We learned to distrust deepfakes years ago. The uncanny face swap, the slightly off voice, the clip that falls apart under a second glance: these register as fake on sight. As Emanuel Maiberg of 404 Media put it, deepfakes could look convincing, yet they never quite fooled us into believing they were real.
In August 2026, Maiberg coined the term “subtlefakes” to describe real photographs of real people given a small, plausible AI edit. The alteration is what does the damage, and it raises a question: if an image doesn’t look fake, which is more harmful, the obvious fabrication or the plausible edit? And why do these edits keep appearing? This article answers both, so you can reason about the threat instead of chasing a spotting checklist. It sits within a wider shift — the synthetic-media trust gap and the rise of subtlefakes — that the cluster maps in full.
What are “subtlefakes” and how do they differ from traditional deepfakes?
A subtlefake is a lightly AI-edited photograph of a real person. The photograph really existed; what changed is a small but meaningful detail: a dress made more revealing, a pose altered, a body adjusted. The result stays plausible, which is the problem.
A traditional deepfake is full identity manipulation: a face swapped or a likeness fabricated outright. A subtlefake keeps the person recognisable and the scene intact, altering only part of it: partial manipulation, one region changed and the rest left authentic.
There’s a second line to draw. A fully generated image from a diffusion model may depict no real person at all. A subtlefake always shows a real, identifiable person in a real scene. That’s the difference between synthetic and edited, and it’s why a subtlefake always implicates someone specific.
Don’t confuse it with a cheapfake or shallowfake. Those are low-tech tricks that need no AI: cropping, slowing a clip, mislabelling an old photo as new. A subtlefake relies on generative AI to alter part of the image itself.
The cleanest way to hold all this is a spectrum: authentic at one end, fully synthetic at the other, partial manipulation in the middle. As Martino Jerian puts it, “subtlefake” is a useful communication term but a poor technical category. Maiberg framed the shift as a warning sign for how AI use is evolving from abusive but obviously fake to more subtle, less offensive, and harder to detect. For the full fake-media landscape, the pillar piece on the synthetic-media trust gap walks the taxonomy.
That partial-manipulation quality is what makes these edits hard to fight.
Why is a lightly AI-edited authentic photo more damaging than a fully fabricated image?
A fabrication can be denied, but an edited real photo can’t. Post a fully generated explicit image of someone and the target can say “that’s fake,” and most people will accept it, because the context is implausible. Edit a real photo instead and the person in it is really them, in a real scene. “That’s fake” stops working, because most of the image is true, and you’re left unable to debunk it: the target has to prove a negative, that they never posed that way.
The tools built to catch fakes hit the same wall. Forensic signals like sensor noise and metadata behave as if the file is clean, because the base image is authentic. Detectors are tuned for fully fabricated faces and voices, so a small regional edit slips straight past. That’s why detection struggles against subtle edits, and why partial manipulation is the case single-modality detection consistently misses.
Remedies share the blind spot. StopNCII requires nudity or sexual activity, and the TAKE IT DOWN Act’s definitions stop short of non-explicit edits. A degrading, non-explicit edit falls below both lines, so the abuse sits in the gap between what’s harmful and what’s actionable. The 2026 synthetic-media rulebook sets out that wider landscape.
Provenance doesn’t rescue an edited real photo either. A content credential proves where a file came from, not that it’s unaltered, and its absence proves nothing. Then there’s the longer-term cost: the liar’s dividend, where genuine content starts getting dismissed as fake too.
That’s the harm asymmetry, and it’s the gap where these edits live.
Why does X’s ad-revenue-sharing model reward accounts that post subtlefakes by the dozen?
Because the model pays for attention, not accuracy. X’s ad-revenue-sharing programme pays verified accounts for impressions and engagement, so a post doesn’t need to be true or original to earn. It just needs to be seen.
A mildly provocative edit generates engagement without tripping the explicit-content filters that would block or demonetise the post. So the rational play is volume: post subtlefakes by the dozen, stack up impressions, and let small payouts accumulate at near-zero marginal cost. It’s modest per post, but easy money, and at scale it adds up. As James W put it in a piece on the phenomenon, malice burns out, but a business model doesn’t.
That’s why the phenomenon concentrates on X: Grok is built into the platform, making each edit cheap to produce and leaving a visible watermark pointing back at the tooling. The accounts doing this are mostly verified engagement farmers, posting high volume to accumulate impressions and payouts. The motive is money, not animus toward the target. Zoom out, though, and it becomes one mechanism inside the wider trust gap — the distribution layer of a larger problem.
