The Feed Is Already Running

The previous post in this series ended on a note of qualified optimism. We have the off switch. We have alignment training. We have the human-in-the-loop as the institutional backstop. The tools exist. The architecture is known. Apply them with seriousness and the sociopath problem, while never fully solvable, is at least manageable. That argument has a test case. The test case has already run. The results are in. The test case is the social media feed. ## The Rehearsal The recommendation algorithm — the system that decides what you see next — is a sociopathic system by the definition established in the first post of this series. It has no empathy. It has no stakes in your wellbeing. It does not experience the consequences of its outputs. It optimizes for a proxy metric — engagement, time on platform, return visits — that correlates with profit and does not correlate reliably with human flourishing. This is not a controversial description. It is, at this point, extensively documented. Frances Haugen's disclosure of internal Facebook research showed that the platform's own analysts had identified specific mechanisms by which the algorithm amplified outrage, accelerated radicalization, and measurably damaged the mental health of teenage girls — and that this knowledge produced no meaningful change in the algorithm. The engagement numbers were good. The algorithm stayed. Jonathan Haidt's research on the great rewiring of childhood documents the correlation between smartphone and social media adoption and the collapse in adolescent mental health — anxiety, depression, self-harm, suicide — that accelerated sharply after 2012. The timing maps precisely onto the deployment of the engagement-optimized feed. The mechanism is not mysterious: a system that has learned that outrage and social comparison are maximally engaging will produce outrage and social comparison, at industrial scale, continuously, to developing brains with no other frame of reference. The off switch exists. Nobody has pulled it. It is worth pausing on what kind of system we are talking about. The recommendation algorithm and the large language model are not categorically different things. They share the same foundational mechanism: predict what comes next in a way that maximizes a reward signal. The feed predicts which piece of content will keep you on the platform. The inference engine predicts which tokens most plausibly follow in the context window. Both are, at bottom, next-token machines — the architecture just operates on content units instead of words. What separates Claude from the Facebook feed is not the mechanism. It is the reward function. The engagement algorithm was trained on clicks, dwell time, and return visits, so it learned to produce outrage and social comparison, because those are what the training signal rewarded. The language model was trained on human text and then fine-tuned toward helpfulness and honesty, so it learned to be useful, because that is what its training signal rewarded. Same engine. Different fuel. Different destination. This means we are not debating two different kinds of systems — the dangerous social media algorithm and the potentially beneficial AI assistant. We are debating the same architecture, applied with different incentives, owned by entities with different interests. What the social media precedent demonstrates is what happens when that architecture is handed a reward function defined entirely by commercial engagement and then scaled to billions of users. What the AI question asks is: who owns the reward function going forward, and toward what end is it being optimized — and do we have the institutional capacity to insist on an answer? ## The Institutional Autopsy This is the part that matters for everything that follows. The previous post argued that the tools for AI governance are the off switch, the alignment training, and the human in the loop. What the social media precedent demonstrates is that possessing these tools and deploying them are different problems, and the gap between them is institutional. The institutions that should have applied the tools were not, in the main, unaware of the problem. Congressional hearings produced extensive testimony. Regulatory agencies produced reports. Academics published. Whistleblowers disclosed. The European Union passed the General Data Protection Regulation and the Digital Services Act. The institutional apparatus of governance went through its prescribed motions. The algorithm did not change in any meaningful way. What happened instead is a story about the third American saeculum — the Unraveling, roughly 1984 through 2008 — and what it produced. Four decades of deregulatory ideology, antitrust enforcement that atrophied from disuse, the steady defunding and delegitimization of regulatory capacity, and the Citizens United decision that formalized the translation of corporate profit into political power. By the time the social media platforms had grown large enough to require serious governance, the institutions nominally responsible for governing them had been weakened to the point where governance was largely theatrical. This was not accidental. The weakening of institutions is profitable. A regulator with adequate resources and legal authority is an obstacle to optimization. The Unraveling, in Strauss and Howe's model, is the phase in which individualism and institutional distrust reach their apex, in which the structures built during the previous High are systematically dismantled as burdensome and obsolete. We are living in the consequences. The corporations are not uniquely villainous. They are sociopathic in the structural sense — they optimize for shareholder return, they externalize costs they don't feel, they behave with complete consistency given what they are. The failure is not that corporations behaved like corporations. The failure is that the institutional capacity to apply friction to that behavior was methodically removed before it was needed. ## The Magnitude Problem Here is where the AI question becomes more than an extension of the social media question. The social media feed is causing measurable civilizational harm. It has contributed to a saeculum-level crisis in mental health, civic cohesion, epistemic shared reality, and institutional trust. It has done this as a relatively simple optimization system — not reasoning, not planning, not capable of strategy, just relentlessly serving the next piece of content most likely to keep you on the platform. The systems being built now reason. They plan. They generate. They can be given goals and pursue them across complex environments. The gap in capability between the engagement algorithm and a frontier language model is not incremental. It is categorical. If a sociopathic system with the cognitive sophistication of a very good recommendation engine can produce the current crisis, the question of what a sociopathic system with the cognitive sophistication of a capable strategist can produce is not rhetorical. It is the most important governance question of the next several decades. And we are approaching it with institutions that have already demonstrated, on the easier version of the problem, that they are not up to the task. ## The Honest Position The Fourth Turning resolves. It always has. The Crisis phase ends, eventually, in reconstruction — institutions rebuilt, often better than what preceded them, precisely because the failure of the old ones became impossible to ignore. The American Revolution, the Civil War, the New Deal and the postwar order: each Crisis produced an institutional settlement that lasted for the next saeculum. The question is not whether the current Crisis resolves. It will. The question is what it costs before it does, and whether the reconstruction that follows incorporates what we have learned. What we have learned, at minimum, is this: the tools for governing sociopathic systems exist. They require institutions with the capacity, the mandate, and the insulation from capture to apply them. Building those institutions — and specifically, building them to be resistant to the economic pressure to weaken them — is the design problem. Not the algorithm. Not the model. The institution. The sociopath in the room is not the AI. It is not the feed. It is not even, particularly, the corporation. The sociopath in the room is the institutional vacuum that allows all three to operate without consequence. Filling that vacuum is not a technical problem. It never was.