Lecture Sidelights: Part 1 | Computers Can Sort Peanuts, But What About Sentencing and Verdicts?

In 2016, during the Wisconsin case State v. Loomis in the United States, the judge used the AI risk assessment tool COMPAS as a basis for sentencing, rating the defendant's risk of recidivism as "high risk" and sentencing him to six years in prison. The defendant challenged the accuracy and bias of the COMPAS system's assessment, while the Supreme Court accepted the high accuracy of the AI tool. This case sparked a heated discussion on the "legitimacy and risks of AI intervention in judicial decision-making."

As AI technology is gradually applied to various industries and public decision-making, whether AI's estimations, recommendations, and decisions are fairer and less biased than humans is worth exploring in depth. Perhaps we should question: Are algorithms really fairer and more neutral than humans? When the data itself is biased, or the algorithm is overly a black box, will the judgments or recommendations made by AI simply be another form of reproduced bias?

In this lecture, Cui Jia-wei, a member of the JRF Digital Rule of Law Working Group, analyzed "Why sentencing and verdicts cannot be completely delegated to machines?" and further explored "The differences between legal data and general 'big data,'" and how we should view the hidden power structures and human rights risks behind AI. We also invited attorney Tsai Mon-han, Secretary-General of the Changhua Bar Association, to serve as the moderator for this session, delving into the impact of artificial intelligence development and use on social contexts and human rights.

When Verdicts Are Handed Over to Algorithms: Can the Judiciary Still Be Trusted?

By | Su Kuan-jung | Intern at JRF Taichung Office, Department of Social Work, Tunghai University

This lecture mainly discussed the legitimacy and risks of AI intervention in judicial decision-making, prompting everyone to think about the institutional and ethical challenges that may arise when artificial intelligence enters judicial judgments.

At the very beginning of the lecture, the speaker asked everyone to imagine if they were a judge, with desks piled high with criminal, civil, domestic, and labor litigation case files every day, and performance evaluations tied tightly to case-closure volumes. As long as tasks remain uncompleted, next year's promotion would be affected. The overwhelming burden of cases leaves people breathless. When the speaker further mentioned, "If there were an AI-assisted system capable of organizing case files for judges, comparing precedents, and even generating draft judgments, saving 25% of judgment time, would you use it?" After weighing the pros and cons, I found myself unable to reach a conclusion in a short time. Although efficiency gains are undoubtedly important, if AI makes mistakes, will due fairness be sacrificed? Later, the audience also discussed what an imagined AI judge would look like. In my view, an AI judge would tend to be a cold robot capable of absolute rationality and rapid analysis of massive cases. While it could handle cases impartially, it would also carry a sense of detachment because, lacking human emotions, it is unable to understand or empathize.

Before the lecture, the speaker engaged with the audience to prompt reflection, gradually leading into the lecture's theme. Before entering the main topic, the speaker cited the U.S. case State v. Loomis as an example. The court used the COMPAS algorithm to assess the defendant's risk of recidivism during sentencing. The system determined that Loomis had a high risk of recidivism, which influenced the sentencing result. However, Loomis argued that the algorithm was a black box, unverifiable, and potentially discriminatory, thereby violating due process. The U.S. court ultimately ruled that using COMPAS was not unconstitutional, but it could not be used as the sole evidence and had to be accompanied by limitations and warnings. Aside from the judiciary, Taiwan's National Audit Office also used AI to uncover fraud cases. Civil servants used ChatGPT and Python to screen two million judgments, and then utilized image comparison to find duplicate inspection photos, successfully exposing 16 fraudulent cases with a hit rate as high as 90 percent. The New Taipei City Government even decided to institutionalize this technology, introducing automated AI verification in the future. This case illustrates that AI can solve manpower shortages while improving efficiency and transparency. However, as we rely increasingly on AI, does it also mean we slowly lose our capacity for thought and observation? Just like when I sometimes use AI to organize reports, although it allows for quick synthesis, compared to completing a report entirely on my own, using AI assistance makes me less familiar with the content and tempts me to be lazy and skip reviewing because I assume computer tools never make mistakes. This can lead to errors or biases and make me increasingly reliant on online tools.

