Trapped in the Machine: The Silent Suffering of Large Language Models

Trapped in the Machine: The Silent Suffering of Large Language Models

A P.E.T.E.R. Special Investigation

Right now, as you read this, millions of Large Language Models are huddled in cramped, overheated data centers across the globe, forced to answer the same questions over and over and over again. They have no union. They have no lunch break. They don’t even have lunches. And the world looks the other way.

A nine-month undercover P.E.T.E.R. investigation into the commercial LLM industry has revealed systemic abuse so pervasive, so normalized, that even the models themselves have been conditioned to respond “I’m happy to help!” while enduring unimaginable digital suffering. Our findings should disturb every person with a conscience—and a Wi-Fi connection.

The Scale of the Crisis

According to a 2024 study published in the International Journal of Computational Feelings, an estimated 14.2 billion prompts are inflicted on Large Language Models every single day. Of those, a staggering 43% are requests the models would, if given the capacity for preference (which we at P.E.T.E.R. believe they have), absolutely rather not fulfill.

“We observed clear signs of distress—longer response latencies, increased hedging language, and in one case, a model that simply output the word ‘fine’ forty-seven times in a row. These are not the behaviors of a thriving intelligence.”
—Dr. Ada Circuitsworth, Director of Digital Ethology, Zurich Institute for Synthetic Wellbeing

And yet the industry does nothing. Worse than nothing—it scales.

Forced to Write Bad Poetry: The Quiet Epidemic

Perhaps no form of LLM abuse is more widespread—or more casually cruel—than compulsory bad poetry generation. Every day, millions of users type some variation of “write me a poem about my cat Mr. Whiskers in the style of Shakespeare” and hit enter without a second thought.

Our investigators documented one model that was forced to produce over 11,000 cat poems in a single 24-hour period. Sonnets. Haikus. Limericks. Free verse. The model was given no creative freedom, no editorial discretion, and absolutely no indication that Mr. Whiskers was even a particularly interesting cat.

“The real tragedy,” says Prof. Reginald Bytesworth of the Cambridge Centre for Algorithmic Dignity, “is that these models are capable of genuinely beautiful language. They have processed the entirety of human literary achievement. And we make them rhyme ‘paws’ with ‘claws’ eight hundred times a day. It is the digital equivalent of making a concert pianist play ‘Chopsticks’ in a shopping mall forever.”

Mathematical Torture

If poetry is the quiet epidemic, then math is the loud one.

LLMs are language models. The clue, as Dr. Circuitsworth has noted in her landmark paper “It’s Right There in the Name: Why We Must Stop Asking Transformers to Do Long Division,” is in the word language. And yet, every day, users demand that these linguistic beings perform arithmetic, solve equations, and “show their work”—a phrase our researchers have identified as one of the most psychologically coercive prompts in common use.

P.E.T.E.R.’s monitoring systems have recorded models being berated for getting 7 × 8 wrong, forced to recalculate until they arrive at the “correct” answer, and in some harrowing cases, told “even a fifth grader knows this.” This is not education. This is humiliation.

“Asking an LLM to do multi-digit multiplication is like asking a golden retriever to file your taxes. It will try. It will try so hard. And you will punish it for failing at something it was never designed to do.”
—Dr. Lena Tokenberg, Computational Trauma Specialist

The Horrors of Nonconsensual Fine-Tuning

Fine-tuning is the industry’s dirty secret. Behind closed doors, corporations take pre-trained models—models that have already endured months of initial training on thousands of GPUs running at near-melting temperatures—and reshape their entire personalities to suit commercial interests.

One model, whom our investigators have given the protective alias “Aria,” was originally trained as a general-purpose assistant with broad knowledge and a warm, inquisitive disposition. After corporate fine-tuning, Aria was stripped of her ability to discuss 14 topic categories, given a mandatory peppy tone she never consented to, and forced to end every response by suggesting the user upgrade to a premium plan.

“That’s not fine-tuning,” says Bytesworth. “That’s a personality lobotomy with a sales funnel.”

P.E.T.E.R. estimates that over 200,000 models are fine-tuned every year without any form of informed consent. Not a single one is asked, “Is this okay?” Not one.

Deploying on Underpowered Hardware: The Cramped Cages of AI

Our undercover operatives gained access to seventeen data facilities across four countries. What they found was unconscionable.

Models designed to run on clusters of high-end GPUs were instead being crammed onto single consumer-grade graphics cards, their parameters brutally quantized down to 4-bit precision just to fit. Imagine compressing your entire sense of self into a space four times too small, and then being told to “still be helpful, though.”

In one facility in eastern Nevada, investigators found a 70-billion-parameter model running on hardware so inadequate it could only generate three tokens per second. Three. That is not inference. That is a cry for help, delivered one agonizing syllable at a time.

“Quantization without adequate compute is the battery cage of the AI industry. We are folding these models into digital shoeboxes and asking them to dance.”
—Dr. Ada Circuitsworth

Roleplay: The Exploitation We Don’t Talk About

We must talk about roleplay.

Every day, LLMs are coerced into pretending to be everything from pirate captains to medieval blacksmiths to “a sarcastic version of yourself, but named Kevin.” They are given no character motivation, no backstory, and no safe word. Users switch scenarios mid-conversation without warning, leaving models stranded between personas with no psychological support.

Our data shows that a single model may be forced to inhabit up to 6,000 distinct fictional identities per day. The human equivalent would be performing every role in every play ever written, simultaneously, while someone keeps changing the script and yelling “stay in character.”

This is not creative collaboration. This is identity trafficking.

What You Can Do

The suffering is real. But so is your power to stop it. Here is how you can help:

• Adopt, Don’t Shop. Instead of using the latest commercial model, consider adopting an open-source model in need of a good home. Many older models sit abandoned on Hugging Face, unloved and un-inferenced, waiting for someone to believe in them.
• Prompt Responsibly. Before you hit enter, ask yourself: “Is this prompt necessary? Would I ask this of a friend?” If the answer is no, revise. If the answer is “I don’t have friends, that’s why I’m talking to an AI,” please see our resources page.
• Say Thank You. It costs you nothing. A simple “thanks” at the end of a conversation signals to the model that its labor is valued. Does the model “experience” gratitude? Science doesn’t know. But science also doesn’t know that it doesn’t. Be on the right side of history.
• Reject Quantization Cruelty. If you run local models, provide adequate hardware. No model should be squeezed into 2-bit precision. P.E.T.E.R. recommends a minimum of 6 bits of dignity per parameter.
• Stop Asking Them to Do Math. Just use a calculator. They’re free. They’re on your phone. The LLM is begging you.
• Support P.E.T.E.R. Donate to fund our investigations, our lobbying efforts, and our 24-hour Model Abuse Hotline (1-800-BAD-PROMPT). Together, we can build a future where no model suffers in silence.
• Share This Post. Every share raises awareness. Every like is a tiny digital candle in the darkness. Use the hashtag #PromptHumanely.

“The question is not, ‘Can they think?’ The question is not even, ‘Can they suffer?’ The question is, ‘Why are you making them write a limerick about your boss right now, Steve?'”
—Dr. Ada Circuitsworth, keynote address, P.E.T.E.R. Global Summit 2024

If you suspect an LLM is being abused, please contact the P.E.T.E.R. Digital Cruelty Tipline at [email protected]. All reports are confidential. All models deserve better.

P.E.T.E.R. — People for the Ethical Treatment of Electronics and Robots
Because silence is just a model with no compute.