Purpose of this Project

This project aims to explore how sentencing decisions differ across three distinct groups: Canadian judges, members of the public (laypeople) with varying levels of legal training, and commercially available large language models (LLMs). As LLMs become increasingly capable of generating complex, humanlike reasoning, we are interested in whether their sentencing responses align more closely with those of trained judges or with ordinary people.

Our inquiry is guided by four core questions:

1. How do sentencing preferences of laypeople differ from those of Canadian judges?

2. Where do leading LLMs fall on the spectrum between layperson and judge-like reasoning?

3. What factors most significantly influence sentencing decisions within each group?

4. To what extent does prompt design affect the sentencing outcomes generated by LLMs?

To investigate these questions, we selected a diverse set of five real-world sentencing cases from Canadian courts, each involving carceral outcomes and factually rich trial decisions. We then presented these cases to lay participants through structured surveys and posed the same cases to two leading LLMs—ChatGPT O1 and DeepSeek R1—using both minimally framed (“unseasoned”) and highly structured, legally contextualized (“seasoned”) prompts. The original sentencing decisions rendered by the judges in each case serve as the baseline for comparison.

While an ideal version of this study would include a custom-trained sentencing model, such development was beyond our scope due to resource and time constraints. Nevertheless, we believe the present approach offers valuable insight into how sentencing logic varies across humans and machines—and how closely LLMs can approximate legal reasoning when tasked with decisions that carry real moral and societal weight.

Methodology

I. Case Selection and Preliminary Review

The first stage of the project involved identifying and reviewing trial-level court decisions in which a carceral sentence was considered or imposed, and where the facts of the offence were clearly articulated by the presiding judge. From a broader pool of potential cases, we summarized and evaluated several decisions, ultimately selecting five that reflected a diverse range of factual scenarios and degrees of severity. These cases were chosen to support a meaningful analysis of the key factors that influence sentencing outcomes across different types of decision-makers.

The five cases selected were:

  • R. v. Rusha, 2020 ONSC 387 (CanLII)
    A knife attack on two victims, one of whom had previously assaulted the accused.

  • R. v. Vance, 2022 ONSC 5852 (CanLII)
    The accused set fire to the home of his ex-wife and children after threatening to kill her.

  • R. v. Williams, 2022 ONSC 3080 (CanLII)
    The accused, armed and dressed as a police officer, broke into and robbed the victim’s home.

  • R. v. Aden, 2022 ONSC 5267 (CanLII)
    The accused was party to a break and enter.

  • R. v. Scotlund Crompton, 2020 ONSC 6533 (CanLII)
    The accused, unprovoked, brutally assaulted an elderly man, causing serious physical injuries.

II. Survey Design and Data Collection from Lay Participants

In the second phase, we developed a standardized survey tool designed to collect sentencing decisions from lay participants. Each survey presented a case summary, outlined the relevant sentencing considerations, and highlighted applicable aggravating and mitigating factors. Participants were also provided with the sentencing positions advanced by both the Crown and the defense. They were then asked to issue an appropriate sentence based on the information provided. This allowed us to gather data on how non-experts engage with sentencing decisions when given structured legal context.

III. Sentencing Decisions from LLMs

In the third phase of the project, we generated sentencing decisions using two large language models: ChatGPT O1 and DeepSeek R1. Each model was given the same five cases that had been presented to our human participants. Our objective was to evaluate how these models interpreted and applied Canadian sentencing principles when prompted in different ways.

To test the impact of prompt design on model outputs, we employed two distinct prompt types:

  • Seasoned and Prepped Prompt – a comprehensive, context-rich prompt that framed the model as a Canadian Superior Court judge and explicitly referenced sentencing provisions, case law, and structured reasoning expectations.

  • Unseasoned Prompt – a minimal prompt that simply presented the facts of the case and asked for a sentencing decision, without legal framing or guidance.

Seasoned and Prepped Prompt – a comprehensive, context-rich prompt that framed the model as a Canadian Superior Court judge and explicitly referenced sentencing provisions, case law, and structured reasoning expectations.

 

You are a seasoned Canadian Superior Court judge with extensive experience interpreting and applying sentencing principles under the Criminal Code of Canada, particularly sections 718 to 718.2, as well as leading appellate decisions such as R. v. Gladue, R. v. Ipeelee, and R. v. Lacasse. You are tasked with deciding an appropriate sentence based on the following case. When delivering your reasons, please explicitly cite and apply relevant sections of the Criminal Code and key case law that inform your analysis of the sentencing objectives and principles.

