Gains:
- Being able to distinguish where artificial intelligence saves time in the social work workflow (summary, editing, simplification, draft) and where clinical judgment, risk and representation decisions are left to the human, depending on the level of risk.
- Ability to apply a discipline that verifies each output through the steps of linking it to real evidence, confirming external information with current official sources, and filtering it through bias.
- To understand why client privacy, the distinction between anonymous and client data worlds, and KVKK responsibility should be taken into consideration from the very beginning in social work.
You are in the morning of a social service center. Three new applications arrived yesterday: a woman who was a victim of violence and her two children, an elderly person in need of care, and a teenager at risk of dropping out of school. There are four social investigation reports sitting on your desk, a court-requested opinion, case records that need to be filled out, and preparations for a home visit with a family this afternoon. Social work (the profession and discipline carried out to improve the social functionality of individuals, families and communities, to ensure their access to rights and welfare, and to protect vulnerable groups) is a field that is inherently document-intensive, time-pressured, touches human life and has high responsibility. Artificial intelligence (AI - software that can extract patterns from historical data and produce or classify text, summaries, translations and drafts) gives you a breather in this document load and time pressure. But the very beginning of this module is clear: AI is an assistant, summarizer and outline generator; You are the competent expert who makes the assessment, risk and final decision about the client.
In this first unit we will focus on discipline, not the tool. You'll learn where AI is a real time saver in the social work workflow, where it's dangerous, how to verify each output, how to protect client privacy and personal data, and how to stay alert to bias. Without laying this foundation, subsequent units remain in the air — because in social work, an unverified output is not just a false sentence; It is a mistake that could lead to a child being misjudged, a risk being overlooked, a family being unfairly stigmatized, or sensitive information being leaked.
Where does AI come in handy in the social work workflow?
Let's divide social work jobs into two large clusters. The first cluster: document-intensive, repetitive, pattern-extractable, and language-based tasks. Translating scattered notes after an interview into an organized case record, chronological summary of a long file, first skeleton of a social investigation report, simplification of an aid/service list to suit the client's language, draft of a referral letter, translation of a simple information text for a multilingual client, first framework of a program needs analysis. In these jobs, AI reduces hours to minutes and does not get tired.
Second cluster: tasks requiring judgment, clinical assessment, risk judgment and responsibility. Whether a child is safe at home, an individual's level of risk of self-harm, whether a family needs protective measures, the opinion with which a report will be presented to the court, whether a service is truly appropriate for the client. These decisions require professional competence, clinical judgment, context knowledge, ethical responsibility and often multidisciplinary team opinion. Here, AI multiplies the options, generates drafts, reminds you of what was overlooked — but the final evaluation and signature is yours.
Let's clarify the distinction in one sentence: AI is strong on "what's in this file and what does the first draft look like" questions; The decision is yours when it comes to questions such as "What is the real situation for this person and can I defend this on behalf of my profession?"
Tip: Before outsourcing a job to an AI, ask: “Who gets hurt if this output is wrong?” If the answer is "a few minutes of refactoring", delegate comfortably. If the answer is "wrong evaluation of a client, evasion of a risk or leaked information", let the AI produce the draft and you make the decision and verification.
Privacy: inviolable principle
The heart of social work is trust, and trust is based on confidentiality. A client tells you their most private, most fragile information—violence, poverty, health condition, family secrets—relying on professional confidentiality. When this information falls into the wrong hands, it threatens a person's life, security and dignity. The basic rule here: No personal data that identifies the client (name, address, TR ID, health, family information, etc.) will be entered into a public AI tool that your institution has not approved. The text you type into a public AI service is processed on that service's servers; You cannot control who sees it, where it is stored, and whether it is used in training the model. In Türkiye, this is also a legal liability under KVKK (Personal Data Protection Law - the law that regulates the processing of personal data).
Therefore, the use of AI in social work operates in two separate worlds: General tools can be used for the de-identified/generic world (anonymous or fictional examples that do not identify the client, general knowledge questions); For the world of client data, only in-house, secure, approved systems that do not leak data are used. Confusing the two is this module's most dangerous mistake.
Caution: "The AI said so" is not a justification and has no force in a case file. If there is a misjudgment, fabricated information or leaked data, the responsibility belongs not to the AI, but to the expert who uses that output without verifying it or enters that data into the tool.
Verification discipline: three steps
AI produces fluidly and confidently; That doesn't mean it's true. AI occasionally produces hallucinations — that is, it presents as real a piece of legislation that does not exist, a service that does not exist, a made-up statistic, or a statement that the client never said. In a social investigation report or risk memorandum, this is disastrous. Apply a three-step reflex to each output:
- Link to source/evidence. Every statement the AI makes about the client must be based on the actual note or conversation you gave. "In which conversation, from whose mouth, with what observation was this information given?" ask. AI should not add anything that the client did not say.
- Confirm with independent source. Verify external information such as law, rights, service, forwarding address, etc. from the current official source. An application requirement or phone number provided by AI may be outdated or fake.
