In this guide · 7 sections
Short answer
No. OpenAI withdrew its own AI text classifier in July 2023 for low accuracy: it caught only 26% of AI text and wrongly flagged 9% of human text. A Stanford-led study found seven detectors flagged 61% of non-native English essays as AI on average. Judge work by its accuracy and sources, not a detector score.
What AI text detectors claim to do
An AI text detector looks at a piece of writing and estimates how likely it is that a language model produced it, usually as a percentage. Most rely on statistical patterns: AI text tends to use very predictable word choices, so text that is "too predictable" scores as AI.
That method has built-in weaknesses, which OpenAI listed when it released its own classifier:
- Very unreliable on short texts below 1,000 characters, and sometimes wrong on long ones.
- Significantly worse in languages other than English, and unreliable on code.
- Predictable text cannot be judged at all (OpenAI's example: a list of prime numbers is the same whoever writes it).
- AI text can be edited to evade the classifier.
What OpenAI found when it built a detector
In January 2023 OpenAI released a classifier to distinguish human from AI-written text. Its announcement was frank about the results:
| Measure | OpenAI's result | What it means at work |
|---|---|---|
| AI-written text correctly flagged | 26% | Three in four AI texts passed as human |
| Human text wrongly flagged as AI | 9% | About one in eleven honest documents accused |
| Status | Withdrawn 20 July 2023 for low accuracy | The company that makes ChatGPT stopped offering a detector |
OpenAI's help article for educators, How can educators respond to students presenting AI-generated content as their own?, puts it bluntly: asked whether AI detectors work, the answer is "in short, not in our experience". It adds that when OpenAI trained a detector, it labelled Shakespeare and the Declaration of Independence as AI-generated, and that ChatGPT has no knowledge of what it wrote: answers to "did you write this?" are random.
Why non-native English writers get flagged
This matters in Singapore, where many staff, candidates and suppliers write in English as a second language. A Stanford-led study, GPT detectors are biased against non-native English writers (Liang et al., published in the journal Patterns in 2023), tested seven widely used detectors:
- On 88 essays by US eighth-graders, the detectors were close to accurate.
- On 91 TOEFL essays by non-native English writers, they wrongly labelled an average of 61.22% as AI-generated.
- All seven detectors agreed on 18 of the 91 TOEFL essays (19.78%) being AI, although humans wrote them all.
- Simple prompting to vary the wording cut the false positives and also let AI text bypass the detectors.
The reason is the same statistical shortcut: writers with a smaller English vocabulary use more common, predictable words, which is exactly what detectors read as AI. Acting on a detector score could therefore fall hardest on staff writing in their second language, an unfair outcome for something the tool cannot actually prove.
Where detector scores cause real problems at work
The risk is not the tool itself but the decisions people make with it:
Reasonable use
- A prompt to look more closely at sources and facts
- A conversation starter about your AI policy
- Checking your own team's drafts for overly generic writing
Risky use
- Rejecting a job application on a detector score
- Disciplining staff based on a percentage
- Refusing a supplier's proposal as "AI-written" without other evidence
- Uploading confidential documents to a free online detector
The last point is easy to miss: pasting a contract, HR document or client report into a free website sends that content to a third party. Under the PDPA you remain responsible for personal data you share, so treat online detectors like any other outside service.
What to do instead: process-based checks
If the real worry is quality, accuracy or honesty, check those directly. Google takes a similar view for websites: its guidance on generative AI content focuses on whether content is accurate, original and useful, not on how it was produced, while warning against mass-producing pages without value.
Write a short AI use policy
Say which AI tools staff may use, for what, with which accounts, and what must never be pasted in (client data, NRIC numbers, contracts).
Ask for disclosure, not secrecy
Let staff say when AI helped draft something. People hide AI use when it is banned, not when it is allowed with rules.
Check facts and sources
For reports and proposals, verify the numbers, dates and citations. Wrong facts are the real risk whether a person or AI wrote them.
Use live tasks for hiring
Replace take-home essays with a short supervised exercise or a discussion of the candidate's own work.
Keep version history
Documents written in Google Docs or Microsoft 365 keep edit history, which shows how a piece of work developed if questions arise.
Our AI policy and staff training checklist has a ready structure, and our comparison of ChatGPT vs Claude vs Gemini for business covers which business accounts keep your data out of model training.
Do it yourself or get help?
Most of this is policy and habit rather than software. A one-page AI use policy and a team discussion go further than any detector subscription.
AI at work: minimum set-up
IT Sorted is a Singapore-based team of developers with over 10 years of experience. Our Security Awareness Training (S$390) sets up a training platform in your account with a 12-month plan of 5-minute lessons, Singapore-specific scam topics, completion tracking and a recorded walkthrough for new joiners, a practical way to roll out the habits above alongside your AI policy.
Frequently asked questions
Not reliably. OpenAI withdrew its own AI text classifier on 20 July 2023 because of low accuracy: it identified 26% of AI-written text and wrongly flagged 9% of human text. OpenAI's help centre says detectors have not proven reliable in its experience. Independent research also found high false positive rates.
No. OpenAI says ChatGPT has no knowledge of what content could be AI-generated and will sometimes make up answers to questions like did you write this or could this have been written by AI. Those responses are random and have no basis in fact, so never rely on them.
Detectors treat predictable wording as a sign of AI. Writers with a smaller English vocabulary use more common words, so their writing looks predictable. A Stanford-led study found seven detectors wrongly labelled an average of 61.22% of TOEFL essays by non-native writers as AI, while US student essays were classified almost correctly.
No. A detector score is not evidence, and the research shows non-native English writers are flagged far more often, which could make your hiring unfair. Assess candidates on a short live task or interview about their own work, and judge written applications on accuracy and relevance.
Google's Search guidance focuses on whether content is accurate, original and helpful, not on how it was produced. It does warn that using generative AI to create many pages without adding value can break its spam policy on scaled content abuse. Review and edit AI drafts before publishing.