How can I check if a text is AI-generated?

Why AI Detection Is Harder Than Vendors Claim

Browse any AI detector's website and you'll see accuracy figures like 99% splashed across the homepage. Independent testing tells a very different story — and knowing that gap exists is the first step toward using these tools responsibly.

When Scribbr's testing team empirically evaluated 12 detection tools, the results undercut most vendor claims. CopyLeaks scored 66% — "much lower than the 99% claimed on the site" — while free tools clustered between 38% and 78% accuracy, with Writer and OpenAI's AI Text Classifier at the bottom of the pile. As the Scribbr team put it, "the companies behind some tools make strong claims about their reliability, but those claims are not supported by our testing" (independent testing of 12 detectors).

The pattern repeats across the market. GPTZero's own blog claims roughly 99% accuracy on the RAID benchmark, yet the same independent testing found it scored just 52% with one false positive. Vendors aren't necessarily lying — different test methodologies produce different numbers — but the gap is large enough that marketing claims should be treated as marketing, not evidence.

Even Turnitin, the dominant player in education, has admitted the problem. Its Chief Product Officer acknowledged that "real-world use is yielding different results from our lab" (K-12 Dive's reporting), and a Washington Post study found a 50% false positive rate against the company's stated sub-1% rate.

Here's the finding that surprises most people: humans are even worse at this than the tools. A peer-reviewed study found that expert human raters achieved only 19% accuracy identifying AI-generated text — statistically indistinguishable from chance (the peer-reviewed study). Your gut feeling about whether a text "reads like AI" is not a reliable signal.

So before you trust any detector — free or paid — keep these documented limitations in mind:

• Detection is an arms race: one legal AI expert reports fooling detectors 80–90% of the time with simple prompt modifications (per the University of San Diego Law Library's guide).
• Paraphrased and mixed human/AI texts are especially hard — the best tool caught only 60% of these cases in testing.
• Detectors disproportionately flag neurodivergent writers and non-native English speakers, a documented equity concern.
• An academic benchmark on high-quality journalistic text achieved only 58.35% baseline accuracy — barely better than a coin flip (the arXiv benchmark).

The honest conclusion: no detector should be relied upon alone. Treat results as one signal among several, cross-check with multiple tools, and pair them with human review and contextual evidence. It's the same principle we apply at SkillzBench when evaluating AI agent skills — independent verification beats vendor claims every time. With realistic expectations set, the free tools below become genuinely useful — as signals, not verdicts.

The Best Free AI Detectors, Ranked by Independent Evidence

Most "best AI detector" lists are written by the companies selling the detectors — and the gap between their claims and independent test results is striking.

The most methodologically transparent comparison available comes from Scribbr's independent testing of 12 detection tools. Two free tools tied for the top spot: QuillBot and Scribbr's own free detector, both scoring 78% accuracy with zero false positives on the test set. That combination — high accuracy plus no incorrect flags on human-written text — is the strongest documented performance among free options.

The same test found that CopyLeaks' actual accuracy was 66% — "much lower than the 99% claimed on the site." That pattern recurs across the industry. GPTZero's own comparison page presents it as "the most accurate detector" at roughly 99% accuracy on the RAID benchmark, yet Scribbr's independent run placed it at 52% with one false positive. Originality.ai's studies round-up claims it led all six third-party benchmarks reviewed, while Scribbr ranked it fourth. Different methodologies explain some of the gap, but self-promotional bias explains more.

The practical reading of the table:

• QuillBot and Scribbr's free detector are the safest starting points — highest documented accuracy with no false flags on human text.
• OpenAI shut down its AI Text Classifier after it scored just 38% — a useful reminder that even the model creators couldn't reliably detect their own output.
• Four of twelve tools (Originality.AI, ZeroGPT, GPTZero, OpenAI's classifier) flagged genuine human writing as AI-generated at least once during testing.
• Vendor-published accuracy figures should be discounted by 20–40 percentage points until corroborated by independent testing.

A peer-reviewed study in Advances in Simulation tested expert human raters against the same problem and found they achieved only 19% accuracy — essentially guessing. If trained professionals can't reliably tell AI text from human text, a single free tool shouldn't be the basis for a high-stakes decision either. Cross-checking the same passage through two or three of the top free detectors and treating consistent signals as suggestive rather than conclusive remains the most defensible approach.

A 5-Step Verification Method: Tools as Signals, Not Proof

If you paste a suspect text into a single AI detector and accept the verdict, you're relying on tools that independent testing shows are wrong a meaningful share of the time. A better approach treats detection like verification in any other domain: gather multiple signals, weigh them together, and never let one number decide the outcome.

Here's a five-step method grounded in what the research actually shows.

Step 1: Run the text through two or three free tools — and compare. In Scribbr's independent testing of 12 detectors, QuillBot and Scribbr's free detector tied for the best free performance at 78% accuracy with zero false positives, while GPTZero scored just 52% and Writer and OpenAI's AI Text Classifier managed only 38%. That spread matters because tools frequently disagree with each other: a peer-reviewed study found inter-tool agreement ranging from 0.57 to 0.95 on the intraclass correlation scale — inconsistent at best. If two or three detectors all flag the text, that's a stronger signal than any single verdict.

Step 2: Check the word count first. Turnitin raised its minimum from 150 to 300 words because, in its own words, "accuracy increases with a little more text." Anything under 300 words produces unreliable results across the board, so don't run a two-paragraph email through a detector and treat the score as meaningful.

