Skip to main content

Back to Original Reporting

Original Reporting7 sources

I Was an AI Skeptic. Then I Tried Searching the Epstein Files.

As I researched my biography of Ghislaine Maxwell, I discovered that Sourcebase could find documents the Justice Department's own search system couldn’t.

Released documents related to court cases against late convicted sex offender Jeffrey Epstein.(Vladislav Nekrasov/SOPA Images/LightRocket via Getty Images)

When I first dug into the “Epstein Files,” the 3.5 million pages that make up the Department of Justice’s public database about all things Jeffrey Epstein, I wanted to know a seemingly simple fact: what was the first email that the notorious Brooklyn-born predator sent to his closest associate, Ghislaine Maxwell, in this cache? But either by design, incompetence, or both, the DOJ’s search function made it impossible to establish even this tiny data point.

I am currently working on a biography of Maxwell for Random House, and in the hopes of peeling back the curtain on Epstein and Maxwell’s relationship, I wanted to recreate her typed interactions with Epstein, in chronological order. But when I attempted to find those communications within the DOJ site, I discovered there was no way to be certain that I’d found them all (I’ll explain why below). Without knowing that I had their complete correspondence, it was impossible to interpret it fully, or even to know when it had begun.

Then I was introduced to Sourcebase, an AI reporting tool that has its own copy of the Epstein Files and makes them easy to navigate. It took Sourcebase 22 seconds to tell me that Epstein’s October 15, 2004, email to Maxwell was the first such communication in this collection. Its poorly punctuated content – “it's fine for now„ if he fels up to it he can come see me at 7.30” – wasn’t earth-shattering (the second email they exchanged was juicier). But the payoff was still huge: for the first time in months, I had confidence I could find everything I needed in the files, and then some.

The powers of Sourcebase only became more valuable to me after Democrats prevailed on November 3, and vowed that in January, after new members of Congress are sworn in, they would move to release of some or all of the remaining 2.5 million pages of documents (the ones that were already required to be made public by the Epstein Files Transparency Act). Like the survivors of Epstein and Maxwell and many other Americans, I’ve been calling for such a release for more than a year. But more files mean more data to sort through. If we finally get the full tranche of files, Sourcebase will be essential to identifying what’s new (and what’s not).

Any reporter who’s ever perused government documents will tell you that the Epstein Files – for all their import and newsworthiness – are nothing special when it comes to their impenetrability. The fact that they are representative of a broader problem, though, doesn’t make navigating them any less frustrating. Trying to find anything inside this huge, disorganized trove can feel like looking for a needle in a stadium’s worth of haystacks. And the enormity of the cache isn’t the biggest problem; the real obstacle is that the DOJ’s search function is lame.

For that reason, Sourcebase isn’t just saving me time; it is saving my sanity. Because now I know I will be able to find every well-hidden needle – and bank statement, and travel itinerary, and email – that I’m looking for.

As a journalist who writes nonfiction books, I’ve been wary of using AI in my reporting. Maybe that’s because AI bots repeatedly confuse me with another writer named Amy Wallace – the one who co-wrote The Book of Lists and whose memoir claims she had an affair with the New Age icon Carlos Castaneda. I did not have an affair with Castaneda, and I am alive, while that other Amy Wallace sadly died in 2013. But still, all too often, AI asserts that we are one and the same person.

Because of that, among other AI-generated mistakes, when I began researching my Maxwell biography, I planned to peruse the publicly available documents the old-fashioned way: one by one. That had worked for me on my previous project, the memoir Nobody’s Girl, which I co-wrote with Epstein survivor and victims’ rights advocate Virginia Roberts Giuffre. That labor-intensive process took a lot of time, but it was worth it to me because I knew the provenance of every document I gathered.

When I first began searching the DOJ’s Epstein website, I naively thought it could be mastered. Yes, the user interface was wonky and glitchy, but I was meticulous and persistent; I thought I’d be able to make it work. The DOJ quickly disabused me of my optimism.

Consider this: When you type “Ghislaine” into the search field, you get 20,185 results. Even if you had the patience to open and read each one of these (which are not displayed in any particular order), the DOJ system has news for you: “Only the first 10,000 can be displayed. Please narrow your search by using more specific or additional keywords.” Since the documents are only vaguely (and mysteriously) organized, were you to review those first 10,000 items that mention “Ghislaine,” you’d have no way of determining how to find the missing 10,185 documents that hadn’t been displayed. And if you tried adding additional keywords, as the site suggests that you should, you’d soon discover that doesn’t work either. The only way to pull up a document that includes two terms you are interested in is to type a snippet of text that includes both terms and reads exactly as it appears in the document. You can’t do this, of course, unless you have the document in front of you – and you don’t, which is why you’re searching for it.

