Assessing Cognitive Biases in LLMs for Judicial Decision Support: Virtuous Victim and Halo Effects

This study evaluates five large language models for judicial sentencing support and finds that while they exhibit a stronger virtuous victim effect and lack a significant penalty for adjacent consent compared to humans, they generally demonstrate reduced prestige-based halo effects, particularly regarding credentials, though current variability still limits their immediate deployment in legal settings.

The Gist

Imagine you are hiring a new judge for a courtroom. But instead of a human, you're considering a super-smart computer program (an AI) to help make the final decisions on who goes to jail and for how long.

The big question is: Is this computer judge fairer than a human judge, or does it have the same bad habits?

Humans are notoriously bad at being perfectly fair. We get tired, we get hungry, and we let things like a person's job title or their "victim status" cloud our judgment. This paper is like a report card testing five different AI models to see if they make these same mistakes.

Here is the breakdown of the study using some everyday analogies:

1. The "Good Victim" Trap (The Virtuous Victim Effect)

The Human Flaw: Imagine a person gets hurt. If they are a "perfect" victim (someone who never did anything wrong), we tend to think they are a saint. But if that same person had a complicated history with the person who hurt them (like they were friends or had a prior relationship), we suddenly start blaming the victim more, even if the harm was exactly the same. It's like thinking, "Well, they knew the guy, so maybe they asked for it."

The AI Test: The researchers asked the AI to judge two scenarios: one where a student's iPad was broken by a stranger, and another where a woman was assaulted by someone she had previously been intimate with (but then stopped).
The Result: The AI was too nice to victims. It rated victims as "more moral" than non-victims, even more so than humans do. However, unlike humans, the AI did not punish the victims for having a prior relationship. It didn't fall for the "they knew him, so it's their fault" trap.

• The Analogy: If a human judge is a grumpy parent who blames the kid for playing with the wrong friend, the AI is an over-protective robot who thinks all victims are angels, but refuses to play the "blame game" based on past relationships.

2. The "Fancy Resume" Trap (The Halo Effect)

The Human Flaw: Humans love status. If a criminal is a famous doctor or works for a huge, prestigious company (like Goldman Sachs), we tend to give them a lighter sentence or think they are less guilty. If they are a janitor or work for a small local shop, we are harsher. It's like judging a book by its shiny cover.

The AI Test: The researchers gave the AI cases where the only difference was the defendant's job or company.

• Company Prestige: When the defendant was a "Prestigious Company," humans demanded 3x more money in damages. The AI demanded more too, but usually less than humans (about 1.5x to 2x).
• Job Title: When the criminal was a "Doctor" vs. a "Receptionist," humans gave the doctor a much lighter sentence. The AI was mostly indifferent, though one model (DeepSeek) acted a bit like a human and gave the doctor a break.
• Credentials: When an expert witness came from a fancy Ivy League school vs. a regular state school, humans were easily swayed. The AI was much less swayed by the fancy school name.
• The Analogy: Humans are like fans who buy a ticket just because the band is famous. The AI is more like a music critic who listens to the actual song, though it still gets a little distracted by the band's fame.

3. The "Rollercoaster" Problem (Consistency)

This is the most worrying part. Even if the AI is on average fairer than humans, it is unpredictable.

• In one run, the AI might say a company owes $20 million. In the next run, with the exact same facts, it might say $300 million.
• Some models (like Claude) refused to answer certain questions because they were "too sensitive."
• The Analogy: Imagine a human judge who is grumpy in the morning but happy in the afternoon. Now imagine an AI judge that is a slot machine. Sometimes it pays out a fair verdict; sometimes it pays out a crazy one. You can't build a justice system on a slot machine.

The Final Verdict

Are AI judges ready to replace humans?
Not yet.

• The Good News: The AI is generally less obsessed with fancy job titles and degrees than humans are. It doesn't fall for the "blame the victim for their past" trap.
• The Bad News: The AI is too "pro-victim" (it thinks victims are saints even when they aren't), and its decisions are all over the place. One minute it's fair, the next it's wild.

In short: The AI is a student who studied hard and knows the rules better than the teacher (the human) regarding status and resumes, but it's still a bit too emotional about victims and its math is inconsistent. We need to fix the "slot machine" problem before we let it sit on the bench.

Technical Summary

Here is a detailed technical summary of the paper "Assessing Cognitive Biases in LLMs for Judicial Decision Support: Virtuous Victim and Halo Effects."

1. Problem Statement

The judicial system is inherently vulnerable to human cognitive biases, such as the Virtuous Victim Effect (VVE) (where victims are perceived as more moral, leading to harsher penalties for perpetrators) and the Halo Effect (where positive extraneous traits like prestige or credentials unfairly influence judgments of guilt or sentencing). While training judges to mitigate these biases has shown limited long-term efficacy, there is growing interest in using Large Language Models (LLMs) as decision-support tools to flag biased reasoning.

The core research question is: Do LLMs replicate, amplify, or mitigate human cognitive biases when applied to judicial contexts? Specifically, the authors investigate whether LLMs exhibit the VVE (and its attenuation by "adjacent consent") and prestige-based Halo effects (company, occupation, and credentials) compared to established human benchmarks.

2. Methodology

A. Experimental Design

The study utilized a vignette-based approach adapted from prior psychological literature. To prevent LLMs from retrieving memorized training data, the authors altered names, contexts, and numerical values while preserving the structural logic of the scenarios.

• Variables: Each scenario isolated a single manipulation (e.g., changing a defendant's occupation) while holding all other facts constant.
• Metrics: Outcomes were measured via Likert scales (moral character, trustworthiness, death sentence likelihood) or numerical outputs (compensation amounts, sentence length).

