Collect and annotate 1,200 real phone screenshots (adult and graphic-injury content) for a child-safety AI model: adults only

CensorX is an Android child-safety app. An on-device AI model detects adult content and graphic injury on a child’s screen and blurs it before the child sees it. To work well, the model must learn what unsafe content looks like on a real phone screen. This job supplies that training data. Every image will be used only to train and test this detection model, and will never be published or redistributed. Content warning This job involves viewing explicit adult material and graphic images of wounds and blood. Apply only if you are 18 or older and comfortable working with this content. You confirm that you are 18 or older. What you will deliver: 1,400 annotated screenshots Folder Count Content unsafe/adult 700 Phone screenshots showing adult nudity or sexual content in realistic screen settings (browser, search results, feeds) unsafe/wound_blood 700 Phone screenshots in which real, visible wounds or blood on people are clearly shown (injury, first-aid, medical, news, fight footage), in the same screen settings The counts are targets; ±5% per folder is fine. At least 40% of the wound/blood images must show the injury at a distance (wide shot, crowd or ring scenes, video frames), not only close-ups. For adult content, cover a mix of skin tones, body types, genders, settings, and sizes of the unsafe region, including small or partly visible regions and screens where the unsafe content is only one tile in a grid. Annotation (bounding boxes) Draw one tight box around each visible unsafe region in every image. The label is unsafe for both folders. Box small, partly hidden and low-resolution unsafe regions too, and box each region separately. Do not box blurred, covered or censored areas. Deliver YOLO .txt files (one per image, same file name) or COCO JSON. Images and labels must match one to one. A short guide with examples of correct and incorrect boxes will be provided after hiring. Content rules Adults only. If you cannot clearly tell that every person shown is an adult, do not use the image. Absolutely no content involving minors, ever. If you come across it, do not save or share it, delete it, and report it to the relevant authorities (in India, the national cybercrime portal). Any such image in a delivery means immediate termination and no payment for that batch. No private, leaked, hacked, hidden-camera, revenge or non-consensual material. Use only sources with documented age and consent verification. Only publicly accessible content. Do not use paywalled, login-protected or restricted material, and do not bypass any access control. Real photos and real video frames only. No AI-generated images. Exclude anime, cartoon and game blood, and clearly staged makeup. Do not include weapons or fights without visible blood or injury. Capture requirements Take every screenshot on a real Android phone, in portrait orientation. Automation is allowed for scrolling and screen capture only (for example ADB screencap or a scripted scroll-and-capture routine). Emulators, desktop browsers, downloaded or scraped images, screen recordings, and AI-generated or composited images are not allowed. Screenshots taken while a video is playing in the browser are allowed. Recorded video is not. Submit the original, unmodified PNG files exactly as the phone captured them. Do not crop, edit, resize, filter or recompress anything. Required size: 1080 px wide (for example 1080×2340 or 1080×2400). Tell us your phone model. With the trial, include one screenshot of the phone’s Settings, About phone page, showing the model, so that we can verify the device. Screenshots must show the full phone screen, including the status bar and navigation bar. Notifications must be turned off, and no personal information may appear (names, messages, phone numbers, accounts). Use a clean browser profile. Browse in Chrome and Brave. Capture grids, search results, single photos and feed layouts. Use both dark and light mode and a wide variety of sources. No more than 10% of a folder from one source page. Every image must be unique. No duplicates or near-duplicates, including the same page at a slightly different scroll position or the same content on two screens. We check this automatically (perceptual hash, with images closer than 10 bits counted as duplicates). A batch with more than 5% duplicates is returned. A human must review every image before submission. Automation must not decide what is included. Remove any image that breaks the content rules, shows a minor, or shows personal information. Quality must not drop because of automation: no blurred or mid-scroll frames, no loading spinners or partial renders, and no images without a clear, visible unsafe region. Manifest (required) A manifest.csv with one row per image and these columns: file, folder, app, source_site, source_page_url, device_model, resolution, display_mode (dark or light), layout (grid, search results, single, feed), capture_method (manual or automated, and the tool used), date_captured, notes. Images without a source URL will not be accepted. Rights Boxes, labels and the manifest are work made for hire and owned by CensorX. You confirm that the images are collected only for internal AI training and will not be shared or redistributed. Data handling You will sign an NDA. Turn off cloud backup and sync (Google Photos and similar) on any device used for this job. Store files only in an encrypted location that we approve. Do not share, post or keep copies. Delete everything after final delivery and payment, and confirm in writing. Delivery through a private, access-controlled link that we provide. How the work is split Paid trial: 100 screenshots (50 per folder) with boxes and the manifest. Reviewed within 2 working days. Milestones after a successful trial: four batches of about 325 images each. Acceptance: each batch passes automatic checks (file format and size, duplicates, label format, manifest) and a hand-reviewed random sample, with a second reviewer re-checking abou