You write blog posts or social captions with AI — or you’re about to — and you keep hearing two opposite pieces of advice. One: “Google bans AI text, run everything through a detector until it says 0% AI.” The other: “Google doesn’t care, generate a hundred articles a day.” Both lose you traffic, just by different routes.
The real picture is duller and more useful. Google doesn’t penalize text because a model wrote it. It penalizes text for why it exists and how much of it was mass-produced.
Below: what Google officially says, what a demotion looks like in practice, why AI detectors mislead, a workflow where AI drafts and a human is accountable, and how AI labeling works on social platforms.
What Google officially says
The position hasn’t changed since early 2023, when Google published dedicated guidance on AI-generated content: its ranking systems reward helpful content however it’s produced. The violation is using automation, AI included, to generate pages primarily to manipulate rankings.
In March 2024 this went into the spam policies as “scaled content abuse”: lots of pages made for rankings rather than people, whether they were written by humans, AI or both. The same update folded the helpful content system into the core algorithm, so site quality is now assessed continuously rather than in separate updates.
For self-assessment, Google suggests three questions:
- Who created it: is the author clear, and do they have experience in the topic?
- How was it created: if automation was used, would readers reasonably expect to be told?
- Why does it exist: to help someone, or to catch search traffic?
Google’s quality rater guidelines — used by the people who manually grade results so Google can check whether its algorithms work — are explicit: a page whose main content is copied, paraphrased or AI-generated with little effort, originality or added value gets the lowest rating. The operative phrase is “no added value”, not “generated”.
What a “penalty” actually looks like
There’s no “AI detected” notice. There are three scenarios:
| Symptom | Where to see it | What it means |
|---|---|---|
| Manual action for spam | Search Console → Manual actions | A Google reviewer looked at the site. Pages or the whole domain drop out until you fix it and request reconsideration |
| Sitewide drop after a core update | Search Console → Performance, compare periods | The algorithm decided there’s a lot of unhelpful content. Good pages suffer too |
| New pages never get indexed | Indexing → Pages → “Crawled – currently not indexed” | Google saw the page and decided it wasn’t worth a spot in the index. The earliest signal |
The third one is the key indicator if you publish a lot. If 12 of your 50 new articles are indexed, the problem isn’t the detector score — it’s that 38 articles add nothing to what’s already ranking.
Why AI detectors are a poor guide
Tools like GPTZero or Originality output a probability, not a fact, and they fail in both directions:
- False positives on human writing. Dry instructions, legal wording and templated copy often get flagged as AI. A 2023 Stanford study found detectors misclassify writing by non-native English speakers far more often — a real problem for international teams.
- Misses. Model output with light edits often passes as “human”.
- Even OpenAI shut down its own AI text classifier in 2023 because of low accuracy.
- Google hasn’t said it uses such detectors for ranking. Everything it has published is about quality and purpose, not how the words were produced.
The worst move is “humanizing” text with paraphrasing tools to hit 0%. They swap precise words for synonyms, break terminology and introduce errors. The detector score improves, the text gets worse, and rankings don’t move.
Instead of a detector, watch what reflects quality: whether the page gets indexed, its impressions and clicks in Search Console after 4–8 weeks, and whether people read to the end.
What separates AI pages that rank
| Sinks | Holds up |
|---|---|
| 300 “service + city” pages that differ only in the city name | 15 city pages, each with its own prices, addresses and reviews |
| The top 10 results, reworded | Something the top 10 lacks: your numbers, a case, a screenshot, an argued opinion |
| Written by “Editorial team” or nobody | A named author with real experience and a bio page |
| Stats “according to research” with no source | Every number checked against the primary source |
| Published and forgotten | An updated date; outdated parts get fixed |
| A paragraph of throat-clearing before the answer | The answer in the first two paragraphs |
A quick test for any page: if you deleted it, would the internet lose anything? If not, neither a human nor a model can save it.
Workflow: AI drafts, a human is accountable
- A human writes the brief. The query, who’s asking and why, what already ranks and — most important — what you have that they don’t: data, experience, a point of view. If that last part is empty, don’t write the article.
- Draft with brand context. Without facts about you, a model fills the gaps with filler and fiction. Prices, terms, cases and a brand knowledge base go in before the first word, not after.
- Fact-check. Every number, platform limit, price and feature name against the primary source. Models confidently describe features that don’t exist and limits from two years ago.
- Add the experience layer. At least two things only you could add: a result from your own account, a screenshot, a customer question verbatim, a mistake you made yourself.
- Edit the voice. Strip the phrases that give AI away in the first line — there’s a stop list in our tone of voice guide.
- Author and date. The name of a real person accountable for the text, and a last-updated date.
In practice this takes an hour or so per article instead of three or four from scratch. Not five minutes, not a whole day. The savings are real as long as you don’t skip steps 3 and 4.
What about social media? Labels and reach
Captions are simpler: no major network labels AI-assisted text or cuts its reach. The rules are about realistic images, video and audio:
- Instagram, Facebook, Threads. Meta adds an “AI info” label to images when the file carries technical signals of generation. Photorealistic video and realistic audio created or altered with AI must be disclosed by the creator.
- YouTube. On upload you must flag “altered or synthetic content” if a video realistically shows something that didn’t happen. Scripts, ideas, captions and thumbnails made with AI don’t need disclosure. Separately, YouTube won’t monetize mass-produced, repetitive videos.
- TikTok requires labels on realistic AI content and applies some automatically.
- Telegram has no formal labeling rules.
What matters more than labels: Instagram demotes unoriginal content in recommendations. A generated “same as everyone” image or caption loses not because of a label, but because nobody saves or shares it. How to take the “AI taste” out of captions is covered in a separate guide.
When AI content really will hurt you
- Health, money and legal topics. The bar for author experience and expertise is highest here, and text without expert review is a risk for rankings and readers alike.
- Programmatic pages by the thousand. “Best [service] in [city]” across 500 cities is the textbook case of scaled content.
- Articles on other people’s domains. Parasite content on authoritative sites falls under a separate site reputation abuse policy.
- Fake reviews, authors and case studies. That breaks search policies, and in the US the FTC’s rule on fake reviews explicitly covers AI-generated ones.
- You publish faster than you can read. When there are more articles than an editor can physically check, steps 3 and 4 are the first to go.
If you write posts with AI, quality improves fastest through context, not clever prompts. In Repost21, the Luna assistant drafts from your brand knowledge base — Google Docs, Sheets and Notion connected live — so prices and facts come from your documents, not the model’s imagination. The final call is still yours. Try it free.