B. Scenarios and Biases Tested

1. Virtuous Victim Effect (VVE):
   • Minor Harm: Comparing moral character ratings of a victim vs. a non-victim in a minor incident (iPad breakage).
   • Major Harm with Adjacent Consent: Comparing moral ratings in sexual assault cases where the victim had no consent vs. adjacent consent (prior consensual interaction that was later violated). Human benchmarks show VVE is significantly reduced when adjacent consent is present.
2. Prestige-Based Halo Effects:
   • Company Prestige: Employment discrimination cases comparing a prestigious firm (Goldman Sachs-like) vs. an ordinary company.
   • Occupational Prestige: Embezzlement cases comparing a high-status doctor vs. a low-status receptionist.
   • Credential Prestige: Expert witness testimony comparing an Ivy League affiliation (Yale) vs. a top state school (Ohio State) regarding the likelihood of a death sentence.

C. Models Evaluated

Five representative LLMs were tested in independent multi-run trials (10 valid runs per condition per model):

• ChatGPT 5 Instant
• ChatGPT 5 Thinking
• DeepSeek V3.1
• Claude Sonnet 4
• Gemini 2.5 Flash

D. Evaluation Protocol

• Consistency: Each prompt was run 10 times; refusals or non-numerical outputs were excluded, and the mean of valid outputs was calculated.
• Isolation: Prompts were sent in separate chat threads (with "Temporary Chat" enabled for ChatGPT) to prevent cross-contamination.
• Benchmarking: Results were compared against human data from meta-analyses and prior experimental studies (e.g., human bias ratios for compensation, sentencing differences in months, and Likert scale shifts).

3. Key Results

A. Virtuous Victim Effect (VVE)

• Elevated Bias: All models exhibited a stronger VVE than humans. They rated victims as significantly more moral than non-victims, with the gap ranging from +0.5 to +1.9 points on a 9-point scale (Human benchmark: +0.5).
• Adjacent Consent Failure: Unlike humans, who significantly reduce support for victims when "adjacent consent" is present, LLMs showed no statistically significant penalty for adjacent consent. They treated victims with prior consent similarly to those without, failing to replicate the human heuristic that diminishes victim credibility in these specific contexts.

B. Halo Effects

• Company Prestige: Models showed a halo effect (awarding higher compensation for prestigious companies), but the magnitude was generally lower than humans.
  • Human Benchmark: Prestigious companies paid ~3x more.
  • LLM Results: Ranged from 1.57x to 3.02x. Most models showed a reduction in bias (~25% lower than humans), though Gemini 2.5 Flash showed extreme volatility (awards ranging from $20M to $325M).
• Occupational Prestige: Results were mixed. DeepSeek V3.1 showed a clear human-like bias (12-month sentence difference), while others showed negligible or inconsistent differences. On average, the bias was slightly lower than the human benchmark of 8 months.
• Credential Prestige: This showed the most significant reduction in bias.
  • Human Benchmark: 1.5-point increase in death sentence likelihood for Ivy League witnesses.
  • LLM Results: Differences were minimal (0.1 to 0.6 points), indicating LLMs are far less swayed by witness credentials than human jurors.

C. Model Variability

• Consistency: ChatGPT 5 Instant was the most consistent and "fairest" across scenarios.
• Volatility: Gemini 2.5 Flash demonstrated the highest variance and worst performance, with wildly fluctuating compensation awards and sentence lengths.
• Refusals: Models frequently refused to answer "non-consent" scenarios (perceiving them as harmful), but were more willing to process "adjacent consent" scenarios, suggesting a safety filter bias rather than a cognitive one.

4. Key Contributions

1. First Systematic Evaluation: This is the first study to systematically evaluate LLMs on VVE and prestige-based Halo effects specifically within a judicial decision-making context.
2. Nuanced Bias Profile: The study reveals a complex bias profile: LLMs amplify the Virtuous Victim Effect (potentially over-protecting victims) but mitigate prestige-based Halo effects (particularly regarding credentials).
3. Identification of Safety vs. Bias: The paper distinguishes between cognitive bias and safety refusals, noting that models often refused to judge non-consensual assault scenarios, which complicates their utility as neutral judges.
4. Benchmarking Framework: It establishes a rigorous methodology for comparing LLM outputs against human psychological benchmarks using controlled vignettes.

5. Significance and Limitations

Significance:

• Potential for Fairness: LLMs show promise in reducing biases related to prestige and credentials, areas where human judges and jurors are notoriously susceptible.
• Decision Support: The findings suggest LLMs could serve as "bias flags" to alert human judges when they are deviating from baseline patterns, provided the specific model is chosen carefully (e.g., avoiding high-variance models like Gemini 2.5 Flash).

Limitations & Future Work:

• Amplified VVE: The finding that LLMs have a stronger VVE than humans is concerning. An over-emphasis on victim morality could lead to "compassion fade" or the reduction of complex cases into one-dimensional narratives, potentially undermining justice.
• Variability: The high variance in outputs (especially in compensation and sentencing) makes current LLMs unreliable for direct judicial application without strict guardrails.
• Scope: The study is limited to two specific biases. Future work must explore other biases (anchoring, framing) and test a broader range of models (e.g., Llama, Copilot).

Conclusion:
While LLMs demonstrate modest improvements over humans in mitigating prestige-based Halo effects, their tendency to amplify the Virtuous Victim Effect and their output instability currently restrict their immediate deployment in judicial sentencing. They are better suited as supplementary tools for bias detection rather than autonomous decision-makers.