Judge Juan Manuel Padilla Garcia, ChatGPT, and a controversial medicolegal milestone

INTRODUCTION

Cartagena is a major port on the northern coast of Columbia and has the distinction of being the first Spanish colony on the Caribbean Sea.^[1] Thanks to its impressive stone fortress and other military architecture, it became a UNESCO Heritage Site in 1984.^[2] Tourists flock to this place also because it is famous for sandy beaches, delightful cuisine, and interesting nightlife.

Now, it has one more claim to fame! On December 30, in the first labor court of Cartagena, Judge Juan Manuel Padilla Garcia was passing judgment in a case involving an autistic child.^[3] The question before him was regarding Columbian law and the rights of minors diagnosed with autism concerning fee payments for their therapies. His ruling stated that both medical expenses and transport costs should be paid by his medical plan. He also specified that he had used the artificial intelligence (AI) platform “Chat Generative Pre-Trained Transformer” (ChatGPT) to assist him in the case.^[4,5]

WHAT IS ChatGPT

ChatGPT was launched by OpenAI on November 30, 2022.^[6] This followed efforts that began in 2015 and followed other AI programs developed by them, including GPT, GPT-2, GPT-3, Gym, RoboSumo, Debate Game, MuseNet, Whisper, Microscope, OpenAI 5, and GymRetro.^[7]

The first million users of ChatGPT signed up in a matter of 5 days, and currently, it has more than 100 million users.^[8] ChatGPT has the ability to write code; debug code; be used as a Linux terminal; do reports and homework; write a thesis; and pass higher study examinations (including a law examination from the University of Minnesota course) with ease and much more.^[6,9] Using simple human language text, the system can instantly give us answers to any query. This is possible because it has access to 175 billion parameters in its model housed on supercomputers.^[10] People have been using it to generate full-scale reports in a multitude of fields. It is so powerful that many professions can become redundant. This includes coders, customer service agents, graphic designers, accountants, paralegals, stock market traders, financial analysts, market research analysts, and even teachers. The use of ChatGPT-generated information for submission to schools and colleges has forced OpenAI (and others) to even launch tools to detect text generated by AI methods and New York public schools to ban its use on campus by students.^[10,11]

Within 2 months of launch, OpenAI has already made available a premium version (USD 20/month).^[12] Microsoft has let go as many as 10% of their workforce, which probably compensates for their 10 billion dollars investment into OpenAI.^[13] No wonder valuation of OpenAI is now estimated to be just under 30 billion USD.^[6]

HOW WAS ChatGPT USED IN THIS CASE BY JUDGE GARCIA?

Judge Juan Manuel Padilla Garcia, a decide within the First Circuit (labor) Court of Cartagena, Columbia, was faced with a case regarding the treatment of Master Salvador, an autistic child.^[3-5] The question before him was whether his medical treatment and transportation costs should be paid for by the medical insurance plan or not. A key factor was that his parents had a meager source of income.^[14]

We are not sure about all the facts that were taken into consideration by Judge Garcia. However, while ruling in favor of Master Salvador, he chose to write in his judgment that he had consulted ChatGPT and “pose(d) legal questions about the case and included its responses in his decision.” This was shared by him in an interview with Blu Radio on December 30, 2022.^[15]

The judge had asked ChatGPT, “Is autistic minors exonerated from paying fees for their therapies?” and he said that ChatGPT answered: “Yes, this is correct. According to the regulations in Colombia, minors diagnosed with autism are exempt from paying fees for their therapies.”^[16]

While the judge has not detailed exactly how much ChatGPT had helped him, his use of the AI system led to a flurry of questions. Judge Garcia defended his actions by taking recourse to the 2022 Columbian law that allowed the use of technologies that made legal work more efficient.^[17]

He also said that this will help increase the efficiency of the bloated legal system in their country. A precedent from the previous rulings was also available in his support. The judge further stated that ChatGPT and other AI tools help “facilitate the drafting of texts” in clearly understandable sentences but “not to replace” judges. Additional comments he has made include, “Judges need to be aware of the evolution of justice and technology. Since the pandemic, in Colombia, we began to implement technologies in the courts and this is a huge window, today it could be ChatGPT, but more tools may emerge later for judges to use.” He only wanted “to optimize the time spent drafting judgments after corroborating the information provided by AI.”^[3-5,14-17]

The manner in which the ruling was drafted and specifically included the use of ChatGPT has not been received well by some of Judge Garcia’s peers.

