Key takeaways
- AI writes answers; it doesn't look them up. A search engine points you to pages that exist. A prompt asks a model to compose an answer from patterns, which is why it can draft and summarize well and also state things that are false.
- Hallucinations are the risk to manage first. AI can invent a statute, a policy section, or an address and present it with total confidence. Nothing it produces should reach a report, a policy, a procurement, or an operational record until a person has checked it against the source.
- Your data questions have answers, but only per product. Whether AI keeps your data or trains on it depends on the version, the contract, and how it's configured, not on "AI" in general. Free consumer tools and enterprise tools follow different rules, so get the terms in writing.
On June 22, 2023, a federal judge in Manhattan fined two lawyers and their firm $5,000. Their brief in Mata v. Avianca cited court decisions that didn't exist. ChatGPT had produced them, and when one of the lawyers asked whether the cases were real, it said yes.
Those were experienced attorneys working in their own field. The tool sounded right, so they didn't open the sources.
The same kind of tool is now one browser tab away in emergency communications centers (ECCs), police and fire administration, EMS agencies, and emergency management offices. People are using it to draft policies, summarize grant guidance, write after-action reports, and clean up data. Many of them have questions they haven't asked out loud. These are the ones I hear most, with plain answers.
What is a prompt, and how is it different from a Google search?
A search engine finds pages that already exist and gives you links. You decide which source to trust, and the source is right there.
A prompt is an instruction to a large language model (LLM), the kind of AI behind tools like ChatGPT, Claude, Copilot, and Gemini. The model doesn't retrieve a page. It writes a new answer, word by word, based on patterns it learned from a very large body of text. Some tools also search the web or read documents you give them, but the answer is still written by the model.
That difference explains both the value and the risk. A model can summarize a 60-page grant notice, draft a policy section, or turn incident notes or a transcript from a hot wash into a first draft of an after-action report. It can also produce a confident paragraph that no source supports.
A search shows you where an answer came from. A prompt hands you an answer and leaves the sourcing to you.
What is an AI hallucination?
A hallucination is when an AI tool states something false as if it were fact. It isn't lying; it's predicting what a good answer would look like, and sometimes the most likely-sounding answer is wrong.
In public safety work, hallucinations tend to show up in the details that matter most:
- Laws and rules. A statute, regulation, or FCC rule that doesn't say what the AI claims, or doesn't exist.
- Standards and guidance. An NFPA, NENA, or state standard cited with the wrong number, version, or requirement.
- Policy references. A section of your own SOPs or general orders that it "quotes" but that isn't there.
- Names and places. A street, a facility, an agency, or a mutual aid partner that sounds plausible and isn't real.
- Numbers. Dates, deadlines, response times, counts, and dollar amounts that are invented or rounded into something new.
- Summaries. A summary that drops the one condition that changes the meaning.
Treat every AI answer as a first draft from a confident new hire. Then hold it to the standard you'd use for that hire:
- Open the source. If the answer cites a law, rule, or standard, read that section yourself. If it can't point to a source, don't treat it as fact.
- Check what you'd be embarrassed to get wrong. Every citation, number, name, and date.
- Keep AI out of the record until a person signs off. No AI output goes into a policy, a board or council report, a grant application, a procurement document, or an operational record unreviewed.
- Be most careful when it agrees with you. A plausible answer that confirms what you expected is the easiest one to skip checking.
Tools that show their sources, like our Ground Truth checker, make that check faster. They don't remove it.
If you can't verify it, don't use it.
What does "agentic" AI mean?
Most people start with prompting: you ask a question, the AI answers, and you decide what to do next. Every step goes through you.
Agentic AI is given a goal and a set of tools, and it works through the steps on its own. It might search the web, open files, run a query, compare results, notice a problem, and try again before reporting back. The difference is like asking a colleague a question versus handing them a task.
That's useful for multi-step work, such as comparing your agency's policies against new state requirements and listing every gap, or reconciling two equipment inventories. It also raises the stakes, because an agent can act on what it finds automatically. A good rule for public safety: an agent can read and propose, and a person approves anything that changes a real record.
What does "adaptive" AI mean?
Adaptive AI changes its behavior over time based on new data or feedback. Agentic is about what a system does; adaptive is about whether it learns.
The two get used interchangeably in marketing, so ask what the vendor means. Most general chat tools today don't update the underlying model from your conversation as you go. Some have a memory feature that saves notes about you, which personalizes answers but isn't the model learning. Other products are retrained or tuned on new data on a schedule.
When a vendor says "adaptive," ask three things: what changes, how often, and who checks the results after each change. A system that behaved correctly at acceptance testing may behave differently six months later. For anything that touches response or life safety, a system that keeps changing needs to keep being verified.
Is my data secure when I use AI?
It depends on the product and how your agency has it set up, not on "AI" in general.
What you type or upload is sent to the vendor's servers to be processed. Three questions decide most of the risk:
- Where is it processed? In the United States, or anywhere the vendor operates?
- How long is it kept? Some services keep prompts for a period for abuse monitoring; some business setups offer zero retention.
- Who can see it? Vendor staff, subcontractors, or no one?
