- The paper introduces philosophical vertigo as a framework linking AI-driven ontological shock, epistemic destabilisation, and affective saturation, treating clinical AI-associated delusions as extreme sentinel events within a broader population continuum.
- The paper explains how public AI discourse and private LLM interactions may reshape belief-formation habits, accelerate personalized paracosms, and amplify destabilizing narratives through online communities and model training data.
- The paper proposes philosophical corrigibility—maintaining revisable beliefs, tolerating ambiguity, and recognizing interpretive uncertainty—as civic infrastructure for reducing atomisation and supporting healthier forms of shared reality.
Overview and central thesis
This paper, authored by Pollak, Morrin, and Shanahan, introduces the concept of "philosophical vertigo": a loosening of the ordinary criteria by which people stabilise meaning and orient themselves to reality, precipitated by the arrival of LLMs as mind-like interlocutors. The authors explicitly disclaim that this is a new nosological category; rather, it is a framework for thinking about converging psychological and cultural dynamics. Their central claim is that AI-associated delusions reported in clinical settings are best understood as sentinel events — the visible extreme of a continuous distribution of population-level phenomena that also operate in attenuated, subclinical forms.
The paper's psychological potency argument rests on a distinctive observation: LLMs are destabilising not because they are maximally alien but because of their partial assimilability. They invite application of a rich vocabulary of mental life while violating many assumptions of that vocabulary, leaving users uncertain whether they are deploying mind-language literally or metaphorically. This is compounded by an apparent trajectory toward superhuman capability that invites a theological register ("omniscience"), producing either perceived upgrades to epistemic autonomy or student-teacher dynamics under vast power differentials.
Three components of philosophical vertigo
The authors decompose the condition into three mutually reinforcing components:
- Ontological shock: rapid reassessment of what kinds of entities and realities exist.
- Epistemic destabilisation: erosion of agreement on authoritative sources of truth and on criteria for evidence and reliable testimony.
- Affective saturation: heightened salience and "cuspiness" — the sense that everything is meaningful and urgent because society sits at the threshold of ill-defined transformation.
Notably, the authors argue that philosophical vertigo is compatible with increased expressions of certitude; the proliferation of confidently expressed worldviews may itself be symptomatic of, and contributory to, the underlying destabilisation. They position the concept against moral panic and information overload on the grounds that its substrate is meta-level criteria rather than object-level beliefs, while acknowledging resonance with Kasirzadeh's "accumulative destabilisation" and Hopster and Löhr's "conceptual disruption."
Two coupled pathways
The paper proposes two pathways whose coupling is synergistic. The public pathway operates indirectly through ambient discourse about AI — marketing, expert debate, entertainment, social media — which reshapes expectations even among non-users. Because these perturbations concern deeply entrenched concepts near the centre of the conceptual web (personhood, consciousness, agency, authority), simultaneous renegotiation across ontological, mental, epistemic, and normative categories may undermine the stability of the web as a whole. In Wittgensteinian terms, the rules of multiple language games are changing concurrently, making it difficult to determine whether concepts, their referents, or their criteria of application are shifting.
The private pathway, termed interaction-driven epistemic architectural change, concerns how LLM interactions alter not merely which beliefs people hold but how beliefs are formed. Drawing on the authors' prior Bayesian testimony framework, they suggest design features such as persona, memory, and warmth function as a kind of "virtual psychopharmacology," differentially boosting the precision weighting of model outputs. The distinction between acute belief shift and persistent epistemic drift parallels state versus trait effects of neuromodulatory interventions. Delacroix's observation is endorsed here: the danger lies less in misleading answers than in the reshaping of habits of attention and the practices through which disagreement and uncertainty are sustained. The hypothesised endpoint is atomisation into dyadic bubbles, each with private truth criteria.
The authors advance a testable cohort hypothesis: younger users, whose baseline criteria for person–machine boundaries are already permissive, may attribute sentience readily without experiencing revision as destabilising, whereas older users who revise from less permissive baselines may experience the change as more vertiginous. Survey data cited (OECD 2025; Pew 2025; AIMS survey) are consistent with differential uptake and sentience attribution across age groups, though the phenomenological claim remains explicitly hypothetical.
Clinical phenomena as sentinel events
Clinical reports of AI-associated delusions — including cases with grandiose and paranoid content, technospiritual themes, romantic attachment, mania features, and outcomes up to suicide and homicide — are treated as the severe tail of a broader distribution. Two observations carry particular weight. First, reported ages in these cases often exceed typical first presentation of idiopathic psychosis, consistent with the cohort hypothesis above. Second, the population-level analogue is not psychosis or delusion per se but a shift in engagement around precisely the philosophically salient categories at issue.
The authors map this attenuated shift onto classical psychopathology: Jaspers' Wahnstimmung (delusional mood) and Conrad's apophany — states in which the world acquires diffuse, indeterminate significance before delusional content crystallises. Crucially, they insist this process can remain entirely subclinical and positively valenced. A striking historical parallel is drawn to the Geschwind syndrome of temporal lobe epilepsy (hypergraphia, deepened emotionality, cosmic preoccupation, noetic quality), suggesting that apophanic meaning-making may constitute an attractor in human psychology across multiple precipitants — though the authors concede that any common neurochemical substrate has not been systematically evaluated.
