Hey everyone, welcome back to My Weird Prompts. I am Corn, and as always, I am joined by my brother and housemate.
Herman Poppleberry, reporting for duty. We have a really fascinating one today, Corn. Our housemate Daniel sent over a prompt that touches on something he has been working on lately. He has been looking into the world of investigative journalism and undercover recordings, specifically how to protect the people who are brave enough to talk on camera.
Yeah, Daniel has been doing some interesting work lately. I know he was mentioning those bus inspector stories in Jerusalem and how difficult it is to get people to speak on the record when they are worried about retaliation. It is a classic problem in journalism, but he is asking about a very modern solution. He wants to know if we can move past the old-school silhouette shots and distorted voices and use artificial intelligence to create entirely new personas for whistleblowers.
It is such a brilliant use case for a technology that usually gets a bad reputation. We spent so much time in episode one hundred eighty six talking about the dangers of deepfakes for misinformation, but Daniel is pointing toward what we might call a white hat application of the technology. Instead of using deepfakes to trick people, you use them to create a digital veil that protects the innocent.
I love that term, the digital veil. It sounds like something out of a spy novel, but it is actually becoming a standard in high-end documentary filmmaking. Before we get into the new tech, though, maybe we should talk about why the old methods are failing. Daniel mentioned silhouette lighting and voice modulation. Most people think those are foolproof, but are they?
Not anymore. That is the thing most people do not realize. Traditional voice modulation often just involves shifting the pitch of someone's voice up or down. But here is the problem: if you know what kind of filter was used, you can often reverse the process. If a journalist just drops the pitch by two octaves, a bad actor can just raise it by two octaves and suddenly they have the original voice. It is remarkably easy to undo if you are not careful.
And the silhouette? I always thought if you could only see a black outline, you were safe.
You would think so, but even a silhouette can give away a lot. You have things like gait analysis, where people can be identified by the way they walk or the way they move their head. Plus, if the lighting is not perfect, you might get a glimpse of a reflection or a specific piece of jewelry. But the biggest problem with silhouettes is not actually security, it is empathy.
That is a great point. It is hard for an audience to connect with a black shadow. You lose the facial expressions, the micro-movements of the eyes, the subtle trembling of the lip when someone is talking about something traumatic. If you want a story to land, you need a human face.
Exactly. And that is where the breakthrough Daniel is asking about comes in. Documentary filmmakers have been experimenting with digital protection techniques for years. One notable example is Welcome to Chechnya, directed by David France, which came out in 2020. This was one of the first times we saw advanced digital protection used on a massive scale. They were documenting the persecution of the LGBTQ community in Chechnya, and the subjects were literally in hiding for their lives. The filmmakers used various digital techniques to protect their identities while maintaining the emotional impact of their testimony.
I remember that one. It was jarring because you could see their faces, but there was something slightly off about them. It was not the uncanny valley, exactly, but it felt different.
Right. What they did was innovative in how they approached source protection. They used digital masking and other techniques to ensure that every single emotion felt by the source was reflected in the final footage. This allowed the audience to see a human being with real emotions, but the person you were looking at was protected by digital means. It was a digital mask that preserved the emotional truth while completely erasing the physical identity.
So if the whistleblower cried, the digital protection ensured that emotional authenticity came through?
Exactly. Every single twitch of a muscle. It allowed the audience to see a human being with real emotions, while the person's actual identity was protected. It was a digital mask that preserved the emotional truth while completely erasing the physical identity.
That is a massive shift. But I am curious about the technical side, Herman. Daniel asked if anyone is doing this now and what tools they are using. If someone like Daniel wanted to produce an investigative piece for YouTube or a documentary today, in early twenty twenty-six, how accessible is this tech?
It is much more accessible than it was five years ago. Back then, you needed a team of visual effects artists and a massive budget. Today, you can do a version of this on a high-end consumer laptop. There are open-source tools like FaceFusion or DeepFaceLab that are incredibly powerful. FaceFusion, in particular, has become a favorite for creators because it is almost a one-click solution. You provide a source image of a fake person, which you can generate using something like Midjourney or This Person Does Not Exist, and then you tell the software to map that face onto your footage.