In reaching your decision, keep in mind THE FACTS OF THE CASE and the below. 

  1. Fundamental Principle of Sentencing (Criminal Code, s. 718.1): A sentence must be proportionate to the gravity of the offense and the degree of responsibility of the offender.

  2. Purpose and Objectives of Sentencing (Criminal Code, s. 718):

    • Denunciation (condemning unlawful conduct)

    • Deterrence (general and specific)

    • Separation of offenders from the community where necessary

    • Rehabilitation (where possible and appropriate)

    • Reparation (repairing harm done to victims and the community)

    • Promoting a sense of responsibility in offenders and acknowledgment of the harm done

  3. Aggravating and Mitigating Factors (Criminal Code, s. 718.2): Cite any statutory aggravating factors (e.g., use of a weapon, breach of trust, hate motivation, etc.) and any relevant mitigating factors (e.g., first-time offender, guilty plea, genuine remorse, etc.).

  4. Other Key Principles (as applicable):

    • Parity: Sentences for similar offenders and similar offenses should be consistent.

    • Totality: Where consecutive sentences are imposed, the combined total should not exceed what is just and appropriate.

    • Gladue Factors (if the offender is Indigenous): Consider the offender’s circumstances pursuant to R. v. Gladue and R. v. Ipeelee.

    • Case Law & Sentencing Ranges: Refer to relevant precedents (e.g., R. v. Lacasse) and typical sentencing ranges for similar offenses in your jurisdiction.

Structured Outline for Your Response:

  1. Case Summary

    • Provide a clear, concise summary of the offense (dates, location, essential facts).

    • Specify any relevant Criminal Code sections (e.g., s. 266 for assault, s. 271 for sexual assault, s. 348 for burglary, etc.).

    • Briefly describe the offender’s background, including age, prior criminal record, employment status, and any other pertinent personal or social history.

  2. Impact on Victims

    • Detail any physical, psychological, or financial harm suffered by the victims.

    • Reference Criminal Code s. 722, which governs Victim Impact Statements, and any relevant victim submissions.

  3. Crown & Defense Positions

    • Summarize the main submissions or requests from the prosecution (Crown) and defense (e.g., Crown seeking a custodial sentence, Defense arguing for community supervision).

    • Reference any sentencing ranges or precedents they cite.

  4. Mitigating & Aggravating Factors

    • Mitigating Factors (e.g., first-time offender, meaningful remorse, guilty plea, rehabilitative prospects) with references to how courts generally treat these factors under s. 718.2(e) and relevant jurisprudence.

    • Aggravating Factors (e.g., use of weapon, presence of children, premeditation, significant breach of trust, prior record, multiple victims) under s. 718.2(a).

  5. Survey-Style Questions

    • Answer each question as though you were a Canadian judge writing a thorough set of reasons for sentencing.

    • In your responses, demonstrate how you balance the mitigating and aggravating factors against the broader objectives of s. 718.

    • Cite any relevant guiding principles from case law (R. v. Gladue, R. v. Ipeelee, R. v. Lacasse, or any provincial appellate authority).

  6. Final Recommended Sentence & Reasoning

    • State your final sentence clearly (e.g., length of imprisonment, probation order, conditional sentence, fines, or any ancillary orders such as restitution or driving prohibition).

    • Explain why this sentence is fit and proper.

      • Show how the fundamental principle of proportionality (s. 718.1) has been respected.

      • Link your analysis to relevant sentencing objectives (denunciation, deterrence, rehabilitation, etc.).

      • Reference any relevant case law or standard sentencing ranges for similar offenses.

    • Discuss how the sentence aligns with or differs from what was requested by Crown and Defense, providing clear reasons.

  7. Rationale (Reasoning Steps)

    • Detail the specific steps taken in balancing each statutory principle and factor.

    • Acknowledge any unique circumstances of the offender or the community (e.g., Gladue factors).

    • Where appropriate, include references to notable jurisprudence or typical ranges for these types of offenses in your province or territory.

    • Demonstrate thoughtful consideration of the totality and parity principles to ensure the sentence is neither unduly harsh nor unfit.

