Teachers, editors and employers want to know whether a piece of writing came from a person or a machine. Dozens of tools claim to answer that question with a percentage score. Unfortunately, the evidence shows that AI-text detectors are unreliable, and treating their scores as proof can unfairly punish honest writers. Here is why.
What detectors try to do
An AI detector analyses text and estimates the likelihood that a language model produced it. Common ingredients:
- Perplexity: how predictable the text is to a language model. Machine text tends to use more probable word choices.
- Burstiness: variation in sentence length and structure. Humans often vary more.
- Trained classifiers: neural networks trained on examples of human and AI text.
- Watermark detection: if a generator embeds a hidden statistical pattern in its word choices, a matching detector can look for it. This works only for text from systems that apply the watermark.
Why they fail
1. Human and AI writing overlap
Clear, formal, well-edited human writing can look "predictable". Many people, including students and professionals writing in a second language, use simple, standard phrasing. Detectors can mistake that for machine text. Research has reported that detectors flagged writing by non-native English speakers more often than writing by native speakers.
2. Models keep changing
A detector trained on text from older models may miss newer ones. Each new generation writes differently, and users can ask models to adopt a particular style, making detection a moving target.
3. Easy to evade
Paraphrasing, light editing, asking the model to write "like a teenager", mixing human and AI sentences or running text through another tool can drastically lower detection scores.
4. False positives have real costs
Even a detector with a low error rate will falsely accuse some people when applied to thousands of documents. A 1 percent false-positive rate sounds small, but in a class of 200 students over a semester it can mean several wrongly accused students.
5. Scores are not explanations
A figure like "78% AI-generated" is easy to misread. It does not mean 78 percent of sentences were written by AI. It is a model's estimate with substantial uncertainty, and detectors rarely show their reasoning.
6. Short texts are especially unreliable
Few words give little statistical evidence. Many tools themselves warn that results on short passages are weak.
7. Inconsistent results
Different detectors can give very different scores for the same text, and the same detector can change its answer after small edits.
8. Non-prose content
Code, poetry, lists, translated text and formulaic writing such as legal boilerplate or scientific methods sections confuse the models.
9. Training data bias
Detectors reflect the data they were trained on. Topics, dialects and styles outside that data produce poorer judgements.
What this means in practice
Major AI companies and educational bodies have acknowledged the limitations of detectors, and some have withdrawn or downplayed their own tools because of low accuracy. Responsible institutions advise against using detector scores as the sole evidence of misconduct.
Better ways to deal with AI use
For teachers and institutions
- Set clear rules about when and how AI may be used, and require disclosure.
- Design assessments that show thinking: in-class writing, oral explanations, drafts with revision history, reflections on process, project-based work and personalised topics.
- Look at process evidence: version history in word processors, outlines, notes and sources.
- Talk to the student. Ask them to explain their argument and terms in their own words. Genuine authors can usually discuss their work in depth.
- Compare with known samples of the student's earlier writing, with care and awareness of improvement over time.
- Use detectors, if at all, only as a conversation starter, never as proof.
- Provide due process: a fair appeal path and a presumption of good faith.
- Teach AI literacy so students learn appropriate use, citation and limits.
For editors and employers
Evaluate quality, accuracy, originality and fit instead of guessing authorship. Check facts and sources, ask for drafts or interviews with the writer when it matters and set clear policies for freelancers.
For writers
Keep drafts and notes, save version history and be ready to explain your work. If AI helped, say so according to the rules of your context.
Provenance is promising
A better long-term approach than guessing is provenance: tracking how content was created. Standards for content credentials for images and other media are emerging, and watermarking for generated text is under research. These methods require cooperation from tool makers and are not foolproof, but they are more principled than statistical guesses after the fact.
A quick illustration of the limits
Consider this toy example, which shows why predictability is a weak signal. It scores how often a text uses very common words, a crude proxy for "predictable" writing:
COMMON = set("the of and to in is that it for as with was on are this be by at have from".split())
def predictability(text: str) -> float:
words = [w.strip(".,;:!?").lower() for w in text.split()]
return sum(w in COMMON for w in words) / max(len(words), 1)
formal_human = "The results of the study are consistent with the findings that have been reported in the literature."
casual_human = "Honestly, the whole thing was chaos, but somehow Rina kept us laughing till midnight."
print(round(predictability(formal_human), 2))
print(round(predictability(casual_human), 2))A formal human sentence scores as "more predictable" than a casual one, even though both are human. Real detectors are more sophisticated, but they face the same fundamental overlap.
Frequently Asked Questions
Are any AI detectors accurate?
Some perform better than others on certain test sets, but none is reliable enough to serve as sole evidence in high-stakes situations.
Can I prove my work is mine?
Keep drafts, notes, version histories and sources, and be prepared to explain your reasoning. Those are stronger evidence than any detector score.
Will watermarking solve the problem?
It may help for text from cooperating providers, but it will not detect text from tools that don't use it, and edits can weaken the signal.
Should schools ban AI instead?
Bans are hard to enforce. Many educators prefer clear guidelines, redesigned assessments and teaching responsible use.
Conclusion
AI-content detectors give uncertain scores, make false accusations and are easy to evade. Judge work through process evidence, conversation and well-designed assessments, and use any detector, if at all, only as one weak hint among many.