Wait, so you do not even need a volunteer actor anymore? You can just use a completely synthetic face?
You can. That is the second-order effect Daniel was touching on. In the past, you needed a human double. Now, you can generate a high-resolution, photorealistic face that has never existed in the history of humanity. You feed that into a tool like FaceFusion, and it will align the landmarks of the synthetic face with the landmarks of your source footage. It handles the lighting, the skin texture, and even the way the eyes reflect the environment.
That sounds almost too easy. But what about the voice? Daniel mentioned fake voices. I assume we are talking about something more sophisticated than just a pitch shifter.
Oh, much more. This is where neural voice cloning comes in. A tool like ElevenLabs is the industry standard right now. Instead of just changing the pitch, you are using a neural network to generate speech that sounds like a specific person. For an investigative journalist, the best approach is something called speech-to-speech.
How does that differ from just typing text and having a robot read it?
That is the key distinction. If you just use text-to-speech, you lose the performance. You lose the pauses, the stutters, and the emotional inflection of the original person. With speech-to-speech, the whistleblower speaks their truth, and the AI takes the rhythm and emotion of that recording and applies it to a completely different voice. It is like a digital voice mask. It keeps the humanity but changes the vocal cords.
So you have a synthetic face from a tool like FaceFusion and a synthetic voice from ElevenLabs. It sounds like a perfect solution, but I have to ask about the security side. If I am an investigative journalist and I have a hard drive with the original, unmasked footage of a whistleblower, isn't that hard drive the most dangerous thing in the world?
You hit the nail on the head, Corn. This is the biggest risk that people often overlook when they get excited about the tech. The AI only protects the final product. It does not protect the production process. If your computer is connected to the internet while you are processing that deepfake, or if you upload the raw footage to a cloud service for editing, you have already failed your source.
So what is the protocol? How would you advise Daniel to actually handle the data?
You have to treat it like a high-security operation. First, you use an air-gapped computer, one that has never been and will never be connected to the internet. You do all your AI processing on that machine. Second, you need to be aware of metadata. Every video file contains hidden data about where it was filmed and what device was used. Even if you swap the face, a smart investigator could look at the file's metadata and find the GPS coordinates of the interview.
Wow. So you have to scrub the file completely.
Exactly. And the most important rule is the one they followed on Welcome to Chechnya: as soon as the digital veil is applied and verified, you destroy the original footage. You do not keep it in a safe. You do not put it on a backup drive. You overwrite that data so it can never be recovered. The only thing that should exist is the masked version.
That takes a lot of discipline. I can imagine a lot of filmmakers being hesitant to delete their raw files, but when lives are at stake, you really do not have a choice. Now, let's talk about the audience for a second. If I am watching a documentary and I see a disclaimer saying the faces have been replaced by AI, does that make me trust the story less?
That is the big ethical debate. Some people argue that by using deepfakes, you are training the audience to believe that video evidence is inherently untrustworthy. It is a bit of a paradox. You are using a tool of deception to tell the truth. But I think the industry is moving toward a standard of transparency. The Archival Producers Alliance, or the A P A, has been working on a set of best practices for this. They suggest that you should not just have a disclaimer at the start, but maybe even a subtle visual tell throughout the film.
Like a watermark or a specific color grade on the fake faces?
Precisely. In the documentary Another Body, which came out in 2024, they used a digital veil for a survivor of deepfake abuse. They were very clear with the audience about what was happening. They showed the process. They explained why they were doing it. By being transparent about the artifice, you actually build more trust than if you tried to hide it.
That is fascinating. It is like saying, we are lying to your eyes so we can tell the truth to your heart. I think our listeners would be interested in the specific advice we would give Daniel if he wants to start doing this for his investigative projects. We have mentioned a few tools, but let's break it down into a workflow.
Alright, if Daniel is listening, here is the Poppleberry protocol for secure source protection in twenty twenty-six. Step one: the interview. Film it in a neutral location with no identifiable background features. No windows, no specific wallpaper, nothing that can be geolocated.
