YouTube will let you ask AI to make a custom video feed

This feels less like a feature update and more like a philosophical shift in how YouTube sees itself, or at least how it wants us to see it. For years, the platform’s identity has been a tug-of-war between a searchable library, a social network, and a broadcaster. This new AI-driven “custom feed” suggests it’s leaning hard into a new role: a responsive listener. The promise is that your vague afternoon whim—”cozy history documentaries about forgotten civilizations”—can be instantly articulated and materialized. It’s the platform bending to the user’s intent, rather than the user bending to the algorithm’s recommendations or the discipline of a search bar.

There’s a seductive simplicity to it. We’re so accustomed to the curated chaos of the main feed, a stream of content we passively consume or actively fight against, that the idea of commanding it feels powerful. It’s the difference between scrolling through a magazine and handing an editor a note that says, “Surprise me, but only with things about mid-century architecture and rainy day jazz.” The control, however illusory, is appealing. It transforms the homepage from a billboard into a workshop, with a new, vaguely magical tool at the center.

But let’s sit with that word: “prompt.” This isn’t just a filter, like clicking on “Sports” or “Music.” A prompt requires interpretation, translation by an AI model we cannot see or fully understand. YouTube is betting its sophisticated recommendation engine can bridge the chasm between a human’s messy, emotional request and its own catalog of metadata, transcripts, and engagement signals. The success of this feature hinges entirely on that translation. When I type “thoughtful sci-fi with a hopeful tone,” will it find Arrival and The Martian, or will it serve me algorithmic approximations—perhaps the trailers for those films, or vaguely similar content that gets high engagement but misses the soul of the request?

This introduces a new, subtle layer of mediation. Previously, the algorithm’s logic was opaque but consistent; it was a single, complex system. Now, we’re interfacing with it through a natural language keyhole. We’re not tweaking sliders or selecting categories; we’re hoping our words resonate with the model’s training. The potential for misalignment is high, and the experience of a “bad” custom feed could feel more personal, more like being misunderstood, than a mediocre set of recommendations ever did. It turns feedback into a conversation with a black box.

Furthermore, this feature cleverly deepens engagement by gamifying curation. Creating and pinning these custom feeds encourages repeated, playful interaction. It’s not just about watching anymore; it’s about designing your watching experience. You become a low-key programmer of your own media environment. This is a brilliant lock-in strategy. Why would you leave a platform that not only holds all the videos but lets you build bespoke, on-demand channels out of them? It makes the entire YouTube library feel personal, like a pantry full of ingredients you’re learning to combine in new ways.

The rollout to English in the US first is telling, highlighting the immense technical challenge of parsing intent across languages and cultural contexts. A prompt’s meaning is deeply idiomatic. But the long-term vision is clear. If this works, it could evolve from a fun novelty into a primary interface. Imagine prompts for work, for learning a skill, for family movie night. It starts to resemble the long-sought “conversational” interface to information, but focused squarely on the entertainment and knowledge space YouTube dominates.

It’s a move that challenges competitors on a new front. While TikTok masters the feed-as-discovery and Netflix perfects the hand-curated playlist, YouTube is introducing a hybrid: AI-powered discovery that starts with a spark of human intention. It’s a confident bet that users are tired of just scrolling and want to be collaborators. Whether that bet pays off depends on how well the machine learns to listen, and whether we find the process of teaching it to be a chore or a charm. For now, it’s a fascinating experiment in giving the audience a direct line to the recommendation engine, one hopeful, weird, or specific prompt at a time.