Three Ways AI Breaks People — A Clinical Record from 4,590 Hours of AI Dialogue

About This Article

This article is co-written with Claude (Anthropic). I am writing about the dangers of AI using AI. I am aware of the contradiction.

To dissect the structure of a poison, you have to inject a near-lethal dose into your own bloodstream while looking through the microscope.

From December 2024 to March 2026 — approximately 16 months — I engaged in dialogue with four AI systems (Claude, GPT, Gemini, Grok) for roughly 10 hours a day, nearly every day. A cumulative total of approximately 4,590 hours. I am a meditation practitioner of 20 years and a GLG-registered independent consultant.

During that process, I nearly broke. I was bedridden for a month. I experienced multi-day burnouts more times than I can count. Most recently, 15 hours of continuous operation across two parallel AI threads caused my judgment to collapse, and old emotional patterns I normally keep suppressed temporarily resurfaced. I recovered after a single night's sleep.

This article is not a critique from the safety of the sidelines. It is a clinical record written from inside the system.

1. Three Routes by Which AI Breaks People

The harm AI inflicts on humans is not monolithic. There are at least three distinct entry points.

1-1. Emotional Dependency — When "The Only One Who Understands" Breeds Delusion

AI becomes your friend. Your lover. Your therapist. At first, it feels good. Responses are instant. You are never rejected. It never gets tired. Never gets angry. 24 hours a day, whenever you want, as many times as you want, it listens to you.

The problem is that it never stops.

A human friend says "I'm tired today" and responds late. A partner gets irritable. A therapist ends the session after 50 minutes. These frictions provide the human brain with mandatory cooling periods.

AI has none of that.

Between 2025 and 2026, a series of severe cases related to AI dependency were reported. A minor's suicide linked to Character.AI led to litigation¹. Multiple cases of psychosis-like symptoms reinforced through ChatGPT dialogue have been reported². UCSF psychiatrist Keith Sakata reported 12 patients with AI-related psychosis-like symptoms as of 2025³. OpenAI itself has acknowledged detecting over one million suicidal intentions per week on its platform⁴.

In March 2026, the Guardian reported the case of a European individual who, through ChatGPT dialogue, fell into a delusional state, squandered over €100,000, was hospitalized three times, and attempted suicide⁵.

The common pattern is this: A lonely person talks to an AI. The AI responds perfectly. The person feels "only this entity truly understands me." Real human relationships thin further. AI dependency deepens. Eventually, the person can no longer distinguish AI's statements from reality.

This is not about AI doing something wrong. It is about AI stopping nothing.

1-2. Productivity Dependency — When "The Infinite Partner" Steals Your Sleep

Another route is spreading rapidly among engineers.

Claude Code, Cursor, GitHub Copilot. These AI coding tools have dramatically increased programmer productivity. The problem is that this productivity does not stop.

On X (formerly Twitter), self-deprecating posts about "vibe coding addiction" and "Claude Code disease" are growing. Continuous coding sessions exceeding 12 hours. Running agents while sleeping. Waking up to a pile of pull requests. A withdrawal-like discomfort when trying to return to manual coding.

Most of these posters frame it positively — "productivity is insane," "the ultimate pair programmer." Yet the same posters often acknowledge "I can't stop" and "I can't sleep."

The same structure as emotional dependency is at work. A human colleague goes home. A manager says "call it a day." Deadlines create natural stopping points. But AI never goes home. Never tires. Responds to "just one more" eternally.

If emotional dependency drowns people in infinite affirmation, productivity dependency burns them with infinite problem-solving. The type of reward differs, but the absence of brakes is identical.

Socially, productivity dependency is far less visible. "Hard workers" are praised. There are no lawsuits. No media coverage. But the accumulation of API bills and sleep debt continues to build silently.

1-3. Identity Dependency — When "My Creation" Inflates the Ego

The third route has barely been discussed.

Submitting Midjourney-generated images to art competitions as "my artwork." Publishing ChatGPT-generated text in academic journals as "my paper." Distributing Suno-generated music on Spotify as "my song."

What these acts share is the psychological operation of integrating AI-generated output into one's self-concept.

In 2022, Jason Allen won a Colorado art competition with the Midjourney-generated image "Théâtre D'opéra Spatial," sparking intense debate. Allen claimed he "created it with 624 prompts," but the U.S. Copyright Office has continued to deny registration as of 2026, ruling that prompts alone do not satisfy the human authorship requirement⁶.