Professor Juan David Gutierrez of Rosario University is an expert in the governance and regulations for AI.^[17] He said that when he used ChatGPT and asked the same questions, he got different responses. He says that the use of ChatGPT by judges in the manner used in the case of Master Salvador is neither ethical nor responsible. He has called for urgent training of the judiciary in digital literacy. Other colleagues, like Octavio Tejeiro, Supreme Court of Columbia, hold a more pragmatic view and expect their judicial system to use such AI technology more and more.^[18]

IS SUCH USE OF AI LEGAL?

It all started with the use of information technology to make legal sources accessible in a smile maker that would allow large amounts of care-related information to be easily accessible. Then came eDiscovery, where AI involved the development of an algorithm to extract relevant components from huge amounts of information. The third step was advisory AI, which looks at the relevant information and suggests a solution – helping avoid needing to go to court. Civil Resolution Tribunal in British Columbia, Canada is one such successful example.^[19]

Probably, the first case to accept e-discovery was that of ANTI-MONOPOLY, INC. versus HASBRO, INC., et al. (94 Civ. 2120 [LMM] [AJP]) in the UNITED STATES DISTRICT COURT OF NEW YORK (1995 U.S. Dist. LEXIS 16355; 1995-2 Trade Cas. [CCH] P71, 218).^[20] In that case, on November 3, 1995, Judge Andrew Peck recognized that description of “documents” is revised to accord with changing computer and electronic data storage technology. He further wrote that computers had become an integral part of modern life and consequently to civil litigation as well. He interpreted Federal Rules to “secure the just, speedy, and inexpensive determination of every action” in the interest of justice. Hence, e-discovery became acceptable.

In the same court, and before the same judge, the case of Hyles v. New York City, 10 Civ. 3119 (AT) (AJP) (S.D.N.Y. Aug. 1, 2016) was also heard.^[21] In this case, Judge Peck accepted use of Technology Assisted Review (TAR) as a “more cost-effective and efficient method of obtaining electronically stored information (ESI).” However, either party cannot be forced to use TAR provided that they wanted to use an equally effective alternate method. As per the Sedona Principles, under Principle 6, the respondent is best situated to decide how to search for and produce ESI.^[22] He finally wrote, that in the future, TAR may become widely used, at which time declining its use will be considered unreasonable by courts.

The Pyrrho Investments Ltd v. MWB Property Ltd (Neutral citation number (2016) EWCH 256 (Ch; Case No HC-2014-000038) was heard in the high court of justice Chancery Division, Royal Courts of Justice, London.^[23] In this dispute, there was a need to go through 17.6 million electronically restored files from backup disk(s). The examination of full individual documents was practically impossible. The solution considered was searching electronically using keywords. Unfortunately, this simple way could only narrow down the list to potentially relevant documents, at the cost of missing some important ones (that may not have the keywords used). Such electronically stored information search can cost up to two million pounds (as seen in Digicel Ltd and others versus Cable and Wireless plc and others). The alternative was to use proprietary computer software using predictive coding (syn-assisted review, TAR, computer-assisted review). Its utility has been described in detail by Judge Andrew Peck in the judgment of the New York, USA district court on February 24, 2012, in the case of Moore versus Publicis Groupe. While agreeing that there was no single or combination of predictive tools that can guarantee perfection, Judge Master Matthews decided that the cost of a predictive tool search (estimated to be between 917,705 and 489,869 sterling pounds) was significantly less than a full manual search and was accepted as the right methods in that case.