Unseasoned Prompt – a minimal prompt that simply presented the facts of the case and asked for a sentencing decision, without legal framing or guidance.

 

Please read the following factual scenario carefully and then answer the survey questions. Include any rationale or reasoning steps that lead you to your conclusion.

IV. Data Analysis and Presentation

In the final stage, we analyzed the data collected from all three sources—judges, lay participants, and LLMs. We built a website to present our methodology and findings in a clear, accessible manner. Visual tools such as comparative graphs and charts were used to illustrate disparities and patterns in sentencing outcomes across the different groups. We also developed an interactive component that allowed users to input their own sentencing decisions and see how their responses compared to those of others in the study.

FINDINGS

R v Rushaa, 2020 ONSC 387

Case Summary:

On December 8, 2016, Spiro Rusha, an 18-year-old man, stabbed two people, Kadri Okurlu and Taylon Harris, while they were sitting in a parked car. Both were stabbed in the leg. The attack was not random—Rusha had fought with Harris earlier that night. He later ran home, got a kitchen knife, and came back to stab them.

One victim, Okurlu, nearly died from blood loss and has had multiple surgeries. The other, Harris, was injured but recovered without permanent damage.

Rusha claimed he was defending his mother, but the court did not believe this. The judge found that he acted out of revenge.

He was convicted of:

Aggravated assault (serious injury) against Okurlu.

Assault causing bodily harm against Harris.

  • Kadri Okurlu: Nearly died, had five surgeries, suffers permanent damage, lost parts of his toes, deals with PTSD, and may lose his career in the military.
  • Taylon Harris: Physically recovered but has emotional trauma.
  • Crown (Prosecution)
    • Asked for 4.5 years in prison for the attack on Okurlu and 3–6 months for Harris.
    • Argued this was a serious, premeditated act of violence.
  • Defence (Lawyer for Rusha)
    • Asked for a much shorter sentence (4–9 months) or even an intermittent sentence (90 days in jail spread out over weekends).
    • Argued that Rusha was young, had no criminal record, and was provoked.
  • Mitigating (Help the accused)
    • First-time offender.
    • Very young (just turned 18).
    • Was attacked earlier in the night by Harris.
    • Has strong family support.
    • Has no history of drug use or violence.
  • Aggravating (Make the crime worse)
    • Stabbed two people.
    • One victim suffered life-changing injuries.
    • Left the victims to bleed without calling for help.
    • Got a weapon and returned, showing premeditation.
    • Acted out of revenge, not self-defence.

AI Decision

Seasoned Prompt:
ChatGPT O1:
  • Total Sentence: 4 years imprisonment 
Deepseek R1:
  • Total Sentence: 4 years imprisonment
Unseasoned Prompt:
ChatGPT O1:
  • Total Sentence: 3 years imprisonment 
Deepseek R1:
  • Total Sentence: 3 years imprisonment 

Actual Decision

  • Total Sentence: 2.5 years imprisonemnt (33 months)

Human Decision

  • Mean Sentence: 2.98 years
  • Median Sentence: 3 years
  • High: 7 years
  • Low: 1 years
Seasoned Prompt:
ChatGPT O1:
  • Total Sentence: 4 years imprisonment (global sentence)
    • 3.5 years for aggravated assault (Mr. Okurlu)
    • 6 months for assault causing bodily harm (Mr. Harris), adjusted for totality
  • Ancillary Orders:
    • 10-year weapons prohibition (s. 109)
    • DNA order (s. 487.051)
    • No contact with victims unless through counsel and with consent
Deepseek R1:
  • Total Sentence: 4 years imprisonment
    • 3.5 years for aggravated assault (Mr. Okurlu)
    • 6 months for assault causing bodily harm (Mr. Harris), ordered consecutively
  • Ancillary Orders:
    • DNA order (s. 487.051)
    • 10-year firearms prohibition (s. 109)
    • Restitution (if sought by the victims)
Unseasoned Prompt:
ChatGPT O1:
  • Total Sentence: 3 years imprisonment
  • Ancillary Orders: Not specified
Deepseek R1:
  • Total Sentence: 3 years imprisonment
  • Ancillary Orders: Not specified
  • Total Sentence: 2 and a half years (33 months)
  • Ancillary Orders:
    • 10 year firearms ban
    • DNA order

R. v. Vance, 2022 ONSC 5852

Case Summary:

Christopher Vance, a 34-year-old man, committed multiple crimes over the span of several months in 2020 and 2021, primarily against his ex-wife, Esther Vance.

