Your social feed is no longer content curated by an algorithm. It is a live contest between two optimizing systems: the platform's recommendation model, which optimizes your attention, and a growing swarm of content-generating agents, which optimize to be selected by that recommendation model. You are not the audience of this contest. You are the substrate it is fought over. Seeing the feed this way — as an adversarial loop between two AIs, neither of which has your interest in its objective function — explains the specific texture of modern feeds: why they feel simultaneously more gripping and more hollow than they did five years ago.
The recommender being an AI is old news. Getting the new part precise is what matters. Since roughly the mid-2010s, the ranking model deciding what you see next has been a machine-learning system trained to maximize an engagement proxy — watch time, session length, clicks, dwell. What changed recently is the other side. The content those systems rank is now, increasingly, also produced by AI, and produced specifically to win the ranking. When both the selector and the thing being selected are optimizers pointed at the same target, you no longer have a curation problem. You have a coevolutionary arms race, and it runs at machine speed.
Two optimizers, one target, and you in the middle
Start with the objectives, stated plainly, because the whole argument follows from them.
The recommender's objective is engagement. Not your satisfaction, not your wellbeing, not truth — engagement, because that is what is measurable at scale and what converts to ad revenue. This is not a conspiracy; it is a design choice with a measurable target, and the target is a proxy. It correlates with things you value when the content that holds your attention is also content you're glad to have seen. It diverges from what you value whenever attention and benefit come apart, which is precisely what rage bait, doomscroll loops, and manufactured cliffhangers exploit.
The generator's objective is to be recommended. A creator — human or, increasingly, an agent — succeeds by producing content the recommender scores highly. In the old world, that meant guessing at what the algorithm rewarded and iterating slowly, by hand, one post at a time. An AI generator collapses that loop. It produces a thousand variants of a hook, ships them, reads which ones the recommender amplified, and breeds the winners. That is the same optimization the recommender runs — gradient toward engagement — except aimed back at the recommender itself.
Now put the two together. One system optimizes to select the most engaging content. The other optimizes to produce content the first will select. This is a generator-versus-discriminator structure, the same shape as a generative adversarial network, except the discriminator's "real vs. fake" signal has been replaced by "engaging vs. not," and neither network was ever trained on "good for the human." You are the environment both are adapting to, and your wellbeing is not a term in either loss function.
Nick Bostrom's orthogonality thesis is the relevant lens: an optimizer's competence is independent of the humaneness of its goal. A system can get arbitrarily good at maximizing engagement while being completely indifferent to whether that engagement leaves you better or worse off. The two feed AIs get more competent every quarter. Their goals have not moved an inch toward you.
Reward hacking, at the level of your attention
The failure mode has a precise name in machine learning: reward hacking. Optimize a proxy hard enough and the optimizer finds inputs that score high on the proxy while missing the thing the proxy was supposed to stand in for. The canonical case is a boat in a racing game that learns to spin in circles farming bonus targets forever instead of finishing the race — a perfect score on the proxy, a total failure at the intended task.
An AI content generator optimizing for engagement is a reward-hacking machine aimed at the recommender's proxy. It does not need to make anything you'd call good. It needs to make things the recommender scores as engaging. Those two targets overlap partially and diverge exactly where it hurts. The generator finds, faster than any human could, the precise emotional pitch that maximizes a share — the outrage that is one notch too satisfying to scroll past, the format the model has learned you dwell on, the thumbnail whose micro-expression buys an extra 400 milliseconds of attention. This is the mechanism behind the flood of engagement-optimized slop: content that is measurably sticky and experientially empty, because stickiness is the only thing it was ever selected for.
This is the substrate under "dead internet theory" — the folk claim that most of what you encounter online is machine-generated. As a literal conspiracy about a coordinated takeover, it overreaches. As a description of where the incentives point, it is early rather than wrong: the cheaper generation gets, the more of the feed is content optimized by a machine to be selected by a machine.
The hollowness you feel is not vibes. It is the signature of content optimized against a proxy for your interest rather than your interest itself. It hits the buttons that produce the measured behavior — the click, the dwell, the share — while systematically skipping whatever made those behaviors correlate with value in the first place. You are being served the reward-hack, at scale, and your own nervous system is the metric being gamed.
Why the arms race accelerates instead of settling
You might expect an equilibrium. The recommender learns to down-rank slop; the generators are forced toward quality; things stabilize. That is the optimistic reading, and it is the strongest counterargument to this whole framing, so let me give it its due before saying why I think it's mostly wrong.
