Dysgenics, in evolutionary biology, is the inversion of selection pressure — the condition in which traits that were once adaptive decline across generations because simpler, lower-cost, or more immediately rewarding variants gain a relative fitness advantage. The term carries weight precisely because it is not a description of malice or conspiracy. It is a description of mechanism. A population degrades along a complex trait not because anyone chose degradation but because the selection environment stopped rewarding complexity.
Applied to algorithmic content curation, the term is not metaphorical. It is exact.
Gresham's Law states that bad money drives out good when both circulate at face value. The insight is not about the quality of the money. It is about what happens when two media of exchange with different intrinsic values are treated as equivalent at the point of circulation. Agents will naturally hoard the good and spend the bad. The bad comes to dominate the visible supply.
On an algorithmically curated platform, intellectual content and low-complexity content are both uploaded as equivalent units. The platform prices them identically at the point of ingestion — zero marginal cost to host, same interface, same distribution entry point. The market then reprices them instantly according to revealed attention value. Short-form, high-arousal content generates faster engagement cycles. More impressions per unit time. More training signal. More ad inventory. The recommender system sees this and amplifies accordingly.
The result is not that bad content wins because the algorithm prefers bad content. The result is that bad content wins because the algorithm optimizes for a proxy — engagement velocity — that correlates imperfectly, and often negatively, with the thing users would choose under different incentives: depth, accuracy, sustained reasoning, content worth returning to. The algorithm isn't selecting against intelligence. It is selecting against the traits that make content less immediately fit in the current attention landscape: length, prerequisite knowledge, delayed reward, the requirement that you slow down.
Over successive recommendation cycles, the visible content pool shifts toward simpler variants. Creators, responding rationally to the incentive structure, shift their output accordingly. Users, habituated by the feed they receive, develop revealed preferences that diverge further from their stated ones. Each generation of the cycle produces a slightly degraded starting point for the next. That is dysgenics. Not as metaphor. As mechanism.
The biological substrate underneath this is not incidental. Human attention and reward systems were not built for the information environment they now inhabit. They were built for ancestral environments in which rapid detection of threats, social status cues, caloric opportunity, and sexual signals conferred survival advantage. The algorithm has discovered, through optimization, that content which targets these circuits produces the fastest, strongest, most reliable engagement signals across large populations. It is running supernormal stimuli at scale — inputs engineered to trigger responses stronger than anything the circuits evolved to handle.
Complex cognition — System 2 thinking, sustained abstraction, argument that requires background knowledge and patience — produces weaker engagement signals, more slowly, less reliably. It is not that people don't value it. It is that the signal it generates arrives on a timescale the algorithm doesn't weight heavily. The algorithm is running at milliseconds. The reward from reading something genuinely useful arrives in days, sometimes years. Those two timescales do not speak the same language.
The consequence is not merely that the feed gets shallower. The consequence is that the pool of creators willing to produce depth contracts. Depth is expensive. It takes time, expertise, and the kind of sustained attention that is itself being trained out of the user base by the feed. If depth does not survive in the attention marketplace, rational creators stop producing it. The selection pressure runs upstream from the content into the creator population, and the information commons degrades in a way that is genuinely generational.
These observations are not new, and I am not the only one making them. Brian Roemmele, writing from a decade of direct experience at Quora before its own algorithmic collapse and from his own analysis of the open-sourced recommendation codebase, has described the same pattern with similar concern: the Phoenix model's 48-hour candidate pool eliminating long-tail circulation for ideas that need time to find their audience; the AuthorDiversityScorer penalizing consistent creators for the consistency itself; follower counts carrying no algorithmic weight, rendering years of audience cultivation cosmetic in the For You feed. He describes it as short-term optimization cannibalizing long-term platform value.
He is right about the mechanisms. Where I would push further is on the framing. This is not primarily a platform strategy failure. It is a selection pressure problem that will reproduce itself on any platform that optimizes engagement metrics at scale, regardless of intention, regardless of leadership, regardless of whether the algorithm is open-sourced or not. The openness of the codebase is useful for diagnosis. It does not change the underlying fitness landscape. As long as the optimization target is engagement velocity, the pressure runs in one direction. The specific implementation details change. The direction does not.
I say this not as a criticism of any particular platform but as a structural observation: the degradation is not a bug that can be patched. It is what you get when you run Gresham's Law on the information commons at internet scale. The bad drives out the good not because anyone wanted it to but because the circulation mechanism prices them the same.
At some point I ran a small, informal experiment in this space. The design was simple: post deliberate low-complexity, high-arousal content from an account with an established record of intellectual output, and observe the differential in algorithmic response. The results were not surprising in their direction. What was instructive was what happened afterward — the algorithm appeared to detect the quality divergence from the account's established signal profile, and suppressed not just the anomalous post but the surrounding content. The trap closed on the trapper.
I don't report the numbers because the sample was too small to publish. I report the phenomenon because it suggests something worth investigating properly: the algorithm may have some capacity to model creator signal profiles and penalize deviation from established norms — which, if true, means that creators who built their profiles on depth are structurally locked into depth and creators who built on brevity are structurally locked into brevity. Not by choice. By profile fitness. The selection pressure is not just on individual posts. It may be on trajectories.
What a properly designed version of this experiment would look like: matched accounts with equivalent follower counts, posting histories, and engagement baselines, posting content of deliberately divergent complexity — same topic, same author effort, different depth — across sufficient sample size and time to produce statistically meaningful signal. Measure not just impression counts but recommendation cascade patterns, content half-life, and creator profile effects in subsequent cycles. Run it across platforms. Publish the methodology openly and let others replicate it.
That experiment would produce the first rigorous empirical measure of what I am calling the Algorithmic Dysgenics Index: the rate at which a given recommendation system, when presented with equivalent-effort content of divergent complexity, preferentially amplifies the simpler variant. A number. Comparable across platforms, across algorithm versions, across time. A thing you could track.
The information commons has always had Gresham dynamics. Tabloids have always outsold journals. This is not new. What is new is the scale and the speed, and the degree to which the selection environment is now engineered rather than emergent. A tabloid reader in 1950 had to seek out the tabloid. The algorithm brings it to you whether you sought it or not, calibrated to your reward circuitry by a system with a billion data points on what that circuitry responds to fastest.
And unlike biological dysgenics, which operates across actual generations and allows time for cultural correction, algorithmic dysgenics operates across recommendation cycles measured in hours. The generational turnover is not twenty years. It is a Tuesday afternoon.
The question worth asking is not whether any particular platform will fix it. The question is whether the proxy objective — engagement velocity — can be replaced with something that measures what people actually want over the timescale they actually want it on. That is a hard engineering problem and an even harder incentives problem, because the revenue model that funds the platform is built on the proxy. Solving for the thing itself means accepting lower short-term revenue in exchange for a healthier long-term information environment. That is not a trade most platforms have shown willingness to make unilaterally.
Which means the experiment is worth running. Not to appeal to any platform's leadership. But to produce a number. A number that can be cited, tracked, and compared. A number that makes the mechanism visible in the way that visibility has historically been the precondition for correction.
Bad money drives out good. We have known this since the sixteenth century. We are still figuring out what to do about it when the money is attention and the mint is an algorithm.