HBO Max has launched a vertical video feed feature called "HBO Max Shorts," using an in-house AI tool to surface compelling clips from its content library for discovery in a short-form, scrollable format. The Verge

The clips powering HBO Max Shorts are generated by a machine learning system built internally at HBO Max. The tool leverages scene-level metadata to scan thousands of hours of titles in the platform's catalog, identifying what it determines to be the most compelling, high-impact moments for discovery. Human editors at HBO Max then review the AI's suggestions and make the final call on which clips best represent the underlying stories. The selected clips are rendered in a new vertical video experience designed for seamless playback within the app. The Verge

The architecture here is worth unpacking. Scene-level metadata, the structured data describing individual scenes within a title, serves as the indexing substrate the AI operates on. Rather than performing raw video analysis frame by frame across the entire catalog, the system can query pre-tagged scene metadata to narrow candidate segments efficiently, then apply ML ranking to surface the highest-impact candidates. This is a hybrid pipeline: machine learning handles scale and initial filtering, human curators handle editorial judgment. The division of labor is deliberate. Automated scene selection at catalog scale requires ML; ensuring that a clip actually captures a narrative arc, or lands without spoiling a pivotal plot point, still requires a person who understands the story.

HBO Max plans to expand Shorts to additional devices and markets over time. Sports content will be incorporated as part of that future expansion. The feature is also expected to be refined as users interact with it, implying that engagement signals, watch-through rates, skip behavior, and replay frequency will feed back into the recommendation and selection pipeline. The Verge

The broader context here is that every major streaming platform is now grappling with the same discovery problem. Catalogs have grown faster than any browse or search interface can serve. Short-form vertical video, the format popularized by TikTok and adopted across YouTube Shorts, Instagram Reels, and other surfaces, has become a primary vector for content discovery, particularly among younger audiences. A streaming service that can auto-generate compelling vertical clips from its own library and feed them into a swipeable interface gains a discovery surface that operates on the same interaction model users already default to on their phones.

The in-house nature of the AI tool matters. HBO Max is not licensing a third-party clip-extraction engine; the ML pipeline, the scene-metadata schema, and presumably the rendering pipeline for vertical output are all built and controlled internally. That gives HBO Max full control over how scenes are scored, how editorial guardrails are applied, and how the system evolves with user data. It also means the engineering investment is entirely on HBO Max's side, with no per-clip licensing or dependency on an external vendor's roadmap.

The human-in-the-loop step is the most consequential design choice in this pipeline. Pure ML scene selection at scale would inevitably surface clips that are visually striking but narratively misleading, emotionally discordant with the tone of the source material, or actively spoiler-laden. An editor who knows the title can reject a clip that spoils a season finale or misrepresents a character's arc. The AI narrows the search space; the editor applies contextual judgment. Whether HBO Max can scale that editorial review as the catalog grows and the feature expands to additional markets and content types, including the faster-moving world of live sports, is an open question. Sports highlights operate on different logic than scripted drama; a system tuned for narrative impact may need retraining for real-time, event-driven selection.

For the technology professionals watching this space, HBO Max Shorts is a concrete example of ML-assisted content discovery moving from recommendation lists to native vertical video, inside a premium streaming app rather than a social platform. The pipeline is not conceptually novel; scene metadata, ML ranking, and human curation are well-established components. What is new is the integration: an end-to-end system that scans the catalog, selects scenes, renders vertical video, and serves it in a TikTok-style feed, all within a single streaming product. The execution, and whether the engagement loop holds, will determine whether this becomes a standard feature across streaming platforms or an experiment that stays niche.