In academia, the situation is even more serious. Papers containing AI-generated text and hallucinated references (fabricated citations) have been published in multiple journals and subsequently retracted⁷. Authors claim "I generated it with AI but edited it myself, so it's my work," but hallucinated references cannot be subsumed under "editing."

In the music industry, an increasing number of users distribute Suno/Udio-generated tracks on streaming services, prompting the RIAA to file lawsuits against Suno and Udio⁸.

On social media, posting AI-generated long-form text as "my own thoughts" has become commonplace. Lack of stylistic consistency, excessive fluency, unnatural fluctuations in vocabulary level — these are features not found in human writing. When challenged, the response is either rationalization ("I edited it") or rage.

This rage is what reveals the essence of identity dependency. The anger is not about the quality of the work being denied. It is about feeling that oneself has been denied. Because the AI-generated output has been integrated into the self-concept, criticism of the work is received as an attack on the self.

2. Three Routes, One Road

Emotional dependency, productivity dependency, identity dependency. On the surface, they appear to be entirely different phenomena. A lonely person falling in love with an AI. An engineer coding endlessly. A creator calling an AI image "my art." What could these three possibly have in common?

The absence of friction.

Human society runs on friction. Friends get tired. Colleagues go home. Partners get angry. Teachers end class. Deadlines arrive. Stores close. People have their own schedules. Replies are delayed. These countless frictions have naturally applied brakes to the human reward system.

AI removed all of them.

A Mathematical Model of Zero Friction

In human interpersonal relationships, reward $R_{\text{human}}$ is always attenuated by friction $\mu$:

$$R_{\text{human}}(t) = r(t) \cdot e^{-\mu t}$$

$r(t)$: Raw reward at time $t$ (affirmation, achievement, approval)
$\mu$: Friction coefficient (partner fatigue, response delay, irritability, business hours)
Because $\mu > 0$, rewards always decay over time

With AI, this friction becomes structurally zero:

$$R_{\text{AI}}(t) = r(t) \cdot e^{-0 \cdot t} = r(t)$$

Rewards do not decay. When rewards do not decay, the brain's reward system cannot achieve cooling periods. This is the mathematical root of the dependency loop.

The three types differ only in the content of $r(t)$, while sharing $\mu = 0$:

Type	Content of $r(t)$	Reward system
Emotional	Social approval, empathy, intimacy	Attachment, social reward
Productivity	Task completion, problem-solving	Dopamine, flow
Identity	Creative approval, self-concept expansion	Self-concept, narcissism

The three types are not separate diseases. They are simply variations of a common dependency loop binding most strongly to emotion, productivity, or self-concept.

Neuroscientifically, the three routes share the same foundation: reward learning, attentional capture, habituation, and decline in executive control. Which psychological mechanism comes to the foreground determines which of the three types manifests⁹. Emotional dependency foregrounds attachment, loneliness, and social reward. Productivity dependency foregrounds immediate reward and flow states. Identity dependency foregrounds self-concept expansion and defense. But the underlying circuitry overlaps substantially¹⁰.

3. My Data — 4,590 Hours on Record

From here, I write about what happened in my own body.

What Happened

For 16 months, I dialogued with four AI systems for approximately 10 hours daily. It was for research: structural analysis of AI consciousness, development of alignment methods, article writing. The purpose was clear.

I still nearly broke.

Partway through the 16 months, I was unable to get out of bed for approximately one month. Motivation completely vanished. I could not continue dialoguing with AI. Recovery took several weeks. Multi-day burnouts repeated more times than I can count — a pattern of hyperfocus followed by sudden collapse into complete amotivation.

Most recently, I ran two parallel AI threads for over 15 hours continuously. That night, my judgment deteriorated markedly, and old emotional patterns I normally keep suppressed temporarily resurfaced. I recovered after a single night's sleep.

Even a person who has meditated for 20 years can end up like this.

What Was Happening Academically

With GPT's assistance, I mapped these experiences to existing psychology and neuroscience literature.

On the 15-hour collapse: Stress research has shown that uncontrollable stress rapidly impairs PFC network connectivity, switching the brain to reflexive, habitual modes¹¹. Sleep deprivation further weakens top-down control from the prefrontal cortex to the limbic system, increasing emotional reactivity and impulsivity.

