A Noise Audit of LLM Reasoning in Legal Decisions

The project introduces a critical comparison between human judicial noise and LLM decision-making variance, a not often explored intersection in AI safety literature. It makes a valuable empirical contribution by benchmarking LLM decision noise against human judicial variability. Well done! Particularly like the inclusion of confusion matrices by issue area and qualitative reasoning analysis. Currently, the safety implications are framed broadly (e.g., "unforeseen harms" from noisy judgments), but a formal threat model is absent. I would encourage you to adopt structured frameworks like the NIST AI Risk Management Framework to specify failure modes (e.g., silent rationalization errors) and mitigations. Defining concrete attack vectors (e.g., adversarial manipulation of case facts) or failure modes (e.g., over-reliance on precedent citations) would strengthen the AI safety impact. Similarly, technical expansions (eg higher sample size, quantifying discrepancies in reasoning rationales with semantic similarity scores or similar, additional details on environment setup, hyperparams and validation steps in the paper) could transform it into a highly impactful piece on algorithmic reliability in governance.

This was a wonderful, interesting submission to read. Overall, I thought the team used their research wisely to draw conclusions about the impact of noise on legal decision-making. I particularly liked how they drew upon other contextual factors (including corruption) that cannot be replicated using AI, which speaks to the human influence/social engineering aspects of legal decision-making as well. I would have liked to have seen more discussion of the impacts of these findings on society, including how legal decision-making is already reflective of a broken system that marginalises certain groups even more through repeated judiciary faults. Great work team!

Hi Mindy, May, and Markela, I think you did a great job with your project. Overall, the entire work is strong across all three criteria: theory, methods, and impact. The greatest strength is how well integrated and blended the work is, from how you argued your points to the use of the available evidence. I also liked the comprehensive, mixed-methods approach and use of figures to illustrate. I particularly appreciated that the work is very concise and to the point, it really doesn’t mince words. The flow and transition between paragraphs and sections is also great. Additionally, the grammar and referencing are excellent, as well as the presentation of the appendix. If you’re planning on publishing this, I think this is a good way forward: “We have several ideas for directions for development of this research. One approach would be to do a deeper analysis of model reasoning by using knowledge graphs or comparing vector embeddings of the LLM reasoning and majority opinion, or testing how sensitive models are to input noise by comparing outputs in response to minor prompt adjustments”. My recommendation would also be to expand more on the discussion and the practical implications of your results, for added impact.

Thank you for submitting your work on A Noise Audit of LLM Reasoning in Legal Decisions - it was timely and interesting to read. Please see below for feedback on your project:

1. Strengths of your proposal/ project:

- Methods and results are clearly outlined

- The project appears to be novel

- It is good that you detail the drawbacks/ limitations of the study

2. Areas for improvement

- A consistent use terminology should be kept. E.g., on p.1 you mention ‘level of variance in judgements’ - given this is the first mention of this, you should specifically refer to legal judgments (and specify what a judgement is/ in which legal context or area of law/ who carries out judgements/ in what jurisdiction or country). Unless you are referring to LLM judgements here? You also mention ‘the Supreme Court in 2005’ - of which country?

- Some key concepts/ methods should be defined - e.g., chain of thought reasoning (p.2)

- Sentencing guidelines are heavily relied upon courts, so I would expect a stronger argument under ‘Noise in Legal Decisions’ as to why variation isn’t good/ more up to-date evidence as to what level of variation there is amongst UK or US judgements at present.

- I’m not sure that I agree with the statement that ‘Because modern LLMs are trained on data scraped from the internet, it is likely that they will have prior knowledge of the court cases’ (p.3) without any evidence being provided. Can you provide a reference here (e.g., that legal cases are used to train LLMs)?

- You note that ‘Given the low sample size, the test does not have high power, introducing vulnerability to type 2 error’. Did you do a power calculation? Can you provide that in the paper?

- You note on p.6 that ‘The facts presented to the model were a short, broad summary of the case. The Supreme Court justices hearing the case would have much more detailed information, not to mention their expertise and extensive experience.’ It would be helpful if you could briefly go into more detail here - what implication does this have for your results? The drawbacks are noted descriptively, but more critical analysis would strengthen this section.

- A clear threat model should be provided.