AI text detectors are showing up everywhere: even if you don’t have a teacher or a workplace directly checking on you, your client might be running your work through a detector. The pitch of the detectors is simple: they’ll tell you whether a human or AI wrote the text. But if you know what they are checking for, it’s not hard to confuse them.
A recent peer-reviewed study testing 805 samples across seven detectors found average accuracy dropped to just 22% once simple human-like edits were added.
Here are the techniques that work – ones that involve a little human intervention. If you’re looking for a simpler solution check out “How to make ChatGPT sound more Human“.
1. Increase complexity of thought and pace (“perplexity” and “burstiness”).
Human writing is jagged: we mix long, winding sentences with abrupt short ones. AI tends to be too smooth. To increase so called “burstiness” ask the AI to
- Write paragraphs with uneven lengths
- Mix long sentences with occasional clipped ones
- Avoid the “balanced essay” structure LLMs default to
AI prioritises readability and stylistic consistency. To increase complexity or “perplexity” ask the AI to:
- include a few advanced or domain-specific terms
- introduce more layered reasoning
- avoid perfectly linear explanations
2. Add small errors, awkward phrasing, and mild redundancy
AI is trained to avoid errors, poor phrasing, fragments, or repetition. But humans do it all the time. Edit your text to:
- Add one or two typos
- A slightly unusual word order. Fragment.
- A phrase repeated in a human, not mechanical, way
Not enough to look sloppy — just enough to nudge the statistical profile toward “human.”
3. Inject originality (content not in the training data)
Detectors fail hardest when you force the model into territory outside its training distribution.
Refer it to a website on the same topic. Give it a specific academic paper to incorporate. Upload some prior writing of yours. Human-brainstorm 10 different half-cooked ideas it can use “if useful” into the prompt. Copy and paste real customer reviews. Give it an anecdote, micro specific detail, or tell it about a prior similar experience.
You are forcing the AI to diverge away from the average of the training data into something new, breaking the “neutral and generic” pattern detectors search for.
4. Adopt opinion
AI is trained to avoid strong political takes, value judgements about groups, blunt emotional reactions or anything that might “risk offending any category of people.”cause harm”.
But more tellingly it probably also
- Avoided any real opinion, any possibility to offend anyone
- Avoided sarcasm, humour, or cynicism.
For the detector to interpret high human-ness explicitly add in some strong opinion, and some blunt or sarcastic statements. Inject a little “real person with preferences”.
5. Make real human edits
LLMs avoid things like:
- using two slightly different terms for the same thing
- shifting synonyms without reason
- stylistic drift
Humans constantly:
- call something “detectors,” then “classifiers,” then “tools that flag stuff”
- use a mix of dialects (“colour/color,” “towards/toward”)
- change sentence structure halfway through a piece
- add a comma, they probably shouldn’t have
This inconsistency destroys AI probability patterns.
The point is that you are human. When you make edits you are unintentionally introducing tone wobble, reasoning gaps, half-opinions, shifts in vocabulary, tiny errors, and (perhaps) a little personality. This is why you need to ACTUALLY human-edit the text.
How the text becomes more human
Divergence from the training data. High perplexity and burstiness. Errors. Opinion. Unpredictable texture.
The result is writing that has micro tone-shifts, unusual connections, a lack of correctness and a divergence from “neutral mode”. Some might even say human.
Once you understand how detectors actually work, it becomes obvious why a little intentional human messiness goes further than any clever prompt ever could.
Perkins, M., Roe, J., Vu, B.H. et al. Simple techniques to bypass GenAI text detectors: implications for inclusive education. Int J Educ Technol High Educ 21, 53 (2024). https://doi.org/10.1186/s41239-024-00487-w
Foltýnek, T., Newton, P.M. What Does YouTube Advise Students About Bypassing AI-Text Detection Tools? A Pragmatic Analysis. J Acad Ethics 24, 8 (2026). https://doi.org/10.1007/s10805-025-09675-3
Leave a Reply