Complicating matters further is that, thanks to Epstein’s terrible spelling habits and various typos, Maxwell’s first name is repeatedly misspelled and mangled in the Epstein files. And because of the way the DOJ site works (or, more often, doesn’t), if you don’t type those terms exactly as written– “ghislianc,” for example, or “Giselle” – they won’t come up in your search.

Sourcebase quickly solved that problem. It found every derivative and misspelling of Ghislaine’s name, as well as “gmax” and “gmax1” from her email handles, and used context to identify when she was being referred to. Because I don’t just want all the emails Maxwell sent or received. I also want instances in which people talked about her. So I posed this inelegant question to Sourcebase: "Can you search for all emails to Epstein from anyone else who mentions Ghislaine or Ghislane or gmax any other mangled spelling of Ghislaine Maxwell?" Within seconds, I had a trove of third-party commenters – documents that had never before appeared in my searches of the DOJ site.

Here’s another example of a rabbit hole that Sourcebase helped me go down. I was interested in Maxwell’s role in refurbishing a 240-foot yacht in 2005 called the Plan B, which was owned by her then-boyfriend, Ted Waitt, the founder of Gateway Computers. On the DOJ website, I’d found a few relevant documents by doing a general search for “Waitt,” and then reading each one while looking for the yacht’s name (on the DOJ site, if you type “Waitt and Plan B” into the search field, you get bupkis). But when I asked Sourcebase for help, it found me several more specific documents that described the renovations in detail. (Later, it also found me an unrelated travel itinerary for Waitt that had never come up in my DOJ search, despite the fact that “Waitt” was spelled correctly in it. What does this suggest? That the DOJ site does not reliably retrieve all the documents that include even a simple five-letter search term).

What really won me over, though, was something I hadn't specifically requested: a reminder about sworn testimony given by Larry Visoski, one of Epstein’s pilots, who also helped manage his planes and cars and other vehicles. I will admit that I had already read this testimony, but years ago, before I knew I was curious about the Plan B. So I hadn’t retained the fact that in the November 30, 2021, transcript of Maxwell’s trial, on page 266, Visoski recalled that Maxwell was involved in renovating a boat owned by Waitt – and specifically, “decorating and building a helipad on it.” All I had to do to refresh my memory was to click a button that took me straight to page 266 of that transcript, with Visoski’s comments helpfully highlighted in yellow.

Then last month, when Maryland Congressman Jamie Raskin, the ranking Democratic member of the House Committee on the Judiciary, released a new report on Epstein’s ties to President Donald J. Trump, I thought I’d give Sourcebase a more analytical query. “Identify the main new claims about Trump and Epstein in the October 2026 House Committee report,” I told it, “then search the wider Epstein collection for underlying evidence that supports, contradicts or adds context to each claim. Cite the specific documents and pages, explain what they actually establish, and flag claims that remain unverified. Do not treat the report itself or repeated versions of the same allegation as independent corroboration.”

The results zeroed in on Beatrice Keul’s expanded account of President Trump allegedly raping her in 1993 — denied by the White House — but then got deep into the weeds of eight other topics, evaluating what was new (and what wasn’t) and providing the underlying sources every step of the way. While I’m pretty well-versed in this stuff, for me to be that definitive would’ve taken hours. The Sourcebase analysis was completed in under a minute.

To state the obvious, my brain isn’t big enough to hold all the data in the Epstein Files. But I’d been wary of using AI because I didn’t want a computer to potentially misinterpret meaning or context, thus introducing potentially untraceable errors or assumptions. Sourcebase reduces that risk by providing the underlying source material for its findings. In other words: I can check the underlying evidence and assess its accuracy.

With apologies to the late Amy Wallace, now that I’ve discovered this particular AI tool, I don’t ever want to write an investigative book without it again.

About the author

Amy Wallace

Amy Wallace is a California-based journalist and author. She is currently working on a book for Random House: a rich, psychological portrait of convicted sex trafficker Ghislaine Maxwell, Jeffrey Epstein’s closest and most notorious associate. She is best known for collaborating with Virginia Roberts Giuffre on her bestselling memoir, Nobody’s Girl: A Memoir of Surviving Abuse and Fighting for Justice. Wallace, a long-time reporter for the LATimes and a member of two Pulitzer-winning teams for the paper in the 1990s, has also been a columnist for the New York Times and contributor the New Yorker, Esquire and Vanity Fair.

Sources cited

Get new stories by email

One email when Original Reporting publishes. No digests, no marketing, unsubscribe in one click.