New Research: People Blame You for What the Codec Changed

Neural compression's most dangerous failure isn't that it loses detail. It's that it silently changes detail, makes the result look flawless, and leaves the viewer blaming a human for the difference. That last part used to be a hypothesis. As of CHI 2026, it's a measured result.

A new peer-reviewed study put the phenomenon in front of 115 people and asked what they saw. The finding should worry anyone shipping images through a neural codec: when the codec quietly alters an image, viewers don't read it as compression. They read it as someone editing the photo on purpose, and they trust the image less because of it.

Key Takeaways

• A 115-subject CHI 2026 study found neural-compression errors raise the perceived risk of misunderstanding, a large effect (Cohen's d = 0.86, p < 0.001).
• Viewers attribute these silent changes to intentional editing, not compression (d = 0.78), and do not recognize them as compression artifacts the way they do JPEG (d = 0.64).
• The reason is the absence of visible tells. Neural output looks clean, so the damage is invisible until it changes meaning, and the blame lands on a person.

What did the study actually find?

It found that silent semantic changes from neural compression measurably erode trust, and that people misread the cause. Researchers Nora Hofer and Rainer Böhme ran a controlled lab study with 115 participants, who compared original images against versions processed by four state-of-the-art neural codecs, conventional JPEG, and uncompressed controls (Hofer and Böhme, When the Codec Hallucinates, CHI 2026, 2026). The study calls these silent changes "miscompressions": unintended semantic alterations introduced when a generative decoder reconstructs detail it didn't actually receive.

Three results, all statistically significant at the p < 0.001 level after controlling for scene and subject effects. Miscompressed images raised the perceived risk of misunderstanding by close to a full step on a six-point scale, a large effect by Cohen's standard (d = 0.86). Participants attributed the differences to intentional editing rather than accident (d = 0.78). And they were significantly less likely to chalk the change up to uncontrollable distortion the way they did with JPEG (d = 0.64).

Put plainly: the codec changed the picture, and the viewer assumed a person did it on purpose.

Why is "silent" the dangerous part?

Because the failure carries no warning label. Conventional codecs like JPEG fail loudly: blocking, ringing, blur, the visible tells that let a viewer judge how much to trust an image. The study's authors note that neural codecs lack these indicators and tend to appear visually flawless, which can create a false sense of trust (Hofer and Böhme, 2026). The damage is real, but the signal that something went wrong is gone.

This is exactly the distinction we drew in when neural compression fails : classical codecs fail legibly, neural codecs fail silently, and the silence is the liability. The CHI study is the empirical confirmation of that argument. Where we described the failure mode qualitatively, 115 people in a controlled study just demonstrated that it changes how humans perceive and trust images.

The specific examples in the paper make it concrete. In one image, the engraved number 8 on a camera became a 6. In another, an arrow-shaped traffic light turned into a round one. A wedding ring and a necklace pendant simply vanished. None of these look like compression damage. They look like the photo was always that way, or like someone edited it.

Why do people blame editing instead of the codec?

Because until now, the only thing that silently changed an image's meaning was a human editing it. People have decades of intuition for retouching, filters, and manipulation, and almost none for a codec that rewrites detail. So when a difference appears with no visible compression artifact to explain it, the mind reaches for the explanation it knows: someone changed this on purpose.

The study spells out why that misattribution is corrosive. Mistaking a compression artifact for editing can poison attempts to resolve a dispute through image comparison, especially when neither party knows the real cause (Hofer and Böhme, 2026). Imagine the difference surfacing in a newsroom, an insurance claim, or a courtroom, where "this was edited" and "this was compressed" carry completely different weight. The paper's opening scenario, a photojournalist whose protest photos arrive subtly altered and spark accusations of misrepresentation, is fiction today and entirely plausible tomorrow.

There's a second-order cost too. As neural compression spreads, the study warns, people may lose their existing ability to spot real manipulation, because the artifacts they once relied on as tells will no longer be reliable signals of anything.

What should teams shipping images take from this?

That perceived quality is not the same as preserved meaning, and a codec optimized to look good can still change what an image says. A neurally compressed image that scores beautifully on perceptual metrics and looks pristine to the eye can still have dropped a digit, a sign, or a face, and the viewer will never know it was the codec. For any pipeline where images carry consequences, that gap is the whole risk.

The practical takeaway is not "avoid neural compression." It's that the decision to use a generative codec has to account for what the content means, not just how it looks. An image whose details are decorative tolerates aggressive neural reconstruction. An image whose details carry facts, a serial number, a medical boundary, a license plate, a face, does not. The study even points toward this as a research direction: future codecs could allocate bits in a risk-adjusted way, preserving the details most likely to be miscompressed into something consequential.

Frequently Asked Questions

What is a miscompression?

A miscompression is an unintended semantic change to an image detail, introduced when a neural codec's generative decoder reconstructs information it did not actually receive. The result looks plausible but differs in meaning from the original, for example a number, sign, or object that has quietly changed.

How is this different from JPEG artifacts?

JPEG fails visibly, with blocking, blur, and ringing that signal compression occurred. Neural miscompressions leave no such tells; the image looks flawless. The CHI 2026 study found people recognize JPEG damage as distortion but attribute neural changes to intentional editing instead.

Is this a real, measured effect or just a concern?

It is measured. A controlled study of 115 participants found statistically significant effects (p < 0.001): miscompressions raised the perceived risk of misunderstanding (Cohen's d = 0.86) and were attributed to editing rather than compression (d = 0.78).

Does this mean neural compression is unsafe to use?

No. It means the choice should depend on what an image's details mean. Decorative content tolerates aggressive neural reconstruction; content where a detail carries facts, such as documents, medical, or evidentiary images, calls for caution or human review.

The bottom line

The scariest failure mode of neural compression is no longer a thought experiment. A peer-reviewed study has now shown that when a codec silently changes an image, people trust it less and blame a human for the change, with effect sizes large enough to take seriously. The output looks perfect. The meaning shifted. And the viewer never suspects the codec.

That is the case for treating compression as a question of meaning, not just appearance. Perceived quality was never the whole story, and a pristine-looking image that says something different than the original is a failure no metric will catch.