Somewhere in the migration backlog sat a task nobody wanted: backfill word_count
and reading_time
on every post we’d ever published. Boring work. Except it wasn’t boring – it was a symptom.
The SEO generator downstream needed those numbers to route content into different distribution logic. Long posts get treated differently than short ones. Reading time isn’t a courtesy to the reader anymore – it’s a signal a ranking system consumes. We weren’t backfilling a column. We were retrofitting curation metadata onto a pipeline that had already moved on to amplification.
That’s the shift, in miniature. And it’s happening in every content system worth building right now.
Two different jobs, wearing the same badge #
Curation is a human (or human-trained) act of selection: pick the best of a corpus, present it, stand behind the choice. Algorithmic curation formalizes this with machine learning models trained on user behavior – past interactions, preferences, engagement – aiming to maximize relevance and retention at a scale no editor could match.
Amplification is a different animal. Algorithmic amplification describes increases in a piece of content’s distribution or visibility, driven by a combination of automated ranking and users’ own sharing. One conceptual model puts it plainly: engagement-based ranking pushes a creator’s post beyond their own follower network, into feeds that never asked for it.
The distinction matters because the two systems optimize for different things. Curation asks “is this good?” Amplification asks “will this spread?” Those questions produce different rankings from the same corpus.
Where we’ve felt this in our own stack #
We run two systems at Glad Labs that sit on opposite sides of this line, and building them side by side made the difference concrete instead of theoretical.
Poindexter is our SEO maturation layer – it takes content-writing output and turns it into a ranking engine using Google Search Console data for keyword enrichment. That’s curation logic wearing an algorithm’s clothes: it’s still asking “what does this specific audience search for, and does our content answer it?” The judgment is encoded in signals, but the intent is editorial.
Jettison is the opposite instinct. It’s a campaign generator that turns a single sentence into ads, social posts, and emails in under 60 seconds. Speed is the entire feature. There’s no curatorial pass built in – the design bet is that fast, high-volume output finds its audience through distribution mechanics, not through hand-picked quality control up front.
Neither approach is wrong. But building a Poindexter and expecting Jettison’s throughput – or building a Jettison and expecting Poindexter’s precision – produces weeks of confusion about why the metrics don’t match the mental model you started with. We did exactly that, more than once.
The mechanics under the hood #
Engagement-based ranking doesn’t care about your content’s provenance. It cares about behavior after the fact: clicks, watch time, replies, shares. The Knight First Amendment Institute has documented this on streaming platforms specifically – recommender systems optimized for engagement, retention, and revenue end up amplifying some genres or creators while quietly overlooking others, even when the underlying catalog hasn’t changed.
That’s the part that should worry technical readers more than the culture-war framing usually applied to this topic. The distortion isn’t ideological. It’s structural. Multiple studies cited in that same research show the amplification effect persists regardless of what the recommender was told to optimize for – it’s an emergent property of ranking by engagement, not a bug you patch with a different training set.
An older piece from Inference makes a sharper version of the same point: the real risk to information quality isn’t exotic generative tech, it’s the mundane recommendation algorithm quietly curating what gets surfaced from search and social platforms. That argument was made in 2021, well before generative AI became the headline concern – which tells you the amplification problem was structural before language models even entered the conversation. It’s a pipeline problem, not a content-quality problem.
We ran into a version of this ourselves while building our RAG pipeline. A retrieval system that ranks by embedding similarity alone will amplify a document that’s superficially close to the query but wrong in substance – confidently, every time, because nothing in the ranking step checks for that. Curation-shaped judgment has to be added back in deliberately, or the system defaults to amplification-shaped behavior by construction.
Why this isn’t just a platform problem #
If you’re building anything that ranks, feeds, or surfaces content – a recommendation engine, a content pipeline, a RAG retrieval layer – you’re not choosing between curation and amplification as an abstract philosophy. You’re choosing it in your ranking function, whether you notice or not. We wrote about a related failure mode in The Echo Chamber Problem, From the Inside: systems that optimize for what a user already engages with tend to narrow what they see, not broaden it. Amplification and echo-chamber formation are two views of the same mechanism. Rank by past engagement, and you’ll get more of what already worked – that’s the amplification logic executing exactly as designed. It’s just not the design most builders think they signed up for.
A piece on art curation makes the case that curators were never just gatekeepers – they were facilitators of dialogue, deciding what conversations got to happen next. Amplification systems inherit that same power without inheriting the judgment. The ranking function decides what conversation happens next, and it does it without anyone in the room asking whether it should.
What to actually do about it #
Instrument the difference. If your system ranks by engagement signals, name that explicitly in your architecture docs – don’t let it hide behind the word “relevance.” Add a curation pass somewhere in the pipeline – a quality gate, a retrieval filter, a human review step – even if it’s thin. Measure what your amplification logic overlooks, not just what it surfaces; the Knight Institute research shows the overlooked catalog is often where the real distortion lives.
The shift from curation to amplification already happened. It happened in recommendation systems, in streaming platforms, in social feeds, and it’s happening quietly inside your own content pipeline right now, one ranking function at a time. The only real choice left is whether you built that shift on purpose.
Sources #
https://en.wikipedia.org/wiki/Algorithmic_amplificationhttps://knightcolumbia.org/content/a-public-service-media-perspective-on-the-algorithmic-amplification-of-cultural-contenthttps://inference-review.com/letter/on-algorithmic-amplificationhttps://medium.com/@mariambrian/the-power-of-curation-why-art-curation-shapes-humanity-more-than-art-itself-ab3b9dac8fce