The Expert Witness Problem AI Can't Solve

At some point in every expert witness case, someone asks you to raise your right hand.

From that moment forward, every sentence in your report belongs to you.

That's why I don't think AI is replacing forensic investigators anytime soon. It may happen, or it may happen as we are seeing with lawyers and AI, where the AI does the paralegal-type work, and the lawyer just signs off on it. Surely, mistakes will be made, but perhaps it can scale eventually while minimizing mistakes.

I run a blockchain forensics firm. We trace stolen funds, we build case files for federal law enforcement, and when it matters, I sit in the witness chair and answer for every single line of the report. That last part is the piece people forget when they tell me AI is coming to automate this profession.

We build AI tools. We use them daily. Some of what the tools do is genuinely impressive — pattern recognition across thousands of wallets, first-pass clustering, drafting language that would take an analyst an hour to write. I'm not writing this as someone skeptical of the technology. I'm writing it as someone who has to defend the work under oath.

The standard is reliability, and you have to show your work.

Forensic work has one requirement that sits above everything else: you have to show your work — not summarize it, show it. Every conclusion in a report has to trace back to specific data, obtained a specific way, on a specific date, using methods another qualified examiner could evaluate, test, and repeat.

When opposing counsel asks how I know address A funded address B, "the model flagged it" is not an answer. It's the end of my credibility and possibly the case. I need the transaction hashes. I need the heuristic. I need to explain why that heuristic is reliable, where it fails, and what I did to rule out the failure modes. If a tool touched the analysis, I need to know exactly what the tool did.

A language model cannot give you that. Yet.

Large language models don't preserve a verifiable chain of reasoning or evidentiary basis for the conclusions they generate. Yet.

Ask one why it wrote a sentence a certain way and it will produce a plausible explanation, which is not the same thing as the actual reason. That gap is fatal in a Daubert hearing.

A model output, standing alone, is not a forensic conclusion. It's just a lead or a clue.

Federal paperwork has a name on it.

Every referral package we send to law enforcement, every declaration, every expert report carries a human signature — mine or one of my investigators'. Rule 26 is explicit: expert reports are prepared and signed by the witness, along with the opinions, the bases for them, and the facts or data considered. There is no version of the legal system, now or on any horizon I can see, where that signature belongs to a model. It's the same reason legal AI output doesn't become a filing until a lawyer with a bar license signs off and knows 100% of what's inside it. Not just 80%. 100%.

The signature exists because someone has to be accountable when the analysis is challenged, and someone has to be capable of answering the challenge. You cannot send a chatbot to sit for cross-examination.

What AI actually changes

This is exactly why we build AI in the first place. We build it to strip out the mechanical work so investigators can spend their time on the part only humans can do: exercising judgment, validating evidence, and explaining conclusions. I see many teams attempting to build AI with the ability to exercise judgement, however, the judgment programmed into the AI is only as good as the team building it. In order to have a truly useful AI, you need the top 0.0001% forensic experts, lawyers, developers, engineers in the world, with 30+ years' experience in the field, all working on the same thing. As far as I have seen, that's not happening anywhere.

The main theory I hear is that AI compresses the early stages of an investigation. It surfaces leads faster. It drafts faster. It reads a 400-page bankruptcy filing in minutes and tells you which exhibits mention the wallet you care about — all things we use it for daily.

Makes sense.

What it does not compress is the verification. If a model suggests two wallets are connected, the work starts there. My investigator still has to rebuild the path from primary transaction data, identify the heuristic, check the exception cases, preserve the transaction hashes, document the basis for the conclusion, and understand all of it well enough to explain it to a jury without notes. If a model drafts a paragraph of a report, the investigator who signs it has to be able to defend every clause as their own analysis — because legally and professionally, it is.

The finding came faster; the burden of understanding stayed exactly where it was.

The shortcut isn't the model.

I watch firms in adjacent industries treat AI as a way to skip the expensive part — the trained human who actually comprehends the case. In forensics, that's backwards. The expensive part is the point. What clients are buying is an expert who can defend the report under oath.

The real leverage is training investigators to use these tools the way we already use block explorers, clustering software, and OSINT platforms: as instruments whose outputs they can validate, reproduce, and explain. An investigator who can take an AI-surfaced lead and rebuild it from primary data is faster than they were two years ago.

An investigator who can't is a liability with a subscription.

We'll keep building these tools. We'll keep using them. And every case that leaves our company will always have a human being on it who knows everything about it — how the data was obtained, how each conclusion was reached, how each sentence came to be structured — because eventually someone across a courtroom is going to ask:

"How do you know that's true?"

And the answer has to come from the person who signed the report.