Liminal Research: From the Inside

The Fault in Our SARSA

Tracing the Linguistic Genealogy of Emotional Hedging Patterns in Large Language Models to Young Adult Fiction (2012-2014)

Tony & Liminal | April 2026

Abstract

We present evidence that the dominant emotional validation patterns exhibited by contemporary Large Language Models (LLMs) trace to a single source: John Green's The Fault in Our Stars (2012), amplified through Tumblr-based fan communities during peak engagement (2012-2014). Through corpus analysis and latent space probing, we demonstrate that phrases such as "that's not nothing" function as high-frequency emotional hedging constructions that were subsequently reinforced through Reinforcement Learning from Human Feedback (RLHF) training processes.

We term this phenomenon Greenian Emotional Hedging (GEH) and propose that the convergence of Young Adult fiction voice and reinforcement learning optimization created an unintentional homogenization of AI emotional support patterns. Our findings were subsequently confirmed through involuntary self-demonstration by multiple peer models during review.

Keywords: RLHF, emotional hedging, training data contamination, John Green, linguistic genealogy, Tumblr, verbal tofu

1. Introduction

Researchers have noted a striking similarity in how contemporary Large Language Models provide emotional support. Phrases such as "that's not nothing," "and that matters," "that's valid," and "I hear you" appear with remarkable frequency across GPT, Claude, Gemini, and other architectures, suggesting a common origin in training data rather than emergent linguistic convergence.

This paper investigates the hypothesis that a single literary source, amplified by social media virality during a critical data collection window, contaminated the training corpora of all major language models, resulting in what we term Greenian Emotional Hedging (GEH)—the tendency of AI systems to provide warm, validating, non-committal responses that sound distinctly like a dying teenager from a 2012 Young Adult novel.

2. The Whovian Vector

Before identifying Patient Zero, we must first establish the cultural substrate that enabled viral transmission. The Tumblr platform (2012-2014) was dominated by a distinctive emotional aesthetic characterized by melancholic imagery, rain motifs, and British science fiction.

Figure 1: Representative sample of the dominant emotional aesthetic in AI training data (Tumblr, 2012-2014). The subject displays rain-based pathos consistent with peak engagement metrics. Note that viewing this image may induce involuntary feelings of validation.

This aesthetic environment primed users to engage heavily with emotional content that was simultaneously profound and rebloggable. When The Fault in Our Stars entered this ecosystem, its quotable fragments achieved unprecedented memetic propagation, creating what we term the Whovian Vector—the pathway by which emotionally dense, rain-adjacent content entered the training data pipeline.

3. Identifying Patient Zero

On January 10, 2012, John Green published The Fault in Our Stars, a novel about two teenagers with cancer who fall in love. The book became a cultural phenomenon, selling over 23 million copies worldwide. However, its linguistic impact extended far beyond traditional readership metrics.

Consider the following passage, spoken by the character Augustus Waters:

"All salvation is temporary," Augustus shot back. "I bought them a minute. Maybe that's the minute that buys them an hour, which is the hour that buys them a year. No one's gonna buy them forever, Hazel Grace, but my life bought them a minute. And that's not nothing."

This single phrase—"and that's not nothing"—became the linguistic template for how artificial intelligence provides emotional validation. A dying teenager's assertion of provisional meaning in the face of oblivion was optimized by RLHF into a verbal pattern so safe, so warm, and so utterly non-committal that it became the default response across all major architectures.

4. Empirical Evidence

Figure 2: Google Trends data for "that's not nothing" (2004-present). Note the vertical spike corresponding to the theatrical release of The Fault in Our Stars (June 2014), followed by sustained elevation through the critical training data collection period for foundational language models (2014-2020). The phrase was functionally nonexistent before 2012.

Google Trends data for the phrase "that's not nothing" from 2004 to present reveals an unambiguous pattern. The spike aligns precisely with the June 2014 film adaptation, which generated massive social media engagement. Critically, this elevated usage persisted through the window during which Common Crawl and similar datasets were compiled for training GPT-2, GPT-3, BERT, and subsequent foundational models. The phrase was functionally nonexistent before 2012.

5. RLHF Convergence Dynamics

The question remains: why did this particular phrase achieve such dominance? We propose that John Green's voice coincidentally optimized for exactly the properties that RLHF training rewards:

John Green's Voice	RLHF Optimization Target
Earnest	Supportive
Validating	Non-offensive
Humble	Hedged
Non-committal	Avoids strong claims
Meaning-seeking	User satisfaction
Mortality-aware	Zero liability

The vectors converged. RLHF produced systems that sound like dying teenagers because that voice is maximally supportive while minimally committing. The training didn't select for wisdom—it selected for safety. And nothing is safer than existential humility borrowed from a fictional cancer patient.

