AI detection tools score text by how predictable its wording is and how much its sentence rhythm varies, then turn that into a percentage. That number is a statistical guess about style, and independent tests show it misfires too often to count as proof of who wrote anything.
- OpenAI’s classifier caught only 26% of AI-written text before retirement.
- Stanford found detectors flagged 61% of non-native essays as AI, versus 5% for native speakers.
- Google says content origin isn’t a ranking factor.
- Under the EU AI Act, human-edited marketing posts generally need no visible label.
How AI Detection Actually Scores Your Text
Most detectors lean on two signals: perplexity and burstiness. Perplexity measures how predictable each next word is to a language model. Burstiness measures how much your sentence length and structure swing around. Some vendors list both openly as score inputs, often next to a third check for tone that feels overly generic or repetitive.
Low perplexity and a flat rhythm push the score toward “AI,” because language models tend to pick the statistically likely word and produce evenly paced sentences. Human writers wander more. They drop in a fragment. Then they follow it with a long, winding sentence and an odd word choice nobody saw coming.
Think of a detector as a predictability meter: it can tell you a sentence sounds like something a model would write, and it has no idea who actually typed it.
Perplexity in plain English: how surprised a language model is by your next word. Simple, common phrasing creates little surprise, and detectors count that as a hint of AI, even when a person wrote every word.
Trained classifiers go a step further. Vendors feed a model large sets of labeled examples (texts known to be human, texts known to be AI) and let it learn its own patterns, which can include perplexity and burstiness alongside subtler features. The output still lands as a probability. None of these signals measure expertise or usefulness, the qualities Google says its ranking systems reward.
Why “99% Accurate” Doesn’t Survive Scrutiny
Vendor accuracy claims come from vendors’ own test setups, and independent results look a lot weaker. The clearest case comes from OpenAI itself, which launched its AI Text Classifier in January 2023 and retired it that July over its low accuracy. When the company building the models can’t reliably spot their output, any third-party percentage deserves a raised eyebrow.
OpenAI’s classifier caught only 26% of AI-written text while wrongly labeling 9% of human writing as AI. In a Washington Post evaluation, GPTZero was the most consistent tool with 5 of 7 correct, and OpenAI’s classifier ranked among the worst. A Stanford study published in Patterns found seven commercial detectors flagged 61.3% of TOEFL essays by non-native speakers as AI, against roughly 5% of native-English essays.
The Stanford authors traced that bias straight back to perplexity: simpler, more predictable vocabulary looks “AI” to the tools. Nearly 20% of the non-native essays were flagged unanimously by all seven detectors. A high score often measures plain language.
Vendor numbers need a footnote too. One widely used detector advertises 95.7% detection with about 1% false positives on the third-party RAID benchmark, plus a de-biasing step for non-native writers. That figure is the vendor quoting a benchmark, with no independent audit behind it, so test any tool on your own content before you rely on it.
Does Google Actually Penalize AI Content?
Google doesn’t penalize AI content. Google’s Search Central guidance on AI content says automation only violates its policies when used “with the primary purpose of manipulating ranking in search results.” What its systems reward, in Google’s own words, is “original, high-quality content that demonstrates qualities of what we call E-E-A-T: expertise, experience, authoritativeness, and trustworthiness.”
Detector scores play no part in that. Neither Google nor Bing uses AI detection tools as a ranking signal, and nobody has shown that rewriting content to “pass” a detector moves rankings at all.
AI search follows a similar pattern. One citation analysis found 97% of Google AI Overview citations include at least one page from the top 20 organic results, so the work that earns rankings also earns AI answers. Perplexity leans more on community sources like Reddit, which makes earned mentions matter more there. For Google’s AI surfaces, classic ranking quality stays the main lever.
When a Flagged Score Actually Matters for Your Team
A score matters when someone has turned it into a rule, through a contract, a platform policy, or your own QA gate. Everywhere else, it’s background noise.
If a client contract requires a specific tool and threshold, agree on both in writing. If a platform you publish on has an AI policy, follow its rules regardless of Google’s stance. Before adopting a detection gate for your team, test it on known human-written pieces and count the false positives. For your own blog with no contractual obligation, a high score is no reason to rewrite a post that already serves readers well.
When a piece does get flagged, the fastest way through is evidence of how it was made. Prepare it once and reuse it every time:
- Pull the version history from Google Docs or your CMS, which shows the piece taking shape over time.
- Collect the brief and outline, plus the sources the writer used.
- Describe in two sentences where AI helped and who edited and approved the final version.
- Run the same text through a second detector and note the result.
Clients and platforms usually name one of five tools. Prices are typical as of 2026, and what moves them is word volume, API access and whether plagiarism checks come bundled in.
| Tool | Where it fits | Watch-out | Access and price |
|---|---|---|---|
| Originality.ai | Agency and publisher workflows that scan client deliverables | Credits drain fast if you rescan every draft | $30 one-time for 3,000 credits, Pro $14.95/month for 2,000 credits (1 credit = 100 words), Enterprise with API $179/month |
| GPTZero | Quick spot checks, built mainly for educators and also used by publishers and marketers | Headline accuracy figures are vendor-reported | Free up to 10,000 characters per scan without an account; paid plans typically in the low tens of dollars a month |
| Turnitin | Academic integrity checks inside schools and universities | Marketing teams can’t buy it directly | Licensed per institution, priced on request |
| Copyleaks | Combined AI and plagiarism checks across many languages, with enterprise integrations | Results vary between independent tests | Entry plans typically around $10–20/month, more for API and volume |
| Winston AI | Publishers and educators who also scan uploaded documents and images | Smaller record of independent testing | Entry plans typically around $12–20/month, priced by word credits |
EU AI Act Labeling: What Actually Applies to Your Blog Posts
For most marketing content, no visible label is required. Article 50 of the EU AI Act, applicable since 2 August 2026, requires providers of generative AI tools to mark their output in a machine-readable format. Companies publishing that output must add a visible label only when AI-generated text covers matters of public interest such as elections or public health, and even then the duty falls away if the content went through genuine human editorial review.