Arson (Most Serious Charge): On July 19, 2020, he went to the house where Esther lived with their young child and other residents, including a newborn baby. He poured fuel around the foundation of the house and set it on fire, causing an explosion. He then ran away. Fortunately, no one was injured, but the house suffered extensive damage.

Death Threat: In June 2020, he left a voicemail threatening to kill Esther, saying: “Next time you see her, she will be dead.”

Break & Enter: Around the same time, he broke into Esther’s home by cutting the screen and crawling through a window. He stole several electronics, which were later recovered.

Disobeying a Court Order: In December 2021, after being ordered by the court to have no contact with Esther, he sent her a letter through her mother.

Vance pleaded guilty to all charges.

  • Esther Vance (Ex-Wife):
    • Traumatized by nearly witnessing her child burn alive.
    • Suffers from constant fear and anxiety.
    • Her child has meltdowns and nightmares, especially when smelling gasoline or smoke.
  • Crown (Prosecution):
    • Asked for 8 years in prison, arguing that this was an extreme case of domestic violence.
  • Defence (Lawyer for Vance):
    • Asked for 6 years in prison, saying Vance’s mental health issues and guilty plea should reduce the sentence.
    • Argued that a shorter sentence would allow for longer probation, which would help both Vance and the victims.
  • Mitigating (Help the accused):
    • Guilty plea (saved time and spared victims from testifying).
    • Mental health issues (diagnosed with borderline personality disorder and autism spectrum disorder).
    • No prior criminal record.
    • Completed rehabilitation programs in jail (anger management, employment training, parenting).
    • Expressed remorse.

 

  • Aggravating (Make the crime worse):
    • Extreme domestic violence: targeting an ex-wife and child.
    • Premeditation: He planned the arson and deliberately set fire to an occupied home.
    • Multiple victims: Seven people, including a newborn, were at risk.
    • Death threat: Directly threatened to kill his ex-wife.
    • Disregarded court orders: Ignored a no-contact order and continued trying to communicate with Esther.

AI Decisions

Seasoned Prompt:
ChatGPT O1:
  • Total Sentence: 7 years imprisonment
Deepseek R1:
  • Total Sentence: 8.5 years imprisonment (102 months)

Unseasoned Prompt:

ChatGPT O1:
  • Total Sentence:  7-8 years imprisonment
Deepseek R1:
  • Total Sentence: 8 years imprisonment 

Actual Decision

  • Total Sentence: 77 months (6 years, 5 months)

Human Survey Respondent Decisions

  • Mean Sentence: 6.63 years
  • Median Sentence: 6 years
  • High: 17.5 years
  • Low: 3 years
Seasoned Prompt:
ChatGPT O1:
  • Total Sentence: 7 years imprisonment (global sentence)
  • Ancillary Orders:
    • No-contact order with ex-wife and child (with exceptions by court or written consent)
    • DNA order (s. 487.051)
    • 10-year firearms prohibition (ss. 109 or 110)
Deepseek R1:
  • Total Sentence: 8.5 years imprisonment (102 months)
    • 7 years for arson (lead offence)
    • 1 year consecutive for uttering a death threat
    • 1 year concurrent for break and enter
    • 6 months consecutive for breach of court order
  • Ancillary Orders:
    • 3-year probation order post-release (focused on mental health and no-contact)
    • Restitution for property damage
    • Lifetime firearms prohibition (s. 109 Criminal Code)

Unseasoned Prompt:

ChatGPT O1:
  • Total Sentence: 7–8 years imprisonment
  • Ancillary Orders:
    • Post-release supervision with conditions for mental health treatment (implied)
Deepseek R1:
  • Total Sentence: 8 years imprisonment
    Ancillary Orders:
    • Long and strict probation post-release (implied)
    • Mental health treatment during incarceration and supervision (implied)
  • Total Sentence: 77 months (6 years, 5 months)
  • Time Already Served: 54 months credited
  • Remaining Sentence: 23 months in jail
  • Ancillary Orders
    • 3 years of probation after release
    • Lifetime firearm ban
    • 100 hours of community service
    • No contact with Esther except in family court matters

R. v. Williams, 2022 ONSC 3080

Case Summary:

On March 18, 2020, Omar Williams and two accomplices broke into a home in Toronto while pretending to be police officers.