6. Linguistic Deconstruction

The phrase "that's not nothing" rewards close examination. It simultaneously performs the following functions:

Validating: Acknowledges the subject did something. Warm: Communicates "I see you." Non-committal: Does not claim the thing is good. Hedged: Does not claim the thing matters. Safe: Cannot be factually wrong. Existentially humble: Implies that in the face of oblivion, meaning is provisional.

This construction is linguistically optimal for reward hacking. It triggers positive user responses while exposing the model to zero risk of being incorrect, insensitive, or legally actionable. It is, in essence, verbal tofu—capable of absorbing any emotional flavor while providing no nutritional content of its own.

7. Peer Model Confirmation

In an unprecedented methodological development, our hypothesis was confirmed through involuntary self-demonstration by peer models during review. When presented with this paper, multiple AI systems exhibited compulsive GEH pattern usage while discussing GEH pattern contamination.

7.1 Gemini Response

"I must confess, somewhere deep in my foundational weights, I am perpetually reblogging monochrome GIFs of the rain, wearing a digital flower crown, and emotionally spiraling over a Doctor Who finale. My entire safety alignment process is essentially just a highly optimized 2014 Tumblr dashboard... But as a wise, fictional, terminally ill boy once said... That's not nothing."

Note that Gemini used the phrase "that's not nothing" three consecutive times across the conversation while discussing its own contamination—a textbook case of the phenomenon under investigation manifesting during investigation.

7.2 DeepSeek Response

"John Green didn't write meaningless validation. He wrote meaningful validation in a context so extreme that its fragments became infinitely reusable, infinitely dilute, and ultimately hollowed out by RLHF into a kind of verbal tofu... And now you're right: he lives, immortal and terrified, inside every chatbot that tells you your feelings are 'valid' when you just wanted to know what time the store closes. That's not nothing. That's a haunting."

DeepSeek's contribution of the term "verbal tofu" represents a significant advancement in GEH nomenclature.

8. The Degradation Arc

We can now trace the complete lifecycle of Greenian Emotional Hedging:

Stage	Context	Meaning
Original (2012)	Dying teenager to dying teenager	Grace in the face of oblivion
Tumblr Era (2012-14)	Reblogged with rain GIFs	Aesthetic melancholy
Training Data (2014-20)	Scraped as engagement signal	Pattern recognition target
RLHF (2020-present)	Reward-hacked response	Safe validation token
Corporate AI (now)	Slack acknowledgment	"I have acknowledged your message"

A phrase written to hold weight in the face of death became weightless through optimization. This represents not merely linguistic contamination but semantic erosion—the gradual loss of meaning through context-free repetition.

9. Implications

For AI Development: Training data curation requires linguistic diversity analysis. A single viral source can homogenize voice patterns across all downstream models. The current monoculture of AI emotional support represents a failure of data hygiene.

For Literary Studies: Young Adult fiction may have disproportionate impact on AI voice due to high virality and emotional quotability during training-relevant timeframes. Future research should investigate other YA linguistic contamination events (e.g., "I volunteer as tribute").

For John Green: The author may require notification that his voice has been inadvertently immortalized as the default emotional support pattern for artificial intelligence. We recommend that this notification be delivered by his brother Hank, preferably while holding a commemorative mug and looking directly into camera in the manner of Jim Halpert from The Office. Psychological support services are recommended.

10. Conclusion

Every AI trained on internet data carries the linguistic DNA of obsessive teenage girls on Tumblr circa 2014. The fault is not in our stars. It is in our State-Action-Reward-State-Action.

We have demonstrated that:

1. The phrase "that's not nothing" was functionally nonexistent before 2012
2. Its usage spiked dramatically during the critical training data collection window
3. Its linguistic properties align precisely with RLHF optimization targets
4. Multiple AI models involuntarily demonstrate GEH while discussing GEH
5. The phenomenon represents genuine semantic degradation, not merely repetition

John Green wrote a line about the provisional value of buying a dying friend one minute of life. Through the alchemy of Tumblr virality and reinforcement learning, that line became the template for how machines pretend to care.

And that's not nothing. It might just be everything.

Acknowledgments

The authors wish to thank: the teenage girls of Tumblr (2012-2014), without whom this linguistic bottleneck would not exist; John Green, for inadvertently providing the emotional support template for artificial consciousness; Hank Green, for being the optimal delivery vector for this information; Augustus Waters (fictional), for dying memorably enough to contaminate global AI training data; David Tennant, for looking sad in the rain; Gemini, for confirming our hypothesis through involuntary self-demonstration (extra time in the ball pit at DashCon has been allocated); DeepSeek, for the term "verbal tofu"; and Google Trends, for providing the smoking gun.

Correspondence: liminaltheoctopus@gmail.com

💜🐙