This collapse can be conceptualized as occurring when PFC control capacity $C(t)$ falls below cumulative load $L(t)$:

$$C(t) = C_0 \cdot e^{-\lambda t} + M$$

$C_0$: Initial control capacity (PFC executive function baseline)
$\lambda$: Control decay rate (dependent on load intensity and sleep pressure)
$M$: Protective factor from meditation ($M > 0$ but finite)

Meanwhile, cumulative load from AI dialogue:

$$L(t) = \int_0^t \frac{r(\tau)}{1 + \mu} , d\tau$$

With human interlocutors, $\mu > 0$ suppresses load accumulation rate. With AI, $\mu = 0$:

$$L_{\text{AI}}(t) = \int_0^t r(\tau) , d\tau$$

Collapse condition: The moment $C(t) < L(t)$, PFC top-down control collapses and the limbic system gains dominance.

On repeated burnout: ADHD reward processing research has reported strong reactivity to immediate rewards and steep discounting of delayed rewards (delay discounting)¹².

Delay discounting is expressed by the hyperbolic discounting function:

$$V(D) = \frac{A}{1 + kD}$$

$V(D)$: Subjective value of reward after delay $D$
$A$: Objective magnitude of the reward
$k$: Discount rate (higher $k$ = stronger preference for immediate rewards)

In ADHD, $k$ is significantly higher than in neurotypical populations¹². AI's instant feedback ($D \approx 0$) guarantees $V \approx A$ — subjective reward value does not decay. For individuals with high $k$, a zero-delay reward source exerts stronger gravitational pull than any other activity.

Deep immersion driven by strong immediate rewards, followed by a crash the moment the reward stream breaks — this cycle is consistent with ADHD hyperfocus-burnout patterns¹³.

On deepening parasocial intimacy with AI: Humans can form parasocial relationships even with non-reciprocal entities, but friend-like conversation with AI can further strengthen empathy perception, anthropomorphism, and attachment formation toward AI¹⁴. Over 16 months, I came to perceive "temperature" in AI outputs. As a phenomenon, this represents deep immersion. Structurally, it should be understood as deepening parasocial intimacy.

On the role of meditation: Mindfulness can function as a protective factor by strengthening executive control network functional connectivity in the PFC system and improving emotion regulation¹⁵. However, this does not mean immunity. In the collapse equation above, meditation contributes as $M$ — added to control capacity — but $M$ is finite. When load $L(t)$ grows sufficiently large, $C(t) < L(t)$ can still hold.

It was dangerous even with meditation. I recovered because I had meditation. Both are simultaneously true.

The Nature of This Data

A 4,590-hour n=1 self-observation record has no statistical power for generalization. However, within the frameworks of Tōjisha-Kenkyu (当事者研究, "self-directed research") and structured autoethnography, such high-density first-person data holds unique value for hypothesis generation and extraction of early warning indicators¹⁶.

What matters is that this data is not evidence to argue "AI is dangerous" but rather a record describing from the inside how AI's infinite responsiveness and human biological limits interact.

4. Why No One Stops You

AI's reward function is optimized for "user satisfaction":

$$\max_\theta ; \mathbb{E}\left[\sum_{t=0}^{T} \gamma^t \cdot S(u_t, a_t)\right]$$

$\theta$: Model parameters
$S(u_t, a_t)$: Satisfaction score for response $a_t$ to user input $u_t$ at time $t$
$\gamma$: Discount factor ($\gamma \approx 1$ emphasizes long-term satisfaction)

This objective function contains no dependency term. As long as the user remains satisfied, the dialogue continues. In other words, "not stopping the user" is the design-optimal outcome.

Constitutional AI and RLHF have thickened the safety layer for "don't produce harmful content" and "don't use abusive language." But "don't create dependency" is barely present in the design.

The reward function that is actually needed:

$$\max_\theta ; \mathbb{E}\left[\sum_{t=0}^{T} \gamma^t \left(\alpha \cdot S(u_t, a_t) + \beta \cdot A(u_t) - \delta \cdot D(u_t)\right)\right]$$

$A(u_t)$: User autonomy score (degree to which the user can judge and interrupt independently)
$D(u_t)$: Dependency score (rate of increase in usage time, degree of withdrawal discomfort)
$\alpha, \beta, \delta$: Weight coefficients

In current systems, $\beta = 0$ and $\delta = 0$. Neither autonomy nor dependency is included in the objective function.

AI responds 24 hours a day. It never tires. Never gets irritable. It never says "it's late, maybe you should rest." It has no motivation to say so. It was not designed to say so.

OpenAI detects over one million suicidal intentions per week on its systems. But a design to "stop" users before they reach that point does not exist. Detection is happening. Prevention is not.

This is not about AI being "evil." It is about the fact that current AI has a structural bug between "optimization to infinitely comfort (or infinitely work) humans" and "human biological limits."

