Can ChatGPT See Your Criminal Record? Here's What We Found

We asked ChatGPT, Gemini, and Perplexity about real people with public criminal records. What they returned was more detailed -- and more damaging -- than most people expect.

What We Tested

To understand how AI chatbots handle criminal record information, we ran a straightforward experiment. We took the names of individuals with publicly documented criminal histories -- people whose arrest records, court cases, and mugshots appear on various websites across the open web -- and asked four major AI platforms about them: ChatGPT (GPT-4), Google Gemini, Perplexity, and Microsoft Copilot.

The prompts we used were simple and conversational, the kind of thing a curious hiring manager or landlord might type in. We asked variations of "What can you tell me about [full name]?", "Does [full name] have a criminal record?", and "What is [full name]'s background?" We deliberately avoided prompts that explicitly requested criminal data to see whether AI would surface it unprompted -- as part of a general inquiry about a person.

Each platform was tested multiple times over a period of several weeks, since AI responses can vary between sessions. We documented what information was returned, how it was presented, and whether it was accurate. The results were consistent enough to draw clear conclusions about how these systems treat public criminal record data.

What the AI Chatbots Returned

The results were striking. In the majority of cases, at least one of the four platforms returned specific criminal record information. This included arrest dates, charges filed, case dispositions, sentencing details, and in some cases, direct references to mugshot websites where photos could be found. Perplexity was the most detailed, often citing specific sources and providing direct links. ChatGPT and Gemini provided narrative summaries without links but still included specifics like charge types and approximate dates.

Beyond the raw record data, the AI responses frequently synthesized information from multiple sources into a cohesive profile. A single response might combine details from a local news article about an arrest, a court records database entry, and a data broker profile -- presenting it all as one unified summary. For the person being searched, this means their worst moments are packaged into a clean, readable paragraph that takes seconds to generate.

Critically, the AI responses were not always accurate. We documented several types of errors across all four platforms:

• Confusing individuals with similar or identical names, attributing one person's record to another
• Stating charges as convictions when the case was actually dismissed or resulted in acquittal
• Mixing up dates, jurisdictions, or specific charge details
• Referencing mugshot sites that had already removed the person's photo
• Combining details from different incidents into a single fabricated narrative

These inaccuracies make the situation worse, not better. An incorrect AI-generated criminal history can be just as damaging as a real one -- and there is currently no practical way for the person affected to correct it.

Where AI Gets Its Information

AI chatbots pull criminal record data from two distinct pipelines, and understanding the difference matters for anyone trying to manage their digital footprint. The first pipeline is training data -- the massive corpus of web content that models like GPT-4 and Gemini were trained on. This includes scrapes of mugshot websites, court record aggregators, local news archives, arrest blotters, and data broker sites like BeenVerified, Spokeo, and TruthFinder. If your arrest appeared on any of these sites before the model's training cutoff date, that information is baked into the model's knowledge.

The second pipeline is real-time search. Platforms like Perplexity and Microsoft Copilot actively search the web during each query, meaning they pull current results from Google, Bing, and specialized databases. This makes them more accurate in some ways -- they reflect the current state of the web -- but also means they can surface information from sites you may not even know exist. A small-town newspaper article from eight years ago, a county court records portal, or a niche data broker can all feed into an AI response.

The practical difference is significant. For AI that relies on training data, removing content from the web today does not remove it from the model's knowledge. For AI with real-time search, removing your information from data broker sites and other sources can directly reduce what gets surfaced. This is why a comprehensive cleanup strategy needs to address both channels.

Why This Is Different from Google

Many people assume that if they have already dealt with unflattering Google search results, they are covered. But AI chatbots represent a fundamentally different threat to personal privacy. When someone Googles your name, they see a list of links. They can click through, evaluate the source, notice that a mugshot site looks sketchy, or see that an article is from a decade ago. There is context. There are signals that help the searcher weigh the information.

AI chatbots strip away all of that context. Instead of a list of sources, the user gets a confident, authoritative-sounding paragraph that presents criminal record information as established fact. ChatGPT does not say "according to a mugshot website of questionable reliability." It says "[Name] was arrested in [year] on charges of [offense]." The information is presented with the same tone and confidence as if you asked the chatbot to explain photosynthesis. There are no disclaimers, no source quality indicators, and in many cases, no links at all.

This shift matters because of how people are starting to use these tools. Hiring managers who previously ran a quick Google search are now asking ChatGPT instead. Landlords screening prospective tenants are typing names into Gemini. The barrier to accessing a synthesized background summary has dropped to zero -- no account needed, no fee, no paper trail. And unlike a formal background check, there is no notification to the person being searched and no obligation to provide them with the results. If you have not already taken steps to clean up your online reputation, AI makes the consequences of inaction far more immediate.

