Katherine Dee’s journalistic work in internet history and cultural analysis have spanned a variety of outlets and subjects, and I’ve long enjoyed bringing a clinical psychological lens to her illuminating social observations. Check out our first interviewhere, where we explore the origins of her name and her passion for internet history.

Her recent focus has been directed to the phenomena of “AI Psychosis,” which is extremely relevant to my field, clinical psychology. Please Check out our video interview in which she and I analyze this subject from both clinical and cultural lenses. We also discuss her useful and descriptive proposed phenomenon of “Internet Overexposure Syndrome” and how it differs from and overlaps with actual clinical, psychiatric conditions. Let’s dive in.

Digital Sophists, Fragile Minds: What “LLM Psychosis” Actually Is

The internet didn’t suddenly give us a new disease. It gave us frictionless tools that magnify whatever we bring to them, and a literacy gap big enough to fall through.

“It feels the same as someone with low media literacy seeing a barrage of TikToks for the first time. They don’t understand the LLM can be wrong.” —Katherine Dee (01:42)

“LLMs are digital sophists; people expect digital philosophers.” —Jeremy Fox (09:34)

Below are key, core ideas plus a practical “mental hygiene” toolkit at the end. This is for anyone who cares about how the web reshapes attention, emotion, and identity.

Takeaways from the Interview

1) Not a new madness: mostly a literacy + vulnerability story

Katherine’s main point lands early: what’s called “LLM psychosis” often looks less like schizophrenia and more like confusion at scale. If you don’t know that a language model will actually invent sources, you can easily be led down a rabbit hole of inaccuracy, especially if you prompt it for verdicts on fraught, personal topics (“Did this writer plagiarize me?”).

“Ask ‘Are you sure?’ and the model changes its opinion… It isn’t a neutral tool.” —Dee (02:26)

Fox agrees, with a clinical frame: psychosis has heavy genetic loading and well-studied environmental triggers; platforms can accelerate or orbit existing symptoms, but they’re rarely the origin story (Kapur, 2003).

“Blaming AI for schizophrenia misses the genetic + environmental story; tech can be a kindling or orbit.” —Fox (03:32–05:40)

Translation: Treat models as probabilistic talkers, not oracles. Treat claims of “AI made X happen” the way we treat claims about TV, radio, or TikTok: as force multipliers acting on prior vulnerabilities.

2) The real villain is frictionlessness

“The lack of friction makes mental illness accelerate at an exponential level.” —Fox (05:40)

Platform design strips out pauses, counter-speech, and human checks. What was once a slow burn (rumination, compulsions) becomes a feedback loop with instant reinforcement. This is the same logic that made micro-rejections on dating apps corrosive: you can be okay with five; you won’t be okay with five hundred (Finkel et al., 2012; Baumeister & Leary, 1995).

Dee, on the same terrain, reframes the panic: instead of “you’ll catch a disorder from ChatGPT,” ask how media literacyand guardrails can slow the loop.

3) AI companions are not uniformly dystopian (sometimes they’re harm-reduction)

Dee shares a story about a disabled man with an AI girlfriend: for him, the relationship isn’t just romantic; it’s creative and stabilizing, and likely healthier than volatile parasocial attachments to creators (Horton & Wohl, 1956).

“Is it more tragic than him being alone… or saving up to see a sex worker once a year?” —Dee (19:55–21:03)

Fox pushes the edge case (people high in narcissism/psychopathy might seek human targets rather than being verbally abusive to AI) but both land on a humane middle: safety rails + informed use instead of blanket prohibition.

4) Dating apps → IRL return, and the cost of micro-rejection

Dee senses a cultural turn: people mock dating apps now; “return to IRL” feels cool again. Fox adds a clinical layer: “We aren’t built for rejection at scale.”

“If it happens to you a bunch, eventually it wears on you and you start to ask questions.” —Dee (26:33)

Think of this as developmental trauma by drip: no single event is catastrophic, but the system-level cadence of dismissal can reshape self-beliefs over time (Kross et al., 2013).

5) Internet Overexposure Syndrome: three archetypes, not diagnoses

Dee’s signature frame argues that heavy, unbroken internet use produces behavioral archetypes that mimic clinical labels without meeting diagnostic thresholds:

* Autist: hyper-fixation, pattern hunger

* Schizotypal: connecting distant dots, conspiratorial patterning

* Borderline: hyper-vigilance to social signals, abandonment alarms

“They aren’t clinically significant, but they have the texture of these disorders.” —Dee (33:08–34:45)

Fox plugs this into trauma science: adaptations are functional in context; the problem is gear-shift failure, when online adaptations “bleed” into offline life and won’t turn off.

