Investigations in the AI Era

What AI changes about evidence, and what it doesn’t

Fabricated recordings, altered documents, and an opposing party that reads faster than ever. A practical guide for attorneys, employers and agencies.

AI has changed two things at once. It has made convincing fakes cheap to produce, and it has made enormous records fast to read. Both change how a matter is investigated. Neither changes the standard: a fact still has to come from a source, and the source still has to hold up.

This section collects what we are seeing in the field and how we handle it. We add to it as the technology, and the rules around it, change.


Authenticity

Is this recording real?

Cloned voices, altered video and generated photographs are now easy to make and hard to spot by looking or listening. A recording that sounds exactly like a supervisor, or a photo that places someone where they never were, can arrive in a complaint, a custody dispute or a lawsuit.

Our approach is to go to the source rather than judge the file by eye. Where did it come from, who made it, on what device, and does anything else corroborate it?

What we check

  • The original file and device, not a forwarded copy
  • File details: creation dates, device information and editing history where available
  • The account or person it supposedly came from
  • Independent corroboration: other witnesses, records, location and time data

Documents

AI-made documents in the record

Screenshots of text messages, emails, receipts and letters are easy to fabricate, and altered PDFs can look identical to the real thing. In a large production, a single invented record can sit among thousands of genuine ones.

When a document matters, we trace it back to the system it supposedly came from. When the original is held by someone else, we identify that up front so counsel can request it by subpoena or order.


Workplace

AI in workplace complaints

Workplace complaints increasingly arrive with screenshots, recordings and chat logs attached. Most are genuine. Some are edited, taken out of context or generated. A finding that rests on an exhibit nobody verified will not survive a challenge.

In our workplace investigations, key exhibits are verified before they support a finding, and the report says how.


Speed

The other side is reading faster too

AI lets anyone pull years of public records and draft cited objections in an afternoon. Project opponents, plaintiffs and regulators can now find the inconsistency in your record before you do.

The answer is to read your own record first. See The Record Is Now Read in Minutes.


Our practice

Built, not bought

When people hear “AI-assisted,” they picture an investigator pasting notes into a chatbot. That isn’t what we do.

Since 2021, Atlantis has been building its own investigative system, MIA. It works inside our case file, under our controls. It takes custody of every document and fingerprints it. It reads every page, including the scanned ones a keyword search skips. It transcribes recordings on our own machines, builds the timeline, and ties every fact to the page it came from. It keeps track of every active case, and it powers the private portals where clients read their own files.

And it doesn’t decide anything. A licensed investigator does, and signs the work. Some of our analysis uses commercial AI services that do not train on client data, and we describe exactly how a matter will be processed before you send it.

Read how MIA came to be →


See it at work in the JFK files demo, or read how Atlantis uses MIA.


The rules

Courts are catching up

Courts have sanctioned lawyers for filing AI-generated citations that did not exist, and many judges now have orders on the use of AI in filings. Rule-makers are also considering how machine-generated evidence should be admitted. We follow these developments and write about the ones that change how evidence is gathered.


From this section

In Michael’s words

Short answers from Michael Braa, founder of Atlantis Private Investigations.

Never accept a screenshot

Read the transcript

Never accept a screenshot as evidence, and here’s why. An original photo file carries hidden information: the date and time it was taken, what phone took it, and sometimes where. A screenshot throws all that away. All it proves is that someone took a screenshot. Many apps strip it out too when you text or post a photo.

This matters. Ask for the original file, straight from the phone that took it. If someone won’t give it to you, that tells you something too.

The fake bank statement

Read the transcript

The fastest way to spot a fake bank statement? Do the math. Take the opening balance, add each deposit, subtract each withdrawal, line by line. A real statement lands on the ending balance to the penny. A doctored one often doesn’t, because people change one number and forget to do the rest.

Then look closely. A font that’s slightly off, or numbers that don’t line up, or something looks like it might have been changed. And it really counts. Get the statement straight from the bank, not from the person.

Check the PDF properties

Read the transcript

Before you trust a PDF, check its properties. Open the PDF, go to File, then Properties. You’ll see when it was created, when it was last changed, and even what program made it.

Now, a bank statement that was made in a photo editor, or changed last Tuesday? That’s a question. It doesn’t prove anything by itself. Files get re-saved all the time. But it tells you exactly where to look harder.

Reverse image search

Read the transcript

Is that profile photo real? Search it. Save the photo, or right-click on it, and run it through a reverse image search, like Google Lens. If that same face shows up under five different names, or maybe on a stock photo site, well, that’s your answer.

Fake profiles run romance scams, business scams, and fake job references, and the photo is usually the first thing they steal. Before you trust a stranger online, or a brand new business partner, check the face. 10 seconds. Search the photo.

The Wayback Machine

Read the transcript

That website changed? You know you can still see what it used to say? Go to the Internet Archive’s Wayback Machine. That’s web.archive.org. Type in the web address and you’ll see a calendar of saved copies, often going back years. Click a date and you’re looking at the page the way it looked on that day.

Prices that changed, promises that disappeared, staff who got quietly removed. Before you rely on what a company says today, check what it said back then.

How old is that website?

Read the transcript

How old is that website? Scam sites are usually brand new. Search for a WHOIS lookup and type in the website’s address. Look for the creation date. That’s when the web address was registered.

A business that says it’s been around since 1995, with a website registered three weeks ago? That’s a question. It doesn’t prove anything by itself. Businesses do move to new sites all the time. But a brand new website, plus a deal that’s too good to be true? Walk away.

The photo remembers

Read the transcript

That photo might be carrying its own date and location. Most phone cameras save hidden information inside the picture. We call it metadata. The date, the time, the phone model, and sometimes even the GPS location. On a computer, right-click the original photo and open Properties. On a Mac, use Get Info. Check the details.

Here’s the catch, though. Texting a photo or posting it online often strips that information out. So always ask for the original file, not a screenshot. The original remembers.

Have a question about evidence that might not be what it seems?

Confidential consultation. Tell us what you are looking at, and we will tell you how we would check it.