Content moderation
Content moderation is the practice of screening what users post, including text, comments, images, audio and video, and removing or restricting anything that breaks a platform's policies or local law. It is a system with four parts: written community guidelines, automated screening with filters and AI classifiers, human reviewers for borderline cases and reports, and an appeals path. Video moderation is the part of content moderation that deals with moving pictures and their soundtracks, and it is the most expensive part to do by hand.
How does content moderation work?
It starts with a written policy that says what is allowed, what is restricted and what is removed. Everything else follows from that document.
Each content type then gets its own screening. Text goes through keyword filters and language classifiers. Images are checked against hash lists of known illegal material and through image classifiers. Audio is transcribed so it can be screened as text. Video combines all three: sampled frames for images, a transcript for speech and sometimes OCR for on-screen text.
Screening produces scores or flags, and a decision layer maps them to actions using the policy: approve, age-restrict, hide from recommendations, remove or queue for a person. User reports feed the same queue. Human moderators handle the queue, and their decisions, along with appeals, are logged so the policy and the thresholds can be adjusted over time.
What are the types of content moderation?
Platforms rarely use just one type. They mix them by content type, by how trusted the poster is and by how much harm a miss could cause. A children's community app leans on pre-moderation; a large creator platform relies on automated and post-moderation with reactive reports on top.
- Pre-moderation: content is reviewed before anyone else sees it. Safest, but every post waits.
- Post-moderation: content goes live at once and is reviewed afterwards. Fast for posters, but harmful items are briefly visible.
- Reactive moderation: content is reviewed only when users report it. Cheap, but depends on viewers seeing harm first.
- Automated moderation: filters and AI classifiers screen everything on arrival and pass the uncertain cases to people.
- Distributed or community moderation: trusted users or volunteer moderators review content in their own spaces.
Content moderation vs. video moderation
Content moderation covers everything a user can post. Video moderation is the subset that handles video files and live streams, and it differs in three ways.
The first is volume. A comment is a few words; a short video holds thousands of frames and a soundtrack, so systems sample rather than inspect everything. The second is time. A harmful moment may last a second inside a long, harmless video, so a video flag needs a timestamp to be useful to a reviewer. The third is that video contains other content types. Speech becomes text once transcribed, on-screen captions become text through OCR, and frames are images. A good stack reuses its text and image checks on those outputs instead of building separate ones.
The parts around a video need moderation too: the title, the description, the thumbnail and the comments underneath. Those usually go through the platform's existing text and image tools.
How should content moderation decisions be logged?
Every decision should leave a record: which item, which version, which rule it broke, which signal triggered the review, who or what decided, when, and what action was taken. For video, the record should include the timestamps of the flagged moments, so a second reviewer or an appeals handler can check the evidence without rewatching the whole file.
Regulation has made this a technical requirement in some markets. The EU Digital Services Act entered into force on 16 November 2022 and has applied to all online platforms since 17 February 2024. It expects platforms to act on illegal content and explain their moderation decisions, which means the reason for a removal has to be stored, not reconstructed later.
Logs also improve the system. Reviewer overturns show where automated flags are too strict or too loose, and appeal outcomes show where the written policy is unclear.
How does FastPix handle content moderation?
FastPix covers the video and audio layer of content moderation. In-video AI runs in parallel with encoding on each upload and returns NSFW and profanity flags based on sampled frames and audio. It does not block publication on its own and has no user-set threshold, so the policy, the decision layer, the review queue and the appeals path stay in your system.
The flow fits a standard moderation stack. A webhook, signed with HMAC SHA-256 so your service can verify it, tells your backend when a video is processed. Your decision layer reads the flags and applies your rules. Videos waiting for review can use a signed playback ID with an expiring token, so only your moderators can play them. The same in-video AI output includes a transcript with word-level timestamps and the on-screen text found in the video, which you can pass to the text classifiers you already run on comments and titles, instead of sending the audio to another service.
Titles, descriptions, thumbnails, comments and still images stay with your existing text and image tools; FastPix screens the video and its soundtrack.
Moderation has its own per-minute rate on the encoding bill, with no tokens involved. Standard encoding is free up to 4K. The free plan includes subtitle and AI sample minutes with no card, which is enough to run a set of past uploads through and compare the flags with your reviewers' decisions.
By the numbers
- The EU Digital Services Act entered into force on 16 November 2022 and applied to all online platforms from 17 February 2024. (Digital Services Act (Wikipedia))
- Companies that breach the EU Digital Services Act can be fined up to 6% of their global annual turnover. (Digital Services Act (Wikipedia))