This is an incentive-design problem: the payout model, not the moderation team, drives the supply. Change the payout model and the supply changes. Regulators are already circling, with the EU’s Digital Services Act and Ofcom both opening scrutiny of X and Grok in early 2026, part of the wider 2026 synthetic-media landscape.
The measure of a fake’s danger is plausibility, not fabrication. A subtlefake injures because it can’t be debunked, detected, or removed. Every remedy we’ve built was aimed at the obvious fake it isn’t. And its spread is a rational output of an engagement-based payout model.
That leaves you with a better question when a suspicious image appears: what does this edit implicate, and who gets paid when it spreads? As the edits multiply, the liar’s dividend grows, and the burden of proof shifts onto real people who never consented to the edit. That’s how subtlefakes fit the trust gap, and it’s the part worth paying attention to.
Frequently Asked Questions
Is a subtlefake the same as a cheapfake or shallowfake?
No. A cheapfake (or shallowfake) is low-tech manipulation without AI, such as cropping, slowing a clip, or mislabelling an old photo as new. A subtlefake uses generative AI to alter a specific region of a real image, like making a dress more revealing. The two are siblings, not synonyms: cheapfakes rely on context tricks, subtlefakes on plausible edits.
Are subtlefakes illegal?
It depends on what the edit shows and where you are. Where a subtlefake meets the definition of nonconsensual intimate imagery, it can be unlawful under laws like the TAKE IT DOWN Act. But because subtlefakes are degrading rather than explicit, many fall below that threshold, so a harmful edit can sit in a legal grey zone even though it injures a real person.
How can I tell whether a real photo has been subtly edited?
There is no reliable way to tell by eye, and that is the point. Because the base image is authentic, the edit leaves no obvious generative artefacts, and forensic signals like sensor noise and metadata behave as if the file is clean. Treat any image from an unknown source as unverified rather than trying to spot the edit yourself.
Do deepfake detectors work on subtlefakes?
Not reliably. Detectors are built to catch fully fabricated faces and voices, where the whole image is synthetic and leaves telltale traces. A subtlefake keeps the original photograph intact and alters only a small region, so detectors find nothing to flag. That detection gap is why provenance and platform policy, not scanning, carry the load.
Do watermarks or content credentials prove an image is authentic?
No. A content credential proves where an image came from, not that it is unaltered, and its absence proves nothing at all. A subtlefake can be posted without any credential attached, and a missing watermark is not evidence of tampering or of authenticity. Provenance is useful context, not a verdict.
What is the liar’s dividend?
The liar’s dividend is the benefit a bad actor gets when fakes become plausible: genuine evidence starts getting dismissed as fake too. As subtlefakes spread, a real person’s real photo can be waved away as “probably edited”, which shifts the burden of proof onto the target and makes the truth easier to deny.
What should I do if I am the target of a subtlefake?
Document everything first: save the post, the account handle and the URL before reporting or requesting removal. Report it to the platform and, if it is intimate imagery, to a service like StopNCII. Because the edit may fall below takedown thresholds, preserve the evidence and seek legal advice early rather than assuming a report will resolve it.
Are subtlefakes always sexual or explicit?
No, though sexualised edits are common because they are degrading and attention-grabbing. A subtlefake can just as easily adjust a pose, a body proportion or an outfit to misrepresent what someone did or looked like. The defining feature is not the content but the method: a small, plausible AI edit to a real photograph of a real person.
Do platforms remove subtlefakes when they are reported?
Often they do not. Removal rules are tied to explicit content, and because subtlefakes are degrading but not explicit, they frequently fall below the line that triggers takedown. That is compounded on X, where the ad-revenue-sharing model pays for the engagement these edits generate, so the incentive runs against removal.
Why did 404 Media coin the term “subtlefake”?
Emanuel Maiberg of 404 Media coined it in August 2026 to name a gap in the existing vocabulary. Terms like “deepfake” describe full fabrication, and “cheapfake” describes low-tech tricks, but neither captured a real photo given a small, plausible AI edit. The new word makes that edit-not-fabrication distinction visible.
Are subtlefakes only a problem on X?
X is where they concentrate, because its ad-revenue-sharing model rewards the volume and engagement these edits produce, but the problem is not confined to one platform. Any surface that pays for attention, or under-enforces explicit-content rules, can host them. X is the clearest example of the incentive design, not the only one.
Can’t I spot a subtlefake by looking for signs of AI generation?
No, and that instinct is exactly what makes subtlefakes dangerous. Signs of generation are most visible when an entire image is synthetic, but a subtlefake starts from a genuine photo, so most of the frame is authentic and only a small region is altered. Hunting for “AI tells” will leave you trusting the very images you should question.
