Slop or Not Help & FAQ

Slop or Not is an AI content detector for text and images. The iPhone and Mac app checks content on-device, so your text and photos stay on your device. The website also has a free online photo checker for quick image checks in any browser. Online submissions are processed on Numen's private Mac server and are never shared with third-party AI APIs or detector vendors.

You can check content in the app or on the web:

• Check text in the app: Copy the text you want to analyze, open Slop or Not, and paste it into the text field. The app checks it on-device.
• Check images in the app: Select any image or photo from your library, or share an image from another app to Slop or Not. The app checks it on-device.
• Check text and photos online: Use the web checker for English text or a suspicious photo in any browser. Online checks run on Numen's private Mac server and are not shared with third-party AI services. Open the AI detector.

The iPhone and Mac app performs AI detection directly on your device. You need an internet connection for the one-time setup that downloads the AI models. After that, you can check text and images in the app offline. We use Apple's CoreML model encryption to protect our on-device transformer models, so your device may occasionally need an internet connection to refresh model encryption keys from the App Store. The online photo checker needs an internet connection because it sends each submission to Numen's private Mac server for analysis. We do not pass online submissions to third-party AI services.

Yes, but the privacy model depends on where you check. In the iPhone and Mac app, Slop or Not processes text and images locally, and your content never leaves your device. Unlike many online content detectors, the web checker does not pass your photo to a third-party AI service or detector vendor. It goes to Numen's private Mac server for analysis, on hardware we own and run. We're based in the EU and comply with GDPR and CCPA.

Use Slop or Not whenever a cloud detector is the wrong place for the content you need to check. It is especially handy for:

• Students & Teachers checking for originality in academic work.
• Journalists & Fact-Checkers verifying the authenticity of sources and images.
• Online Daters spotting fake "catfish" profiles.
• Shoppers checking if product reviews are real or AI-generated.
• Anyone wanting to identify misinformation and deepfakes on social media.

Accuracy by content type:

• Images that carry a content credential or a SynthID watermark: These reads do not come from a model at all. C2PA and IPTC credentials are either in the file or they are not, and a signed C2PA credential from OpenAI, Microsoft, or Adobe is checked against a bundled trust list, so a forgery fails instead of passing as proof. A SynthID watermark read is a calibrated confidence score, so it is reported as a likelihood rather than a yes or no.
• All other AI-generated images: 90% accuracy.
• AI-generated text: Independently scored on RAID, the public AI-text-detection benchmark we do not run: 99.5% of AI text caught at a 5% false-positive rate, adversarial attacks included (top 10 of 39 detectors as of August 2026).

*Note: Image accuracy is based on our internal tests. Text detection is scored on the public RAID benchmark. The image figure is measured on AI-generated images, so it describes how often those are caught. It is not a false-positive rate and does not tell you how often a real photo is flagged. Results can vary with new AI models and advanced methods designed to trick detectors.*

Read the full RAID benchmark results

Image metadata is attached to a particular file, not permanently bound to its pixels. A screenshot creates a new file, so it keeps information about the screenshot instead of the original image. Editing, export, re-encoding, and re-uploading through a website or messaging app can also remove or rewrite EXIF, IPTC, XMP, and C2PA records. Some apps do this deliberately because those records can reveal location, device, and camera details. The original download or camera file usually preserves the most information; screenshots and re-saved copies preserve the least. Before uploading, open the image in Photos on iPhone or iPad and tap Info, or open it in Preview on Mac and click Info to view the Inspector. Those views are a quick check, not a complete inventory; Slop or Not may recognize records they do not display, and it can still inspect the pixels. Slop or Not is the only AI image detector that reads Google DeepMind's SynthID watermark from both Google and OpenAI images. Missing metadata or a missing readable watermark does not prove that an image is human-made or AI-generated.

AI-generated text often carries hidden characters and swapped punctuation that act as mechanical fingerprints. Text Cleanup strips them in one tap: zero-width spaces and other invisible Unicode, curly quotes and em dashes, and homoglyphs (letters shaped like other letters). The browser checker sends the passage to Numen's server, which runs the same scan in memory, hands back a cleaned copy in one click, and discards the text with the response. The iPhone and Mac apps add two things the browser cannot: spelling conversion between US and UK English, and a scan that runs with no network at all. Cleanup removes mechanical traces only, and no word is rewritten, so the AI check still reads your own prose. Non-breaking spaces are swapped for ordinary ones rather than deleted, so 100 % keeps its separation; check the spacing before you paste the clean copy over the original.

When you check text in Slop or Not, the app shows a readability score next to the AI result. The formula depends on the language. English uses Flesch Reading Ease and Flesch-Kincaid Grade Level; German uses Wiener Sachtextformel; Spanish uses Flesch-Szigriszt; Italian uses GULPEASE; Swedish, Danish, and Norwegian use LIX. That works across seven languages today, with more planned. Readability tells you how hard the text is to read, not who or what wrote it. The result is an estimate, so use it as a quick gauge. The score is still available when language detection confidence is low, as long as the app identifies a supported language.