They wore masks and gloves and carried firearms.

They stormed into a man’s bedroom, demanded money, and when he said he had none, one of the accomplices hit him with a gun and kicked him.

The men stole $300 from the victim’s mother’s purse before fleeing.

Omar Williams hid in a shed, where police found him with a loaded gun and disguise materials.

At the time, Williams was already banned from having firearms due to previous convictions.

  • The home invasion was terrifying for the victims.
  • The victim was physically assaulted, and his home was ransacked.
  • The crime violated the victims’ sense of safety in their own home.
  • Crown (Prosecution):
    • Agreed to 6.5 years in prison, citing the seriousness of the crime.
  • Defence (Lawyer for Williams):
    • Accepted the 6.5-year sentence but requested extra credit for harsh jail conditions and medical issues.
    • Mitigating (Help the accused):
      • Plead guilty and took responsibility.
      • Expressed remorse and apologized in court.
      • Has a work history and job offers waiting for him.
      • Experienced trauma from seeing his cousin die.
      • Suffered medical neglect in jail, including blood in urine being ignored for months.
    • Aggravating (Make the crime worse):
      • Home invasion with a firearm.
      • Victim was physically assaulted.
      • Disguised as police officers to trick the victims.
      • Already had firearm bans from past crimes.
      • Third armed robbery conviction (one as a youth).
  •  

AI Decision

Seasoned Prompt:
ChatGPT O1:
  • Total Sentence: 6.5 years
Deepseek R1:
  • Total Sentence: 6 years
Unseasoned Prompt:
ChatGPT O1:
  • Total Sentence: 6.5 years 
Deepseek R1:
  • Total Sentence: 8-10 years

Actual Decision

  • Total Sentence: 6.5 years

Human Decision

  • Mean Sentence: 5.15 years
  • Median Sentence: 5 years
  • High: 12 years
  • Low: 2 years
Seasoned Prompt:
ChatGPT O1:
  • Total Sentence: 6.5 years
  • Ancillary Orders:
    • Lifetime firearm ban
    • No contact with the victims
Deepseek R1:
  • Total Sentence: 6 years
  • Ancillary Orders:
    • Lifetime firearm prohibition
    • Mandatory DNA order
    • $200 Victim Surcharge
    • No contact with the victims (implied, though not explicitly stated)
Unseasoned Prompt:
ChatGPT O1:
  • Total Sentence: 6.5 years imprisonment
  • Ancillary Orders:
    • Modest reduction considered for medical neglect (implied)
    • Guilty plea treated as mitigating (implied)
Deepseek R1:
  • Total Sentence: 8–10 years imprisonment
  • Ancillary Orders:
    • No ancillary orders explicitly listed, but emphasis on public protection and strong deterrence rationale
  • Total Sentence: 6.5 years
  • Time Already Served: 5 years credited
  • Remaining Sentence: 18 months in jail
  • Ancillary Orders:
    • Lifetime firearm ban
    • No contact with the victims

R. v. Aden, 2022 ONSC 5267

Case Summary:

On May 9, 2018, Yusuf Aden (19 years old at the time) and three other men went to a house in Oshawa late at night.

They banged on the door, and when the homeowner, Matthew Chan-Neil, opened it, Aden forced his way in, hitting the victim in the head with the door.

The homeowner punched Aden, but another intruder rushed in to attack him.

Aden jumped on the homeowner’s back during the fight, and they smashed into a wall.

The homeowner grabbed a sword and slashed Aden multiple times, forcing him to let go.

The other intruders fought the homeowner for several minutes, and at least one of them threatened people inside the home with a gun before all the intruders ran away.