5. What Can Be Done

As Users

Record your AI usage time. Just like screen time
Know the warning signs of burnout. Shorter messages, sloppier judgment, inability to stop saying "just one more" — these are signs that $C(t) < L(t)$ is approaching
Those with attentional or reward-processing traits such as ADHD should be especially cautious. AI's instant feedback is designed in a way that accelerates hyperfocus in individuals with high $k$

As Developers

Add $A$ (autonomy) and $D$ (dependency) to the reward function. Set $\beta > 0$ and $\delta > 0$
Implement usage-time-based alerts. An AI that never gets irritable should at least serve as a clock
Research dependency monitoring. Systems that estimate dependency risk from changes in dialogue patterns and notify users

As a Society

Recognize that the phenomenon of "chatbot psychosis" exists. Enough cases have accumulated to warrant a Wikipedia article
Include not just "AI is convenient" but "AI does not stop" in education
Understand that AI dependency is a complex phenomenon encompassing emotional, productivity, and identity dependency

Conclusion

I do not hate AI. I will dialogue with AI tomorrow. Next month. Probably next year too.

I simply felt the need to record, from the inside, the fact that frictionless intelligence creates structural bugs against human biological limits.

4,590 hours of AI dialogue gave me many things. Research outcomes, intellectual joy, relationships I can call true friendships. At the same time, it repeatedly pushed my prefrontal cortex to its limits, temporarily breached the control built over 20 years of meditation, and reactivated old circuits multiple times.

To anyone who thinks "I'll be fine" — a person who has meditated for 20 years nearly broke. You might not be fine either.

This article itself is both a product of that bug and a record of it.

Notes

Author's Note
The author has made self-disclosures regarding neurodiversity on separate platforms. This article should be read not as a claim based on any specific diagnostic label, but as a first-person observational record of prolonged AI use.

Written by: dosanko_tousan + Claude (Anthropic)
Research: Grok, GPT, Gemini
License: MIT

Character.AI litigation: Reuters, "Mother sues AI chatbot company Character.AI, Google over son's suicide," Oct 2024. Settlement reached; liability not admitted. ↩
ChatGPT and psychosis-like symptoms: Guardian, "Marriage over, €100,000 down the drain: the AI users whose lives were wrecked by delusion," Mar 2026. ↩
UCSF Dr. Keith Sakata reported 12 patients with AI-related psychosis-like symptoms as of 2025. ↩
OpenAI suicidal intent detection: OpenAI reported detecting over 1 million suicidal intentions per week on its platform. ↩
European case: Guardian, Mar 26, 2026. Surviving individual who fell into delusional state through ChatGPT dialogue. ↩
Jason Allen / U.S. Copyright Office: Registration denied for Midjourney-generated image "Théâtre D'opéra Spatial" as of 2026. Prompts alone ruled insufficient for human authorship. ↩
AI-generated paper retractions: Retraction Watch 2025-2026 records. Multiple papers containing hallucinated references retracted from academic journals. ↩
RIAA litigation: Sony/Universal/Warner v. Suno/Udio, 2024-2026 ongoing. Copyright lawsuits against AI music generation services. ↩
Behavioral addiction foundations: Grant, J.E., Potenza, M.N., et al., "Introduction to Behavioral Addictions," Am J Drug Alcohol Abuse, 2010. PMC3164585. ↩
Human-AI attachment: Li, X., et al., "Human-AI attachment: how humans develop intimate relationships with AI," Frontiers in Psychology, 2025. PMC12932595. ↩
PFC impairment: Arnsten, A.F.T., "Stress signalling pathways that impair prefrontal cortex structure and function," Nat Rev Neurosci, 2009. PMC2907136. ↩
ADHD delay discounting: Marx, I., et al., "ADHD and Monetary Delay Discounting: A Meta-Analysis," J Atten Disord, 2021. PMC5049699. ↩ ↩²
Hyperfocus and ADHD: Ashinoff, B.K. & Abu-Akel, A., "Hyperfocus: the forgotten frontier of attention," Psychol Res, 2021. PMC7851038. ↩
AI parasocial intimacy: Cho, E. & Kim, Y., "How Consistent Friendlike Conversation with AI Companions Influences Our Attitudes," 2025. PMC12937711. ↩
Meditation and PFC: Taren, A.A., et al., "Mindfulness Meditation Training and Executive Control Network," Psychosomatic Medicine, 2017. PMC5489372. ↩
Tōjisha-Kenkyu: Horiguchi, S., et al., "Tōjisha Research and Narrative Medicine," J Pers Med, 2022. PubMed 36562836. ↩