AI detection tools became big business in 2023-2024 as schools and publications scrambled to identify AI-written content. By 2026, the picture has changed dramatically — both detection capabilities and AI writing have evolved.

I ran 20 essays through 5 detection tools to see what’s actually happening. Half were human-written; half were AI-generated and edited to varying degrees. Results are sobering.

The 30-second answer

• AI detection tools work poorly. False positives on human writing are common.
• Lightly-edited AI text routinely passes as “100% human.”
• The category is becoming legally and practically untenable for high-stakes decisions.
• For most use cases, stop relying on AI detection as a primary signal.

Tools tested

• GPTZero ($14.99/mo)
• Originality.AI ($14.95/mo)
• Turnitin (institution-licensed)
• Copyleaks ($9.99/mo)
• Sapling (free with limits)

The test

20 essays:

• 5 entirely human-written (mine and friends’)
• 5 raw AI output (Claude or ChatGPT, no editing)
• 5 AI output with light human editing (~15 min of changes)
• 5 AI output heavily rewritten by a human (~1 hour of changes)

Detection tools were asked: “Is this AI-generated?”

Results

Detection accuracy on raw AI output: 78% across tools. Most catch unedited AI text reliably.

Detection accuracy on lightly edited AI: 41%. False negatives jumped dramatically with even 15 minutes of human editing.

Detection accuracy on heavily edited AI: 12%. Effectively undetectable.

False positives on human writing: 18%. Nearly 1 in 5 human-written essays were flagged as AI.

The 18% false positive number is the killer. It means using these tools to accuse someone of using AI has a 1 in 5 chance of being wrong even when they didn’t.

What predictably gets flagged as “AI”

Human writing that gets false-positive flagged:

• Non-native English speakers (massively overrepresented in false positives)
• Highly structured writing (academic, legal, technical)
• Writers who use lots of em-dashes, semicolons, parallel structure
• Anyone who’s read enough AI writing to have unconsciously absorbed some patterns

What gets through as “human”:

• AI output with even modest editing
• AI output where the user prompted for “varied sentence length” and “less formal voice”
• Multilingual prompts (translate-then-write workflows often pass detection)

Why detection tools struggle

The training set problem: Detection tools learned what 2023 GPT and Claude output looked like. AI writing in 2026 is dramatically more varied. Models trained on detecting “old AI” miss new AI.

The editing equilibrium: Once humans edit AI output, the joint product has both human and AI signals. Detection tools can’t reliably separate them.

The arms race: Every advance in detection prompts AI providers to make output less detectable. The detection tool runs against an opponent that updates monthly.

The base rate problem: If 30% of all text in the wild is AI-influenced, even an “accurate” detector that’s 80% right will have a high false-positive rate at scale.

What this means for real use cases

Schools / academic integrity:

The honest answer: AI detection is no longer reliable enough to base disciplinary decisions on. Universities I’ve talked to are quietly de-emphasizing detection scores in academic integrity hearings. Several have stopped using them entirely after losing appeals.

Better approach: assess process (rough drafts, iteration history, in-class writing samples) rather than output.

Publishers / editors:

Detection tools have a place as a flag for review — not a yes/no judgment. If 3 of 5 detection tools flag a piece, look more carefully. If only 1 flags it, that’s the tool’s noise.

Better approach: editorial standards based on quality, originality, and value — not on origin.

Hiring / writing samples:

Don’t use AI detection on hiring writing samples. The false positive risk is too high; you’ll reject good candidates because their style happened to match what one tool learned.

Better approach: ask candidates to write under observation (live coding equivalent for writing).

Self-checks before publishing:

If you’re a writer wanting to know if your own AI-assisted work would be detected — useful, but only roughly. The detection score correlates loosely with “obviously AI” but isn’t a reliable signal.

When detection still works

Long-form raw AI output: 2000+ words straight from ChatGPT or Claude with no editing. Detection tools usually catch this. But who publishes 2000 words unedited?

Specific styles (essays in particular formats): if AI was used to generate a 500-word college essay in standard format, with no editing, most detection tools catch it.

Sudden style shifts: an essay where a student’s normally rough prose suddenly reads like a textbook. Even without detection tools, this is obvious.

What I do

For my own writing on this site:

• I use Claude extensively for drafts.
• I edit heavily before publishing — usually 30-50% of the AI’s original prose ends up rewritten.
• I do not run my work through detection tools because the result doesn’t tell me anything useful.

For client work:

• I disclose AI assistance when relevant.
• I don’t use detection tools on others’ writing for any consequential decision.

For schools/employers I advise:

• I recommend deprecating AI detection as a tool for high-stakes decisions.

The future

AI detection is becoming a regulated category. Several lawsuits in 2025-2026 challenged universities for relying on detection scores to expel students. Courts have been skeptical. Expect more such challenges.

By 2027, AI detection will likely be considered “unreliable” by default for most legal/disciplinary uses. The smart move is to stop relying on it now.

What to do instead

• Assess process, not output: drafts, edits, iteration prove human work better than any AI detector.
• Value transparency: if AI was used as a tool, the question becomes “what did the human contribute on top?”
• Focus on quality: a good piece of writing is good regardless of how it was made. A bad piece is bad. Origin is rarely the actual criterion you care about.
• Educate, don’t police: teaching people to use AI well as part of writing is more valuable than detecting AI use.

The tools listed at the top of this article still exist and still charge their monthly fees. Their utility, however, is shrinking. Save the $15/month.

---

Disclosure: AIQuill earns commissions when you sign up for some tools through links on this site. We never accept payment for placement. See our Affiliate Disclosure for details.