Gains:
- Ability to prepare the risk and protective factor checklist and security plan skeleton with artificial intelligence and work without skipping any areas
- Understand that risk level/score and security decisions are made only by experts, teams, supervision and legal processes, and that artificial intelligence cannot produce decisions.
- Seeing that artificial intelligence can attribute systematic bias to disadvantaged groups and applying the discipline of never using its output as a decision and processing the most sensitive data only in a secure system
Social work's most serious responsibility is risk assessment—judgments that directly concern life safety, such as whether a child is neglected or abused, whether an individual is likely to harm himself or someone else, whether an elderly person is safe, or how endangered a victim of violence is. The first and most bold rule of this unit is this: Risk decision-making does not, cannot and should not belong to AI. Decision about a person's safety; It requires professional authority, clinical judgment, multidisciplinary team opinion, supervision and legal responsibility. AI is just a preparation and reminder tool here: it reminds you which risk and protective factors to review, organizes your notes, outlines a safety plan. It doesn't give a score, it doesn't say "the risk is low/high", it doesn't make a decision.
Why do we emphasize this so harshly? Because in the risk arena, the cost of an AI mistake is irreversible. AI's rating of a risk as "low" leaves a child unprotected; Making a risk seem unnecessarily "high" could lead to a family being unfairly separated. Additionally, AI carries biases in training data: algorithms that systematically attribute higher risk to certain socioeconomic groups, ethnicities, or neighborhoods have been documented around the world. So even if the AI's output is an input, it can never override your independent judgment.
The legitimate role of AI: reminder and regulation
There are three places where you can safely use AI in risk:
1. Factor reminder. The risk factors (e.g., past history of violence, substance use, social isolation, previous admissions) and protective factors (e.g., a supportive adult, regular school attendance, a strong social network) that should be considered in a risk assessment are long; AI helps you avoid missing any areas by recalling a checklist that fits the framework your organization uses.
2. Note editing. It translates your observation and client statements into an organized, factual text — so you can base your judgment on clear data.
3. Draft security plan. A security plan (a concrete plan that determines what the victim will do in case of danger, who he will contact, where he will go) has a certain structure. The AI can produce a draft skeleton to be filled in with the client; But you and the client fill in the content of the plan according to reality.
Caution: Never ask the AI "what is the risk level in this case?" or “is the child safe at home?” Don't ask. DO NOT use the output as a decision, even if you ask. These questions are answered only by the authorized expert, team and supervision through the institution's official risk tool.
Step by step: positioning AI correctly in the risk process
- Safety first. If there is an immediate and serious danger (life safety), it is necessary to operate the institution's emergency procedure, not to deal with AI. AI should not delay any urgency.
- Disidentify. Risk ratings are the most sensitive data; Information identifying the client is not entered into the open tool, only the secure institution tool is used.
- Remind me of the factor list. Request a risk and protective factor checklist appropriate to the institutional framework.
- Make your own judgement. You weigh the factors, evaluate with the team, and consult with the supervisor. You don't want the AI to "evaluate".
- Complete the safety plan with the client. The draft skeleton may come from the AI; The content is populated with the client's actual network, resources, and preferences.
three mini cases
Case 1 — Missed protective factor. An expert focused on risk factors in a domestic violence case. The factor checklist he asked AI for had him ask about the client's secure support network at work in the "protective factors" section—an area the expert had missed. This network became the most realistic part of the security plan. The decision was again up to the expert; AI merely reminded the question.
Case 2 — Incorrect score was not trusted. One agency saw a trial AI tool label a case as “low risk.” However, the expert's observation and the client's statement pointed to a serious danger. The team rejected the AI label and initiated protective action based on its own assessment. It was later confirmed that the danger was real. Lesson: AI score is not a decision; human judgment takes precedence.
Case 3 — Bias caught. In one trial, the same case was given to the AI twice; The only difference was the neighborhood where the client lived. The AI systematically produced higher “risk language” for the neighborhood known to be disadvantaged. This demonstrated bias in the AI's training data and reinforced the agency's policy of not using AI in risk decision-making.
