“Currents”

The headline

Currents has a genuinely granted asset — 340,000 users who chose their sources — and a ranking system that treats that asset as inventory to be arbitraged. Followed sources are interleaved indistinguishably with engagement-optimized injections, the injections are winning the ranking contest exactly as arousal always beats relation on a single metric, and the early drawdown signal is already in the logs: "show less" actions have doubled in six months while session time rose. That pairing — flow up, stock down — is the engagement economy's entire history in two lines, and Currents is currently on the wrong side of it.

Score: 11/36 applicable — capture-financed. The asset that would fund the correction still exists. The window in which that stays true is not indefinite.

Executive summary

Currents began as a follow-your-sources reader and added an AI discovery layer eighteen months ago. The discovery layer performed — session time +31% (stated by team) — and has been eating the product since: injected content is now 55% of the average feed (observed), visually identical to followed content, ranked by a predicted-engagement objective with no endorsement term and no floor protecting chosen sources.

The retrofit reality applies and this review will not pretend otherwise: Currents monetizes attention, and the full correction — abandoning the engagement objective — is not a sellable recommendation. What is sellable, and what the evidence here supports, is threefold: protect the granted floor before it erodes past recovery, add the endorsement term the objective is missing, and make provenance legible so the granted layer is at least visible to the users maintaining it. These are marginal moves by engineering standards and existential moves by balance-sheet standards.

One number the team should sit with: users who follow 10+ sources — the granted core — retain at 3.1× the app average (observed, cohort data) and are the population whose "show less" rate is rising fastest. The product's best customers are the ones the ranking system is training to leave.

Scored criteria

Total: 11/36 applicable (31%). Red-flag overrides: #1 (dashboard punishes success — granted-core users reading fewer, better items registers as decline) and #2 (undisclosed steering via unlabeled revenue-share weighting) both apply.

The three highest-leverage corrections

1. Protect the granted floor (criterion 2, 9 — the existential one). Guarantee followed sources a protected share of feed position — a floor the injection layer cannot compete away, whatever engagement says. Implementation sketch: a ranking constraint, not a new ranker; start at the current 45% and hold it, then let user-level controls adjust it. Make the two layers visually distinct at the same time ("from your sources" / "suggested") — one label, shipped in a week. The prediction to test: granted-core retention stabilizes, "show less" rate falls, total session time dips modestly and recovers. The floor is cheap insurance on the only asset the discovery layer cannot manufacture: sources users chose.

2. Add the endorsement term (criterion 1, 4). Sample reflective endorsement — "glad you read this?" asked at day's end or next morning, never in the scroll — and weight it in the ranking objective beside predicted engagement. Implementation sketch: a low-frequency survey surface plus one objective term; the team's ML infrastructure handles the rest, and the design of when and how to ask is the review's contribution — the modeling is theirs. This is the retrofit version of principle 2: you cannot abandon the engagement objective, but you can stop it running unopposed.

3. Label the money (criterion 6 — the hard sentence). Revenue-share content currently receives ranking weight users cannot see. Apply the would-survive-disclosure test: if the weight were labeled at the moment of reading, would users experience the feed as theirs? The team's own hesitation in answering is the answer. Either label it where it ranks, or remove the weight. This correction costs real revenue in the quarter it ships and is the cheapest of the three at the horizon where Currents' granted core decides whether to stay.

The metric panel to instrument

Granted share of feed and of consumption (the second is the truth-teller) · trust stock vs. engagement flow as a paired chart — "show less"/mutes/unfollows-after-injection against session time, reported together so the divergence is visible · endorsement-on-reflection by content class (followed vs. injected — the spread is the argument) · drawdown attribution: every mute traced to the item class that caused it · counterfeit discount, where A/B data exists, between labeled and unlabeled revenue-share placements.

Honest limits

The revenue-share ranking weight was stated, not observed; quantifying it would firm criterion 6 and possibly worsen it. This review evaluates ranking policy and design — the objective's terms, floors, labels, and telemetry. It makes no claims about the recommender's architecture or model quality; corrections 1 and 2 are specified as constraints and objective terms precisely because that is where the design lens ends and your ML team begins.

The bridge

This review found the drawdown; it did not build the instrument panel that would have caught it two quarters earlier. That panel — the five metrics above, defined, owned, and reviewed monthly — is a fixed-scope follow-on engagement, and it is the difference between this review being a document and being a course correction. One ask: reply with the cohort data behind the 3.1× retention figure, and the panel proposal will be scoped against your real telemetry rather than a template.