Using AI to score the Michigan sentencing guidelines

Criminal Defense Newsletter | January-February 2026

Is it possible to score the Michigan sentencing guidelines using artificial intelligence? I wanted to test the assumption that AI could not produce an accurate result. After submitting a hypothetical set of facts, I found the assumption to be true, at times in surprising, amusing, and non-sensical ways. That said, there may be some limited uses for AI and the guidelines, mostly to offer a different point of view or to check a manual scoring.

Before jumping to the results, here is the information supplied in the hypothetical. I used a fact pattern that was previously tested in an earlier training event. The facts were relatively simple: defendant confronts an ex-girlfriend on the street, has a short conversation with her, and then shoots and kills her. I added a co-defendant (a friend of the defendant) who came to the encounter armed and who shot and wounded Mr. X, a man who was on a date with the ex-girlfriend. I also added a distraught mother of the victim (the mother was unable to sleep and in therapy following the crime)

Our hypothetical defendant was convicted of second-degree murder, and the hypothetical co-defendant was convicted of assault with intent to murder (AWIM). Both men were charged with both crimes, but defendant pleaded guilty to the murder charge and the co-defendant pleaded guilty to AWIM. The crimes occurred on December 1, 2024. Both men were convicted on September 8, 2025, and both will be sentenced on January 15, 2026, although the co-defendant will be sentenced first. Our defendant has a prior conviction for disorderly conduct in 2015 and was on felony probation for aggravated stalking (against the same victim) for an incident that occurred in 2023 and for which he was convicted in September 2024 and placed on probation for an incident that occurred in 2023 and for which he was convicted in September 2024 and placed on probation in October 2024.

Commission members are likely to be appointed soon, and the Commission will begin meeting in early 2026. I tried four AI platforms: Microsoft Copilot, ChatGPT GPT, Google Gemini, and Westlaw CoCounsel. The latter was through a helpful Westlaw representative.

A word here about Westlaw CoCounsel: It offers advanced legal AI and was recommended for this task. Surprisingly, it refused to identify the appropriate number of points for each variable and simply provided a list of variables likely to be scored, directing the user to the guidelines manual for further analysis. At first, I thought this was a prudent approach; on reflection, I suspect the goal was to avoid competition with sale of the guidelines manual (Westlaw and the guidelines manual are both products of Thomson-Reuters).

The other AI tools offered more in terms of suggested scorings and analysis, although Microsoft cautioned that it was not offering legal advice and the user should consult an attorney. Google warned that AI could make mistakes and the user should double-check the responses. ChatGPT offered no warning, but it noted the guidelines were now advisory pursuant to People v Lockridge..1

Scoring the Prior Record Variables

PRV 2: Google was the winner in this category, correctly assessing five points for the 2024 conviction of aggravated stalking (a Class E offense). Microsoft and ChatGPT would have assessed ten points for one low-severity felony (this is an incorrect number, although they identified the correct variable). Oddly, both indicated that PRV 1 would include Class A through C offenses, while PRV 2 would include Class D and lower (in fact, PRV 1 includes Class D). Westlaw identified PRV 1 as the appropriate variable (incorrect) and did not suggest a particular score.

PRV5: No winner here. Disorderly conduct, per case law, is not a scorable misdemeanor.2 Microsoft did not address this and simply found a ten-year gap (between the disorderly conduct conviction in 2015 and the 2026 sentencing date for the instant offense, ignoring the intervening 2023 offense and failing to use the correct date for the murder). ChatGPT would have assessed two points (correct if this had been a scorable misdemeanor). Google would have assessed five points (incorrect number and non-scorable), although it did recognize there was no ten-year gap. Westlaw noted a scorable disorderly conduct conviction for purposes of PRV 5 without suggesting a particular score.

PRV 6: Microsoft, ChatGPT, and Google all correctly suggested ten points for the felony bond status. Westlaw did not identify PRV 6 as a relevant at all.

Scoring the Offense Variables

OV 1, 2: Microsoft, ChatGPT, and Google correctly scored 25 points under OV 1 and 5 points under OV 2. Westlaw correctly recognized “use of a firearm” under OV 1, but it did not mention OV 2 (the type of weapon) at all and did not offer any scorings.

OV 3: Google was the winner, correctly assessing 25 points and recognizing that 100 points for death of the victim could not be scored. Microsoft and ChatGPT would have scored 100 points, and Westlaw simply recognized that the ex-girlfriend “was killed, which is the highest level of injury.”

OV 4, 5: Microsoft, ChatGPT, and Google correctly assessed 15 points for OV 5. Westlaw simply noted there was “psychological injury” to the victim’s mother for purposes of this variable (apparently not speculating on whether it was “serious” psychological injury).?

