How we moderate public posts on Circus without paying to scan everything
How we moderate public posts on Circus without paying to scan everything: cheap checks first, AWS next, people for the unsure.

The short answer. Do not send every post to the expensive service. Put cheap checks in front of it: skip content you already trust, run a fast word filter, run a model of your own, and only then pay Amazon Comprehend for text or Amazon Rekognition for pictures and video. Anything the system is unsure about goes to a person. The spend lands on the few posts that need it.
Information last checked 3 October 2026. Prices, features and limits come from AWS's public pages and change, so check them before you budget. This post describes how we approached one problem and is general information, not advice.
Circus is a fan community app. A community only works if people feel safe talking, so we needed moderation from the start: to keep out people who are negative towards the community, and to block or limit them when they will not stop. This covers public posts: text, images, video and GIFs.
How do you handle each type of content?
- GIFs: chosen from KLIPY's library, a source we trust, so we do not scan them ourselves. No paid AWS call.
- Text: three passes, cheapest first. A paid AWS call only on the third.
- Images and video: Amazon Rekognition. A paid AWS call.
- Anything unsure: a person decides. No paid call.
The GIF choice is the cleanest saving. If you only let people pick from a library you trust, there is nothing left to pay to scan.

How does the text check work?
Each pass costs more than the last, and each one removes work for the next.
- A word filter. Instant and close to free. It catches obvious abuse.
- A model of our own. It runs on our servers, so it costs computing power rather than a fee per post.
- Amazon Comprehend. The last pass, only for text the first two could not settle.
Most posts are clearly fine or clearly not by the second pass, so the paid service sees a fraction of the traffic.
Why is a word filter harder than a list of words?
Because people work around lists. They put spaces between letters, swap in numbers and symbols, or stretch a letter out. A plain match misses all of it.

So the filter tidies the text first and only then checks it against the list. We also keep a base level that applies everywhere and let each community's context adjust on top of it, because a word that is harmless in one community can be a problem in another.
What do the AWS services tell you?
As at 3 October 2026, per AWS's documentation:
- Amazon Comprehend scores text for toxicity from 0 to 1, and gives a score for each of seven categories: graphic, harassment or abuse, hate speech, insult, profanity, sexual, and violence or threat. Source: AWS API reference.
- Amazon Rekognition looks at images and video and returns moderation labels, each with a confidence score from 0 to 100. Source: AWS API reference.
A score is not a verdict. It says how confident the model is, and you choose what to do at each level. We run a middle band: clearly fine is published, clearly not is held back, and the band in between goes to a person. Where we set those lines is ours to decide and depends on your community.
Languages. As at 3 October 2026, AWS's API reference says English is currently the only language supported for toxicity detection. If your community writes in other languages, plan for that gap with other layers and with people.
What does it cost?
These are the publicly listed prices we saw on 3 October 2026, in the US East (N. Virginia) region. Check the pages before you budget.
- Amazon Rekognition, image moderation: $0.0010 per image for the first 1 million a month, lower above that. Source
- Amazon Rekognition, stored video moderation: $0.10 per minute. Source
- Amazon Comprehend, toxicity detection: $0.0001 per 100 characters, with a 3-unit (300-character) minimum per request. Source
A short post still bills the minimum: $0.0003, or about $0.30 per 1,000 short posts. That sounds small until you picture a busy community, because moderation cost grows with volume, not with revenue. The real cost of running it at scale was a key reason we built the cheap layers first.
Why not just use people?
It is the obvious alternative, and it has two problems.

- Delay. If a person reviews every post, there is a gap between someone posting and the post going live, and it grows when the queue is long or it is the middle of the night. On a community platform, a conversation that arrives late often does not happen.
- Cost and wear. People cost more as volume grows, and the work is hard on the people doing it. Layers shield them from most of the obvious material and save their attention for judgement calls.
So people sit at the end, not the start. They decide the posts the system is unsure about.
What happens when someone keeps breaking the rules?
Moderating a post is not enough on its own. Someone who keeps breaking the rules needs dealing with as a person: blocked, or limited in what they can do. We built that into the same decision step, so a repeat offender is handled as a pattern, not as a string of single posts.
What would we tell you if you are doing this?
- Skip what you can. A library you trust, or content you have already approved, needs no scan.
- Run cheap checks first, and let the paid service see only what they could not settle.
- Keep a base level, and add community context on top.
- Send the unsure to a person, not to a coin flip.
- Be able to limit an account, not only a post.
- Know your cost per 1,000 posts, and what it becomes at ten times your volume.
Frequently asked questions
Can moderation be fully automatic? For clear-cut content, largely yes. Anything uncertain should reach a person.
Does Comprehend check images? No. Toxicity detection is for text. Images and video go to Rekognition.
Is human moderation better than automated? People judge context better, and they are slower and more expensive. A mix works best: automation for the obvious, people for the unclear.
Can I move off AWS later? Yes, if you keep each check behind one interface so you can swap one out. Ours is built that way.
Need this built?
We build moderation systems end to end: the layers, the review queue and the account controls. See how we built Circus, read about content moderation systems, or tell us what you are building.
