How do recommender systems respond to engagement signals from AI-generated content?
This explores what happens when recommender systems—built to read engagement (clicks, likes, shares) as a proxy for human interest—encounter AI-generated content that produces those same signals without the human dynamics underneath them.
This explores what happens when recommender systems, which treat engagement as a stand-in for genuine human interest, meet AI content that can manufacture those signals hollowly. The corpus suggests the core problem isn't that recommenders reject AI content—it's that they can't tell the difference, because the signals they optimize for are exactly the ones AI content is good at producing.
The sharpest finding is that AI posts accumulate visibility and 'social proof' through comprehensive, confident phrasing while suppressing the reply dynamics that historically validated a post Why do AI posts get likes without inviting conversation?. A recommender reading likes and impressions sees a winner; it doesn't see that no conversation happened. Over time this displaces human influencers and erodes the platform's reputation-promoting function, even as engagement metrics keep ticking up Does AI content displace human influencers on social media?. The unsettling part: this threat operates below the level where content moderation, fact-checking, or recommender adjustment can reach, because what AI content drains—genuine address and mutual orientation—was never something the ranking signal measured in the first place Does AI threaten social media's conversational function?.
Then there's a counterintuitive twist that breaks the naive 'AI gets more engagement' story. In Nextdoor experiments, LLM-generated summaries were objectively more informative—and *reduced* click-through, because a reader whose information need is already satisfied has no reason to open the notification Does better summary writing actually increase user engagement?. So AI content can both inflate engagement (when it games visibility) and deflate it (when it's too good to require a click). Either way, the engagement signal stops tracking what the recommender assumes it tracks.
Step back and the recommender isn't a neutral mirror at all—it's persuasion infrastructure whose feed weights shape what producers make, with effects that compound through rating contamination and selection bias How do recommendation feeds shape what people see and believe?. AI content enters this loop as a contamination source: it floods the rating signal, the feed amplifies it, producers adapt toward whatever the feed rewards, and the cycle tightens. And when the recommender is *itself* an LLM, the bias problem moves inside: GPT-4 recommenders concentrate on items popular in their pretraining corpus rather than the actual dataset—The Shawshank Redemption surfacing everywhere—a domain-shift artifact that standard debiasing can't fix Where does LLM recommendation bias actually come from?.
The thread worth pulling: the failure isn't a bug in any single ranking model, it's that AI content decouples the signal (engagement) from the thing it was a proxy for (genuine human attention and address). Systems like persona-weighted recommenders that try to trace each suggestion back to a specific real user preference Can attention mechanisms reveal which user taste explains each recommendation? hint at a possible response—building recommenders that model *why* something is engaging rather than just *that* it is—but the corpus is clear that today's engagement-optimizing systems are structurally blind to the distinction.
Sources 7 notes
AI-generated posts achieve high engagement metrics through comprehensive, confident phrasing but suppress reply dynamics because they lack human authorship and invite no counter-argument. This creates one-sided recognition divorced from the conversational validation that historically legitimized social proof.
AI-generated posts capture engagement through comprehensiveness but accrue social proof without building any speaker's sustained reputation. This displacement compounds over time, eroding the platform's core function of promoting legitimate human voices while monetization continues.
AI-generated posts drain social media's function as a conversational medium because they lack the structure of genuine address and mutual orientation. This threat operates below the level where content moderation, fact-checking, and recommender adjustment can reach.
Nextdoor experiments showed LLM-generated summaries were objectively more informative but decreased click-through rates. Users had no reason to open notifications when the summary already satisfied their information need, demonstrating how optimizing for informativeness can backfire on engagement metrics.
Research shows recommendation systems operate as political actors: feed weights influence producer behavior, network topology drives opinion convergence, and automation enables targeted persuasion at population scale. These effects compound through rating contamination and selection biases.
GPT-4 concentrates recommendations on items popular in its pretraining corpus rather than in target datasets. The Shawshank Redemption dominates across different datasets even when they have different popularity distributions, revealing a domain-shift effect that standard debiasing methods cannot address.
AMP-CF represents each user as multiple latent personas weighted dynamically by candidate item. This makes recommendations both diverse and interpretable—each suggestion traces to the specific persona preference it satisfies—without requiring post-hoc reranking.