Can AI Writing Be Detected?

The Short Answer

AI writing cannot be reliably identified from the text alone.

Once a piece of writing exists without a verifiable record of how it was produced, there is no method that can determine with certainty whether it was written by a human, an AI system, or some combination of both.

Why Detection Feels Possible

Detection tools exist, and they sometimes appear to work.

They analyse features such as:

From these signals, they assign probabilities: likely AI, possibly human, or similar classifications.

In controlled conditions for example, unedited outputs from known models, these tools can show reasonable accuracy.

This creates the impression that authorship can be inferred from the text itself.

The Structural Limitation

The difficulty is not primarily technical. It is structural.

A piece of text does not contain a record of its own origin.

The observable properties of writing — vocabulary, grammar, style — describe what the text is, not how it was produced. Those properties can be generated by:

Because these processes can produce overlapping outputs, there is no feature or combination of features that uniquely corresponds to one origin.

Detection systems therefore rely on correlations, not direct evidence.

And correlations are:

They can indicate patterns, but they cannot establish origin with reliability.

Output Origin Uncertainty

This condition can be described as Output Origin Uncertainty (OOU).

OOU refers to a simple structural fact:

Once the chain of production is not preserved, the origin of a text cannot be recovered from the text alone.

This is not a limitation of current tools that will be resolved by scale or refinement. It arises from the relationship between production and observation.

The output does not encode its own history.

Why “Better Detection” Doesn’t Solve It

It is reasonable to expect detection systems to improve.

They may:

However, these improvements do not remove the underlying constraint.

As models evolve, and as human and AI writing continue to overlap, any pattern used for detection can be reproduced, avoided, or altered through editing.

This means detection remains probabilistic and conditional, rather than definitive.

What About Watermarking and Provenance?

There is an important distinction between inferring origin from text alone, and recording or embedding origin during generation.

Techniques such as watermarking or cryptographic provenance aim to attach origin information to outputs at the point of creation.

These approaches can be useful in controlled environments.

However, they depend on:

Once text is edited, paraphrased, or re-used, these signals can be weakened or removed.

Crucially, these systems do not change the core condition:

The text itself does not inherently contain reliable evidence of its origin.

They introduce external or injected signals, rather than enabling origin to be inferred from the text as such.

Practical Implications

In practice, this means:

In contexts where origin matters — education, research, professional work — the output alone can no longer serve as unambiguous evidence of human capability.

This limitation becomes more visible in practice when writing is produced through mixed human and AI workflows. In these cases, the connection between a finished text and its origin weakens further.

For a deeper explanation of how this affects authorship itself, see: The End of Clear Authorship in AI Mediated Writing .


This explanatory note relates to the formal definition of Output Origin Uncertainty and the wider EntityWorks Standard.

Last updated: April 2026