The EU Commission has published three optional icons: “Partially AI-modified,” “Fully AI-generated” and a basic AI icon for general involvement. A routine post about product features or SEO tactics, edited and owned by a person on your team, generally needs none of them. The rules don’t apply retroactively to content created before 2 August 2026. Fines follow the AI Act’s general tiers, and transparency breaches sit in the middle one:
- Large companies: up to €15 million or 3% of global annual turnover, whichever is higher.
- Startups and SMEs: whichever amount is lower, capped at 3% of turnover.
What Applies in Austria
Article 50 applies directly in Austria, with no separate national law needed. Your official point of contact is the KI-Servicestelle at RTR, the AI service desk of Austria’s broadcasting and telecom regulator in Vienna, which publishes guidelines, infographics and FAQs on these duties.
RTR sets a clear bar for the human-review exemption: someone has to check the content substantively, and a person or company has to take editorial responsibility for publishing it. A spell-check or grammar pass doesn’t qualify. RTR also notes that routine product or how-to posts often fall outside the public-interest scope in the first place.
Austrian SMEs get simplified documentation duties and mostly free access to regulatory sandboxes. No official statistics on Austrian AI Act fines exist yet, and Austrian legal advisors expect authorities to focus on guidance during the first months before moving to sanctions.
Orientation only: This summary reflects the Article 50 text and RTR guidance as of 2026 and doesn’t replace legal advice. If you publish public-interest content at scale, check your setup with counsel or the RTR service desk.
Why Two Detectors Disagree on the Same Text
Each detector learned from its own training data and was built against a specific generation of language models. Each one also sets its own threshold for what counts as “AI.” Paste one paragraph into two tools and you get two separate guesses from two separate models, so a low score in one and a high score in the other can both be honest outputs. A single score proves nothing, and a pattern across several tools still only describes style.
Editing shifts scores fast. In the Stanford study, rewriting the same non-native essays with more elaborate vocabulary cut the false-positive rate from 61.3% to 11.6%. Same writers, same ideas, different words. Humanizer tools exploit exactly this, which says a lot about how shallow the signals are. Swapping synonyms or shuffling sentences can move the number without making the piece any better, while adding your own examples, numbers and opinions changes both.
Building a Content Process That Doesn’t Depend on a Score
Skip detection gates. A score can’t tell you whether a post answers the reader’s question, and that answer is what both Google and your readers reward. A human editor who substantively reviews and owns every piece also keeps you inside the AI Act’s editorial-review exemption. Run these three checks before anything goes live:
- Does it answer the searcher’s question within the first few lines?
- Does it include specifics a generic draft couldn’t produce, such as real numbers or clear opinion?
- Has a named person fact-checked and approved it?
If you want the research, writing and publishing loop automated while keeping that human sign-off, Claire is an autonomous AI content marketing agent that writes and publishes blog content, then keeps optimizing it, offered in Autopilot or Copilot mode. For service businesses, the page on SEO and AI visibility for coaches shows how Claire writes articles you approve, and contractors or hotels can apply the same approval routine.
The Question That Actually Matters More Than Any Score
The question worth asking about any post is whether a person would want to read it to the end. Google says it rewards that, AI Overviews mostly cite pages that already rank for it, and your readers decide on it in seconds.
Pick your last five published posts and run them through the three checks above. Fix what fails, write down who edited and approved each one, and drop any detection threshold from your publishing workflow unless a contract requires it.
FAQ
Can an AI detector tell the difference between ChatGPT output and heavily edited AI-assisted writing?
Rarely. Detectors have no reliable category for “AI draft, human edit.” They score style, and heavy editing changes the style, so scores for edited pieces land almost anywhere on the scale.
Should we tell a client or boss we used AI if a detector flags our content?
Usually, yes, if AI was part of your process. Being upfront early protects your credibility far better than arguing about a percentage. Show the version history and walk them through the quality checks the piece passed.
Do AI detectors get better over time as models like GPT-5 improve, or do they fall further behind?
They tend to lag behind. Each new model generation writes with more varied rhythm and word choice, and detectors have to retrain to catch up. OpenAI retired its own classifier for low accuracy, which shows how hard keeping pace is even for the model maker.
Is there any detector with a genuinely low false-positive rate we can trust more than others?
No, independent research hasn’t confirmed one. The lowest figures you’ll see, around 1% false positives, come from vendors quoting their own benchmark runs. Test any tool on a handful of your own human-written posts before trusting its verdicts.
Where does an AI label go if a post does need one?
Directly on the content itself. RTR states that a note in your imprint or terms and conditions doesn’t meet the transparency duty. The Commission’s optional icons, paired with plain text like “AI-generated,” are designed for exactly that spot.