The Singapore Law Gazette

AI and Professional Judgment: What Must Be Strengthened and Taught

When I was invited to write this contribution, the first thing that came to mind was how the past is always relevant to the future. Being “bilingual” in computer science and law for more than 30 years, I have watched and taught legal professionals in Australia and Singapore, and seen how technology has permeated legal research and legal practice. The current pace may seem furious, but relatively speaking, it is not. When I was finishing my 5-year double degree programme, I was concurrently the Computer Systems Manager at the Law School library. The personal computer had just come on the market, so legal research was for the first time on computers and CD-ROMs, not paper volumes.

Do we need a pause? AI systems and vendors will not pause, so it is the legal profession which must take stock and re-centre on what is important, what is essential. The explosion of Generative AI (“Gen AI”) in recent years has indeed impacted upon how lawyers work and make decisions. However, it is imperative to recognise both the transformative potential of AI systems, and to understand the inherent limits and boundaries of Gen-AI systems.

Professional Values and Ethics

As legal institutions overseas and the Ministry of Law in Singapore have categorically stated in their respective AI ethics Guides¹Singapore Ministry of Law, Guide for Using Generative AI in the Legal Sector (6 Mar 2026). https://www.mlaw.gov.sg/files/Guide_for_using_Generative_AI_in_the_Legal_Sector__Published_on_6_Mar_2026_.pdf and Practice Directions,²Federal Court of Australia, General Practice Note on the Use of Generative Artificial Intelligence (GPN‑AI), released on 16 April 2026. https://www.fedcourt.gov.au/law-and-practice/practice-documents/practice-notes/gpn-ai nothing about legal professional responsibility and ethical judgment has changed. Lawyers now more than ever must continue to be accountable and take responsibility for their work and to uphold integrity. Duties of confidentiality and data and information security also remain.

All the values of the legal profession remain intact and there is possibly now an even more urgent call for these values to be not just preserved, but strengthened. The intense commercialisation and globalisation of the profession mean values of integrity, accountability and independence of the profession may fade into vague memories, along with the moral, and ethical dimensions of the law. This must not happen.

Professional Skillsets

The essential skillsets of legal professionals remain the same and are more valuable than ever to ensure sound professional judgment at every stage of lawyering. What are these skills? These competencies can be categorised into fundamental skills and advanced, higher-order skills.

The fundamental skills would include the ability to read and understand well. Other fundamental skills would include the ability to write, and to write well in a logical, concise, precise and persuasive manner.

Some examples of higher order essential skills would include (advanced) analytical reasoning, higher order reasoning, rigorous critical thinking, comprehensive research expertise, ability to synthesise the vast amounts of facts and legal knowledge (what I call complexity navigation), strong written and verbal communication abilities, evaluating evidence, anticipating opposing arguments, ethical reasoning, empathy and adept negotiation capabilities. All these form the foundation of building the legal professional’s capacity to understand and interpret well, leading to sound legal judgment.

How many of these skills do the current Gen AI systems, or those in the immediate future, possess? I will answer this question based on how current AI systems function, not based on hype. This knowledge is critical.

AI Background

Back in the 1980s when I was a double-degree computer science and law student, my computer science classmates and I were building Knowledge-Based AI Systems as part of our Final Year Projects. I was also a Research Assistant under one of my law professors and we built an AI copyright expert system funded by a large external grant – the system allowed users, including non-lawyers, to determine if their creation was protected by copyright law. These AI systems were reliable and if programmed properly, did not contain errors. I strongly dislike the euphemism “hallucinations” because as lawyers, we should call a spade a spade, and these so called “hallucinations” are nothing more than errors and inaccuracies. Also, “hallucinations” is a term which unnecessarily anthropomorphises AI systems – a sticking point in AI ethics – because anthropomorphism readily gives children and those new to AI the impression that AI is somehow human, which is factually incorrect.

These Knowledge-Based AI Systems function by drawing on a Knowledge Base of facts and logic rules, that is verified by experts, and running it into an inference engine to deduce insights, analyse situations, derive conclusions and answers queries. The results are trustworthy and reliable if programmed accurately.

There are currently many different AI techniques, and many types of AI systems – the full range of which cannot be covered here due to space constraints. I will cover the most used and talked about in legal circles: text-based Gen AI. Most text-based Gen AI comprise Large Language Models (LLMS) which, as the name suggests, is focussed on constructing language. It does this through machine learning techniques, which depend largely on statistical pattern analysis of language.

