Effective AI Policy for Probation

Probation's AI Dilemma

Probation leaders are at a crossroads. Your officers desperately want to use AI and they have real examples of tasks it can save them time on.

You totally understand where they're coming from so you want to give them access to tools that will make their work more efficient.

But you're concerned about data security, and rightfully so! So you've probably told your officers they can't use tools like ChatGPT, Claude, Gemini, or Grok to help them with their work.

This sucks, because much of the work your officers do can be dramatically accelerated with AI.

You know it will improve morale, you know this will help improve outcomes, and you know it's going to revitalize the work your team does.

The reality is - you can't take the risk of having officers use consumer AI without well thought out policies. So you're stuck.

There's good news though. CJIS compliant options do exist. And, even if you're not in a position to afford a CJIS compliant AI solution, there's steps you can take to create policies that keep your data secure while using consumer AI applications.

In this blog we'll discuss your options and how you can start implementing AI with your probation officers.

Data Security + AI

Before we talk about what your options are for using AI securely with CJI, I think it's important for you to understand how the data you share with AI providers is used and what CJIS requirements really are so you can start thinking about robust policies for your team to be able to use AI safely.

How AI Providers Use Your Data

There's a common misconception about "Open" vs "Closed" AI. It all stems from the company "OpenAI".

"Open" refers to a software engineering philosophy called "open source". Open source is all about the code - not the data. Open source means the code written to power the model is publicly accessible. So if you want to download the code and run it on your computer you could.

Open source does not mean your data is shared with anyone that uses the AI model you use. This is where the common misconception about your data being shared publicly comes from - people misconstruing open source with open data sharing.

That being said, the concern about how AI providers use your data is real. While AI providers don't just take your data and share the raw details with anybody who wants it, they will do two things with it:

• They will use your data to train their new models
• They will evaluate your conversations with AI to measure performance

This means by default any conversation you have with a consumer AI chatbot will result in your data being used to improve the performance of the AI you're using.

To be clear though, this does not mean if you tell an AI system, "John Doe lives at 123 main st with his mom and sister" that another user of the AI system will be able to ask a question about where John Doe lives and get the address you put in.

AI models are trained on trillions of lines of text. While your data may be a piece of that, the model turns that into generalized parameters for how to answer questions in the future.

Essentially it learns that if someone asks where a person lives, it should respond with an address - not necessarily an exact address.

So when you use one of these AI providers, remember that your data is not shared with the public. BUT, that does not mean your data is private to you and being accessed in a CJIS compliant manner.

What Does CJIS Compliance Really Mean?

We keep talking about CJIS but what does CJIS really mean? CJIS is one of the most highly misunderstood security and data compliance standards that exists.

The CJIS (Criminal Justice Information Services) not only specifies the technical requirements of software systems, it specifies the operational and access requirements for systems AND agencies that store and process CJI.

That's because the data being protected is incredibly sensitive. It includes the details of someone's personal criminal history and it also contains information about victims, property, and biometrics. This is the kind of information that inherently deserves the highest degree of protection.

While we won't go into all of the CJIS details here (there are over 450 "controls") we'll cover the major principles of CJIS and what they mean for you.

Warning: I'm going to be overly simplistic here. If you want to understand the details of the CJIS policy check it out here.

CJIS covers a few critical areas for defining compliance:

• Technical Data Protection: Encryption methods, firewalls, multi-factor authentication, passwords, etc.
• Access Controls: Only authorized people can access CJI (fingerprint + background check).
• Auditability: A clear way to show who accessed what and when.
• Physical Protection: Unencrypted CJI must be accessed in a physical environment isolated from public access.
• Incident Response: Plan to respond to actual or suspected data breaches.
• Data Dissemination: Only share CJI with authorized parties using approved channels.

Again, this is an oversimplification but the goal of the CJIS policy is to limit the exposure of CJI as much as possible.

But, there's a common misconception that any data related to corrections or community supervision qualifies as CJI.

This is not the reality. And this is where you can take steps to securely use consumer AI tools to help you with your work as a probation officer.

