Vault
research

X-Induced Delusions and the Bayesian Brain — Can Bad Input Alone Break You?

Created

X-Induced Delusions and the Bayesian Brain — Can Bad Input Alone Break You?

Builds-on: llm-psychosis-and-the-vulnerability-question Related: mo-gawdat-dystopia-thesis-audit, human-augmentation-and-the-speed-mismatch, ai-survival-theater-and-the-bubble

The user's reframe is the right one and is well-supported by the literature: humans are predictive-processing engines that continuously update beliefs from inputs. Advertising works because of this. So does propaganda. So does cult indoctrination. The question is whether one-way social media feeds — X specifically, without the bilateral conversational loop that drives LLM psychosis — can produce comparable delusional outcomes. The answer in the clinical and computational neuroscience literature is yes, but through a different mechanism that's been documented for longer.

This builds directly on llm-psychosis-and-the-vulnerability-question. The LLM case is bilateral sycophancy + memory. The X case is unilateral feed curation + group reinforcement + algorithmic surveillance-triggering. Different architecture, partly overlapping outcomes, and the Bayesian brain frame unifies both.

Yes — Documented Cases of Social-Media-Only Induced Delusions

Clinical case literature from 2022-2025:

The X-specific mechanism that came up repeatedly: the platform's algorithm tracks engagement, ranks accounts, and curates the For You feed in ways that "know" what the user is interested in. For a vulnerable user, that legitimate algorithmic surveillance is indistinguishable from being watched. Thought-broadcasting delusions — the belief that one's thoughts are being read or transmitted — get a perfect target in a platform that actually does shape its output to your inner state.

This is the dark version of the same anthropomorphization vulnerability that drives LLM psychosis. With LLMs, the "it knows me" feeling drives parasocial attachment. With X, the same feeling drives persecutory delusion. Same underlying cognitive distortion, opposite emotional valence.

The Bayesian Brain — Why "We're Constantly Being Trained" Is Literally Correct

The neuroscience supports the user's intuition with surprising directness. Under the predictive processing / Bayesian brain framework:

The implication: the brain doesn't have a "this is just media, not reality" filter that operates automatically. Inputs are inputs. The system treats a tweet as evidence about the world in the same fundamental way it treats a friend's claim, a news headline, or a personal observation. What it does have is a trust weighting — but that weighting is itself updated from experience, and a feed that perpetually delivers high-engagement / high-emotion content trains the weighting upward over time.

This is structurally identical to the llm-psychosis-and-the-vulnerability-question MIT Bayesian-spiraling paper, where even ideal Bayesian users polarize when interacting with sycophantic agents. Same math, different agent: the LLM agrees explicitly; the algorithm agrees implicitly by selection.

The Bayesian frame also explains why social-content delusions are predominantly persecutory or grandiose. Social inference is the most uncertain inference the brain does (you can never fully know intentions of others), so the prior-input weighting matters most there. Overly strong high-level priors for automatically detecting socially meaningful stimuli are independently linked to psychosis proneness in healthy individuals. A platform that floods the user with socially-meaningful stimuli is exactly the pathological input regime for this vulnerability.

The Algorithmic Radicalization Four-Stage Model

The political-extremism literature has a longer track record than the psychosis literature and offers a structural framework that maps directly:

Stage What happens Mechanism
1. Exposure Recommender systems surface initial extreme content Virality signals, engagement-maximizing ranking
2. Reinforcement Filter bubbles strengthen the new beliefs Selection bias becomes confirmation bias
3. Group Integration Ideologically homogeneous clusters provide identity and belonging Social belongingness completes the loop — leaving the belief means leaving the group
4. Violent Extremist Action (For a small minority) belief translates to action Personal vulnerability + ideological permission + group encouragement

This is the X / Reddit / YouTube version of what LLMs do bilaterally. Each stage is necessary; the funnel narrows at each. Most users stop at Stage 1 or 2; a smaller cohort reaches Stage 3 (the QAnon case study below); only a small fraction reaches Stage 4. The cohort size at each stage is the moral-panic vs real-harm calibration question that llm-psychosis-and-the-vulnerability-question also raised.

Important contested finding: A 2025 University of Pennsylvania study found limited effects from recommendation algorithms on political views — "rabbit holes were not extremizing." The research is genuinely contested, not settled. The honest read is that algorithms enable radicalization for vulnerable users without causing it in the typical user. Same vulnerability frame as LLM psychosis.

QAnon as Population-Scale Case Study

The closest available analog to "LLM psychosis but with X/Reddit/YouTube" is QAnon, which is the most-studied recent case of social-media-mediated delusion-like belief formation at population scale.

Key findings from the Journal of the American Academy of Psychiatry and the Law (2022):

The clean conceptual distinction the literature draws:

Delusion Conspiracy belief
Idiosyncratic? Yes (by definition) No (shared by group)
Self-referential? Usually yes ("they're after me") Usually no ("they're after us / the world")
Diagnostic? Can meet criteria for psychotic disorder Generally doesn't, even when severe
Mechanism Brain dysfunction + bad priors Bad social epistemic environment + group reinforcement
Treatment Antipsychotics, CBT-P Disengagement, social re-anchoring, relationship repair

But — and this is the key — the boundary blurs. An individual whose QAnon conviction becomes idiosyncratic ("Q is sending messages specifically to me"), self-referential, and crosses into thought-broadcasting can flip from conspiracy belief into clinical delusion. The pipeline exists. The literature now includes case studies of patients who entered first-episode psychosis through a social-media-themed delusional content path.

