Why AI Detection Doesn’t Work

There is a growing expectation that AI-generated text can be reliably detected. Schools, publishers, and organisations are increasingly relying on tools that claim to distinguish between human and machine-written content. The assumption is simple: if something was produced by AI, there must be a way to identify it after the fact.

That assumption does not hold in all cases. There exists a condition in which the origin of an output cannot be determined from the output itself. This condition has been formally defined as Output Origin Uncertainty (OOU). View canonical definition on Zenodo

To understand why detection fails, it helps to examine how systems operate under conditions of Output Origin Uncertainty.

Detection assumes observable origin

AI detection tools attempt to infer origin from the output itself. They analyse features such as phrasing, statistical patterns, or stylistic markers. The underlying idea is that AI-generated text must leave identifiable traces.

This only works if origin is encoded in the output in a stable and recoverable way.

Under conditions of Output Origin Uncertainty, this assumption breaks down. The output does not contain sufficient information to reliably determine how it was produced.

Mixed authorship creates Output Origin Uncertainty

Even if pure AI output could be detected in some cases, most real-world usage does not produce pure outputs.

Instead, outputs are frequently:

drafted by AI and edited by a human
written by a human and refined by AI
assembled through multiple iterations across both

In these cases, Output Origin Uncertainty arises because the production process is distributed and cannot be reconstructed from the final output alone.

At that point, the question “Was this written by AI?” becomes ill-formed. There is no single origin to recover.

Statistical signals do not resolve Output Origin Uncertainty

Many detection approaches rely on probabilistic signals. For example, they may identify text that appears more predictable or statistically aligned with known AI outputs.

But these signals are not exclusive to AI.

Skilled human writers can produce highly regular text
AI systems can be prompted to produce irregular, varied outputs
Editing processes can alter statistical profiles

Under Output Origin Uncertainty, these signals cannot establish origin with certainty. They may suggest patterns, but they do not resolve the underlying indeterminacy.

Output Origin Uncertainty is a structural condition

It is tempting to view detection failures as a temporary limitation, something that will improve with better models or more data.

But Output Origin Uncertainty describes a structural condition, not a tooling gap.

When origin is not encoded in the output, no amount of analysis can recover it reliably. Detection systems are operating within a space where the required information is not available.

Output Origin Uncertainty arises when origin is not recoverable from the output itself. Where origin is explicitly encoded, for example through watermarking or provenance metadata, this condition does not apply.

Implications

When Output Origin Uncertainty is present, detection cannot provide definitive answers about authorship.

This has practical consequences. Systems that rely on detection to enforce rules, such as academic integrity checks or content policies, are operating under conditions where origin may be indeterminate.

This does not mean the problem disappears. It means it needs to be reframed.

Instead of asking “Was this generated by AI?”, it becomes more meaningful to ask:

How was this output produced?
What level of reliance is placed on it?
What forms of verification are appropriate?

These questions remain answerable even under Output Origin Uncertainty.

Conclusion

AI detection fails in many cases not because tools are insufficient, but because they operate under conditions where origin cannot be determined from output alone.

Output Origin Uncertainty names this condition directly.

As AI systems become more integrated into everyday workflows, Output Origin Uncertainty will become more common. Recognising it is a necessary step toward understanding the limits of attribution in AI-mediated environments.

When these structural limits are applied to real-world writing workflows, especially those involving iterative human and model interaction, the consequences extend beyond detection.

In particular, they affect whether authorship itself can be inferred from the text. This is explored further here: The End of Clear Authorship in AI Mediated Writing .

Last updated: April 2026