Cheapfakes: Why the Low-Tech Version Still Works

For all the alarm about AI deepfakes, the manipulated media that actually spreads most often is not generated by AI at all. It is the cheapfake: a real photo or video, altered with ordinary tools or not altered at all, just relabeled, slowed down, trimmed, or pulled out of context to tell a lie. Cheapfakes are low-tech, nearly free, and, awkwardly for an industry pouring resources into deepfake detection, they routinely defeat the very tools built to catch synthetic media, because there is nothing synthetic in them to catch.

This piece explains what cheapfakes are, why the low-tech version still works so well, why deepfake detectors miss them, how they differ from deepfakes as a threat, and what actually defends against them. It is written for anyone who assumes that solving deepfakes solves manipulated media, because that assumption leaves the more common threat wide open.

The honest framing is that "is this synthetic?" is only one of the questions that matter. A cheapfake forces a second one, "is this authentic and in context?", and the two require completely different answers.

What a Cheapfake Is

A cheapfake, also called a shallowfake, is media that has been edited or recontextualized using ordinary, non-AI tools, or sometimes not edited at all. The term was coined by researchers at Data & Society to name the whole family of low-tech manipulations that predate and outnumber deepfakes. The techniques are mundane and that is the point: adjusting the speed of a video to make a speaker appear impaired, as in the widely shared 2019 clip of a US politician slowed to seem slurring; splicing or selectively trimming real footage to change its meaning; adding a false caption to a genuine photo; relabeling old or unrelated footage as a current event; basic photo editing like cropping or text overlays; or layering a misleading voice-over onto authentic video.

The most important and most dangerous category involves no editing of the media at all. Out-of-context media takes a genuine, unaltered image or clip and simply attaches a false claim or caption, so the visual content is completely real and only the framing is a lie. This is what separates cheapfakes from deepfakes at a fundamental level: a deepfake fabricates new pixels with a neural network, while a cheapfake works with real pixels, misused. That difference, invisible in the content itself, is exactly what makes cheapfakes so persistent.

Why Cheapfakes Still Work

The simplest reason cheapfakes endure is economic. They cost almost nothing, take minutes, require no technical skill, and can be made with tools everyone already has or with no tools at all. That accessibility means anyone can produce them and they can be churned out at volume, which is why they remain, in the words of the research community, known to be more prevalent than deepfakes. A cheapfake does not need to be perfect, either; it only needs to be good enough to provoke a reaction before anyone checks, and it is engineered to exploit confirmation bias and the speed of social sharing, where a misleading clip travels far faster than its correction.

There is also a deniability that makes cheapfakes uniquely sticky. Because the footage is genuine, a cheapfake is harder to debunk than an obvious fabrication: you cannot simply prove the image was never real, because it was, and the argument shifts to the murkier ground of context and intent. Fact-checkers consistently report that even as AI-manipulated media rises, these simpler manipulations remain the more widespread threat and continue to dominate the misinformation they flag, precisely because the ease and speed of making and sharing them outpaces everything built to counter them.

‍

Why Deepfake Detectors Don't Catch Them

This is the uncomfortable part for the detection industry. Deepfake detection works by finding the statistical fingerprints that generative models leave behind, frequency-domain artifacts, temporal inconsistencies, the absence of a real pulse, the traces of synthesis. A cheapfake has none of them, because no synthesis happened. The pixels are real camera output, so a detector scanning for generation artifacts finds nothing to flag and correctly reports that the media is genuine, which in the case of a cheapfake is both true and completely beside the point.

Out-of-context media is the extreme version of this problem and the hardest of all to detect automatically. As researchers put it, detecting out-of-context misuse is especially arduous because the visual content is unchanged and the misinformation is conveyed only through the image-and-text combination. There is literally nothing wrong with the image; the lie lives in the pairing of a real picture with a false claim, which no amount of pixel analysis can surface. Cheapfakes are caught, when they are caught, by metadata analysis, reverse-image searching, and contextual fact-checking, an entirely different toolkit from the one that catches deepfakes. Both can deceive, but only deepfakes require AI-based detection methods.

Two Different Threat Domains

It helps to stop treating cheapfakes and deepfakes as competitors for the same title and see that they increasingly dominate different arenas. In misinformation and disinformation, especially around elections and viral news, cheapfakes still rule: they are more common, faster to produce, and often just as persuasive, and research has not conclusively shown that deepfakes mislead more effectively than simpler fakes or even plain text. In fraud and identity verification, the picture is reversed. The notion that deepfakes are mostly a misinformation problem is now outdated, because by 2024 to 2026 their dominant use by volume and dollar value shifted to financial fraud, identity-verification bypass, and synthetic-identity attacks, where fabricating a convincing face or voice is the whole game and a cheapfake would not do.

