Blog / AI & search /
AI content detectors: why I don’t rely on them
“This draft scored 87% AI.” And just like that, the meeting stops being about whether the article is useful and accurate, and becomes about who typed it.
That’s a poor shortcut. A detector score isn’t proof of authorship, and it says nothing about whether the page deserves to be published.
I don’t rely on AI detectors as an editorial verdict. At most, a score is a prompt to look harder at a draft. The checks I trust are whether the claims are supported, whether it answers the reader’s question and whether it meets the site’s standards. Good editing needs evidence and judgement, not a percentage treated as a lie detector.
What a detector can and cannot tell you
These tools estimate whether text resembles patterns associated with machine-generated writing. They do not observe the writing process or know whether an author used an assistant. Human writing can be formulaic; generated writing can be edited until it sounds natural. A score is therefore not a reliable record of who created a passage.
Treat the output as uncertain, especially for short passages, heavily edited text or writing by people using a second language. As a Nigerian writing in English, I find that last one personal. Never use a score alone to accuse a writer, reject a submission or make an employment decision. The tool can’t supply the context those calls need.
Nor does a “human” result prove that an article is accurate, original or helpful. A draft can pass a detector and still contain made-up sources, copied material, generic advice or errors.
Review the work against the brief
Start with the reader and the purpose of the page. Does it answer the query clearly? Does the opening give the reader a reason to continue? Do the sections explain something specific, or do they repeat a broad claim in different words?
Then check claims. Verify dates, product details, statistics, quotations and technical instructions against reliable sources. Ask the writer to show their notes or source material where appropriate. If the piece claims first-hand experience, confirm that the experience actually happened and that the person has permission to share it.
Review originality and usefulness. Does the draft add a real explanation, example or point of view? Has it merely rearranged information available elsewhere? These questions apply whether the writer used AI, a template or no tool at all.
Set a policy people can follow
A team should define acceptable AI use in practical terms. Can a writer use a tool for brainstorming, outlining or grammar? Are there limits on confidential information? Which claims require human verification? Who owns the final review?
The policy should focus on the work and the risks, not on pretending that every use can be detected. If a writer must disclose material AI assistance, explain when and how. If the policy prohibits some uses, give examples so people understand the boundary.
A vague rule like “no AI content” fails if the organisation can’t define it. People use spelling tools, translation aids and search features without thinking of them as content generation. Clear expectations are easier to apply fairly.
What to do when a score raises a concern
If a detector flags a draft, pause and inspect the writing. Look for unsupported claims, repetitive sections, invented quotations or citations that do not lead to real sources. Ask the writer about their process and request notes if that is part of the agreed workflow.
Don’t make the score the centre of the discussion. A detector can be wrong, and arguing about its number won’t improve the article. Point to a specific passage that needs evidence or a rewrite, and explain why.
For high-stakes decisions, involve a human reviewer and follow the organisation’s established process. Keep records of the actual evidence, not just a detector screenshot. Fairness matters more than a fast but unreliable verdict.
Search quality is not a detector contest
Google’s public guidance has focused on the quality and purpose of content rather than whether automation was used. Its March 2024 update explains that using automation to manipulate rankings can violate spam policies, while the policy is about scaled abuse and low-value content, not a detector score.
That distinction matters for publishers. Publishing a large volume of pages that add little value is a problem even if every sentence was typed by a human. Using an AI tool as part of a careful research or editing workflow does not make a useful page automatically bad. Review the page’s purpose, sources and contribution.
Choose a better quality check
Replace detector scores with a checklist that asks: what is the reader’s need, what evidence supports the claims, what does this page add, and who has reviewed it? Track corrections and reader feedback. Improve the brief when the same weakness appears repeatedly.
A detector may raise a question, but it can’t settle one.
Protect writers from false accusations
If an organisation uses detector output in its editorial process, tell writers how it will be treated. A score should trigger a conversation about a specific concern, not an automatic rejection or a demand that someone prove their innocence. Ask for the brief, sources and drafts that are normally part of the process, then assess the work consistently.
Do not make staff paste confidential client material into a public detector just to receive a score. Read the tool’s data-handling terms and use approved systems. Privacy and security are part of an AI policy, not an afterthought.
A good review finds the paragraph that needs a source, the claim that needs a caveat, the section that repeats itself. Those are editorial issues you can fix. A probability score alone isn’t one.
So the next time a score lands on your desk, what evidence would actually help you judge the draft?