Back to Blog
Why Prompt Engineering Can't Make AI Text Undetectable
AI Research

Why Prompt Engineering Can't Make AI Text Undetectable

Tech Research Team6 min readOctober 10, 2026

There is an entire subgenre of internet advice devoted to making AI text undetectable. Write like a human. Vary your sentence length. Use a conversational tone. Avoid the word "delve."

Some of these instructions improve the writing. Almost none of them fool detection. We tested the most popular ones against our own detector to see what actually moves the score.

The Premise of Prompt Evasion

The logic sounds reasonable. If detectors look for certain patterns, remove the patterns. If AI writes in a uniform register, prompt it for uneven prose.

The flaw is that prompt instructions change the surface of the text, and detection does not primarily read the surface. It reads how predictable the language is, and predictability is a property of the generating process, not of the instructions it was given.

What We Tested

We started with a baseline: five prompts, each generating a 600-word essay with no special instructions. Then we regenerated each essay with a different evasion technique applied.

  • Instructing the model to vary sentence length dramatically
  • Instructing the model to write in a casual, conversational voice
  • Banning a list of "AI words" including delve, tapestry, and testament
  • Asking the model to include a personal anecdote
  • Asking the model to add a deliberate typo or two
  • Asking the model to write like a specific human author
  • Running the output through a paraphrasing tool
  • Running the output through a "humanizer" service

Which Tricks Moved the Number

Two techniques produced a meaningful drop.

The first was the personal anecdote, and the reason is simple: the model has to invent specifics, and invented specifics are irregular. Names, places, and small details break the statistical uniformity that triggers detection.

The second was the aggressive paraphrase, which restructured sentences rather than swapping synonyms. It lowered scores somewhat, at the cost of making the prose notably worse. Several paragraphs became vaguer and one lost its actual argument.

Everything else moved the number by a few points, which is noise when scores naturally vary between runs.

The typo trick did nothing measurable. Neither did banning words. The conversational voice produced text that read like a casual model, which is still a model.

Why the Rest Don't Work

Every language model generates text one token at a time, selecting from a probability distribution. That process has a signature: locally smooth choices, low variance in structural complexity, and a tendency toward the most conventional phrasing available.

When you tell a model to vary sentence length, it varies sentence length in a way that is itself conventional. When you tell it to sound casual, it reaches for the same casual markers every model reaches for. The instruction changes which distribution the model samples from. It does not change the fact that it is sampling.

Human writing is different in kind, not degree. People write a sentence because of something that happened to them that morning. That is not a distribution a prompt can reach.

What Actually Changes the Score

If you want text that reads as yours, write it as yours. The parts of a document that reliably score low are the parts only you could have produced:

  • A specific number from your own work
  • A named person, place, or project
  • An opinion you would defend in an argument
  • A sentence that is a little awkward because that is how you think

None of these are prompt tricks. They are just writing.

The Takeaway

Prompt engineering is genuinely useful for producing better output. It is useless as a disguise, because the thing detectors read cannot be prompted away.

There is also a practical argument against evasion. Text engineered to beat a detector reads like text engineered to beat a detector: slightly off, oddly paced, and stripped of whatever the original was trying to say. Even if it worked, it would not be worth publishing.