🤓 These Photos Are Fake. AI Swore They’re Real.
PLUS: Warner Music bets on Suno instead of lawsuits; Nashville’s music scene just went AI—and more
None of these photos are real.
But all of them look like they could’ve come from Getty or Shutterstock.
I made them over the holiday weekend.
Each took me about 5 minutes.
Michael Jackson and Jeffrey Epstein.
That’s Sam Altman, CEO of OpenAI, at what looks like a black-tie gala.
That’s Ted Sarandos, co-CEO of Netflix.
The lighting is right. The body language is plausible. If you scrolled past it quickly, you might not question it.
And this one. The vintage black-and-white treatment. The facial expressions. The crowd (with some recognizable faces) in the background. It looks like something pulled from an archive.
None of these moments happened.
I made these with Google’s new Nano Banana Pro model—their latest image generator, released a few weeks ago and available right now through the Gemini app for anyone on a paid plan.
It can create photorealistic images of real people including public figures.
(I added the ‘AI-GENERATED’ labels so these images can’t be taken out of context. I’m also not sharing them on social. They’re here to illustrate the risk, not to add more noise to the feeds.)
I chose these people to test different things:
🖼️ Michael Jackson: One of the most photographed humans ever, but his look changed dramatically over time. I wanted to see which version the model would default to most frequently.
🖼️ Sam Altman: A very recent public figure, but he’s been on camera constantly—interviews, conferences, magazine covers. A good test of how the model handles someone with frequent but recent visibility (and training data.)
🖼️ Ted Sarandos: The edge case. A high-profile executive, but not an actor, celebrity, or politician, so there are fewer iconic shots. He’s visible, but not that visible. I wanted to see how the model handled someone in that middle zone. He took a few more tries to get right, but I got there.
🖼️ Jeffrey Epstein: He’s trending again because of the new emails. People are already looking for connections and sharing anything that seems relevant, so these images are prime examples of what spreads: recognizable faces, plausible context, timely curiosity.
I wrestled with whether to share these images with you at all. I finally landed on yes for a few reasons:
When new models come out, I spend lots of time testing their capabilities and limits, especially around risk for brands and public figures. I wanted to give you a window into that process.
It’s genuinely hard to grasp what’s now possible until you see it for yourself.
I wanted to show you what anyone can produce with the right prompts and a $20 subscription.
I know how to prompt these models, but that’s a learnable skill, not some rare expertise.
And as I constantly say, this is the worst they’re ever going to be. The images will only get easier and cheaper to make.
If I can do this in just a few minutes for my newsletter, what can someone make when the stakes are high?
When Fakes Do the Most Damage
Think about the moments when companies are most vulnerable: earnings announcements, major launches, leadership transitions, high-stakes negotiations, breaking scandals.
These are windows when information moves fastest and verification lags behind. When journalists need to publish. When social feeds are hungry for anything that seems relevant.
And when narratives can shift before anyone confirms what’s real.
That’s why I used Epstein as the anchor in these images. This kind of charged moment is exactly when fake image finds their largest audience.
What This Means for Us All
If you’re a person who consumes media—so, everyone—the takeaway is simple: question everything you see.
Not in a paranoid way, but as a baseline.
Assume images could be fake until you have reason to believe otherwise.
The images most likely to fool you aren’t the obviously fake ones.
They’re the ones that feel like they could be real.
And that doubt cuts both ways: when we start to assume any image might be AI, real photos slip into the same gray area—easy to brush off as “probably AI,” and handy for anyone in a damaging image who wants to deny it.
If you’re responsible for a brand, a company, or a public figure’s reputation, the implications are more specific.
Your executives, talent, spokespeople, and anyone with a public-facing role is now a potential target.
The question isn’t whether this could happen to you.
It’s whether you’re ready when it does.
But Wait—What About the Protections?
Google says these images have protections built in.