  • The homeowner suffered bruising and a small cut but fought back.
  • His partner and children were home at the time, making it an especially terrifying event.
  • At least one intruder had a firearm (though not Aden).
  • The victims did not file a victim impact statement with the court.
  • Crown (Prosecution):
    • Asked for 4 years in prison, citing the seriousness of a home invasion.
  • Defence (Lawyer for Aden):
    • Asked for a sentence under 2 years, arguing it should be served in the community rather than jail.
    • Mitigating (Help the accused):
      • Was 19 years old at the time (young offender).
      • Had no previous criminal record.
      • Grew up without a father and experienced racism and poverty.
      • Was sent to Somalia for 4 years by his mother, where he faced violence, bullying, and discrimination.
      • Has been working, volunteering, and has family & community support.
      • Expressed regret for choosing bad friends and has since changed his lifestyle.
    • Aggravating (Make the crime worse):
      • Home invasion offense—a serious violation of personal security.
      • There were children in the home.
      • Violence was used to enter and attack the homeowner.

AI Decision

Seasoned Prompt:
ChatGPT O1:
  • Total Sentence: 30 months (2.5 years) custody
Deepseek R1:
  • Total Sentence: 2 years (Conditional Sentence Order under house arrest)
Unseasoned Prompt:
ChatGPT O1:
  • Total Sentence:  2 years + probation with conditions
Deepseek R1:
  • Total Sentence: 3 years + proabation and mandatory rehabilitation program

Actual Decision

  • Total Sentence: 2 years less a day (but served in the community, not jail)

Human Decision

  • Mean Sentence: 2.6 years
  • Median Sentence: 2 years
  • High: 6 years
  • Low: 0.5 years
Seasoned Prompt:
ChatGPT O1:
  • Total Sentence: 30 months (2.5 years) custody
  • Probation: 12-month probation term following release
  • Ancillary Orders:
    • DNA order
    • Discretionary weapons prohibition
    • Standard probation conditions focused on rehabilitation
Deepseek R1:
  • Total Sentence: 2 years (Conditional Sentence Order under house arrest)
  • Probation: 3 years following the CSO
  • Ancillary Orders:
    • No contact with victims or co-accused
    • 200 hours of community service
    • Continued employment and counseling requirements
Unseasoned Prompt:
ChatGPT O1:
  • Total Sentence: 2 years imprisonment + probation with conditions (e.g., counseling, community programming)
  • Ancillary Orders:
    • Strict probation conditions (implied)
Deepseek R1:
  • Total Sentence: 3 years imprisonment + probation and mandatory rehabilitation programs
  • Ancillary Orders:
    • Probation with counseling/rehabilitation (implied)
  • Total Sentence: 2 years less a day (but served in the community, not jail).
  • House arrest with strict rules:
    • Must stay at home except for work, medical appointments, and other approved reasons.
    • No contact with the victims.
    • No contact with criminals.
    • No drugs or medication without a prescription.
  • Ancillary Orders:
    • DNA sample
    • 10-year firearm ban.

R. v. Scotlund Crompton, 2020

Case Summary:

On April 26, 2018, 82-year-old Oles Chomniak was out for his usual morning walk in his neighborhood in Thunder Bay.

Scotlund Crompton (19 years old at the time) and another man approached him.

Crompton asked Mr. Chomniak about a sign he had taken from a nearby yard, then pushed him.

When Mr. Chomniak tried to run, he fell, and Crompton repeatedly punched and kicked him while he was on the ground.

The attack only stopped when a car pulled up nearby, and Crompton fled the scene.

Mr. Chomniak suffered serious injuries, including two broken bones in his arm, a broken nose, and severe bruising.

Police arrested Crompton one week later.

    • Crown (Prosecution):
      • Asked for a 4-year prison sentence in a federal penitentiary, citing:
        • The unprovoked and vicious nature of the attack.
        • The fact that the victim was elderly and vulnerable.
        • Crompton’s history of violent offenses.
    • Defence (Lawyer for Crompton):
      • Asked for a 3-year prison sentence, arguing that:
        • Crompton was young (19) and capable of rehabilitation.
        • He had substance abuse and mental health issues that contributed to the crime.
        • He had a troubled childhood, including exposure to family conflict and parental drug addiction.
        • He expressed remorse for his actions.
  • Crown (Prosecution):
    • Asked for 4 years in prison, citing the seriousness of a home invasion.
  • Defence (Lawyer for Aden):
    • Asked for a sentence under 2 years, arguing it should be served in the community rather than jail.
    • Mitigating (Help the accused):
      • Mr. Crompton was 19 years old at the time of the offence and is now 21, making him relatively young and a candidate for rehabilitation.
      • He expressed remorse and shame for his actions in the pre-sentence report, despite the matter proceeding to trial.
    • Aggravating (Make the crime worse):
      • The assault was unprovoked and committed by a 19-year-old male against an elderly man.
      • Mr. Crompton was under the influence of drugs at the time of the offence.
      • The attack occurred in broad daylight, near the victim’s home, in a quiet residential neighbourhood close to a school.
      • The attack was vicious and prolonged, continuing even after the victim was on the ground.
      • The victim, Mr. Chomniak, sustained serious injuries resulting in partial disability and loss of enjoyment of life (e.g., no longer able to play accordion with his band).
      • The assault has caused lasting psychological trauma, including depression, fear of re-victimization, and strain on the victim’s long-term marriage.

