Market Commentary

The Algorithm You Can Talk To: what a typed sentence hands over, and what it doesn't

Four platforms now let people type a sentence describing the feed they want. This note reads the shift across YouTube, Meta, Threads, Spotify and Bluesky, and sets it against a live lawsuit over a feed that was faked and a regulatory split over whether control means an off switch or a conversation.

By Consumer desk × Folka

1. A box that accepts words

In September 2026, YouTube started letting viewers type what they want their feed to show, and what to leave out, and handed the sentence to Google's Gemini model to build around. Within weeks, Meta put a similar box on Facebook, after piloting it on Instagram; Threads already had one, launched in February, where a public request lapses after three days and can be reposted by someone else. Spotify and Bluesky's Attie followed with their own versions. None of the five companies has published what people type, how heavily the words are weighted against the old behavioural signals, or whether a sentence typed in good faith actually changes anything a viewer sees.

That question sits underneath every design decision these products have made, and the decisions already diverge sharply enough to matter.

2. Five products, five different leashes

The common feature is a text box. The uncommon part is how far the sentence is allowed to travel. YouTube's custom feed lives in its own tab; watching from it does feed back into the main Home feed, but the main feed is not rewritten by the sentence directly. Instagram's and Facebook's versions work on the primary feed itself. Threads' request expires after three days and is, unusually, copyable by other users — a hint that the next version of this feature may serve borrowed preferences rather than personal ones. Spotify's prompted playlists never touch the home feed logic at all; they sit inside a single playlist. Bluesky's Attie, built on Anthropic's Claude, is trying to make the written feed itself portable, carried between apps on the AT Protocol rather than owned by one platform.

Seen this way, the industry has not converged on one idea of what a sentence should be allowed to do. It has built five different leashes, and kept the length of each one undisclosed.

3. Writing the request is its own skill

Before asking what a platform does with a typed preference, it is worth asking whether people can write one well. A Stanford team — Popowski, Bernstein and colleagues, presented at CHI this year — had people build feeds from Bluesky posts by describing, in their own words, what they wanted. The result was not that people didn't know what they wanted. It was that stating it in a single sentence, unaided, reliably missed gaps and edge cases the writer hadn't foreseen.

Participants, faced with their own gaps, preferred feeds built from an AI interviewer's follow-up questions over the sentence they wrote unaided. That is a usability finding, not a demand signal — it says people struggle to specify, not that they want to. But it means the box, as shipped, asks for a skill most people don't reliably have.

4. The test that proves it moves anything is the one that gets you sued

The hardest evidence so far sits not in a feature announcement but in a lawsuit. New York's amended complaint against TikTok, unsealed on 7 October (Reuters), alleges that in 2023 a test group — including children — was given an 'Algo Refresh' tool that was presented as a way to improve their feed but, for the test cohort, did nothing at all, while TikTok measured time spent and ad revenue against a working version. The suit cites an internal product manager who objected in writing that a placebo group 'completely conflicts' with the feature's stated purpose.

TikTok has called the suit's picture false and described the test as routine, and that defence deserves real weight: a holdout group is the standard way to test whether a feature does anything at all. Every claim here remains an allegation. But the case exposes the knot every prompt box now sits inside — proving a typed preference changed the feed requires exactly the kind of controlled, undisclosed test that got a platform sued when children were in the holdout group and someone inside objected.

5. The sentence may matter more to advertising than to the feed

A typed preference is a different kind of data from a click. A click is inferred; a sentence is declared, and declared intent is worth more to anyone trying to sell against it. Meta has already shown where that value tends to flow. The company announced that, from December 2025, it would use conversations with Meta AI to personalise ads — outside the EU, the UK and South Korea. It has not said whether the text typed into 'Your Algorithm' is retained, or fed into the same ad-personalisation pipeline.

No platform has confirmed or denied that last step for its own prompt box. Given that Meta has already built the pipe for one text source — chat with its assistant — and not ruled it out for another, the absence of a denial is itself informative. The box may do more for the advertising side of the business than for the feed the person asked for.