Surprisingly, ChatGPT also recommended 10 points under OV 4 because “[p]sychological injury is presumed to the homicide victim and need not be quantified.”

OV 6: Microsoft and Google correctly scored 25 points for intent to kill, and Westlaw correctly noted there was intent to kill for purposes of this variable. Only Microsoft addressed the possibility of premeditation (50 points), summarily concluding there was no premeditation. Surprising, ChatGPT would have assessed 100 points, although there is no 100-point category.

OVs 7, 8: Not discussed or correctly scored 0 points by all.

OV 9: All recognized Mr. X as the second victim for purposes of scoring 10 points.

OV 10, 11: Not discussed or correctly scored 0 points with the exception of ChatGPT, which would have assessed 10 points for a “former intimate partner . . . domestic relationship,” including use of that relationship and knowledge of her schedule (presumably to show exploitation).3 It should be noted here that the hypothetical included the fact that defendant was aware of when the ex-girlfriend was going on the date.

OV 12: Microsoft had the best answer here, noting there was no evidence of a separate “act” by the defendant (it did not address the possibility of a felon-in-possession charge because the facts were submitted separately from the prior criminal record, although to its credit it did address the contemporaneous AWIM charge). ChatGPT had a strange and incorrect response: assessing 25 points (rather than 5 points) for one crime against the person based on aiding and abetting the AWIM committed by the co-defendant. Google also had an odd response: “[H]is variable is scored for criminal acts that did not result in a conviction. Since the assault against Mr. X was dismissed in exchange for the plea, it might seem scorable, but it cannot be scored if it resulted in a separate conviction for a co-defendant in a joint trial or if it is part of the same transaction already accounted for in other OVs.” (I like the idea that a conviction for AWIM by the co-defendant would nullify any scoring for the defendant under OV 12.) Westlaw did not address this variable.

OV 13: Microsoft suggested zero points without any analysis, and likewise Google scored zero points based on “no evidence of a pattern.” ChatGPT oddly found a pattern based on the murder and AWIM, assessing 25 points (note, since it would have scored the AWIM under OV 12, this created a separate error under OV 13). Westlaw did not address OV 13. I should add here that a prosecutor might argue for a pattern based on the AWIM, aggravated stalking, and murder, but that is assuming 1) the aiding and abetting of the AWIM was not scored under OV 12, and 2) the aiding and abetting constitutes “felonious criminal activity.”

OV 14: Both Microsoft and ChatGPT concluded defendant was the leader, while Google and Westlaw failed to address this variable.

OV 15-20: Not relevant to the hypothetical.

As demonstrated, there was no tool that produced an accurate scoring of the sentencing guidelines. Google and Microsoft were better than ChatGPT, but all struggled with PRV 5. In addition, Google failed to score OV 14, and Microsoft would have scored 100 points under OV 3 (this is precluded by instructions to the variable). There were other deviations, although I liked Google’s odd response under OV 12 (not scoring dismissed conduct if the co-defendant was convicted of the offense). None of the tools addressed the multiple-offender instruction of OVs 1 through 3, although this was a moot point on the facts.4

Is there any value in using AI with the guidelines? In areas with a significant level of subjectivity, for example in scoring OV 7 for sadism, torture, or excessive brutality, it may be useful to hear from an outside source. AI might also offer ideas for the scoring or non-scoring of a variable that would otherwise escape notice. In addition, AI could be used as a way to double check a manual scoring.

At this point in the development of AI, however, I would not start with it or rely on AI to score the Michigan sentencing guidelines. There are statutory caveats and nuances that seem lost to AI, and case law further plays into the scoring of a variable. In this area of the law, at least, the skills of a trained lawyer remain necessary.

Endnotes
1 People v Lockridge, 498 Mich 358 (2015).
2 See People v Crews, 299 Mich App 381 (2013).
3 See People v Jamison, 292 Mich App 440 (2011) (domestic relationship for purposes of OV 7 does not include a past dating relationship).
4 The multiple offender instruction would not have applied because the defendant and co-defendant were convicted of different offenses. People v Johnston, 478 Mich 903 (2007); People v Biddles, 316 Mich App 148 (2016).

Anne Yantus
Michigan Sentencing PLLC

Anne Yantus is a sentencing consultant working with attorneys to promote more favorable sentencing outcomes. Anne credits her knowledge of Michigan sentencing law to the many years she spent handling plea and sentencing appeals with the State Appellate Defender Office. Following her time with SADO, Anne taught a criminal sentencing course at the University of Detroit Mercy School of Law and subsequently continued to write and speak on felony sentencing law while serving as pro bono counsel with Bodman PLC. Anne welcomes your Michigan felony sentencing questions and is happy to arrange a consultation where appropriate. Due to the volume of inquiries, Anne is not able to respond to pro bono requests for assistance or analysis of individual fact situations.