shell.
On AI, Szabadkai referred to Pope Francis's address to the G7 and to Pope Leo XIV's teaching on protecting the human person in the age of AI, and insisted that the final decision must remain with humans. In his view, AI can substantially ease the administrative burden of ESG and CSRD reporting, while also helping to forecast supply chain risks and track regulatory change.
The next event of the series will take place in January 2027.
_________________________________________________________
This report was prepared with the support of the ELTE CSS Algorithmic Constitutionalism Lendület/Momentum Research Group (LP2024-20/2024), funded by the Hungarian Academy of Sciences.
_________________________________________________________
The views expressed above belong to the speakers and do not necessarily represent the views of the Centre for Social Sciences.

Berkes Rudolf