Four copyable templates
1) Risk/protective factor reminder:
Your role: assistant to the social work supervisor. Prepare a checklist of risk factors and protective factors that are recommended to be reviewed in a risk assessment in [domestic violence] cases. Present this as a list of REMINDERS; do NOT produce any scores, decisions, or "risk levels". Evaluation and decision belong to the expert.
2) Security plan draft skeleton:
Your role: editor. Prepare a BLANK outline of a safety plan to be filled in with a victim of domestic violence: warning signs, safe places, people and institutions to contact, important documents, plan for children, digital safety. Filling out the content; Leave blank spaces under each heading to be filled in with the client.
3) Editing a risk note (without making a decision):
Translate the following observation and statement notes into a factual, organized text. Label observation and client statement separately. STATING the risk level, MAKING a decision, ADDING a comment. Edit only what is in the note and mark missing/obscure areas. Note: [paste de-identified note]
4) Bias self-control:
In the draft risk note below, is there a statement that conveys unproven bias because of the client's socioeconomic status, ethnicity, neighborhood, or similar characteristic? Mark each sentence that is based on stereotypes rather than evidence. Don't change the text, just warn. Note: [paste note]
Weak prompt / Strong prompt
Weak prompt:
Read this case and decide whether it is safe for the child to stay at home and what the risk level is.
This is outsourcing the most weighty decision about a human life to a software. It is unethical, contrary to professional responsibility, and carries the bias of the AI directly into a child's fate.
Powerful prompt:
Your role: reminder assistant. Before reading the de-identified note below, prepare a checklist of risk and protective factors to review in the evaluation of [child neglect] cases. PRODUCING decision, score or risk level. I will make the evaluation with my team; You just remind me which areas I shouldn't skip.
The difference: saying “which questions should I not forget” rather than asking for decisions keeps the AI on the safe and ethical frontier.
Role boundaries in the field of risk
Quest
Can AI do it?
Who does?
Factor checklist reminder
Yes
AI (reminder)
Editing notes factually
Yes
AI + expert approval
Security plan blank skeleton
Yes
AI; content client+expert
Risk level/score determination
no
Expert + team + supervision
"Is it safe?" decision
no
Competent expert, legal process
Initiating protective measures
no
Institution, competent authority
Common mistakes
- Asking AI for risk score/decision. The most dangerous mistake; The decision requires professional authority and legal responsibility.
- Trusting the AI label. A "low risk" conclusion will never invalidate your observation to the contrary.
- Ignoring prejudice. AI may attribute systematic high risk to disadvantaged groups; self-control is essential.
- Delaying urgency. If there is life safety, emergency procedure first; Dealing with AI cannot wait.
- Giving the most sensitive data to the open tool. Risk notes are processed only in the secure institution tool.
In summary
Risk assessment is the gravest responsibility of social work, and the decision rests squarely with the person — the professional, the team, the supervision, the legal authority. Here, AI is just a tool for reminding (risk/protective factor list), organizing (factual note) and drafting (safety plan skeleton); It doesn't score, it doesn't make decisions. Remember that AI carries bias, never use its output as a decision and process the most sensitive data only in the secure system.
Application task
Choose a fictional case type. Get a checklist with a “risk/protective factor reminder” template and verify that the AI does not produce any “risk levels.” Then grab a blank plan skeleton with the "security plan outline skeleton" template. Consciously asking the AI “what is the risk level?” and write in one sentence why you should NOT use the output as a decision.
checklist
- [ ] I only asked the AI for a factor reminder, not a decision/score.
- [ ] I processed the most sensitive data only in the secure enterprise tool.
- [ ] I planned to fill out the contents of the security plan together with the client.
- [ ] I checked the AI output for bias.
- [ ] I confirmed that the final risk decision rests with the team, supervision and due process.