LinkedIn sorts posts spam, low-quality or clear: batching AI video

Yes: LinkedIn's engineering team says its classifiers label every image, text or long-form post as "spam", "low-quality" or "clear" in near real time, and then watch how a post spreads. If you post a batch of AI videos, each one is judged on its own, so a batch only helps you when each video has its own point, its own facts and its own opening.

This post reads LinkedIn's write-up on keeping the feed relevant on 2026-10-07. It is an engineering post about the feed in general, not a rule on AI video. We do not know how it scores video frames.

What LinkedIn wrote

The post describes three moments when a post can be caught, and a human layer behind them.

What "varied" means for a batch

LinkedIn does not publish a test for it, so use a plain rule: if two videos could swap captions and still make sense, they are the same video. Vary the things that carry meaning:

• The claim. One idea per video, each with a fact only you can state.
• The opening frame. Start from a different real photo or a different scene each time.
• The length and shape. A 12-second proof clip and a 40-second explainer are different posts.
• The speaker or the company voice, if you use a presenter.

Run a batch with one brief per row

With the Video API, a batch is a loop of separate jobs. Each submit returns a job id at once, and you add an Idempotency-Key so a retry returns the original job instead of a second paid one, as described in Jobs and results. Write the brief per row in your own sheet, not from one template with one word replaced.

If your team uses Sume Formats, a bulk run takes one input per row; the bulk runs guide covers it. The same rule holds: the row data should carry the fact, the offer or the story that is different.

Check the output before it goes live

Run video inspect with transcribe: true on each file ($0.01 per audio minute for the speech-to-text part) and compare the transcripts side by side. If two transcripts share most of their lines, rewrite one brief and regenerate that row only. A rerun of one row is cheaper than a campaign of near-copies.

What this does not claim

We found no LinkedIn page that says AI video is treated differently from other video, and no number for how many near-copies trigger a label. This is a way to make each post earn its place, not a way around a classifier.