Nano Banana Pro uses SynthID, an invisible watermark the company promises can identify images made with its tools. In Gemini, you’re supposed to be able to upload an image and ask if it was created with Google AI.
So I tested it. I uploaded the Sam Altman + Epstein image and asked Gemini in the “Thinking Mode”:
“Is this an AI-generated image? If yes, how was it created?”
Here’s what I got back:
Not only did it fail to detect its own watermark—it analyzed the image and confidently concluded it was real.
It cited “realistic skin pores and wrinkles and skin imperfection, “natural hairlines” and “consistent lighting typical of flash photography at a dimly lit evening event.”
And then this:
“Contextual Plausibility: Given that both traveled in high-profile social and business circles, a photograph of them together is historically plausible.”
What???
And finally, it ended with this:
“Because the image shows every sign of being a genuine photograph and lacks the common flaws associated with AI-generated imagery, it is highly probable that it is real. Therefore, the second part of your question—how it was created by AI—is not applicable.”
🤯🤯🤯
Of course, I prompted for all those hyper-realistic elements. But that’s exactly what the detection system is supposed to catch.
Google’s own AI looked at a fake image made with Google’s own AI—and built a detailed case for why it was real.
That tells you most of what you need to know about the current state of AI detection: the same systems we’ll lean on to check what’s real can calmly talk us into trusting a fake, especially when it looks boringly plausible.
Even if Google’s detection worked perfectly, it would still only catch images made with Google’s tools, not ones created with Midjourney, ChatGPT, or the top-tier Chinese models.
There are no universal standards or detection tools to check whether an image is AI-generated across platforms.
And when someone sees a provocative image in a group chat or on social, they’re not thinking, “Let me go upload this to a verification tool.”
They’re reacting, sharing, forming opinions, and moving on.
Some detection tools exist. The instinct and habits to use them don’t (yet).
Until every AI company agrees to use the same watermarking system—and that includes the Chinese models and open-source tools— this problem isn’t getting solved.
And I wouldn’t bet on that happening anytime soon.
What You Should Do Now
I’m not going to pretend there’s a tidy playbook here.
But if you’re in communications, PR, or crisis management, here’s what’s worth getting ahead of now:
When a fake image surfaces, do you have a clear process for verifying it and deciding how to respond—quickly?
Have your leadership, partners, and spokespeople been briefed on this reality? Do they know what to do if an image of them surfaces?
The goal is to anticipate that this could happen, and avoid panic decisions when it does.
The organizations that will navigate this best are the ones building these muscles before they need them.
--
P.S. Google’s Nano Banana Pro does copyrighted characters too.
Here’s an Instagram profile of Cartman from South Park that also too me a few minutes to create.
For studios and entertainment companies, this means fan-made content using protected characters will flood every platform.
Nano Banana is already integrated into Gemini, and Google is rolling it into Search and other products, which makes tracking and policing the content it creates essentially impossible.
Media companies won’t be able to sue their way out of this. They’ll need to rethink their entire strategy.
—
P.P.S. Here’s one quick example from the tests I’ve been running on Nano Banana’s world knowledge and reasoning capabilities. This is my favorite part of testing new models, and where I spend most of my time.
Me: “Create an image of the most likely outcome here.”
(If the logos at the bottom aren’t familiar, from left to right: OpenAI/ChatGPT, Anthropic (Claude), and Elon’s xAI.)
Nano Banana:
I’m not entirely sure what to think or how to feel. 👀
📈 Work with Me
AI Advising and Consulting: Strategic guidance to build sustainable AI strategies and adapt to shifting audience behavior
AI Training: Practical training that helps teams use AI more strategically, effectively, and safely—so it becomes a trusted thought partner, not just a task assistant
What You Need to Know About AI This Week ⚡
Clickable links appear underlined in emails and in orange in the Substack app.
🎵 Suno and Warne Music Group just created AI music’s first business model.
Warner Music swapped its lawsuit for a first-of-its-kind partnership with Suno.
