The conversation does not end when a video finishes. For creators trying to keep that conversation welcoming, YouTube is testing optional AI comment moderation intended to adapt to a channel’s moderation style over time. The company described the experiment in its 23 September 2026 Made on YouTube announcement. Moderation here means reviewing comments against community expectations.
A review queue, with the creator still deciding
TeamYouTube’s detailed announcement says it learns from manual removals, holds similar new comments in Held and provides a seven-day summary. It does not delete without consent. A desktop Studio experiment was planned for later in 2026; universal access was not verified on 3 November.
That distinction matters. Putting something aside for review gives the creator a decision to make; deleting it means the decision has already been carried out. A useful reading of this announcement keeps those actions separate rather than treating every filtered comment as permanently removed.
What changes from the filters already in Studio?
AI-assisted filtering predates this experiment. YouTube’s current comment-settings guide explains that Basic and Strict moderation already use AI to hold comments, with Strict holding a wider range. Those systems can make mistakes. The new announcement puts the emphasis on adapting to an individual channel’s approach.
Ruby’s take: Make room for disagreement, too
A tidy comment section is not automatically a good conversation. The more useful question is whether the right conversations still have room to happen. A frustrated viewer explaining a problem and someone repeatedly disrupting a discussion may call for different responses, even when neither message sounds cheerful.
If you receive access, our suggested starting point is a short written moderation rule. Describe the behaviour you want to remove and the criticism you want to keep. Then manually review a chosen sample of comments and record where your decisions differ from the tool’s. Counting those disagreements gives you something concrete to examine instead of relying on an impression that the channel feels quieter. This is a proposed review method, not an experiment we have conducted.
We would look for both unnecessary holds and disruptive comments that were missed. Neither measure alone tells the whole story: reducing one kind of mistake can leave the other unnoticed. Keeping examples also makes a team discussion about channel tone easier to follow.
There is no measured speed-up, learning period or accuracy claim in this report. The editorial goal is a conversation worth joining, with enough room for a viewer to ask a difficult question and for the creator to answer it.
