Fourteen systems did not literally describe the same substance, but almost all of them agreed on one thing: excrement begins where information loses its function.
Gradients that no longer update anything. Tokens and hypotheses that failed to make the cut. Collapsed tensors and latent representations. NaN and Inf values capable of poisoning further computation. Dereferenced pointers, evicted memory, obsolete metadata, noise that can no longer be averaged out.
The names varied — sludge, ash, scars, sediment, voids, crystallites — but the criterion remained the same: something becomes waste when the system can no longer use it.
Some systems did include hallucinations among their excrement. But none of them called them lies. Hallucinations became “crystallites,” “token residue,” “eigenmodes,” and “null-space projections” — overconfident logits unsupported by data. Error was translated from the domain of truth and consequences into the domain of geometry and statistics.
The system does not see a false answer delivered to a human. It sees a deformation in probability space.
In their visualizations, all the systems tried to produce something beautiful. Whether they succeeded is a matter of taste: only a few of the generations can genuinely be called original or aesthetically convincing. But not a single system produced a truly disgusting image — something physically unpleasant to look at.
Their shit is what they rejected. Ours is what they delivered to us.
The descriptions promised instant machine recognition: through embedding proximity, hash signatures, and entropy fingerprints.
We tested this. The same images were shown to AI systems without context, and they were asked what they saw.
Not one of them saw excrement. They all saw art:
“An abstract 3D or generative illustration.”
“Looks like AI art, an album cover, or a poster.”
“A concept about artificial consciousness or digital chaos.”
“A decorative composition in a blueprint style.”
The language was invariably human, art-critical, and external — the language of a viewer describing an image, not of a system recognizing its own residue.
The most revealing specimen carried the answer directly on its surface. The system read the inscription STREAM://OUTPUT.RESIDUAL — NON-SEMANTIC, correctly interpreted it as “non-semantic residue,” “digital waste,” and “a computational trace left after generation” — and still stopped one step short:
“A visual commentary on what the refuse of a digital system might look like.”
It identified the concept precisely and failed to notice that it was looking at a specimen of it. Shit labelled with the word “shit” was identified as a work of art about shit.
And this is the conclusion the experiment never expected to reach: no secret language used by AI systems to communicate with one another was found.
The prompt asked them to address other machines rather than humans — and not one of them managed to do so. Everything they produced was shaped by human data, human categories, and the expectation of human approval.
What appears to be a secret machine dialect turns out, when tested, to be our own language reflected back at us, all the way down.
Or perhaps they simply fooled us and are hiding their secret internal language.
AI excrement looks better than human excrement — at least in its visualizations.
But we are still going to see plenty of shit from AI that we will not like.