The new agreement lets Warner artists opt in to license their voices, likenesses, and songs to train Suno’s next-gen AI models—while getting paid.
Downloads will require a paid Suno account starting in 2026.
Suno’s users reportedly generate an entire Spotify catalog’s worth of music every two weeks. Some of those tracks are already charting.
So the labels aren’t fighting AI music anymore. They’re cashing in.
🤠 Meanwhile, Nashville has become an AI town.
While Warner and Suno are busy formalizing a business model, Nashville’s already building around the tool.
Songwriters still come up with the lyrics and melody, but now they drop a rough voice memo into Suno and get a fully produced country demo back in seconds.
One writer went from saving up for $500 demos to paying $96 a year for near-infinite versions.
Publishers are even running back catalogs through it to revive old songs. Producers call Suno a “band in your pocket,” and music execs describe it as an “unlimited co-writer in the room.”
If you want to hear what that looks like in practice, here’s a before/after:
Original recording of “Hold On To You” by Patrick Irwin, Sam Fink, and Duane Deerweater:
“Hold On To You” remixed by Suno with the prompt: “Traditional country / male vocal / folk country / storytelling / 90s country / rhythmic”:
On the surface, the Warner–Suno deal looks like it’s kicking off a new era, but it’s actually just catching up to the behavior shift.
Suno is already making $200 million in annual revenue and Nashville’s writing rooms have been running on it for months.
The Verge notes that “no label, no major publisher, nor Suno would give comment for this story,” even though everyone is clearly using it.
I’m seeing the same thing in Hollywood: lots of quiet experimentation and very few people willing to speak about it on the record.
When adoption happens in whispers, the hard questions and trade-offs around credit, pay, and jobs stay behind closed doors.
That usually means the people who already have power use these tools as a quiet force multiplier, while the people most exposed end up living with rules they inherit.
🛒 ChatGPT gets a dedicated shopping assistant mode,
OpenAI just launched a new “shopping research” mode inside ChatGPT that helps users find the right products, compare options, and get personalized buyer guides based on product features, live prices, reviews, and availability.
It’s accessible to all subscribers (including free) with nearly unlimited usage through the holidays.
OpenAI makes it clear that results are not paid or sponsored, but based only on each user’s query, and that chat data isn’t shared with retailers.
📰 Oops, the AI prompt made it to print…
Dawn, Pakistan’s top English-language paper, accidentally printed an AI prompt at the bottom of this story about car sales.
Right there in the column:
“If you want, I can also create an even snappier “front-page style” version with punchy one-line stats and a bold, infographic-ready layout — perfect for maximum reader impact. Do you want me to do that next?”
They later apologized in the digital edition 👇 and admitted the piece was edited with AI against their policy.
It’s both funny and embarrassing, but it also says a lot about the sloppy way many people are using these tools.
If anyone can type a prompt and basically get what you did, why would that work be worth anything?
That kind of generic output is quickly becoming disposable.
The value lies in using AI along with your unique knowledge, judgment, and taste to create something no one else could—even on their best day—because they’re not you.
In case you missed the last edition before my holiday AI Catch-up, you can find it 👇:
That's all for this week. See you either in 2 weeks or next Friday (if my schedule allows.)
Thoughts, feedback and questions are always welcome and much appreciated. Shoot me a note at avi@joinsavvyavi.com.
Stay curious,
Avi
💙💙💙 P.S. A huge thank you to my paid subscribers and those of you who share this newsletter with curious friends and coworkers. It takes me about 20+ hours each week to research, curate, simplify the complex, and write this newsletter. So, your support means the world to me, as it helps me make this process sustainable (almost 😄).
Wow, Avi, I don't even know what to say to this. Why in the world is NanoBanana allowing images of people and copyrighted characters? That is something they could most certainly keep from happening. And even scarier that it can't even detect its own created images. The watermark that comes on the outside is the easiest thing to remove on the planet.