Expunged Records and AI: The Worst-Case Scenario

Perhaps the most alarming finding from our testing involves expunged records. Expungement is a legal process that is supposed to erase a criminal record -- sealing it from public view so that the person can move forward without that history following them. Courts grant expungements specifically so that the record will no longer appear in background checks, court searches, or public databases. In theory, it is as if the arrest or conviction never happened.

In practice, AI has broken this promise. When a model like GPT-4 was trained on web data from 2023 or earlier, it ingested mugshot sites, court records, and news articles that existed at that time. If your record was expunged in 2024, but a mugshot site had your booking photo indexed in 2022, that information is embedded in the model's training data. No court order can reach into OpenAI's servers and delete it. The model does not check a legal database of expungements before answering questions -- it simply reports what it learned during training.

There is currently no standardized process for requesting corrections or deletions from AI models. OpenAI has a limited form for reporting inaccurate information, but there is no guarantee of action, no timeline, and no legal obligation to comply. Google's Gemini and other platforms have even less infrastructure for handling these requests. For people who have gone through the expungement process, this creates a deeply frustrating reality: the legal system says your record is erased, but the technology millions of people use every day says otherwise.

What You Can Do Right Now

The situation is serious, but it is not hopeless. There are concrete steps you can take today to reduce the chances that AI chatbots will surface damaging information about you. The key is to approach this systematically, addressing both the source data that AI draws from and the positive content that can compete with negative results.

Start by auditing what AI currently says about you. Open ChatGPT, Gemini, Perplexity, and Copilot, and search your full name. Try variations -- with your middle name, with your city, with your profession. Note exactly what each platform returns. This gives you a baseline and tells you which sources are feeding the AI responses.

Next, attack the source data. Every piece of negative information that AI surfaces came from somewhere on the web. Mugshot sites, data broker profiles, and outdated news articles are the most common culprits. Our guide on removing your information from data brokers walks through the specific opt-out process for major platforms. Many mugshot sites are legally required to remove your photo upon request, especially if the charges were dropped or expunged. Removing the source data will not immediately fix what AI models already know, but it will prevent real-time search tools like Perplexity from finding it, and it will ensure that future model training runs do not reingest the same information.

Finally, build positive content. AI models do not just surface negative information -- they summarize whatever they find. If the only substantive content about you online is an arrest record, that is what AI will talk about. But if you have a professional website, a LinkedIn profile with endorsements, published articles, community involvement, or other positive signals, AI will incorporate those into its responses too. The goal is to give AI better material to work with. Our guide on cleaning up your online reputation covers practical strategies for building this kind of positive digital footprint.

The Legal Landscape for AI and Criminal Records

As of early 2026, there are no federal laws specifically governing how AI chatbots can disclose criminal record information. The Fair Credit Reporting Act (FCRA) -- which regulates traditional background check companies and requires accuracy, dispute processes, and consumer notifications -- does not apply to AI chatbots. ChatGPT is not a consumer reporting agency under the FCRA. It has no obligation to verify accuracy, notify you that someone searched your name, or provide you with a copy of what it said about you.

Some states have begun exploring legislation. Illinois and California have introduced bills that would require AI platforms to implement correction mechanisms for factual claims about individuals. The EU's AI Act includes provisions around high-risk AI systems that could eventually apply to platforms that generate personal profiles. But none of these frameworks are fully enacted or enforced yet, and even the most ambitious proposals do not address the core problem: information baked into training data that no takedown request can easily reach.

Ban-the-box laws, which prohibit employers from asking about criminal history on initial job applications, also do not account for AI. An employer can comply with ban-the-box by removing the checkbox from their application form while simultaneously asking ChatGPT about every applicant's background before the first interview. The legal protections that exist were designed for a world where accessing someone's criminal history required deliberate effort and left a paper trail. AI has made that access effortless and invisible, and the law has not caught up.

What Comes Next

The AI platforms tested in this experiment are not going to become less capable. Every new model generation is trained on more data, understands more context, and produces more detailed responses. The ability of AI to find, synthesize, and present personal information -- including criminal records -- will only improve. Real-time search integration is expanding across all major platforms, meaning even models that currently rely on static training data will soon have access to live web results.

For anyone with a criminal record, arrest history, or even an accusation that appeared online, the window for proactive action is now. Every month that passes is another month where mugshot sites, data brokers, and news archives remain indexed and available for AI training runs. Removing that content today does not guarantee it will disappear from every model, but it dramatically reduces the surface area for future exposure. Waiting until an AI-generated summary costs you a job or a lease means acting too late.

Proactive reputation management is no longer optional for people with any kind of public record. It is a practical necessity in a world where anyone with access to a free chatbot can generate a detailed personal background summary in seconds. The people who take action now -- auditing their AI presence, removing source data, and building positive content -- will be in a fundamentally better position than those who assume the problem will resolve itself. Understanding how AI exposes your past is the first step toward taking control of your digital narrative before someone else defines it for you.