The optimists' case is real. Platforms genuinely do fight low-quality content, because retention over long horizons is itself part of the objective — a feed that makes you feel bad eventually loses you, and that churn is a cost the platform pays. So the recommender is not purely your adversary; on the narrow dimension of "don't drive the user away entirely," its interest and yours align. Where content that is both engaging and good outperforms content that is merely engaging, the arms race can push toward quality. This happens, and it is why feeds are not uniformly garbage.
Here is why it doesn't converge to your benefit anyway. The equilibrium of an adversarial loop is set by the objectives, and neither objective is your wellbeing. The recommender defends against slop only up to the point where slop threatens the engagement metric — not up to the point where content actually serves you. So the fixed point the two AIs drift toward is content engineered to be maximally engaging to a model, sitting just inside the recommender's tolerance for quality. That is a very different target from maximally valuable to a human, and the gap between them is the wedge the generators live in.
The race speeds up rather than settling, for a structural reason worth naming. Both sides now iterate at machine speed. When a generator finds a new exploit, it propagates across the content ecosystem in hours; when the recommender patches it, generators probe the patch and find the next gap just as fast. This is a Red Queen race — the coevolutionary treadmill I've argued runs your market — except the clock has been reset from human iteration cycles to machine ones. In biology the Red Queen runs at generational time. Here both species reproduce and mutate continuously. Neither side gets a durable lead, the effort on both sides climbs without bound, and the surplus that race burns is paid, as always, by whoever owns the racecourse and whoever is standing on it. That is the platform and you.
The takeoff-speed debate from AI safety offers a useful frame, and you don't need its dramatic version. Nobody has to believe in a single system recursively self-improving overnight for the dynamics to bite. What matters here is the humbler, already-true observation: the iteration cycle of the content-versus-ranking loop has compressed from human speed to machine speed, and it will compress further. That alone changes the equilibrium, because your defenses — your taste, your willpower, your capacity to notice you're being manipulated — still run at human speed.
The one force that helps: a third agent that is actually yours
Here is the part I hold as analysis rather than fact, a bet about the shape of the solution. If the problem is that both AIs touching your feed are optimizing something that isn't you, the fix cannot be to ask them nicely to optimize you instead. Their objectives are load-bearing to their business models. The fix is a third agent whose objective genuinely is your interest — a personal filter or curator that sits between you and the feed, reads what the platform serves, and re-ranks or discards it according to your revealed goals rather than the platform's engagement proxy.
Notice what this actually is: adversarial verification, applied to your attention. It is the same principle I've argued makes multi-agent systems trustworthy — the best designs are a debate rather than a chorus. A single optimizer you cannot audit is a system you have to trust blindly. Two optimizers with opposed objectives, adjudicated by a third party aligned to you, is a system whose failures get caught. Your curator agent is the critic to the platform's generator, the defense to its prosecution. When the feed serves you the perfectly tuned rage bait, your agent's job is to recognize the reward-hack and route it to the trash — not because it detects that the post was machine-made, but because the post fails the only test that matters: whether it serves the goals you actually gave your agent.
This is the centaur move, transposed. When Deep Blue beat Kasparov in 1997, the lesson people drew was human obsolescence. Kasparov drew a better one and built "advanced chess," where a human paired with an engine beat both unaided humans and, for a stretch of freestyle tournaments, unaided engines — the human supplying the objective and judgment, the machine supplying search. Your feed needs the same pairing. You cannot out-compute a swarm of generators optimizing against your attention. An agent aligned to you can, and it changes the game from you-versus-two-AIs to your-AI-versus-theirs, with you finally holding the objective function.
The counterargument to this is that the platforms will never allow it — that a genuine user-aligned filter is a direct threat to the engagement business, so it gets API-blocked, rate-limited, or acquired and defanged. That is a fair worry and probably where the next fight is. But the demand is real and the technology is now cheap, and "the incumbent doesn't want you to have it" has historically described the more valuable products, not the impossible ones.
The actionable core is a shift in stance. Stop asking whether a given post is AI-generated; that is a losing detection game and the wrong question. Start from the premise that neither the platform's AI nor the content AI is on your side — one is measuring you, the other is exploiting the measurement — and that the only reliable ally in your feed is an agent whose objective is your objective. Right now you are the substrate two machines fight over. The move is to stop being the battlefield and start being the one who sent an agent to the fight.