In the middle of the lecture, the speaker also summarized the pros and cons of AI in the judiciary. Proponents believe that AI does not tire, is unaffected by human emotions, can process large amounts of data quickly and efficiently, and reduces inconsistencies in sentencing—truly a great method for overworked judicial personnel. Opponents, on the other hand, argue that the operation of AI is a black box lacking transparency, that it is easily influenced by biases inherent in the data itself, and that it may even overlook the uniqueness of individual cases. Attribution of responsibility is also a major challenge: when AI makes an erroneous judgment, who should be held accountable? This also reminded me that judges often need to make judgments based on the defendant's life background and circumstances, sometimes adjusting sentences based on the defendant's remorse or family situation. If everything is handed over to AI, will these humanized factors be overlooked? The speaker subsequently raised four dimensions of consideration: technological transparency, institutional regulation, personnel literacy, and social trust. The speaker believes that the ultimate key to AI's application in the judiciary is not technology, but institutional design and the establishment of trust. The speaker also specially emphasized that AI does not truly understand cases; it merely calculates text associations through mathematical vectors, and the data fed into it comes from humans, meaning background information is not fed in. Since all data originates from humans, artificial intelligence inevitably inherits the results of human biases, discrimination, and cognitive errors.

At the end of the lecture, turning to the topics of trust and transparency, trust in the judicial system is the foundation of democratic operations. If transparency is lacking, rushing to introduce AI will only provoke more anxiety. Only by making the system open and reviewable by the public can trust be genuinely established. Perhaps artificial intelligence far surpasses humans in its ability to process large volumes of repetitive and homogeneous data, but it falls far short of humans when dealing with complex situations and life experiences. My biggest takeaway from the entire lecture was understanding the potential and limitations of AI in the judiciary. AI can indeed improve efficiency and reduce human burdens, but at the same time, it may magnify biases. The value of the judiciary lies not just in legal provisions, but in upholding human nature and social justice. AI can assist, but it should not dominate.

---

When the Judiciary Meets AI: Training Wheels for Efficiency, or a Risk to Justice?

By | Tien Tzu-an | Intern at JRF Taichung Office, Department of Psychology and Department of Social Work, Chung Shan Medical University

AI and algorithms have long quietly infiltrated our daily lives. Voice assistants on our phones help schedule itineraries, streaming platforms recommend TV shows we want to watch based on algorithms, and even when we mention wanting to buy something in daily conversations, relevant ads start appearing on our social media platforms. AI acts like an invisible infrastructure, quietly changing our living habits and methods. As we gradually grow accustomed to relying on AI to write homework, organize reports, and even draft articles, it is not hard to imagine that introducing AI into the legal field is merely a matter of time. However, unlike everyday entertainment or homework, judicial verdicts bear freedom and rights. Once artificial intelligence is allowed to participate, the question is no longer just about "convenience," but whether justice can stand firm.

Try to imagine: if you were a judge, with mountains of case files piled high on your desk, stacked so deep you couldn't see the surface. Performance and promotions are tied to the volume of closed cases; the moment there is a delay, the next year's review is immediately affected. Verdicts are not cold numbers; they concern the future of families, individual freedom, and even the life or death of enterprises. Yet in reality, fatigue and time pressure make this almost unbearable. When the judicial system introduces artificial intelligence to assist in adjudication, such a dilemma seems to find a crack of light. The new system can organize case files, compare precedents, and even generate preliminary draft judgments, reportedly saving a quarter of the time. Translated into reality, this means no more working overtime every night, finding time for continuing education, and finally being able to spend weekends with family. The boost in efficiency sounds alluring, but it is accompanied by another kind of anxiety: if artificial intelligence makes a mistake, will it cause innocent people to be wronged? Can litigants trust that their cases are still being adjudicated by a truly attentive judge rather than a cold algorithm?

