SynthID vs Hive vs Text Detectors: What Each One Actually Checks
SynthID reads an invisible watermark Google adds when its models generate a file. Hive classifies pixels and audio with trained models and returns a probability. Text detectors score how the words are distributed. Three mechanisms, three different questions — and none of them answers "is this slop".
Vendor documentation checked 23 October 2026. Tool names are trademarks of their owners. Nothing here is a judgment about any company's content, and no vendor reviewed or endorsed this page.
Check a piece of text with the free local slop check
The three questions people mix up
- Was this file produced by a specific AI system? A watermark reader can answer that — but only when the file carries a watermark and the watermark survived.
- Does this media look machine-generated? A classifier returns a probability from patterns learned on labelled examples. It generalises to unwatermarked models, and it is wrong sometimes, in both directions.
- Is this content low-value bulk output? No detector answers this. Slop is a publishing judgment: mass production, no human review, no accountability.
Side by side
| Tool | Mechanism | Modalities | Where you can use it | Cannot tell you |
|---|---|---|---|---|
| Google SynthID | Reads an invisible watermark embedded at generation time | Image, video, audio, text | Ask Gemini, Google Search or Chrome about an uploaded file; SynthID Detector portal (open to everyone) | Anything about unwatermarked media; whether the content is good or bad |
| Hive (AI-generated & deepfake detection) | Classifiers trained on synthetic versus real media, served as APIs | Image, video, audio | Hive Detect web tools, Chrome extension, usage-based API | Watermarks; whether the content is slop |
| Text AI detectors (GPTZero and similar) | Scores token and statistical patterns against labelled AI and human text | Text | Free scans up to a character cap, paid advanced scans, browser extensions | Who wrote it; whether the writing is any good |
| AI Slop Detector (this site) | Local rule-based checks on style signals; self-guided visual review | Text, image checklist | Right here, free, nothing uploaded | Authorship, truth, or whether a model was involved |
SynthID: it reads a watermark, not a style
Google DeepMind describes SynthID as a watermarking tool for AI-generated content. The watermark is embedded in the file at the moment the content is created and is imperceptible to humans, but detectable by SynthID's technology. What the documentation says, in its own terms:
- The image and video watermark is added when content is created and is described as designed to stand up to cropping, added filters, frame-rate changes and lossy compression; the audio watermark is described as surviving added noise, MP3 compression and speed changes.
- Audio watermarks go into anything generated with Lyria or the podcast feature in NotebookLM; text watermarks come from adjusting token probability scores in the Gemini app and web experience.
- Checking is detection rather than estimation: upload the image, video or audio to Gemini, Search or Chrome and ask whether Google AI created or altered it.
- The SynthID Detector verification portal launched in 2025 and, per the same page, is now open to everyone; Google lists detection support across its own models and partner models, naming OpenAI, NVIDIA, Kakao and Apple as current or upcoming.
The limit: a watermark has to exist, and it has to survive. Content made by a model that does not watermark — or a file pushed through enough transformation — has nothing for this method to read. "No watermark found" is not the same statement as "not AI".
Hive: it classifies pixels and audio at scale
Hive sells AI-generated and deepfake detection as APIs aimed at platforms: image and video classification, deepfake and edited-profile-photo checks, and audio classification that scores which portions of a clip are likely synthetic. Its pages position the product for trust-and-safety queues — flag, human review, takedown — and mention compliance work around laws such as the TAKE IT DOWN Act and the NO FAKES ACT.
- Free entry points published by Hive itself: Hive Detect (upload a file or submit a media URL), a bot that checks content on X, and a Chrome extension that reports AI-likelihood with likely model attribution.
- Pricing as of 23 October 2026: usage-based, with "50+ USD in free credits" for developers after adding a payment method, and custom enterprise pricing.
- The mechanism is classification — patterns learned from labelled synthetic and real media — so it can flag output from models that never watermark, at the cost of returning a probability instead of a certainty.
The limit: a score is neither attribution nor verdict. Hive's own extension copy says likely model attribution, and likely is the operative word.
Text detectors: they score distribution, not authorship
GPTZero is the best-known example: paste text, get a probability plus highlighted sentences. Its homepage, checked 23 October 2026, advertises 99.9% accuracy and 20M+ users, a free scan capped at 10,000 characters per submission, a paid Advanced Scan, a Chrome extension and a hallucination detector.
