Humanity in Low Resolution

Débora Kaz ·

For a long time, good technology was technology that made defects disappear.

Less noise in photography, less hiss in audio, corrected skin, controlled light, the right framing. We edited breaths, cut hesitations, fixed sentences. In advertising, everything seemed to come from a reality where no one sweated, aged, or touched the products before the photograph was taken.

It was polish. It was quality. It was professional.

Then it became too perfect.

Today, a slightly blurred photograph can seem more truthful than an impeccable one. Hard flash, grain, motion blur, and a shaking camera have returned. Bodies appear only partially, someone looks outside the frame, the video begins before we are ready for it. We leave the breath in the voice.

Even text has entered this regime of suspicion. An argument that is too organized, sentences with the same rhythm, a conclusion that is too precise and there it is: “it looks like AI.”

After decades of trying to erase signs of imperfection, we have begun putting those signs back in.

Not because of a sudden love of defects. What changed was their meaning.

Error began to suggest that someone was there.

Skin that is too smooth became suspicious. So did a voice without fluctuation. In some contexts, the bad photograph acquired a credibility the perfect one lost.

But this did not remain our exclusive territory for long.

Today, we can ask an image generator for an out-of-focus photograph, with blown flash, poor framing, pores, wrinkles, grain, and that feeling of a moment caught by accident. Current image-generation guides already recommend precisely this kind of instruction when the goal is to produce a photograph that feels more natural and less staged.

Synthetic voices no longer need to sound like airport announcements either. Current systems allow control over pauses, rhythm, and reactions such as sighs, laughter, and changes in intonation.

A text can be instructed to break its rhythm, avoid certain constructions, and sound less organized.

AI is not only learning to produce perfection. It is learning which defects we read as human.

Then a slightly irritating question appears:

when error can be commissioned, are we still looking at an error?

One image is blurred because someone moved at the instant of the photograph. Another receives blur because it was requested in the prompt.

The result may be practically the same.

The difference lies in what happened before the image.

And that changes the conversation considerably.

Because the manufacture of spontaneity did not begin with artificial intelligence. Advertising already did this very well.

The model looks distracted. The photograph looks casual. The product is lying on a bed as though no one spent forty minutes deciding exactly where that package should fall.

There was probably a photographer, art direction, styling, retouching, a client, and someone arguing about whether the glass should be three centimeters farther to the left.

Few things require as much work as making it look as though no work was done at all.

Researcher Crystal Abidin called this deliberate fabrication of a spontaneous, intimate, and unprofessional appearance among influencers “calibrated amateurism.” Improvisation, in this case, is not exactly improvisation. It is a language.

On social networks, this logic found ideal terrain. The more easily we recognize the face of an advertisement, the more the advertisement tries not to look like advertising. Campaign imitates behind-the-scenes footage. Sponsored content imitates conversation. Professional production imitates the phone camera.

Authenticity became a visual language.

And language can be learned.

This is where things become stranger, because we have begun making the opposite move.

Humans are learning how to look human.

The sentence sounds absurd until we notice that we do this.

“Make this text less polished.”

“This photo looks too produced.”

“Do not write like that, it sounds like ChatGPT.”

“Do not retouch the image so much.”

We begin preserving small accidents because they have started functioning as signs of presence.

There is something like an inverted Turing test in all of this. For decades, we fantasized about machines trying to convince people that they were human. Now we find people worried about convincing other people that they are still people.

And we turn to defect as identification.

I err, therefore I exist.

Except the document can already be forged.

If tremor, grain, hesitation, asymmetry, wrinkles, and crooked framing can be reproduced, they cease to function as certificates of origin. Not because they have lost aesthetic value, but because they have lost exclusivity.

The first generative images still gave us some comfort. There were impossible hands, too many teeth, deformed letters, anatomies that abandoned their careers halfway through the image.

We recognized the machine through failure.

That boundary is no longer so comfortable.

In a study published in 2025, researchers tested synthetic images of fictional faces and also of familiar people. Participants were unable to reliably distinguish generated images from real photographs. Neither familiarity with the faces nor access to other genuine photographs for comparison solved much of the problem.

But technical boundaries age quickly. The hand improves, text appears correctly, movement becomes convincing, and we go looking for another clue.

There is a romantic hope that some impossible-to-reproduce detail will remain, a secret flaw, the final hiding place of humanity.

I doubt it.

And it would be a somewhat depressing fate for the species to discover that its last line of defense was a convincing dark circle under the eyes.

Perhaps the problem lies elsewhere.

We are treating appearance as though it had ever been proof.

It never was.

A photograph of someone crying never proved sadness. It could be an actor, a staged scene, a reconstructed memory, a gesture performed for the camera. A visible scar does not, by itself, tell us how it came to exist. An account can be true, invented, exaggerated, or remembered differently many years later.

Surface has always required trust, context, and interpretation.

It simply worked well enough as a shortcut.

It looked human, so it was probably human.

Artificial intelligence has not necessarily destroyed that certainty. Perhaps it is showing that the certainty was never as solid as we imagined.

What changes now is the scale of doubt.

Because we are no longer merely facing an image that can stage an emotion. We may be facing a face that never aged, a voice with no biography, a scene whose camera was never anywhere.

And yet all of it can carry the codes we learned to associate with experience.

The wrinkle is there.

The pause is there.

The tired gaze is there too.

What we do not know is what existed before them.

It is at this point that lived experience, origin, and body return to the discussion, but not as visible attributes.

Having a memory is not the same as appearing to have a memory. Having been wounded is not the same as representing a scar. Feeling fear involves a body exposed to the possibility of some consequence.

Intention and responsibility enter as well.

Who chose to produce this? From what? Who answers for the image, the text, the decision to circulate them? Was someone trying to say something, or was a system simply fulfilling a request? And even when there was someone, where does the experience of the person who created it end and the experience merely represented by the work begin?

These questions become more important precisely because the surface can no longer resolve them.

Perhaps this is the most interesting shift.

For a while, we imagined we would recognize machines because they would be too perfect. When that began to fail, we went looking for humanity in noise, hesitation, and grain.

Except machines learned to stumble too.

They learned the pause, the wrinkle, improvisation. Even chance can receive art direction.

When looking human becomes an aesthetic, appearance continues to produce impression, affection, desire, and trust. It simply stops telling us, by itself, where all of it came from.

And perhaps the question “does this look human?” begins to lose some usefulness.

The next question is far less comfortable.

What life, what experience, what intention, and what responsibility exist before this trace?

Perhaps the problem is not only that machines have learned to look like us.

Perhaps it is discovering that we were never very good at recognizing who was really there.

References

ABIDIN, Crystal. #familygoals: Family Influencers, Calibrated Amateurism, and Justifying Young Digital Labor. Social Media + Society, vol. 3, no. 2, 2017.

KRAMER, Robin S. S.; JONES, Alex L.; FITOUSI, Daniel et al. AI-generated images of familiar faces are indistinguishable from real photographs. Cognitive Research: Principles and Implications, vol. 10, art. 70, 2025.

OPENAI. GPT Image Generation Models Prompting Guide. Technical documentation, 2026.

ELEVENLABS. Text to Speech Best Practices. Technical documentation.