And use a high-quality microphone so the AI has a clean signal to work with later.
Correct. Step two: the visual mask. I would suggest using FaceFusion. It is open-source, it runs locally so you do not have to upload anything to a server, and the quality is stunning. You will need a source image. My advice? Do not use a real person's face. Go to a site like This Person Does Not Exist and generate a few thousand faces until you find one that fits the age and vibe of your source.
That prevents the weird situation where a volunteer might later regret having their face associated with a controversial story.
Exactly. Step three: the voice. Use speech-to-speech. ElevenLabs is great, but make sure you are using their professional settings where the data is not used for training. Or better yet, use an open-source model like R V C, which stands for Retrieval-based Voice Conversion. Again, you can run that locally on your own hardware. You can clone a voice that sounds generic but retains all the emotional weight of Daniel's source.
And step four is the most important: the security.
Yes. Use a dedicated hard drive for the project. When the edit is done and the masking is rendered, you use a secure wipe tool to overwrite the original footage multiple times. And throughout the whole process, that computer stays offline. No Wi-Fi, no Bluetooth, no ethernet.
It sounds like a lot of work, but compared to the risk of someone being imprisoned or worse because they spoke to a journalist, it is a small price to pay. I am curious about the future of this, Herman. We are in January of twenty twenty-six right now. Where do you see this going in the next two or three years?
I think we are going toward a world where this tech is integrated directly into the cameras. Imagine a camera with a toggle switch for anonymity. As you film, the A I chip inside the camera is already swapping the face and the voice in real-time. The unmasked footage never even hits the S D card. That would be the ultimate protection because the dangerous data would never even exist.
That would be a game-changer. It would also make undercover work much safer in the moment. If you are caught with a camera, the footage on it already has the protection applied.
Exactly. And we are also seeing research into biometric-resistant masks. These are digital veils specifically designed to fool facial recognition algorithms. Even if a government has a massive database of every citizen, the digital veil is engineered to have a biometric signature that points to a completely different, non-existent person. It is basically an invisibility cloak for the digital age.
It is amazing how we have gone from blurry boxes over faces to this level of sophistication. It really empowers the whistleblower. It takes the power away from the people who use intimidation to keep the truth hidden.
It really does. And I think it is important to remember why we do this. We talked about this a bit in one of our earlier episodes when we were discussing the ethics of surveillance. Technology is often used as a tool of control, but when used creatively by journalists and whistleblowers, it can be a tool of liberation. Daniel is looking at the right things here.
I agree. It is about leveling the playing field. If the bad actors have advanced surveillance and biometric tracking, the good actors need advanced anonymization. It is a digital arms race, but for once, the tools of the arms race are being used to protect human rights.
Well said, Corn. I think we have given Daniel a lot to chew on. From the digital veil in Welcome to Chechnya to the practical use of FaceFusion and ElevenLabs, the path is there. It just requires a lot of technical care and a very high standard of ethics.
Absolutely. And to our listeners, if you are finding these deep dives into the intersection of tech and society useful, we would love to hear from you. Daniel is the one who sends us these prompts, but we know many of you are out there using these tools in your own fields.
Yeah, and if you have a moment, please leave us a review on your podcast app or on Spotify. It genuinely helps other curious people find the show. We are a small, housemate-run operation here in Jerusalem, and every rating makes a difference.
It really does. You can find all our past episodes and a contact form at our website, myweirdprompts.com. We also have the R S S feed there if you want to subscribe directly.
I think that wraps up our look at AI personas for whistleblowers. It is a heavy topic, but a hopeful one. The technology of deception can, in fact, be the technology of truth.
Thanks for joining us for another episode of My Weird Prompts. I am Corn.
And I am Herman Poppleberry. We will see you next time.
Wait, Herman, before we go, I just thought of one more thing. What if Daniel uses a deepfake of you to ask his next prompt?
Then I would finally have a version of me that actually cleans the kitchen, Corn.
One can only dream. Thanks for listening, everyone. Bye!
Goodbye!