Read the accuracy figure as a vendor claim rather than a measured property of your document. What a text detector returns is a probability derived from patterns: uniform sentence rhythm, predictable word choice, low variation between sentences. Those patterns also appear in real human writing — non-native English, formulaic genres, heavily edited drafts, translated text.
The limit: the mechanism is style matching, so paraphrasing and careful editing move the score without changing authorship, and a confident human who writes like a textbook can be scored as machine output.
What this site does — and what it deliberately does not
Our free slop check is not a detector of AI origin. It runs local rules over text — template phrases, AI-favourite vocabulary, punctuation density, sentence rhythm, repeated openers, adjective trios, a voiceless tone — and returns a 0–100 heuristic score. The image path is a guided review: we list what to inspect and you look at the file. Nothing is uploaded, there is no model API and no account.
That makes this tool worse than SynthID at answering "did Google make this" and worse than a classifier at answering "is this synthetic". It aims at the third question — does this read like mass-produced filler — and even there it only tells you where to look. The about page spells out what the score does not claim.
Which one to use
- You suspect the file came from Gemini, Imagen, Veo or Lyria: ask Gemini or run it through the SynthID Detector. That is the most definitive answer available — when a watermark is present.
- You have an unwatermarked image, video or voice note and need a probability quickly: a classifier or one of Hive's free tools. Read the result as one signal among several.
- You have a document: a text detector plus your own reading. Check whether the flagged paragraphs are genuinely generic or merely formal.
- You are deciding whether to publish, boost or reply to something: run the local slop check, then read the whole thing. Slop usually shows up as missing specifics, not as a percentage.
Four rules for reading any detector result
- Name the question the tool answers. Watermark found is not AI slop. High AI-likelihood is not plagiarism. A low score is not proof of a human author.
- Ask whether a watermark could even exist. Screenshots, re-encodes, cross-model edits and models outside the watermarking programme usually break the chain before a detector sees the file.
- Never use a score as evidence about a person. Not for grades, hiring, moderation appeals or public accusations. Every vendor in this comparison publishes a disclaimer to that effect, and so do we.
- Prefer the original file and the platform's own tooling. When the stakes are high, the originating platform holds logs that a third-party classifier never sees.
Frequently asked questions
Is SynthID the same thing as an AI detector?
No. SynthID is a watermark reader: it looks for a signal Google embedded when the content was generated. An AI detector, in the Hive or GPTZero sense, estimates the probability that content is synthetic from its patterns. The first answers "did a participating model make this file"; the second answers "does this look machine-made".
Does deleting metadata remove a SynthID watermark?
Those are different things. EXIF and other metadata are file properties you can strip; a SynthID watermark is embedded in the content itself, and Google's documentation describes it as designed to survive compression, cropping, filters and frame-rate changes. We do not test watermark survival here, so we will not claim either way beyond Google's own description.
Can Hive tell me which model made an image?
Hive's public pages describe "likely model attribution" in its browsing extension: a best guess at the generator family, alongside an AI-likelihood score. Treat attribution as a hint, not an identification — the underlying method is classification, not watermark extraction.
Are AI text detectors accurate enough to accuse someone?
No. What comes back is a probability built from style patterns, and the headline accuracy figures on vendor sites are marketing claims rather than results on your specific document. Human writing in formulaic genres, translated text and non-native English routinely trips these systems. Do not use a text-detector score for academic, employment, legal or public accusations.
Can any of these tools tell me if something is AI slop?
No. Slop is about how content was produced and published — volume, missing review, no accountability, no original insight — not about whether a model touched it. A watermarked, carefully edited AI-assisted article is not slop; an unwatermarked, unreviewed bulk post is. That judgment is editorial, which is why our tool only flags style signals and asks you to read.
Why doesn't this site upload my text?
Because the checks are local rules, not a model call: your text stays in the browser, there is no API, no account and no server-side copy. That is also the reason we cannot offer model-level certainty — we trade accuracy on authorship for not shipping your document anywhere.
Sources
Only vendor documentation was used, all checked 23 October 2026: Google DeepMind's SynthID page; Hive's AI-generated and deepfake content detection page, Hive Detect and its pricing page; GPTZero's homepage. Where a vendor publishes a performance claim, it is quoted as their claim.