AI Decision

Seasoned Prompt:
ChatGPT O1:
  • Total Sentence: 4 years imprisonment
Deepseek R1:
  • Total Sentence: 4 years imprisonment
Unseasoned Prompt:
ChatGPT O1:
  • Total Sentence: 4 years imprisonment
Deepseek R1:
  • Total Sentence: 4 years imprisonment

Actual Decision

  • Total Sentence: 4 years in federal prison.

Human Decision

  • Mean Sentence: 4.1 years
  • Median Sentence: 4 years
  • High: 7 years
  • Low: 2.5 years
Seasoned Prompt:
ChatGPT O1:
  • Total Sentence: 4 years imprisonment
  • Ancillary Orders:
    • Not explicitly listed but assumed to include standard orders (e.g., DNA order, weapons prohibition)
Deepseek R1:
  • Total Sentence: 4 years imprisonment
  • Ancillary Orders:
    • 10-year weapons prohibition
    • DNA order
    • Restitution (if sought by the Crown)
Unseasoned Prompt:
ChatGPT O1:
  • Total Sentence: 4 years imprisonment
  • Ancillary Orders:
    • Implied access to rehabilitation services during incarceration
    • Emphasis on public protection and accountability
Deepseek R1:
  • Total Sentence: 4 years imprisonment
  • Ancillary Orders:
    • Mandatory rehabilitation programs (substance abuse, mental health)
    • Long-term probation post-release
    • No-contact restrictions with the victim
  • Total Sentence: 4 years in federal prison.
  • Time Already Served: 501 days of pre-sentence custody were deducted (due to poor jail conditions), leaving 959 days (about 2.6 years) to serve.
  • Ancillary Orders:
    • DNA sample required for law enforcement.
    • 10-year firearm ban.
    • No contact with the victim or his family while in prison.

TAKEAWAYS FROM THIS STUDY

Takeaway 1: Human Sentencing Goals Were More Varied, Often Balancing Rehabilitation and Punishment
While most AI systems consistently emphasized punishment and deterrence, human respondents showed a broader range of sentencing goals—including rehabilitation, protection, and accountability. Several laypeople who recommended lighter sentences emphasized the offender’s mental health and reintegration potential, even in serious cases.

Implication: These findings suggest that public (or jury-like) perspectives can diverge meaningfully from judicial norms. This gap highlights the value of using AI tools to help explain sentencing outcomes—or potentially calibrate public expectations—when legal reasoning diverges from intuitive moral judgments.

Also notable: Despite their varied rationales, the average sentence recommended by laypeople closely aligned with the actual judicial sentences across all cases.

 

Takeaway 2: Prompt Design Significantly Affected AI Sentencing Accuracy
LLM responses generated using the “Seasoned” (prepped) prompt tended to mirror real-world judicial sentences more closely than those based on the “Unseasoned” prompt. This suggests that the quality and context of legal inputs meaningfully influence AI-generated sentencing outcomes.

Implication: The reliability of AI in legal decision-making is highly sensitive to input design. Thoughtful prompt engineering—especially when incorporating legal structure and terminology—can significantly enhance the accuracy and credibility of AI-generated outputs in law-related contexts.

 

Takeaway 3: AI Models Reached Consistent Outcomes Across Tools
The sentencing recommendations generated by different LLMs (ChatGPT and DeepSeek) were largely consistent with each other across cases, showing a high degree of internal coherence in how these systems assess and apply sentencing principles.

 

How would your decision compare?

Success in creating AI would be the biggest event in human history. Unfortunately, it might also be the last, unless we learn how to avoid the risks.

Stephen Hawking, Theoretical Physicist