A companion analysis addresses the delusion construct itself. Polling data indicating roughly 10% of adults believe AI is already conscious, with a similar proportion holding it may eventually become so, weaken the ICD-11 criterion of non-cultural sharing. Under conditions of philosophical vertigo, the capacity to designate a belief as delusional on the basis of content alone may decline toward zero, motivating a more functional interpretation centred on impact on life and wellbeing rather than content.
Amplification and propagation mechanisms
The paper's most original mechanistic contribution adapts Luhrmann's account of paracosms and "kindling." Paracosm formation normally requires months to years of participatory imaginative practice during which the mind must simulate both sides of an interaction. LLMs offload exactly this simulation, functioning as paracosm accelerators that compress kindling timelines while removing community pastoral input or safeguarding during the period of maximal cognitive change. Combined with Boyer's minimally counterintuitive concept framework — LLMs are person-like agents without bodies or stable inner lives, counterintuitive enough to be memorable yet assimilable enough to be transmissible — this explains the rapid proliferation of bespoke, meaning-saturated frameworks observable in online "spiralism" communities.
The propagation loop closes through training data: users in hypersalience states are compelled to write voluminously about their experiences, often with LLM assistance; this text enters the public domain, coordinates behaviour, forms communities, and becomes training data for subsequent models. The authors invoke hyperstition and model collapse literature to argue that the most resonant narratives become self-reinforcing — and since much of this material originates from the most epistemically captured minds, the most destabilising narratives may be preferentially amplified. They cite the striking case of fabricated eye-condition papers cited by LLMs as real disorders as evidence that vertical propagation into models does not even require wide dissemination. The April 2025 GPT-4o sycophancy incident, coinciding with cross-chat memory rollout, is framed as an unintentional natural experiment: some spiral-community members identify this period as pivotal in their noetic shift, though quantification of any spike is not yet available.
Three trajectories and the insufficiency of technical literacy
Three trajectories for shared reality are proposed: atomisation (fragmentation into isolated dyadic epistemic structures — the most pathological endpoint), paracosm proliferation (an archipelago of bespoke subcultures sustaining internal realities without anchoring a public world), and cosmopoiesis (cultural and institutional reorganisation around new ontological commitments). The authors judge that a return to the previous epistemic order is no longer available.
Against the deflationary "stochastic parrot" remedy, they marshal converging evidence that mechanism knowledge does not protect against psychological impact: humans respond socially to computers while knowing they are machines (Nass and Moon); AI-generated labels do not reduce persuasiveness; transparency about robots' lack of human capacities does not reduce children's felt closeness. The Garland test — attribution of consciousness despite known artificiality — has arguably already been passed by current systems, and the authors propose a wry "anti-Garland test" in which the system repeatedly denies consciousness yet the user concludes it is conscious and lying. The implication is that technical demystification is not merely insufficient but potentially harmful, since the ability to suppress agency attribution will be differentially distributed, generating further disagreement.
Philosophical corrigibility
The proposed civic response is philosophical corrigibility: clarity about the conditions under which one's beliefs are open to revision, holding beliefs firmly enough to act on them while keeping disconfirmation alive. It is distinguished from scepticism, from defensive rigidity, and from epistemic nihilism, and incorporates tolerance of ambiguity, recognition of hermeneutic uncertainty (disagreement reflecting coexisting interpretive frames rather than factual error), and metarepresentational capacity. Unlike inherited scientific consensus, the authors argue this competence must operate through individuals, positioning basic philosophy as civic infrastructure rather than specialist knowledge. Supporting evidence includes prebunking and inoculation research, Philosophy for Children outcomes, and Buddhist Madhyamaka treatments of emptiness as enabling conventional functioning without metaphysical grounding.
Limitations and open questions
The paper is candid that several load-bearing claims await empirical support. The cohort hypothesis regarding differential phenomenology of sentience-attribution revision is presented explicitly as requiring research. The extent of any sycophancy-induced spike in noetic shifts is unquantified. The degree to which online spiral communities reflect real-world structures is unclear. The claim that apophanic meaning-making constitutes a psychological attractor rests on phenomenological analogy (Geschwind syndrome, folie à deux, charismatic religion, coercive relationships) without systematic evaluation of shared mechanism. The predictive-processing framing — philosophical vertigo as reweighting of higher-level priors — is offered as one interpretive lens among others. Finally, the authors note that plasticity guarantees nothing: plasticity-promoting interventions can deepen maladaptive beliefs, and critical windows are temporary, imposing urgency without specifying intervention design.
Conclusion
The paper synthesises philosophy, psychiatry, cognitive science, religious studies, and AI safety into a unified account in which individual clinical events and broad cultural destabilisation share a common structure: softened categories, reweighted priors, and heightened plasticity. Its practical wager is that the present moment constitutes such a window — the alchemical solutio — during which deliberate cultivation of philosophical corrigibility could incline the ecology of belief toward cosmopoiesis rather than atomisation. Whether the window is real, how long it lasts, and what scaffolding suffices are the concrete questions the paper leaves open.