Some readers may have come across a post by a lawyer using Gen AI that went viral on social media in early 2026. What I set out below may dissolve some of the “wow factor” in people’s minds, and lawyers should pause and consider professional conduct rules and confidentiality before feeding entire documents into any Gen AI, or allow any Gen AI to run loose on computer or cloud drives. One may think contractual clauses will cover legal liability, but when there is a data ingestion, data leak, Gen AI malfunction, or a cyber security breach, the reputational risk is still present, especially when commercially sensitive information and documents are sold on the dark web or worse, freely available. There are, of course, also issues of legal professional privilege.

Only earlier this year, in January and February, due to a programming error, the AI system of a big tech company had ingested, and summarised confidential files and emails of its subscribers when it was not supposed to. This is a clear breach of confidentiality and data protection laws in many jurisdictions. I am not privy to the clauses in the licence agreements, but I would not be surprised if there were disclaimer clauses built in, safeguarding the liability of the big tech company. If this is the case, this would leave legal professionals high and dry.

The Workings of Gen AI

Text-based Gen AI, like many computer programs and algorithms are excellent at processing data, including keyword searches, finding patterns on data, finding sequences on data, finding patterns on data, comparing data, matching data and combining data, and all at great speed, far faster than any human can eyeball.³Hannah YeeFen LIM, Autonomous Vehicles and the Law: Technology, Algorithms, and Ethics (Edward Elgar Publishing, UK Dec 2018) https://www.e-elgar.com/shop/gbp/autonomous-vehicles-and-the-law-9781788115100.html?srsltid=AfmBOoruoxd7hg_p0iVSPX53sZz1kkfci055-gfKEO5oo2q_62rJmNj- Add some useful statistical analysis and mathematics, and we have semblance of intelligence, but without true intelligence. Gen AI systems are often designed and used in combination with other AI and non-AI systems to boost its efficacy and use cases. Further, as the purpose of Gen AI is to produce text in the form of sentences, it is usually used to produce the results of the document processing.

Without getting bogged down with too much technical detail, in essence, many text-based Gen AI systems have been trained on millions of datasets, comprising vast amounts of text – over 45 Terabytes of text,⁴GPT-3, the predecessor to ChatGPT, was trained on nearly 45 terabytes of text data, see IEEE Transmitter, How Big will AI Models Get? https://transmitter.ieee.org/how-big-will-ai-models-get/ – so that if given a word like “thank”, it can statistically calculate that the probability of the next word should be “you” is extremely high. Similarly, if asked to finish the sentence “the cat sat on the”, the most probable word to finish the sentence would be “mat” because it has encountered this sentence many millions of times in the training datasets.

Immediately, we can already see bias and inaccuracy issues. If the training datasets are largely from a particular country (or replace this with discipline, subject matter, language etc), which uses particular legal terms with particular meanings, the Gen AI will have no idea about anything outside of the boundary. Which is why if you are a baker wanting to learn more about cookies, you would need to explicitly tell the Gen AI system that, otherwise, it might search its system based on webpage cookies, and tell you about how you can use cookies on your webpage.

Thus, Gen AI would be highly efficient and accurate at comparing documents and drafts of documents – think of the “compare” function in your computer software, and multiply that a hundred times in terms of speed. Thus, Gen AI can review hundreds and thousands of documents very quickly, and would be a useful tool for eDiscovery. It can identify relevant documents, duplicates, and based on statistics, identify key themes or words in the documents. Some systems use predictive coding technology or TAR (Technology Assisted Review) together with the Gen AI system to achieve more accurate results. Predictive coding is essentially a form of AI technique called supervised machine learning.

AI systems also perform strongly at extracting particular types of clauses in contracts, such as termination, indemnity, and liability limits. Due to its search and compare and statistical analysis abilities, it is efficient at flagging missing or unusual terms. Needless to say, it excels at comparing contracts against templates.

In short, AI and Gen AI systems in general excel at repetitive and structured or rule-based tasks, such as due diligence checklists and certain types of regulatory compliance, and when documents follow predictable formats.

Given the above, I was hardly surprised that Gen AI can pass Bar exams that comprise multiple choice and essay questions. That is what Gen AI has been trained on, to put sentences together that it has seen before. So far, no Gen AI has passed any of my hypothetical legal problem questions, and this is what legal educators need to focus on for assessment purposes.

