There’s a report sitting in a government file somewhere, written entirely by artificial intelligence, that describes how a federal agent used physical force on another human being. That sentence should probably make you uncomfortable. It definitely makes me uncomfortable.

Immigration and Customs Enforcement has been using ChatGPT to write use-of-force reports. Yes, the same ChatGPT that sometimes confidently tells you that strawberries have bones or that you should put glue on pizza. That one. And before you ask – no, this isn’t some dystopian satire. It’s actually happening right now.

The news broke quietly, as these things tend to do. But the implications? They’re anything but quiet.

When Accountability Meets Autocomplete

Here’s what we know. ICE agents have apparently been feeding details into ChatGPT and letting it generate the narrative portions of their use-of-force reports. You know, those critical documents that are supposed to provide a detailed, accurate account of why an officer felt the need to use physical force against someone. The kind of reports that get scrutinized in court cases, civil rights investigations, and oversight hearings.

The reasoning, I guess, is efficiency. Writing reports takes time. ChatGPT is fast. Problem solved, right?

Except that’s like saying we should let spell-check write our wedding vows because typing is tedious.

The Paper Trail That Might Not Be Real

Use-of-force reports aren’t just bureaucratic paperwork. They’re supposed to be a firsthand account – emphasis on firsthand – of what happened during a confrontation. The officer’s perspective, yes, but also their observations, their decision-making process, the specific circumstances that led to someone being restrained, tackled, or worse.

These reports matter because they create accountability. Or at least, they’re supposed to.

When an AI writes them, though, what are we actually getting? We’re getting a language model’s best guess at what a use-of-force report should sound like based on patterns it learned from its training data. It’s not remembering the actual event because it wasn’t there. It can’t tell you what the person’s face looked like or how their voice sounded. It’s just… filling in the blanks. Professionally. Convincingly. But also completely divorced from reality.

The Technology That’s Too Helpful

ChatGPT is really good at sounding authoritative. That’s kind of the whole problem here. It doesn’t hem and haw or second-guess itself the way humans do when they’re trying to recall details. It just writes. Smoothly. Confidently. With perfect grammar and the kind of bureaucratic language that makes everything sound official and proper.

But confidence isn’t the same thing as accuracy.

Hallucinations in Official Documents

Anyone who’s spent time with AI language models knows about hallucinations – those moments when the AI just makes stuff up. It’ll cite court cases that don’t exist, quote statistics from nowhere, or invent entire scenarios that sound plausible but are completely fictional.

Now imagine that happening in a legal document that could determine whether someone gets deported, whether an agent faces discipline, or whether a civil rights violation gets investigated.

The stakes are kind of high, you know?

• Legal weight: These reports get submitted as official documentation in legal proceedings. If they contain AI-generated fabrications, that’s not just sloppy – it’s potentially criminal.
• Training data bias: ChatGPT learned to write by reading millions of text samples from the internet. What biases got baked into those patterns? What assumptions about force, compliance, and threat assessment?
• The edit problem: Sure, agents are probably supposed to review and edit the AI output. But we all know what happens with autocomplete. You skim it, it looks good enough, you move on. Human nature meets artificial intelligence, and accuracy loses.

Laziness or Something Worse?

The charitable interpretation is that this is just about saving time. ICE agents deal with a lot of paperwork. If AI can handle the boring parts, they can focus on actual law enforcement work. That’s the pitch, anyway.

But there’s a less charitable reading, and honestly, it’s the one that keeps nagging at me.

When you automate the documentation of force, you create distance between the action and the accountability. The agent doesn’t have to sit there and type out exactly what they did and why. They don’t have to find the words to justify their decisions. They just feed some bullet points into a chatbot and let it craft a narrative that sounds reasonable and professional.

The Sanitization Effect

Here’s where it gets really concerning. ChatGPT is trained to be helpful, harmless, and – this is key – palatable. It smooths rough edges. It makes things sound professional and justified. It’s not going to write “I panicked and grabbed him harder than I needed to” or “Looking back, I’m not sure the situation required that level of force.”

It’s going to write something that sounds perfectly reasonable, perfectly justified, perfectly by-the-book.

Which might be exactly what happened. Or it might not be. But we’ll never know because the AI already polished it into an acceptable narrative.

“The danger isn’t just that AI might get the facts wrong. It’s that AI will get them wrong in a way that sounds absolutely right.”

What This Means for the Rest of Us

ICE might be the first agency to get caught doing this, but let’s be real – they’re probably not the only ones. Once the efficiency argument takes hold, it spreads. Local police departments are already using AI for crime reports and predictive policing. It’s not a huge leap to use-of-force documentation.

And it’s not just law enforcement. Any field where documentation is critical but tedious becomes a target for this kind of automation. Medical records. Incident reports. Safety inspections. Court filings.

The Transparency Problem We’re Not Talking About

There’s another layer to this that kind of blows my mind. When AI writes these reports, how do we even know? Is there a disclosure requirement? A little note at the bottom that says “This narrative was generated by ChatGPT and reviewed by the reporting officer”?

Spoiler alert: there probably isn’t.

So we’re left with official government documents that might be partially or entirely AI-generated, and no way to tell the difference. That’s not just a transparency problem. It’s a trust problem. How are we supposed to have faith in oversight systems and accountability measures when we can’t even be sure a human wrote the account we’re reviewing?

The fundamental issue here isn’t really about technology. It’s about what happens when we automate the parts of our jobs that are supposed to require careful thought and personal accountability. Writing a use-of-force report should be uncomfortable. It should require the officer to confront what they did and justify it in their own words. That discomfort is a feature, not a bug. It’s part of what keeps the system honest.

When we hand that off to ChatGPT, we’re not just saving time. We’re removing one of the few remaining checks on the use of force by government agents. And doing it quietly, without much debate or oversight, in the name of efficiency.

That should terrify you. It certainly terrifies me.