· ai regulation · 5 min read
Virginia Just Told the Criminal Justice AI Industry to Sit Down
Virginia HB 1642 requires a human decision-maker for every criminal justice determination. AI can recommend; it cannot decide. The implications for recidivism scoring and risk assessment tools are significant.
Photo by Hunters Race on Unsplash
Virginia’s HB 1642 is a short statute. It’s one substantive section. It took effect June 6, 2025, and it may have done more to constrain a specific industry than any other state AI law I’ve read.
Here’s what it says, in full:
All decisions related to the pre-trial detention or release, prosecution, adjudication, sentencing, probation, parole, correctional supervision, or rehabilitation of criminal offenders shall be made by the judicial officer or other person charged with making such decision. No such decision shall be made without the involvement of a human decision-maker. The use of any recommendation or prediction from an artificial intelligence-based tool shall be subject to any challenge or objection permitted by law.
That’s the whole thing. Clear, direct, unambiguous. Human in the loop — mandatory.
Why This Matters
The criminal justice AI industry is not small. Tools like COMPAS, PSA (Public Safety Assessment), and dozens of smaller risk assessment products are used across the country to inform bail decisions, sentencing recommendations, parole determinations, and supervision conditions. These tools generate scores — risk of flight, risk of recidivism, risk of violence — and judges, magistrates, and parole boards use those scores when making decisions that determine whether a person goes home or goes to jail.
The academic literature on these tools is mixed, to put it charitably. The ProPublica investigation of COMPAS in 2016 found that the tool’s error rates differed substantially by race. The company disputed the methodology. Courts have largely upheld the use of these tools as not violating due process, with the reasoning that a judge is still the one making the final call.
Virginia took that reasoning and turned it into a legal requirement. Yes, the judge is still making the call — and the statute now ensures that’s actually true, not just theoretically true.
What the Statute Prohibits
The prohibition is on AI-made decisions without human involvement. An AI system cannot:
- Automatically set bail conditions
- Issue a detention order
- Determine sentencing
- Make a parole determination
- Set probation terms
What it can do is generate a recommendation that a human then reviews. That human has to actually be exercising judgment — not rubber-stamping a score.
The last sentence of the statute is the one with practical bite: any AI recommendation or prediction is subject to any challenge or objection permitted by law. This means defendants and their counsel can challenge the AI’s output. They can request the underlying methodology. They can argue the tool is biased or inapplicable to their client’s circumstances.
This is significant because several courts have previously declined to require disclosure of proprietary algorithmic scoring, treating the vendor’s methodology as a trade secret. Virginia’s statute makes the challenge a legal right. Whether courts require disclosure of the underlying model in response to that challenge is the next fight.
The Practical Impact on Risk Assessment Vendors
If you’re selling a risk assessment tool to Virginia jurisdictions, your product just became advisory-only with legal force behind that characterization. That was probably already your pitch — “we generate recommendations for human review” — but now it’s the law, and a jurisdiction that lets an AI determination go unchallenged by a human decision-maker is violating the statute.
This creates a documentation problem for vendors. You need to be able to demonstrate that your tool’s outputs are being used as inputs to human decisions, not as final determinations. That means your clients need to maintain records showing that a named judicial officer reviewed the AI’s recommendation before acting on it.
The statute defines “artificial intelligence-based tool” as “any machine-based system or algorithm, including machine learning models, predictive analytics, and decision support systems, that analyze data and generate recommendations or predictions.” That’s broad enough to cover almost anything algorithmic. If your product has a model in it that affects criminal justice decisions, you’re in scope.
The Constitutional Dimension
The due process case against automated criminal justice decisions has been building for years. In State v. Loomis, the Wisconsin Supreme Court upheld the use of COMPAS but required courts to consider its limitations when relying on it. The court noted that the judge was still the decision-maker.
Virginia codified that reasoning and added teeth. The statute doesn’t resolve the constitutional questions — it sidesteps them by just requiring what due process arguably already demanded.
The interesting next question is whether Virginia courts start requiring meaningful engagement with AI challenges. If a defendant says “this risk score is methodologically flawed and my lawyer wants to cross-examine the model,” does the court give them that? The statute says the recommendation is “subject to any challenge or objection permitted by law.” That phrase will get litigated.
The law is right on the merits. AI should not be making criminal justice decisions without a human accountable for the outcome — both practically and constitutionally. Virginia put that principle in the code. Other states should follow.
You can find the original text of Virginia HB 1642 on the Virginia Legislature’s website.