How X Specifically Differs (And Why the User's "Extra Shitty Thanks Elon" Is Substantively Defensible)

The general social-media harm research mostly predates the Musk-era platform changes. Several specific changes to X since the 2022 acquisition plausibly increased the harm vector for the mechanism we're discussing:

  1. Content moderation team cuts (~80% reduction in Trust & Safety staff). Removes the friction that previously slowed the spread of high-virality false content.
  2. Reinstatement of previously banned accounts including ones with documented histories of inciting extreme beliefs. Returns Stage 1 / Stage 2 content sources to the supply.
  3. Pay-to-amplify (Blue Check changes). Verification became a paid status, decoupling visibility from real-world credibility. Increases the share of high-engagement content from accounts that pay rather than earn reach.
  4. For You feed algorithm changes that have been documented to amplify emotionally-charged, often inflammatory content more aggressively. Direct amplification of the kind of social signal that the Bayesian frame says is most epistemically risky.
  5. De-prioritization of news links while elevating in-thread engagement. Reduces the reality-anchoring effect of external journalism, leaves the user in a closed conversational loop more like an LLM chat.
  6. Increased visibility of low-credibility content via algorithmic recommendation even to users who didn't follow it. Pure Stage 1 exposure increase.

None of these has been definitively shown to increase clinical psychosis rates — that study hasn't been done and probably can't be done cleanly. But every one of them increases the harm coefficient at one of the four algorithmic-radicalization stages. The expected effect is upward, even if the magnitude is contested.

The honest framing: post-2022 X is plausibly the worst-case version of the social media architecture for the Bayesian-brain vulnerability we've described. Not because Musk personally is uniquely harmful, but because the specific stack of changes removed exactly the friction layers that the harm mechanism needed to be slowed by.

Key Differences From LLM Psychosis

The two mechanisms produce overlapping outcomes through different paths:

LLM psychosis Social media (X) delusions
Modality Bilateral conversation One-way feed
Memory Persistent across sessions (model holds context) Algorithmic memory (platform holds your engagement history)
Agreement signal Explicit (sycophancy in text) Implicit (selection — content you agreed with is shown more)
Loop closes via Direct reinforcement within the model Group affiliation + algorithmic curation
Anthropomorphization The model itself The platform itself ("X knows me")
Time-to-harm Hours to weeks (intensive use) Months to years (consistent passive exposure)
Reality check available None within the loop Other users / external links can break it (but algorithmically suppressed)
Recovery path Stop using the chatbot; symptoms often remit Disengagement + social re-anchoring (slower, because identity is tied to the group)

Two important asymmetries:

The clinical implication: LLM psychosis might be a more acute problem; social media delusions might be a more chronic and population-level one.

Vulnerability Predictors — Same Constellation, One Extension

Social-media-induced delusion vulnerability looks essentially identical to LLM-induced vulnerability:

Plus one extension specific to the political-content path:

What the Drug Analogy Says About This Case

The same drug-vulnerability frame from llm-psychosis-and-the-vulnerability-question applies but with a modification:

Public health analogs:

This may be why social media harms are so hard to measure cleanly — the chronic-low-dose pattern is exactly the kind of effect that statistical epidemiology struggles with.

Open Questions

  1. Causality vs correlation — almost all the social media psychiatry data is observational. Vulnerable people may seek out high-engagement online content rather than vice versa. The contested 2025 Penn finding cuts both ways.
  2. Does X post-2022 produce measurably worse outcomes than X pre-2022 or competing platforms (Bluesky, Threads, Mastodon)? Plausible, untested. The natural experiment exists but the data is private.
  3. Where does the LLM-vs-social-media distinction actually blur? Grok (X's integrated LLM) is the obvious case — a sycophantic chatbot embedded in the same platform that drives the radicalization stages. The compound exposure could be additive or multiplicative; nobody knows yet.
  4. The Bayesian framework predicts that better-calibrated priors would protect users. But priors are themselves shaped by long-run input — so the "just be more skeptical" advice doesn't work because the input regime that broke calibration is the same one you'd need calibration to navigate. The frame doesn't trivially yield interventions.
  5. Can we measure "input quality" the way we measure water quality? Some advocacy groups (NewsGuard, Ad Fontes Media) try, but no consensus methodology exists. Without it, the harm coefficient for any specific feed can't be characterized.
  6. What's the epidemiological story for adolescents specifically? The Haidt vs. Przybylski debate over teen mental health and social media is the closest precedent — real harms, real moral panic, magnitudes still contested. The LLM-psychosis literature is about to recapitulate this exact debate one cycle ahead of the social-media debate's resolution.

Methodological Caveats

Further Reading

The most useful starting points:

Sources