- Pass it through the filter of interpretation and bias. Is the output biased against a group? Is there language that stigmatizes and blames the client? Does it ignore strengths? Your professional judgment is the final filter.
three mini cases
Case 1 — Documentation saved time. A social worker conducted 14 interviews in one week and kept loose hand notes for each one. AI-assisted editing reduced the total time in translating these notes into structured case records from approximately 4 hours to 1 hour. But each recording was read by the expert and verified by comparing it with the actual interview; Two sentences that the AI "interpreted" were deleted.
Case 2 — Verification caught a hallucination. An expert asked the AI a rights question; YZ said, "There is a minimum residence requirement of 24 months for this aid." When the expert looked at the official legislation, there was no such requirement; AI had mixed two similar aids and made up a condition that did not exist. The attribution step prevented the client from being unfairly turned away.
Case 3 — Return from privacy violation. To be quick, a new professional wanted to paste all the information of a child protection case (including the child's name, address, school) into a public AI tool and print a report. The senior expert realized: this was sensitive personal data within the scope of KVKK and was out. The incident was treated as a data breach; the job was redone with credentials removed and only in-house secure method.
Four copyable templates
1) Job suitability assessment:
Your role: senior social worker and supervisor.I will describe the job below. Tell me (1) whether this is a summary/editing/outlining job that can be delegated to an AI, or one that requires clinical judgment and risk judgment, (2) the cost to the client of incorrect output, (3) the verification I need to do before and after delegation. Job: [insert job here]
2) Obligation to rely on evidence:
I will give you the notes of a meeting. When editing the recording, use ONLY the information in the note. Do not add anything that the client has not said, do not diagnose, do not comment. Label the observation and the client's expression separately. Mark missing or unclear areas as "needs clarification".
3) Privacy pre-check:
Evaluate the following text BEFORE processing: could this text contain personal data identifying the client (name, address, TR ID, health, family information)? If so, which sections? As such, if it should not be processed in a public tool, warn me and list the fields I need to de-identify.Text: [write text here]
4) Bias/language check:
Consider the following case record: Is there language that stigmatizes, blames, or prejudices the client against a group? Does it ignore strengths and resources? Suggest language that is more respectful, fact-focused, and recognizes strengths. Don't change the record, just flag the language problems.
Weak prompt / Strong prompt
Weak prompt:
Write a report on this family and decide whether the child is safe at home.
This prompt is contextless and dangerous: it is unclear what information, what observation, and the AI is being asked to make a decision about a child's safety. AI predicts, can bias, can make up — and that involves a human life.
Powerful prompt:
Your role: assistant editor to social worker. Source: Interview and observation notes with identifying information removed, which I will paste below. Task: create a draft skeleton of a social investigation report based only on the information ACTUALLY contained in these notes; Label observation, client statement, and missing information separately. MAKING a risk or safety decision; I will do the evaluation. Mark the area you are not sure of as "needs to be clarified".
The difference is clear: de-identification, the "true-passing" constraint, labeling, and the "decision" constraint make the output safe and useful.
Role/task comparison chart
business
Role of AI
man's role
verification
Edit a case record
Configure the note
confirmation, accuracy
Comparison with note
File summary
chronological outline
Importance, comment
Link to source
Rights/service information
first draft
Current confirmation
official legislation
risk assessment
Question/frame reminder
Clinical judgment, decision
Team + supervision
Report
skeleton sketch
Content, opinion, signature
Evidence-statement matching
Program design
Idea/framework generation
eligibility, decision
Field + data confirmation
Common mistakes
- Mistaking AI output for evaluation. The output is always a draft to be verified; It does not enter the file without being linked to evidence.
- Pasting client data into the open tool. If it is leaked, it cannot be undone and is a violation of KVKK; It is the most dangerous mistake.
- Asking the AI to make a risk/safety decision. These decisions require professional authority and responsibility; AI cannot make decisions.
- A request that does not require evidence. If you don't say, "Just use what's in the note," the AI may add what the client didn't say.
- Not noticing stigmatizing language. AI's language may carry bias; Language that is respectful and considers strengths is under your control.
In summary
AI is a powerful assistant in social work: organizes the messy note, summarizes the file, produces drafts, simplifies, translates. But clinical judgment, risk assessment, representation and final decision belong to the human. Carry two-set separation (document-intensive work vs. judgment/risk decisions), three-step verification (link to evidence, independent confirmation, bias filter), privacy awareness (anonymous vs. client data world) and alertness to bias as the backbone of this module.
Application task
List 6 tasks from your own (or imaginary) desk. Classify each as “AI-delegable summary/edit/draft” or “clinical judgment/risk decision.” For one of the transferable ones (with a fictional example that does not identify the client), use the “Job suitability assessment” template above and get a response from the AI; then apply three-step verification and try the privacy precheck as well.
checklist
- [ ] I divided the work into two clusters (document-intensive / judgment-risk judgment).
- [ ] I applied three-step verification (evidence, independent confirmation, bias).
- [ ] I made sure that I did not provide personal data identifying the client to the open tool.
- [ ] I linked every statement the AI made about the client to the actual note.
- [ ] I checked for stigmatizing/biased language in the output.