Limits and Boundaries of Gen AI

Gen AI struggles with legal judgment and interpretation. Gen AI systems cannot reliably interpret ambiguous clauses, nor can they assess commercial or other risks, nor are they able to understand client objectives. Gen AI really does not understand law at all – its main task and ability is to predict language and the relationship between words.

Hence, Gen AI will miss subtle but crucial wording differences and can often misinterpret bespoke language and drafting. Gen AI may think a clause looks standard, but the clause could be highly detrimental within a particular context. Being poor in legal judgment, Gen AI systems will often fail to detect strategic issues, let alone be able to advise on strategies. In short, Gen AI would be unreliable in the provision of legal advice.

In order to demonstrate the workings and limits of Gen AI for a seminar I was presenting in March 2026, I had asked ChatGPT a question. I asked: “If pythagoras had failed to install a critical security update on his computer?”

The answer that ChatGPT gave was as follows:

“Then the Pythagorean theorem might have had a serious vulnerability. Hackers could exploit the system and rewrite it as a² + b² = c² + malware

Or worse – right triangles would start leaking the data from the hypotenuse.

Honestly, Pythagoras would’ve been forced to release a patch:

Version 2.0: Now with stronger proofs and fewer exploits in acute angles.

Moral of the story: always update your system… or your theorems might go irrational.”

It was clear ChatGPT had ingested information about the Pythagorean theorem and also about critical security updates on computers. As for true understanding or cognition, it had none because it spewed garbage about leaking data from the hypotenuse, amongst other non-sensical sentences. It was interesting too that it had been trained on trending words such as “Version 2.0”, possibly from “Web 2.0”, and it even emphasised it in bold, but it clearly had no understanding and no ability to reason.

This example is to demonstrate that on the whole, text-based Gen AI systems are designed for building language and sentences, not for legal interpretation, although it can largely imitate some semblance of understanding.

Legal Education and Training

Gen AI is fundamentally limited by its inability to exercise legal judgment because it does not reason like humans reason, making human lawyers and their learned skills indispensable. Thus, whilst Gen AI can perform some of the fundamental skills of reading by comparing documents, searching, and finding patterns and summarising and so on, Gen AI is clearly lacking in most of the advanced, higher order skills outlined above because it has no cognition. Whilst Gen AI can write, the language and style is often verbose, superfluous, bombastic, repetitive and is often not in any strong persuasive writing style, nor does it present arguments precisely, logically and strongly.

A legal education should impart far more than learning rules and the ability to apply rules; it fosters adaptability, as adaptable as the common law, critical insight, and human judgment – versatile and resilient skills that technology cannot replicate.

Does legal curriculum need to change? Yes. Do all law students need to learn how to code? Absolutely not. What is more essential is for law students and lawyers to understand at conceptual levels how the different AI technologies function, so they have better understanding of the abilities and limits of the technologies and apply the use cases appropriately – so, a form of AI-legal fluency. Without this skillset, lawyers may be mystified by AI or place unwarranted trust in its capabilities.

Lawyers must acquire a practical familiarity with various AI models, appreciating that AI encompasses a spectrum of technologies rather than a monolithic entity. It is thus vital for law schools to introduce students to diverse AI tools and promote analytical evaluation through critical comparison.

Pedagogy may also need to change. With AI at the fingertips of all students, rigorous critical thinking skills, and higher order complexity navigation and distillation skills need to be stressed, so that students can translate complex information into legal arguments, weigh precedent with nuanced contextual insights, navigate legal strategies, appreciate jurisdictional boundaries, present persuasive arguments, and so on. Seminar groups of 40 to 66 that many law schools had transitioned to for engendering learner engagement may no longer be effective to achieve this. Many years ago, I had the privilege of teaching Torts at NUS Law School when I was a full-time Visiting Professor there for four years. The tutorial groups I taught were around 12 to 15 in size with plenty of student interactions – there was really nowhere for them to hide! It is this kind of condition that is needed to ensure each student is able to develop higher order reasoning skills, so that they will be able to learn to see what is missing in the big picture, ask the right questions, and reach the correct conclusions or solutions. And perhaps this should be done in more than just the first year compulsory courses, so that students do not forget the core higher order skills as they move through law school.

Faculty staff workloads and headcounts need not increase; but much depends on the configuration of the set number of hours of each tutorial and lecture. I call this a modified Oxford-style of legal education for the age of AI.