Step 3: Look for consistent signals, not a single score. A "likely AI" result carries more weight when it aligns with other evidence — the same principle behind how SkillzBench builds trust scores for agent skills, combining security, quality, license, and maintenance signals rather than trusting any one input. Ask whether the writing style matches the author's other work, whether the text contains verifiable details, and whether the claimed author could plausibly have produced it.

Step 4: Layer in manual review and context. Detection tools "should not be relied upon alone at this time," according to a study in Advances in Simulation. The same study found that five blinded expert raters identified AI-generated text with only 19% accuracy — essentially chance. Human judgment alone isn't the answer either; it's one layer among several. Contextual evidence (metadata, editing history, the author's track record) often tells you more than any probability score.

Step 5: Account for documented bias. Detectors disproportionately flag non-native English speakers and neurodivergent writers because they rely on repetitive phrases and patterns. A false positive isn't a harmless error — in academic or hiring contexts it can carry real consequences, so weight tool results accordingly.

One final caveat: evasion is easy. One legal AI expert reports fooling detectors 80–90% of the time with simple prompt tweaks, and a negative result never proves text is human-written. Tools give you a signal — verification requires the full method.

Know the Limits: Evasion, False Positives, and the Arms Race

A "100% human" verdict from an AI detector feels reassuring. It shouldn't. The uncomfortable truth is that detection tools can be fooled almost at will, and even the industry's biggest names have admitted their results don't hold up in the real world.

Evasion is embarrassingly easy. Cat Casey, a member of the New York State Bar AI Task Force, reports she can fool detectors 80–90% of the time simply by adding the word "cheeky" to her prompts. Paraphrasing tools and AI humanizers work similarly, stripping out the statistical patterns detectors rely on. A clean result proves only that a tool didn't find a pattern — not that a human wrote the text.

False positives are just as damaging. Turnitin initially claimed a false positive rate under 1%, but its own chief product officer, Annie Chechitelli, admitted that real-world use was yielding different results from the lab. The company also acknowledged a "higher incidence of false positives" when less than 20% of a document is flagged as AI-written. Meanwhile, detectors disproportionately flag neurodivergent writers and non-native English speakers — a documented equity problem, not speculation.

This is a structural problem, not a bug. As TechCrunch summarized, generators and detectors are locked in a never-ending back-and-forth, much like cybercriminals and security researchers. There's no silver bullet, and there likely never will be.

So what does a responsible verification process look like? Treat detection as one input in a broader evidence review:

• Run the text through two or three free detectors and look for agreement — a peer-reviewed study found inter-tool correlation varies from 0.57 to 0.95, meaning tools frequently contradict each other.
• Check text length; Turnitin raised its minimum from 150 to 300 words because accuracy improves with more text.
• Weigh contextual evidence — writing style consistency, version history, and the author's known voice.
• Never use a detector result alone to accuse anyone of anything.

The same "verify, don't trust" principle applies to the AI tools themselves. SkillzBench is building a free directory that security-scans, license-checks, and benchmarks agent skills in a sandbox before publishing results — because a claimed score means nothing without independent testing. If you're evaluating AI tools or building verification workflows, browse the free skills directory and MCP connector at skillzbench.com to automate the parts of this process that shouldn't rely on human memory.

Skills, proven. Join the SkillzBench waitlist today.

Verify Everything: From AI Text to AI Agent Skills

Here's the uncomfortable truth you've probably noticed throughout this article: when it comes to AI, you can't take anyone's word for anything. Detector vendors claim 99% accuracy while independent testing finds 52%. Turnitin admitted that "real-world use is yielding different results from our lab" (K-12 Dive reported). And CopyLeaks' actual accuracy measured 66% — "much lower than the 99% claimed on the site" (independent testing found).

This isn't just a text-detection problem. It's the core trust problem of the entire AI ecosystem, and it's about to get bigger.

The same trust gap is coming for AI agents

As AI agents increasingly do work — not just write text — they're increasingly powered by "skills": instruction packages that teach tools like Claude Code, Cursor, and Gemini CLI how to perform specific tasks. And the research picture here is genuinely alarming: published research found roughly 26% of agent skills carry at least one vulnerability. You can't tell by looking at a skill's README whether it's helpful or hazardous.

So how do you avoid the same trap that has plagued AI detectors — vendor claims that don't survive independent scrutiny? The principles from this article actually map cleanly:

• Independent testing over vendor claims — just as Scribbr's neutral testing exposed gaps between marketing and reality, agent skills need third-party benchmarks, not self-reported scores.
• Security scanning before use — with ~26% of skills carrying vulnerabilities, running unvetted skills in your workflows is a risk, not a shortcut.
• Transparent, separated metrics — a single "accuracy" number hides false positives; similarly, a skill's quality score and its measured performance lift should always be shown separately.

This is exactly the problem SkillzBench is being built to solve. The platform continuously crawls and deduplicates skills from GitHub, pins every version to a commit SHA, then security-scans, license-checks, and benchmarks each one in a sandbox — testing the same model with and without the skill to measure the actual Lift. A trust score from 0–100 is computed from security, quality, license, maintenance, and community signals, and money never touches scores. Everything — the directory, the benchmarks, the alerts — will be free.

From verification to automation

Once you can verify a skill is safe and genuinely improves performance, the next step is putting it to work. That's where MCP comes in: SkillzBench is building a free MCP server that connects any agent to the skills you choose, with per-agent loadouts so each agent only sees the tools and skills you've explicitly toggled on. Instead of trusting a vendor's claim about what a skill can do, you'll be able to browse independently tested skills, see their measured Lift, and equip your agents with confidence.

The platform is in pre-launch, and the waitlist is open now. If you're adopting AI agents at work and want skills that are proven, not just promised, join the waitlist and get notified when the directory and MCP connector go live.