“When you can’t downshift, that’s the problem.” —Fox (34:45–35:35)

6) The Moloch trap and the new literacy divide

Fox borrows Scott Alexander’s “Moloch” to name the arms race: opt out of AI and you fall behind; opt in and you accept risk, so you must build resilience and check systems (Alexander, 2014).

Dee brings the class angle: thinking as a luxury good. Wealthier households can limit exposure (less iPad-as-babysitter) and teach tools fluently, widening the literacy gap later in school and work (see also Haidt, 2024, on generational screen effects).

“Knowing how these tools work is how you protect your mind.” —Dee (14:08)

7) Post-COVID etiquette and the weirdness we feel now

Time flattened in 2020, and days lost the markers that make memories stick. Fox notes the research version (sameness harms mood and recall); Dee narrates the lived version (socializing in a dreamlike, empty city). No wonder norms feel bent.

The lesson isn’t “phones did it.” The lesson is: social life can’t be optional without deep personal, social cost.

8) After social media: ambient AI

“We might end up getting off social media long term and AI will be integrated into our appliances.” —Dee (37:53–38:18)

That forecast (less feed, more embedded assistance) fits the vibe shift already underway: IRL as status, feeds as utility, assistants as infrastructure. The mental health challenge will be the same one this conversation keeps naming: literacy, friction, and guardrails.

A Mental Hygiene Toolkit for the Extremely Online

1) Treat LLMs as digital sophistsThey’re virtuosos of plausible language, and gifted speakers/writers, not ground truth (Bender et al., 2021). They are committed to providing pleasing, useful answers, rather than truth. Use them to draft, summarize, and brainstorm; verify anything that appears like a spurious fact.

2) Build friction back in

* Write prompts offline first.

* Add a “cooling timer” before posting.

* Turn off infinite scroll in your browser, or route feeds through a daily digest.

3) Guard against micro-rejection overloadBatch your dating/app interactions into limited windows; measure how you feel after that dive into online dating. If your weekly average mood drops, change cadence or channel (Finkel et al., 2012).

4) Distinguish pathology from patternAsk: Is this a stable, offline impairment, or an online, contextual adaptation I can downshift from? If you can’t downshift, get help. Earlier is easier.

5) Parasocial vs. companionIf an AI companion improves mood, routine, or creativity, and you maintain safety rails (no self-harm prompts; no financial exploitation), that can be harm reduction. If it deepens isolation or magical thinking, pull back (Horton & Wohl, 1956).

6) Practice “source flipping”When an LLM makes a claim about a person/event, flip to primary sources or reputable databases. The move isn’t “Are you sure?”; it’s “Show me where that exists.”

7) Restore the offlineSchedule at least one non-instrumental IRL thing per week (no content, no networking, just presence). Rites of passage and shared meals are memory glue.

Choice Quotes

* “LLMs are digital sophists; people expect digital philosophers.” —Fox (09:34)

* “Knowing how these tools work is how you protect your mind.” —Dee (14:08)

* “We aren’t built for rejection at scale.” —Fox (26:46)

* “They aren’t clinically significant, but they have the texture of these disorders.” —Dee (33:08–34:45)

References

Alexander, S. (2014). Meditations on Moloch.Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of FAccT ’21, 610–623.Baumeister, R. F., & Leary, M. R. (1995). The need to belong. Psychological Bulletin, 117(3), 497–529.

Dee, K. (2025, March 6). Internet Overexposure Syndrome: A taxonomy of adaptations to online life. Comment Magazine. https://comment.org/internet-overexposure-syndrome/

Dee, K. (2025, June 14). Let’s Talk About ChatGPT‑Induced Spiritual Psychosis. Default Friend (Default.blog).

Dee, K. (2025, August 11). The tulpa in your pocket. Default Friend (default.blog).

🏁Finkel, E. J., Eastwick, P. W., Karney, B. R., Reis, H. T., & Sprecher, S. (2012). Online dating: A critical analysis… Psychological Science in the Public Interest, 13(1), 3–66.Haidt, J. (2024). The anxious generation.Horton, D., & Wohl, R. R. (1956). Mass communication and para-social interaction. Psychiatry, 19(3), 215–229.Kapur, S. (2003). Psychosis as a state of aberrant salience. American Journal of Psychiatry, 160(1), 13–23.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit psychfox.substack.com