We would just be cooked. We'd be like yeah, I can see how you think it's fake. You should find a vendor that can help with that problem."
-Director of AI, global private investigations firm
MEET THE INVESTIGATOR
Their job is to figure out whether evidence is real. A file comes in, they look at the metadata, run a spectrogram, listen to the audio, form a view, and tell counsel what they think. For years, that process worked well enough.
Then the audio started getting harder to distinguish from the real thing.
“I think that ambiguity is where we really started to see firms like y’all rise up,” they said. “Because the human gut check wasn’t working anymore. And now I’m at a point where I don’t trust my gut.”
They recently moved into a new role as Director of AI at one of the world’s leading private investigations firms. The firm works across litigation support, financial forensics, and digital evidence, with a team that includes former federal prosecutors, lawyers, cyber specialists, and investigators. Its clients come to them when something is wrong and they need to establish what happened which has become much harder when the evidence itself may have been generated or manipulated.
THE PROBLEM
A client came to the firm with voicemails and text messages they believed had been fabricated. The team pulled the files and ran spectrograms, where they saw repeated waveform spikes, evidence that audio had been copied and pasted, and the same frequencies appearing in places where they should not have been.
The spectrograms pointed to a problem, but bringing a spectrogram and a gut read into a room full of attorneys was not going to be enough. “Our reasons were a little bit more qualitative than an actual tool would be,” they said. “I can’t really do much off the vibes of that.”
Counsel wanted something more concrete: a report with numbers behind it, an explanation of why the audio was suspicious, and documentation of what specifically was wrong. They needed something they could reference with a client or partner and defend if the finding was challenged.
The firm did not have a dedicated tool for that, and they were not seeing many options elsewhere in the industry either. Without one, a suspicious file could mean hours of manual analysis, often pulling investigators away from other cases without knowing whether they would end up with something they could confidently put in front of counsel.
The cases were not always sophisticated, either. In another investigation, someone inside a company had been fabricating invoices and using AI to manipulate supporting documents. It was one person, a fraudulent expense claim, and evidence that needed to be authenticated quickly.
That kind of case has become increasingly familiar: individual fraud that is targeted and opportunistic, with people using whatever tools are available to them faster than many investigation teams can adapt.
THE SOLUTION
The firm built Resemble Detect into its digital investigations workflow. When counsel suspects that evidence may be synthetic or manipulated, investigators now have more to work with than their own assessment of the file.
Resemble Detect runs audio through multiple models and shows which parts of a file appear synthetic, which sections pass cleanly, and what signals contributed to each finding. For investigators who had been relying heavily on spectrograms and their own judgment, that gave them another way to examine the evidence and explain what they were seeing.
Just as important was what they could hand to counsel. The team can now bring a heat map visualization alongside a written analysis with specific findings that can be reviewed, referenced, and discussed with a client or partner.
“The red color means it’s probably fake. The green is good,” they said. “That gives us something tangible to hold on to. We can go back to our client or partner and explain the outcome.”
Deepfake detection is probabilistic, and no tool can provide a guaranteed verdict. For the team, the value has been having the reasoning behind a finding documented rather than having to explain it from memory in a meeting.
“Counsel is pretty binary. They really want us to say 100%, yes, we know it’s a deepfake. The report has made it really helpful to walk through with them and be like, here’s why we think it is, based on these metrics.”
-Director of AI, global private investigations firm
WHY RESEMBLE
Before Resemble, the firm evaluated other options, including Hive AI. The multimodal capability was appealing, but the integration came with more process overhead than the firm wanted.
Their investigations move quickly and tend to be highly customized, so the team needs to be able to adapt its approach from one case to the next.
“It was built for big corporations,” they said. “Our firm is really not very large. The kind of projects we do are very fast-paced. Very customized.”
When they first met with Resemble’s co-founder, the conversation focused on API access, customization, and how the technology could fit into the firm’s existing workflow.
“It was nice to know that even though we’re a small firm, we’re still treated as partners.”
“Y’all’s ability to jump in and partner with us has been really helpful. Something I don’t think we would have been prioritized for from some of these other vendors.”
- Director of AI, global private investigations firm
THE RESULT
Cases that once ended with a qualified verbal assessment can now include a report, a visual, and documented analysis. Counsel has something they can review with a client or partner, while the firm has a capability it can bring into investigations where the authenticity of evidence is in question.
“Fighting fire with fire,” they said. “Models create this. There are models detecting it. That language really helps them understand how it works conceptually. Explaining technical things to attorneys can be pretty tricky but when you take it down to that form, and you’re like, look at this report, you can see the spikes here, it gives them something tangible.”
The reports have also raised a different question about how this kind of evidence will be presented in court. In one case, the team showed a law firm partner the heat map visualization, and the partner immediately objected to how compelling the finding was.
“Judge, this is prejudicial. You can’t display this evidence.”
Their response: “A jury would love that, though, I can tell you that much.”
The legal standards around AI-generated and manipulated evidence are still developing and vary across states and jurisdictions. For investigation firms, that makes clear documentation increasingly important.
The investigator has since spoken at an industry security symposium, advising CISOs and security professionals on detection best practices. Deepfake detection is now a capability the firm brings into investigations, and their new role as Director of AI exists in part because of that work.
WHAT THEY ARE BUILDING NEXT
Most of the firm’s Resemble usage has been in the US. In their new role, they are now bringing deepfake detection into conversations with the firm’s international offices.
They do not expect the human baseline to get any more reliable as the technology improves.
“I think we’ll just have even less faith as these models continue to increase in ability,” they said.
At the same time, the team is seeing more evidence that falls somewhere between authentic and fully generated. A signature may have been altered, a single element in an image may have been replaced, or a document may have been edited with AI without the entire thing being generated from scratch.
Those cases matter just as much in an investigation. Resemble Detect now surfaces these distinctions in its reporting, with separate sections covering live detection, fraud analysis, and AI manipulation, and the firm is already using that capability.
PRODUCTS USED
Resemble Detect / Resemble Intelligence