Seeing this split matters because it prevents two mistakes. One is assuming a deepfake detector solves manipulated media in general; it does not touch the cheapfakes that dominate misinformation. The other is assuming cheapfake-era instincts, like looking for crude edits or checking context, are enough for identity fraud; they are not, because a real-time face swap in a KYC flow is genuinely synthetic and needs synthetic-media detection. The right defense depends on which arena you are in.

‍

How to Defend Against Cheapfakes

Because cheapfakes are a context problem more than a pixel problem, the defenses are correspondingly different and layered. Out-of-context media is addressed by reverse-image search, which can reveal that a photo is older or from a different event, and by provenance and context verification that establishes when and where something actually originated. Edited or spliced media is the territory of media forensics and tamper detection, which look for the seams of ordinary editing rather than the fingerprints of AI. Mislabeled footage calls for source checking and fact-checking of the claim attached to it. And content provenance standards such as C2PA help across the board, because a credential establishing an asset's true origin and history undercuts both a miscontextualized real image and a synthetic one.

Where does deepfake detection fit, and where does it not? Honesty is the credible position here: dedicated deepfake detection is the right and necessary tool for the synthetic and AI-manipulated layer, and it is not, on its own, a cheapfake solution, because a genuine image with a false caption is invisible to it. This is the layer DuckDuckGoose, based in Delft, focuses on, with DeepDetector analyzing images and video for the signatures of synthetic media, and the useful overlap is that tamper and manipulation detection catch some edited cheapfakes even when no AI is involved. But the broader lesson is the layered one: synthetic-media detection answers whether something was AI-generated, while context, provenance, and forensic checks answer whether real media is being used honestly, and defending against the full range of fakes, cheap and deep, requires both. For the clean distinction between the two, see our guide to what a deepfake actually is, and for the human side, how to spot a deepfake.

Frequently Asked Questions

What is a cheapfake?
A cheapfake, or shallowfake, is media manipulated or misrepresented using cheap, low-tech methods rather than AI, including speed changes, selective editing, mislabeling, crude photo edits, and, most commonly, taking a real image or clip out of context with a false caption. The term distinguishes these simple manipulations from AI-generated deepfakes.

How is a cheapfake different from a deepfake?
A deepfake is synthetic media generated or manipulated with neural networks, creating new pixels. A cheapfake uses real media, altered cheaply or not at all, with the deception coming from editing or false context. They require completely different detection: cheapfakes need context, provenance, and reverse-image checks, while deepfakes need AI-based detection of generation artifacts.

Why are cheapfakes still so common?
Because they are cheap, fast, require no skill, and are good enough to work. Research consistently finds cheapfakes more prevalent than deepfakes, and fact-checkers report that simpler manipulations still dominate the misinformation they flag. They exploit confirmation bias and the speed of sharing, and because the footage is often genuine, they are harder to debunk.

Why can't deepfake detectors catch cheapfakes?
Because deepfake detection looks for the statistical fingerprints of AI generation, and a cheapfake has none, since no synthesis occurred. The pixels are real, so the detector correctly finds no synthetic artifacts, which misses the point when the lie is in the editing or the context. Out-of-context media is especially hard, because the image is entirely genuine.

Are cheapfakes more dangerous than deepfakes?
It depends on the arena. In misinformation, cheapfakes are more prevalent and often just as persuasive, so they are arguably the bigger problem there. In fraud and identity verification, deepfakes dominate, because fabricating a convincing face or voice is the point. They are different threats in different domains, and neither cancels the other.

How do you detect a cheapfake?
With a different toolkit than for deepfakes: reverse-image search to find a photo's true origin, context and provenance verification, media forensics to spot ordinary editing, and fact-checking of the attached claim. Content provenance standards like C2PA help by establishing an asset's real history, which counters both miscontextualized and synthetic media.

Does deepfake detection help against cheapfakes at all?
Partly. Dedicated deepfake detection targets the synthetic and AI-manipulated layer and does not, by itself, catch a genuine image with a false caption. However, tamper and manipulation detection can flag some edited cheapfakes even without AI. The practical answer is layered: synthetic-media detection for the AI layer, and context, provenance, and forensic checks for the rest.