The Anti-Slop Fiction Contest: How to Make AI Write Like It Isn't AI

When AI Writing Meets an Anti-Formula Challenge

The results of a contest called the "Unslop AI-Written Fiction Contest" were recently announced, sparking widespread discussion in the tech community. The premise of this competition is genuinely subversive — rather than showcasing how beautifully AI can write, it was designed specifically to combat a pervasive problem in AI-generated content known as "slop" (empty, formulaic filler).

"Unslop" represents a collective reckoning with the writing flaws endemic to today's large language models (LLMs). Anyone who has used ChatGPT, Claude, or similar tools for creative work will recognize the telltale signs: an excess of stacked adjectives, overused parallel structures, forced grand conclusions tacked onto endings, and a rotating cast of tired clichés. This contest set out to identify works that had successfully escaped that "AI voice."

What Is "Slop" in AI Writing

The Hallmarks of the AI Voice

To appreciate what this contest was trying to accomplish, you first need to understand what "AI slop" actually means. Experienced readers can now spot AI-generated text at a glance, because it exhibits a remarkably consistent set of patterns.

The most obvious tells include: heavy reliance on constructions like "not just… but also…", opening sentences like "In this rapidly changing world…", and the obligatory philosophical summary jammed into the final paragraph. AI also has a fondness for certain words — "tapestry," "delve," "testament" — terms that appear rarely in everyday human writing but have become verbal tics of large language models.

Why Does AI Writing Become "Slop"?

This tendency toward slop is rooted in how large language models are trained. At their core, LLMs are probability-prediction systems built on the Transformer architecture. When generating each token, the model computes a probability distribution across its entire vocabulary and samples from that distribution. This mechanism naturally biases models toward high-frequency, "safe" expressions — because those appear most often in training data and carry the highest probability weights. The temperature parameter controls how sharp or flat this distribution is: as temperature approaches zero, the model almost always picks the highest-probability token, producing extremely conservative output; as temperature rises, lower-probability tokens get more chances to appear, making the text more creative but also less stable. This is the mathematical root of AI writing being "safe but dull."

The RLHF (Reinforcement Learning from Human Feedback) process introduces another layer of systematic bias that further reinforces this homogeneity. RLHF typically works in three stages: first, fine-tuning the base model with supervised learning; second, training a "reward model" to predict human preferences; third, using reinforcement learning to maximize the reward score on model outputs. The problem is that human annotators, when faced with subjective creative judgments, tend to rate responses higher if they are well-structured, politely worded, and comprehensive — rather than rewarding genuinely inventive but potentially "strange" expressions. This bias is amplified across tens of thousands of iterations, ultimately training models to internalize a writing template optimized for pleasing evaluators. In AI alignment research, this phenomenon is considered a variant of "reward hacking."

The result: in its pursuit of "never being wrong," AI has sacrificed the distinctiveness and personality that literary writing prizes most.

The Human-AI Creative Philosophy Behind the Contest

AI Creation Under Human Guidance

The value of this contest lies not in crowning "the best AI," but in exploring how humans can actually harness AI for genuine creative work. Participants had to guide AI out of its comfort zone through carefully engineered prompts, multiple rounds of iterative refinement, and even line-by-line manual intervention — producing content with real literary merit.

This represents a mature paradigm for AI-assisted creation: AI is not a replacement for human writers, but a "paintbrush" that requires careful handling by the artist. True creators actively push back against a model's default tendencies, injecting their own aesthetic judgment and stylistic intentions.

Redefining Quality Standards for AIGC

From an industry perspective, the emergence of competitions like this signals an evolution in how people evaluate AI-generated content (AIGC). Early on, people marveled that "AI can actually write." Now attention has shifted to "Can AI write well — and write distinctively?"

Behind this shift lies a sobering ecological reality. Research institutions have found that since ChatGPT ignited the AI writing wave in 2022, detectable AI-generated content online has grown exponentially, creating a "content inflation" effect: when vast quantities of stylistically similar articles flood search results and social platforms, readers' attention is diluted and genuinely original content becomes harder to find. A deeper technical concern is the risk of "Model Collapse" — if future models are trained heavily on AI-generated content, they will gradually lose the diversity and nuance of human writing, falling into a degenerative cycle of self-replication.

As AI-generated content proliferates across the internet, "anti-slop" is becoming a new quality standard. Content platforms, publishers, and ordinary readers alike are developing aesthetic fatigue toward homogeneous machine-generated text, and turning instead toward writing that has a genuine soul.

Practical Implications for the Content Industry

The Next Frontier of Prompt Engineering

For content professionals, this contest offers highly valuable practical lessons. Prompt Engineering has evolved significantly since GPT-3's release in 2020: from simple zero-shot descriptions, to few-shot prompting with examples, to Chain-of-Thought techniques that ask models to show their reasoning step by step.

What "anti-slop" demands is the latest stage in this evolution — "negative constraint" prompt engineering. This means explicitly instructing the model to avoid clichés, specifying stylistic references, restricting high-frequency vocabulary, and even providing negative examples for the model to actively steer away from. This approach echoes Anthropic's "Constitutional AI" concept — shaping model behavior through explicit rules rather than pure reward signals, intervening at the root level of the model's probabilistic preferences.

This skill demands that practitioners know not only what they want, but precisely what they don't want — a higher cognitive bar for creators.

The Long-Term Trend of Human-AI Collaboration

As AI writing tools become ubiquitous, finding the right balance between efficiency and individuality will be a challenge every content creator must confront. The highest-quality content of the future will likely be neither purely human nor purely AI, but a deep fusion of human creativity and machine capability.

The "anti-slop" spirit is a reminder that no matter how powerful the technology becomes, what ultimately moves people is distinctiveness and authenticity.

Conclusion

The "Unslop AI-Written Fiction Contest" may be a small community event, but it reflects an important shift in values within the AI writing space — from pursuing "the ability to write" to pursuing "writing well," from accepting formulaic output to championing distinctiveness.

For everyone using AI in creative work, this is a signal worth sitting with: the real challenge isn't getting AI to speak, but getting it to say something genuinely valuable and individual. And that requires far deeper involvement and far more skillful guidance from human creators. With the dual pressures of model collapse risk and a deteriorating content ecosystem, "anti-slop" is no longer just an aesthetic pursuit — it is an active, deliberate effort to preserve the value of human creativity.