Before You Trust AI With a Threat Assessment: Seven Questions to Ask Any Vendor

Artificial intelligence has arrived in threat assessment. In the last two years, a wave of new tools has appeared: AI note-takers, AI intake forms, AI decision support, AI chat screeners. Some are serious. Some are a text box with a disclaimer.

If you lead a school district, a clinic, or a threat assessment team, you will be pitched one of these tools this year. The names will differ; the claims will sound alike. These seven questions will tell you, in one meeting, whether you're looking at an instrument or an interface.

1. Is it validated in peer-reviewed research?

Ask specifically whether the tool has been validated: studied in a real population, measured against an outcome, published, and reviewed by independent researchers. Many vendors will use phrases like "built on established frameworks" or "informed by best practices," which describe their starting point, not evidence for their product.

Ask for the citation. A vendor with real validation will hand you a journal reference before you finish the question. A vendor without one will talk about the research their tool is "grounded in" (meaning someone else's research, about something else).

For context: independent validation research found that Homicide Zero's screening instrument identifies individuals with active homicidal ideation at 27 times the general-population base rate. That number exists because the study exists. Ask every vendor for their number.

2. Does it measure anything, or does it organize paperwork?

A surprising share of AI threat assessment products never assess anything. They collect intake forms, summarize documents, and export tidy case files. That's useful clerical work, but a well-organized folder is not a measurement, and it will not tell you whether the person in front of you needs intervention today.

Ask the vendor plainly: does your tool produce a result, or a document? Both have a place. Documentation is not assessment.

3. Was it built for lethal violence, or borrowed from somewhere else?

Most risk instruments in circulation were designed for general violence, recidivism, or self-harm, then stretched to cover targeted lethal violence. The research is clear that these are not the same phenomenon: the warning behaviors, base rates, and relevant risk factors are different.

An instrument purpose-built for lethal-violence risk asks different questions than a repurposed general screener, because it was designed around what actually precedes an attack.

4. Can you audit the score?

If a tool produces a risk category, ask the only question that matters in a courtroom, a board meeting, or an IEP dispute: show me how it got there.

Transparent, rule-based scoring can be walked backward item by item: this answer, this weight, this norm table, this result. A black-box model cannot. When a lawyer or a journalist asks why your team acted (or didn't), "the algorithm decided" is not a defensible sentence. Choose a tool whose every result can be explained by a human, to a human.

5. Does it weigh protective factors, or only count red flags?

An assessment that only tallies warning signs tends toward overreaction. Real structured assessment weighs what pushes risk down (connection, support, engagement, stability) alongside what pushes it up.

If a vendor's demo never mentions protective factors, you've learned what their instrument thinks a person is.

6. Who does it listen to?

A threat is never seen whole from one chair. The parent sees something the teacher doesn't; the clinician sees something the parent can't. A serious instrument gathers multiple raters and knows they differ, scoring each perspective against norms for that kind of rater rather than pretending a mother and a school counselor mean the same thing by "often."

Single-perspective tools aren't wrong; they're just one chair.

7. Where does the AI stop?

This is the question that separates the serious from the dangerous. In a defensible system, AI carries the logistics (the conversations, the languages, the scheduling, the structured capture) at a scale no human team can match. The measurement comes from a validated instrument. The decision belongs to a trained professional, every time.

Ask the vendor to draw that line. If they can't say precisely where the AI stops and the human begins, walk away. If the honest answer is that AI does the judging, walk away as well. And be equally wary of the opposite: an AI that merely chats, attached to nothing validated at all. A conversation without an instrument behind it is a transcript, not a screening.

The questions share a focus

None of these seven questions ask what the AI can do. They ask what the instrument can prove, and where the human sits.

AI intake is becoming a commodity. Within a few years, every tool will talk to people by phone, text, and chat, in any language. What will still be rare is what was always rare: a validated instrument with transparent scoring and results a professional can defend.

That's the standard we hold ourselves to at Homicide Zero: a screening and assessment platform purpose-built for lethal-violence risk, with peer-reviewed validation, auditable rule-based scoring, protective factors weighed alongside critical items, multiple raters with rater-specific norms, and a structured path from determination to intervention. The final call always belongs to a person. Ask us these seven questions; we'd enjoy it.

Ask them of everyone else, too.

Homicide Zero provides the HTS screener and the VSRA (Violence Safety and Risk Assessment) to schools, clinicians, and threat assessment teams.