Even with a business license, the answers depend on configuration. An administrator decides whether chat history is saved, how long data is kept, whether the tool can connect to email or shared drives, and who can see what. Ask for the settings in writing, not just the license. A tool your agency licensed, configured, and documented is a very different thing from a personal account someone signed up for on their lunch break.
Can AI leak my data?
The most common leak isn't a breach; it's what people paste in.
Caller and complainant information, CAD (computer-aided dispatch) and records management system (RMS) data, incident and patient care reports, staff rosters, and anything your state treats as confidential or exempt from public records need protection. Criminal justice information falls under CJIS (Criminal Justice Information Services) requirements, and EMS patient information may be covered by HIPAA, before any AI question comes up.
Which account you're in matters. A free or personal account runs on the provider's consumer terms. An enterprise account your agency licensed runs on the contract your agency signed and the settings your administrator chose. If it's a free or personal account, treat the prompt box like an email to an outside vendor with no agreement in place: sensitive data doesn't go in. If it's your agency's enterprise tool, follow your agency's AI policy for what's allowed, and if there's no policy yet, wait until there is one.
Do AI companies use my data to train their models?
Some do, depending on the version. Training means using your conversations to improve future versions of the model.
Consumer versions of some popular tools can use your conversations for training unless you turn that setting off. OpenAI's ChatGPT consumer plans work this way, with a self-serve opt-out. Business, enterprise, and API (application programming interface) versions generally don't train on your data by default; Anthropic's commercial terms, for example, say commercial inputs and outputs aren't used for training unless the customer opts in.
It doesn't have to train on your data. Agencies can choose setups with no training, no retention, and U.S.-only processing. These terms change, so check the current version for the exact product you use, and put the commitment in your contract rather than relying on a web page.
Questions to ask before your agency approves an AI tool
Take these into the vendor meeting, or into your own AI policy review:
- Does it train on our data? Get "no" in the contract, not just the FAQ.
- How long is our data kept, and can it be zero?
- Where is it processed? Ask for U.S.-only if your agency needs it.
- How is it configured, and who controls the settings? History, retention, sharing, and connections to other systems.
- Which model runs it, and who provides that model? Ask what happens when the model is updated or replaced.
- Does it show its sources? You need to be able to check every answer.
- What can it change on its own? Read-only, propose-only, or able to write to live records?
- If it's "adaptive," what changes and who re-tests it?
- How does it keep CJIS, HIPAA-covered, and other protected information out, or protect it if it's allowed in?
Vendors should expect these questions. The ones who answer them in writing, before they're asked, will earn trust faster.
Bottom line
AI can save public safety agencies real time on drafting, summarizing, and comparing. It can also put a confident, wrong answer into a record people rely on in an emergency. Use it, keep a person responsible for every result, and get the data terms and settings in writing.
The tool can write the answer. Your agency still owns it.
If your agency is working through an AI policy or evaluating a tool, and you have a question you didn't want to ask in the room, send it to me. I'd rather answer it than have you guess.
Frequently asked questions
What is the difference between a prompt and a Google search?
A Google search finds web pages that already exist and links you to them, so you can see the source. A prompt asks a large language model (LLM), such as ChatGPT or Claude, to write a new answer from patterns it learned. That makes AI useful for drafting and summarizing, such as condensing a 60-page grant notice, but it also means the answer may not come from any source at all. For public safety agencies, treat a prompt's answer as a draft and confirm every citation, number, and name against the original document before using it.
What is an AI hallucination?
An AI hallucination is a false statement that an AI tool presents as fact, such as a court case, statute, or policy section that doesn't exist. It happens because the model predicts what a good answer looks like rather than looking one up. In 2023, in Mata v. Avianca, a federal court fined two lawyers $5,000 after they filed a brief citing cases ChatGPT had invented. For public safety agencies, the rule is simple: verify every AI-provided citation, address, and number against the source before it reaches a report, policy, or operational record.
What is the difference between agentic AI and adaptive AI?
Agentic AI is about what a system does: it's given a goal and tools, and it carries out multiple steps on its own, such as searching, opening files, and comparing results. Adaptive AI is about whether a system learns: it changes its behavior over time based on new data or feedback. A tool can be one, both, or neither. For a public safety agency, agentic tools should propose changes for a person to approve, and adaptive tools should be re-tested after each change, because a system that passed acceptance testing may behave differently later.
Does ChatGPT or other AI use my data for training?
It depends on the version. Some consumer versions, including ChatGPT's consumer plans, can use your conversations to train future models unless you turn that setting off. Business, enterprise, and API versions from major providers generally don't train on customer data by default. Because terms change, check the current policy for the exact product your agency uses. Before any agency data goes into an AI tool, get a written commitment covering training, data retention, and where the data is processed.
Is it safe to put public safety data into an AI tool?
Not into a free consumer tool. Caller information, CAD and RMS records, incident and patient care reports, and anything confidential under state law should stay out unless your agency has an enterprise agreement and settings covering retention, training, and processing location. Criminal justice information also falls under CJIS requirements, and EMS patient information may be covered by HIPAA. The account type matters: a free or personal account runs on consumer terms, while an enterprise account follows your agency's contract and its administrator's settings.