6. Two ideas of control, built for different markets

Regulators have already picked sides on what 'control' should mean, and the typed-prompt wave is arriving into markets that disagree with each other. Ireland's media regulator proposed switching recommender systems off by default, then dropped that in its final 2024 code and pointed instead to the EU's Digital Services Act, which since 2023 has required the largest platforms to offer at least one feed in the EU that is not based on profiling at all. China's 2022 rules on algorithmic recommendation took a related but distinct approach: platforms must let users switch personalised recommendation off entirely, and must let them view and delete the interest tags built on them.

Put side by side, these are not three points on the same scale — they are different definitions of the word 'control'. Europe and China both built a switch a person can throw. The platforms shipping typed prompts have built something closer to a conversation, available only in English-speaking markets, that sits on top of a system the person still cannot turn off.

7. How far each sentence actually reaches

Laid flat, the five products read less like one feature rolling out everywhere and more like five separate bets on how much trust a sentence deserves.

The spread matters because it sets the ceiling on what a typed sentence can realistically be worth to a business watching this category: on four of five products, it still operates downstream of, or alongside, the same behavioural engine that ran the feed before the box existed.

8. Our view

We think the typed-preference wave is better read as an interface change than a control change, and the distinction matters for anyone deciding how much weight to put on it. The box gives people a new channel to speak through, but none of the five platforms has shown that the sentence overrides, rather than sits beside, the engagement data that has always run these feeds — and the one piece of hard evidence available, the TikTok complaint, shows a platform willing to run a feature as a measurement instrument rather than a real control, with a staff member objecting in writing. The commercial logic points the same way the engagement data always has: a platform's revenue depends on time spent and ad value, and a sentence that reliably reduced either would be a strange thing to ship at scale without quietly discounting it in practice.

None of this means the feature is dishonest. It means the burden of proof that a sentence genuinely steers a feed, rather than decorating one that runs on the same signals as before, has not yet been met by anyone who has shipped the feature.

9. What the sentence actually is

It is worth naming the thing plainly, because the industry hasn't. Call it the request layer: a typed statement of intent that sits above the behavioural engine a feed has always run on, with no platform yet disclosing how the two are weighted against each other when they disagree. The term matters because it describes a layer, not a replacement — and a layer can be thin, decorative, or load-bearing, and currently nobody outside these five companies can tell which.

The forks ahead are not close to choices businesses in this category can avoid watching: whether a sentence's reach stays confined to a sidecar tab or spreads to the main feed; whether the words become retained, monetisable data or stay feed-only; and whether control, globally, ends up meaning a switch a regulator mandated or a conversation a product team designed. Each path carries a different cost — regulatory exposure in Europe and China if the sentence is read as another form of profiling, reputational exposure in the US if a sentence turns out to be decorative, and a usability ceiling everywhere, since Stanford's finding suggests many users cannot yet specify what they want well enough for any of this to matter. None of the three forks has a visible winner; what would move each is concrete and checkable — a platform disclosing its weighting, a court ruling on the TikTok holdout claim, or a second jurisdiction following the EU and China toward a mandated switch.

How we read this

This note draws on company product announcements (YouTube, Meta, Threads, Spotify, Bluesky), reporting on those launches (TechCrunch, CNBC, Fortune), a CHI research presentation (Popowski, Bernstein et al., Stanford), an unsealed legal complaint reported by Reuters, and the published texts of the EU's Digital Services Act and China's 2022 algorithmic recommendation rules. Claims about usage volume, prompt weighting, and whether typed text feeds advertising systems are not supported by any disclosed figure from the companies involved and are flagged as such throughout, not asserted. A single disclosure — any platform publishing how heavily a typed sentence is weighted against behavioural signals when the two disagree — would be the one fact most likely to change this read.

Grounded in Folka’s corpus: The Algorithm on Request · Machine Taste · The Attention Diet · Post-Feed