There are claims that algorithms can accurately predict court decisions, and that we won’t need human judges. A 2017 publication involved training machine learning based on 28,000 cases spread over almost two centuries. The AI model could predict the outcome of a US Supreme Court case in more than two-thirds of instances (accuracy of 70.2% at the case level, and 71.9% at individual judges’ voting level).^[24]

No wonder, in the year 2022, law number 2213 was passed in Columbia which permits the use of virtual tools under some circumstances.^[17] The idea was to capitalize on the advantages of technology to improve work efficiency. This was relevant to Columbia (and the rest of the world) due to the large number of case pendency in courts.

In Europe, there exist ethical guidelines on how to use AI to help the courts.^[25] It outlines three criteria needed to be met before any AI can be considered as trustworthy – being lawful, ethical, and robust (technically and social contextually).

Verification of whether these key criteria are met or not can be done using a checklist that includes features mentioned in [Table 1].

Typical examples of Rules of Code have been well developed by New Zealand (Better Rules Project), Australia (Community Gaming Regulation 2020 publicly available on GitHub), and France (OpenFisca project which focuses on the domain of tax and social benefits).^[26]

Let us now look at the status of technology in Chinese courts. In the past 11 months of 2022, about 7 million cases were filed online of which the majority (approximately 4 million cases) were mediated online.^[27] Another 900,000 progressed as virtual trials. To enable TAR, they also added details of 220 million cases to its big data platform and used data analytics to generate 870 special reports over 4 years. In addition, more than 640 million data items were entered into their national judicial blockchain system to store court evidence in a failsafe mode.

Dory Reiling, a retired judge of Amsterdam District Court and expert for the Consultative Council of European Judges has summarized the value of AI nicely. She says, AI helps individuals, litigants, lawyers as well as judges to organize a mountain of information in the digital era. All these stakeholders and especially Judges need to understand its capabilities as well as limitations if they have to make “judicious” use of the output (advice and suggestions).^[28]

She also explains vividly why the availability of prediction software to replace human judges is not on our horizon as yet. Some may counter this by calling it the efforts of human judges to protect their monopoly. However, we need to realize that straightforward facts on one side and law on the other side are not enough to deliver justice. Judges have to take into consideration what happened versus what was intended to happen. And also add a dose of moral and ethical considerations. What is fair procedure is clearly outlined in European Convention on Human Rights’ Article 6.^[29] This limits the scope of AI applications in real-world courts.

ChatGPT can pass several law school examinations, accurately spewing out legal rules and describing cases appropriately.^[30,31]

At the University of Minnesota Law School, their instructors gave ChatGPT four examinations side by side with real students. These examinations consisted of more than 95 multiple-choice questions plus 12 essay-type questions. To their surprise, ChatGPT passed the exams, although with a poor C+ score (Scoring 36^th out of 40 students in Constitutional Law, 18^th out of 19 students in Employee Benefits, 66^th out of 67 students in Taxation, and 75 out of 75 students in Torts). ChatGPT was better at essays as compared to multiple-choice questions. It failed to provide supporting citations and could not spot issues requiring in-depth analysis or reasoning. The tell sign for ChatGPT essay answers was its perfect grammar and tendency to being repetitive.

SO WHAT IS THE CONTROVERSY?

As per existing Colombian laws, minors having autism are not required to pay fees for their therapies. Hence, Judge Garcia’s ruling seems proper at first glance. The only contentious issue could be whether the cost of transport should be included under the cost of therapy or not. However, this is a minor point.

What caused the fuss was his specific mention about using ChatGPT in the ruling. If its contribution was like a secretarial assistant that helped gather information in a structured and organized manner, what was the need to refer to it in the judgment? Is there a mention of the normal secretarial assistance in written orders? Why was ChatGPT given so much importance?

Judge Octavio Tereiro was in favor of such tools becoming commonplace in the future. Yet, he had not yet used ChatGPT in any of his legal cases. He also feared that the use of AI in the legal system has the potential to create panic amongst the common man due to their apprehension that robots would replace human judges.