There's 5 categories of data the FBI classifies as CJI:

• Biometric Data: Data derived from intrinsic physical or behavioral traits used to identify people, including fingerprints, palm prints, iris scans, and facial recognition data.
• Identity History Data: Textual data corresponding with biometric data, providing a history of criminal and/or civil events for the identified individual.
• Biographic Data: Information about individuals associated with a unique case, not necessarily connected to identity data. The policy says this does not provide a history of the person, only information related to a unique case.
• Property Data: Information about vehicles and property associated with crime when accompanied by personally identifiable information.
• Case / Incident History: Information about the history of criminal incidents.

Most of the data your officers will handle in day-to-day operations probably won't qualify as CJI.

Some examples of non-CJI data your team interacts with include:

• A risk assessment (assuming you do not include case details in it)
• UA results
• GPS tracking data
• Interlock reports
• Chronological notes (assuming CHRI is not added)
• Progress reports

This absolutely does not mean that officers can do whatever they want with this data. It still deserves to be protected as much as possible. But it is important to understand that this data is not inherently CJI.

How to Create AI Policy for Your Agency

So how can you create policies that let your team use AI tools while keeping your agency data secure? Well frankly, it's hard. It's hard because you have to rely on trust. And it's hard because it adds manual steps to your officers work.

But, it can be done, and it does have massive benefits for your team!

If you're able to afford new software and want to explore AI systems created specifically for probation, options like CaseBase are CJIS compliant out of the box. That means you get all of the benefits of AI without having to create new security policies for your officers or having to trust that they'll always follow them.

CaseBase is an AI assistant that does things like turn office visits into case notes, help write reports, personalize case plans, and even assess things like interlock reports for violations.

The agencies using CaseBase across the US save an average of 5+ hours per officer per week on documentation and assessment tasks.

If you're not in a position to pay for new software, let's talk about what you can do to allow your officers to use free tools in a secure manner.

Safety Settings

When officers use an off the shelf AI tool, the first thing you need to make sure they do is create an account and change their data sharing settings.

By default, when you log into a consumer tool like ChatGPT any data you send the AI will be used for training. But the good news is, if you create an account you can turn this off.

To turn this off, have your officers go to "settings" and then look for something like "improve the model for everyone". This is usually code for, "We'll use everything you send us to train a new model".

Typically there's just a switch to turn this off. This is only possible if your officers have an account!

Data Restrictions

The second most important thing you must do in order to protect your agency data is enforce that officers cannot share any CJI into a consumer AI system.

No matter what, if your officers use the consumer version of ChatGPT, Gemini, Claude, Grok, etc, these systems are not CJIS compliant.

That means even the act of adding a piece of CJI into one of these systems is an automatic violation of the CJIS policy established by the FBI.

The good news is, it's somewhat straightforward to identify the pieces of CJI your agency accesses and provide that list to officers as a "do not share list".

While your list may include more than the following, it should always include at least:

• Criminal history records
• Judgements or court filings
• Police reports
• Vehicle or property data like license plates involved in a crime
• Fingerprints, DNA samples, or pictures of your clients
• Case/cause numbers

Again, this is the smallest set of things you must restrict but your list may include more.

Generally, the next most important thing you must enforce is that officers do not share PII. This one is somewhat simple. If your officers are going to put anything into AI they must always redact names, addresses, birthdates, and government identifiers.

Use Cases

So you're probably wondering, if I have to restrict my team from entering any of that information into an AI system, what value will my officers even get from it?

You're in luck! Last week I wrote a blog about that very thing here. I give you 3 examples of workflows your officers can set up from start to finish that keep your data secure, abide by the policies we suggest above, and still help them save tons of time while getting better results.

Some examples include evaluating interlock reports, writing case notes (chronological notes), and building a case plan. These are things your officers do frequently that AI can help them save a significant amount of time.

Wrapping Thoughts

Your officers are going to use AI whether you tell them they can or they can't. It's the most transformational technology change for probation officers maybe ever. And they can feel it.

What you need to prevent is misuse. And you can do this! I know you can! You can help shape the work your officers do while assuring they do it securely.

If you're able to get ahead of this you'll see incredible results. Less burnout, better outcomes, and a safer community.

If you have any questions about using AI safely in probation or how you can start building policy to support your team's use of AI don't hesitate to reach out to me at tyler@casebase.co.