AI assistants can save investigators and attorneys hours. They can also leak confidential case material. A practical guide to drawing the line.
Every case team has had the thought by now. The file is enormous, the deadline is close, and somewhere a general-purpose AI assistant is sitting in a browser tab, promising to summarize anything in seconds. Paste in the incident report, the witness statements and the lab results, and ask for a tidy summary. What could go wrong?
Quite a lot. Case files contain some of the most sensitive information any organization handles: names of victims and witnesses, people who are never charged, juvenile information, medical details, criminal history, sealed material and investigative strategy. Once that information is pasted into a commercial AI service, the agency may no longer control where it is stored, who can access it, or how long it is retained. Depending on the material and the tool, that can create legal, ethical and practical problems that no amount of saved time is worth.
This article explains where the line usually sits, why it matters and how case teams can still get real value from AI tools without crossing it. It is general information, not legal advice. Your agency’s policies, applicable law and your own professional obligations always come first.
Why case data is different
Most workplaces worry about AI tools and trade secrets. Case teams carry a heavier load. Their files can include:
- Criminal justice information covered by strict security requirements, such as the FBI’s CJIS Security Policy in the United States
- Personal information about victims, witnesses and bystanders who never asked to be part of a case
- Information about juveniles, which is often specially protected
- Medical records, mental health information and other sensitive categories
- Sealed, expunged or court-protected material
- Investigative techniques and strategy that could compromise a case if disclosed
- Privileged attorney work product on the defense or prosecution side
Handling these categories carelessly can harm real people, jeopardize prosecutions or defenses, violate law and policy, and damage public trust in the justice system.
The basic rule
A simple rule covers most situations: if your agency has not approved a tool for a category of data, that data does not go into the tool. Approval typically means the tool has been reviewed by IT, security and legal staff, that appropriate contracts and security controls are in place, and that its use fits applicable policies. A personal account on a public chatbot almost never meets that standard for real case material.
Caseprep itself takes this approach seriously: highly sensitive media is handled through a separate, locally controlled pathway rather than a general commercial service, and agencies review data handling before any real case material is used. That same mindset is useful whenever you consider any AI tool.
What usually should not be pasted
Unless a tool is specifically approved for it, avoid pasting:
- Names, addresses, dates of birth, phone numbers or identification numbers of anyone involved in a case
- Incident reports, statements, transcripts or evidence descriptions from real cases
- Photos, video, audio or screenshots from case files
- Criminal history records
- Medical or mental health information
- Juvenile records
- Sealed or protected material
- Internal strategy memos, charging discussions or privileged communications
- Login credentials, system screenshots or internal network information
“I removed the names” is not always enough. Details like locations, dates, unusual facts and descriptions can identify people even without names, especially in smaller communities or high-profile cases.
What is often reasonable
AI tools can still help case teams with a large amount of work that involves no confidential data at all. Examples include:
- Drafting templates for reports, checklists and memos, filled in later inside approved systems
- Creating training scenarios using clearly fictional facts
- Explaining general concepts, terminology or procedures, with verification against authoritative sources
- Improving the clarity of writing that contains no sensitive details
- Brainstorming interview question categories for a generic type of case
- Organizing personal productivity tasks, like scheduling or to-do lists, without case details
- Drafting grant applications, policy summaries and public outreach materials
Even here, follow agency policy. Some agencies restrict or prohibit certain tools entirely.
Use fictional or synthetic examples
When you want help with a type of task, like structuring a case summary, build a fictional example that captures the pattern without real details. Caseprep’s own demo uses a fictional 437-item case for exactly this reason. You can learn how a tool behaves, design a good template or refine a workflow, then apply the approach to real material only inside approved systems.
Understand the tool’s data practices
AI services differ widely in how they handle data. Questions your agency’s IT and legal teams typically consider include:
- Is data used to train the provider’s models?
- Where is data stored, and for how long?
- Who at the provider can access it?
- What security certifications and controls are in place?
- Is there a contract with appropriate data protection terms?
- Can the service meet requirements like CJIS or other applicable standards?
- How are logs, backups and deletions handled?
Individual users are rarely in a position to answer these questions alone. That is why centralized review matters.
Personal accounts are a common risk
One of the most common problems is well-meaning staff using personal AI accounts for work tasks. It feels harmless, especially late at night when a deadline looms. But personal accounts sit outside agency controls, and whatever is pasted may be retained under consumer terms. If your agency offers an approved tool, use it. If it does not, raise the need through proper channels rather than improvising.
Build a simple team policy
If your team does not have clear guidance yet, advocate for one. A practical policy usually covers:
- Which tools are approved, and for what data
- What categories of data are prohibited in which tools
- How to request approval for a new tool
- How outputs must be reviewed before use
- Who to contact with questions or incidents
- Training requirements
Keep it short enough that people actually read it, and include concrete examples.
Review every output
Even when data handling is appropriate, AI outputs need careful human review. Tools can misread facts, invent details, miss context and present guesses as conclusions. In case work, an error in a summary can mislead a supervisor, a prosecutor, a defense attorney or a court. The person using the tool remains responsible for the accuracy of anything that goes into a file or a filing.
If something goes wrong
If you realize that sensitive information was pasted into an unapproved tool, report it promptly according to your agency’s incident procedures. Do not try to quietly fix it. Early reporting allows the agency to assess the situation, contact the provider if needed and meet any notification obligations.
A quick checklist before you paste
- Is this tool approved for this kind of data?
- Does this text include names, identifiers or unique details?
- Could someone identify a person or case from it?
- Is any of it sealed, protected, privileged or juvenile information?
- Could I achieve the same goal with a fictional example?
- Will I review and verify the output before using it?
If any answer gives you pause, stop and check.
Questions case teams ask
Is it okay if I remove names and case numbers?
Sometimes, but not always. Unique facts can still identify people. Follow your agency’s guidance on de-identification, and when in doubt, use fictional examples.
What about public information, like a published court opinion?
Publicly available documents are generally lower risk, but check policy, and remember that combining public information with internal notes can create sensitive material.
Can AI tools ever be used on real case data?
Yes, when the tool and the use have been properly reviewed and approved by the agency, with appropriate security, contracts and oversight in place.
Key takeaways
Treat case data as the most sensitive information you handle. Use only approved tools for real material, avoid personal accounts, use fictional examples to learn and design workflows, review every output, and report mistakes promptly. AI can assist the people working a case; it should never put the case, or the people in it, at risk.
A short exercise for your team
One of the fastest ways to make these rules stick is a fifteen-minute team exercise. Print a fictional case document, such as a mock incident report, and ask each person to mark everything that would need to be removed or generalized before the text could go into a public AI tool. Compare answers. Most teams discover that people disagree about indirect identifiers, such as a distinctive vehicle, a small-town business name or an unusual job title, and that conversation is exactly the one worth having before a real file is involved.
Getting value from AI the safe way
The good news is that a large share of everyday case-team work, like templates, training materials, writing clarity and checklists, does not require any confidential data at all. That is where many teams start getting real value from AI assistants like ChatGPT, Claude and Gemini today, within agency policy.
The difference between a useful result and a useless one is usually the prompt. A vague request produces generic filler; a well-structured prompt that specifies the role, the format, the audience and the constraints produces something you can actually use. Learning to write those prompts well, with fictional placeholders instead of real details, is one of the most practical skills a modern case team can build.
General information only. Follow your agency’s policies on data handling and approved tools, and never place confidential case material in a tool that has not been approved for it. Caseprep does not provide legal advice.