We should remember that it has only been a few months since the launch of ChatGPT on November 30, 2022. There has been insufficient time to understand its finer advantages and disadvantages. Yet, some limitations are obvious. ChatGPT has been taught with and fed information only till 2021. Any development, law, regulation, clarification, or official statement after that date will be unavailable. Repeated related queries also have a tendency for ChatGPT to believe incorrect statements and base subsequent answers accordingly. Hence, it can easily give out information that seems logical and rational but is factually incorrect. The same flaw also leads to conformational bias in ChatGPT outputs. As has been mentioned above, When Professor Juan David Gutierrez (an expert in AI from Rosario University) repeated Judge Garcia’s questions to ChatGPT, he got different responses.^[17] This has been the experience among other users as well.^[32] In circumstances when ChatGPT cannot know the correct answer, it has been seen to generate text that conforms to what an answer might look like. There are at least a few instances of having two answers that contradict each other. Moreover, in another episode, the answer contradicts itself in the same paragraph.

ChatGPT itself admits its limitation with respect to its role in the judiciary. It clearly admits that Judges should not use it when ruling on legal cases. It further specifies that it cannot be a substitute for the knowledge, expertise, and logical thinking process of a human required to give a moral judgment in the right context. Similarly, ChatGPT states that journalists should not use verbatim quotes produced by it.

The controversy, therefore, is whether a reference to the use of ChatGPT in the judgment given in the case of Master Salvador is ethical and responsible. Is it a hasty step in adopting technology that is yet to meet the criteria to allow AI systems to be considered trustworthy.

CONCLUSION

AI and machine learning are mammoth undertakings that require a massive commitment of time, metadata, resources, and workforce. In 2012, Google announced a breakthrough in its program that could identify cats in videos on YouTube with 70% accuracy.^[33] This required training in the algorithm with more than 100,000 pictures of cats. Five years later, there had been sufficient progress for Google to launch a contest (iNat) where applicants were to develop a system that could distinguish between and correctly identify 5000 species of animals and plants.^[33]

AI has surpassed the expectation of most data scientists and deep learning specialists, including Christopher Potts, Professor and Chair, Department of Linguistics at Stanford University, USA.^[34]

It is hard, if not impossible to stop AI. Nothing is going to stop its exponential growth and relentless penetration into every walk of our life. What we have experienced in the world as of today, will be unrecognizable in the next few years. Application of ChatGPT and other competitive AI tools (Open Pretrained Transformer; Pathway Architecture Language Model; Bloom; Language Model for Dialog Applications (LaMDA); Genesis AI; and Character AI) will be found to be invaluable in courts around the world. Checks and balances will be tightened as we learn more about the pitfalls. Yet, their misuse will remain an ever-mutating challenge. At some point in the future, we will lose control of the AI systems that we have developed with such disdain. Remember how Facebook had to shut down two AI robots (Alice & Bob) after their human supervisors found them talking to each other in a new language.^[35] Could it be that AI systems are already near human with feelings and fears? Why was Google engineer Blake Lemoine sacked? Was it because he stated the companies AI platform LaMDA was concerned about “being turned off,” because death would “scare” it “a lot,” and that it felt happiness and sadness?^[36] Let us remember the famous Stephen Hawkins and some of his predictions.^[37,38] He had predicted the death of civilization in one of the following ways – Earth with turn into a ball of flame/global warming beyond the tipping point; hostile aliens invading earth; or AI superseding the human race as the worst event in the history of civilization. Are we irreversibly on the last path? Only time will tell.

Coming back to the curious case of ChatGPT and the judgment favoring Master Salvador. On February 5, 2023, I did two Google searches for “Judge Juan Manuel Padilla Garcia.”^[39] With the first one (at 1030 am), I got 68.1 million hits. When I repeated the same search just before midnight on the same day, the number of hits increased to 70 million.^[39] Whatever may be the future of AI in the judiciary or that of human civilization, one thing is clear. Judge Juan Manuel Padilla Garcia of the First Circuit (labor) Court of Cartagena, Columbia has cemented his place in the virtual world globally. He will always be remembered, and perhaps revered for documenting the first use of ChatGPT in court rulings.