They are fun iMessage sticker packs included with the app! When you find AI-generated content online or in chats, you can use our stickers to call it out in a lighthearted way.

Turn on Incognito Mode when you don't want a check to appear in the app's history. The content still stays on your device; the app simply leaves no local history entry.

The agentic AI Humanizer skill lets Claude, Codex, Hermes Agent, OpenClaw, Cursor, Gemini CLI, or OpenCode rewrite a draft with a five-pass workflow. It works without Slop or Not for core rewriting and voice matching. When Slop or Not is installed on Mac with Pro active, it supplies the on-device AI detector, readability, Text Cleanup, and cleanup stats through its CLI or MCP server.

Start with the install guide, the CLI guide, the MCP guide, or the voice-matching guide.

Yes, if you provide a writing sample. The skill can extract a style fingerprint for register, sentence rhythm, contractions, hedges, and paragraph shape. It should use style traits only, not private facts from the sample.

Yes. The slop CLI and slop mcp server are Pro features in the Mac app. The agentic AI Humanizer skill can still run its full five-pass rewrite workflow and voice matching without Pro, but on-device AI detector scoring, readability, Text Cleanup, and cleanup stats require Pro.

Slop or Not detection, readability, cleanup, and image checks run on your Mac. Rewriting happens in the AI assistant you choose, such as Claude, Codex, Hermes Agent, OpenClaw, or Cursor. If that assistant is cloud-based, its own privacy policy applies to the rewrite and to any voice sample used for fingerprint extraction.

No. Slop or Not returns a probability verdict, not a guarantee. The CLI, MCP server, and agentic AI Humanizer loop are quality-control tools that help you inspect and revise drafts. Results can vary with short samples, new models, and heavy human editing.

GPTZero and Originality.ai are cloud AI-content suites that people often use from a web dashboard. Slop or Not focuses on local checks on iPhone and Mac. The new CLI and MCP server let your existing AI assistant use the local Slop or Not score instead of sending every check through another web humanizer workflow.

We keep the App Store download small so it is faster and takes up less space. The detection models arrive through Apple's On-Demand Resources the first time you open Slop or Not. Usually, that is the only download. Apple's service can occasionally interrupt it, though. Check your connection, wait a few minutes, and reopen the app. The download should pick up where it left off.

Yes. When you use a low-resolution thumbnail or a screenshot, the model still runs but may return a lower confidence score because fine details are lost during our normalization step. Always try to use the original or highest-quality version of the picture. More pixels mean more clues for the detector, so results are stronger and more stable.

The text detector needs at least two full sentences to work. Very short snippets give too little context for the model to spot AI fingerprints. Whenever possible, paste a whole paragraph or the entire document. The longer the passage, the higher the detection confidence.

New AI models and deliberate evasion methods can be harder to detect. We update our models as those methods change.

• The image you checked might be a low-resolution thumbnail or a heavily compressed screenshot. Try again with the highest-resolution version available.
• The text sample could be too short. Make sure you provide at least two sentences; more text improves the results.

These checks solve most false-negative reports. If it still happens, send us the example. It can help improve a future model update.

The percentage is how confident the check is, not how much of the picture was generated. A middling number usually comes from the AI image detection model and means it landed close to undecided. A watermark read can also report a middling confidence of its own.

Slop or Not names the result after whichever side is ahead, so a photo at 55% is named for the AI side even though the two numbers are nearly even. Treat a result near the middle as a reason to look closer rather than as an answer.

When the score came from the image model, genuine photos land in that range for ordinary reasons:

• Filters, heavy edits, upscaling, and aggressive compression smooth away the fine detail the image model reads.
• Modern phone cameras assemble one photo from several frames and reconstruct detail computationally, so the file sits further from a plain sensor read than a photo shot ten years ago.
• A tight crop leaves fewer pixels to judge, and the model works from what survives.

A verified content credential or a SynthID watermark read is evidence about where a file came from. A mid-range score from the image model is not. Slop or Not reports a probability and leaves the decision with you.

Give the detector the original file. Slop or Not checks content credentials and watermarks before it scores the pixels, and each re-save can strip away a layer of that evidence.

• Share the original camera file or the original download rather than a screenshot. A screenshot is a new file and carries none of the original's records.
• Pick the largest version you have. A thumbnail or a heavily recompressed copy gives the image model less to work with.
• Skip copies that have passed through a social or messaging app. Many strip metadata on upload, so the signed content credentials usually go even when a readable watermark survives.

SynthID sits in the pixels, so it can survive a screenshot when content credentials do not. If a result still looks wrong, report it from the result itself and attach the file when the form offers you the option. The report brings the verdict and the scores along, and the file is the part that lets us run the check again.

See also: Does image resolution affect detection scores? · Why Is Image Metadata Missing?