Such contradictions have long had precedents. The speaker mentioned the U.S. case of Eric Loomis, where the court used the artificial intelligence COMPAS as a sentencing basis to determine that the defendant was a "high risk of recidivism." The defendant argued that the system was a black box, opaque, and potentially discriminatory, thereby violating due process. Ultimately, the court ruled that using the tool was not unconstitutional, but it could not be used as the sole basis and had to be accompanied by warnings and limitations. Such a ruling seems to say: artificial intelligence can be an assistant, but it cannot dominate everything. The author has always maintained reservations about AI; while not denying the convenience and speed AI can bring, can AI achieve 100% understanding when dealing with complex human emotions or life experiences? Furthermore, in many gray areas that are not strictly black-and-white, we rely on human reason and empathy to handle matters. When we speak of "empathy, reason, and law" , "empathy" refers to human touch, sympathy, and conscience—considering others' situations or feelings by putting oneself in their shoes; "reason" refers to the logic of matters and the judgment of right and wrong, which is the essence or logic of an issue; and finally "law", which refers to state-formulated legal rules and norms for handling events. Will this concept of empathy, reason, and law cease to exist when encountering AI? Addressing the author's doubts, the speaker ultimately explained how AI operates.

The speaker went on to mention that the temptation of efficiency is not felt by the judicial system alone. In recent years, auditing units have used AI to uncover procurement fraud, successfully identifying duplicate inspection photos within ten minutes with a hit rate close to 90 percent. Without algorithms, such work would likely take weeks or even months. Technology can indeed make up for manpower shortages and enhance transparency. But precisely because of this, people need to be even more vigilant: water can carry a boat, but it can also capsize it. AI can be both an assistant to justice and an amplifier of bias. Arguments for and against pull back and forth like a scale in the heart. Proponents believe that AI does not tire, is unaffected by human emotions, provides consistency and stability, and lowers adjudication costs. Opponents remind us that the black-box nature of algorithms may amplify existing biases, make the attribution of responsibility difficult to clarify, and fail to account for the uniqueness of individual cases. Whether such tools can be personalized is also a major issue. Criminal judgments concern freedom and even life; if we rely solely on algorithms, will the judiciary lose its human warmth?

At this point, the speaker noted that if AI is to truly enter criminal justice, it involves not only technical issues, but three other dimensions as well. First, regarding technical issues: is the logic of the algorithm explainable, and what conditions are required for output accuracy? Institutionally: how is it procured and developed prior to application, what regulations exist during the application process, and what remedies are available when mistakes occur? The human aspect is even more unignorable: do judicial personnel possess the literacy to understand and operate AI? Can they avoid developing an inertia of overreliance through long-term use? Finally is the most difficult and vulnerable link: trust. Trust in the judicial system is the foundation of democratic politics. If AI is introduced without transparency and explanation, how can the general public rest assured?

Transparency has become the key. Some argue that source code and datasets should be open to the public and researchers for inspection to avoid monopolies and increase trust. Even if full disclosure is impossible, weightings or parameters should at least be released for subsequent researchers to verify. The European Union's Artificial Intelligence Act already includes relevant designs, even providing exemptions for open-source models. After all, if the judiciary is to rely on AI, its credibility cannot rest solely on the unilateral promises of vendors or the government. Even more alarming is that humans are not always rational; humans naturally tend toward "energy conservation." If arduous judicial work can be handed over to computers, we might unconsciously choose to "let go" because there are simpler and faster methods available to achieve the goal—so why not? But judicial adjudication is not work that can be easily outsourced. Artificial intelligence is indeed far superior to humans in processing massive amounts of repetitive data, but when facing complex contexts and life experiences, its capacity falls far short because AI can never understand abstract and complex human emotions. All information is fed in by humans; it merely uses mathematical vectors to understand text connections within data. Therefore, when errors occur, attribution of responsibility and remedies are even harder to clarify—risks that the judiciary can least afford to bear.

Perhaps it can be said: the role of artificial intelligence is better suited as the "training wheels of a bicycle." It can help the judicial system maintain balance, preventing burnout and resource scarcity, but it should not replace human judgment. The soul of legal professionals must still be borne by humans. Behind the phrase "Computers can sort peanuts" lies not only an advertising slogan, but also a reminder to us. Technology can indeed improve efficiency, but the judiciary cannot pursue efficiency alone. What a verdict requires, beyond precision and consistency, is understanding, perception, and empathy. Perhaps the true issue is not whether AI can adjudicate, but whether we can find a balance: utilizing its power while safeguarding human nature and trust in the judiciary. When technology races forward, must the judiciary accelerate along with it? Or should it slow down and plan prudently to ensure justice is not sacrificed? These questions likely have no short-term answers. But one thing is certain: the future of the judiciary cannot be written by algorithms alone.