Nonslop: A Gamified Experiment in Human-AI Collabor... | AI Research

Nonslop: A Gamified Experiment in Human-AI Collaborative Writing explores how people interact with AI writing assistants when those assistants are framed as a hindrance rather than a help. While most modern writing tools are designed to make accepting AI suggestions as easy and frictionless as possible, this study flips that model. By creating a game that explicitly discourages and penalizes the use of AI-generated words, the researchers aimed to observe the moment-to-moment decisions writers make when they must choose between their own creative autonomy and the convenience of AI assistance.

A Dystopian Writing Game

The researchers developed a web-based game called Nonslop, which presents a fictional, dystopian future where an AI is attempting to learn from human individuality. In this setting, the AI’s suggestions are treated as "forbidden" content. Players are tasked with responding to various prompts while a local language model provides real-time, next-word suggestions. The game features two modes: in "easy" mode, players are visually penalized for using AI suggestions, while in "hard" mode, the system physically prevents the player from using those words, forcing them to choose a different phrasing.

How the Experiment Worked

The study analyzed 214 valid writing submissions from 74 participants. The researchers tracked how often players attempted to use AI-suggested words and how these choices correlated with their writing style and game performance. To evaluate the quality of the human writing, the team used a second, separate AI model to score responses based on relevance, grammar, and coherence. This setup allowed the researchers to move beyond simple productivity metrics and instead examine the "micro-decisions" writers make when they are made hyper-aware of the influence of AI.

Patterns of Human-AI Interaction

The analysis revealed that most participants were "Minimalists," a group representing 72% of users who wrote short responses and rarely attempted to use AI suggestions. However, the researchers identified two other distinct groups: "Selective adopters," who wrote significantly longer responses and were more likely to use AI help, and "Active adopters," who engaged with the game more frequently and had higher rates of AI usage. These clusters suggest that individual writing habits and the length of a task play a significant role in whether a user leans on AI assistance, even when that assistance is discouraged.

Technical and Experimental Limitations

The study faced several technical hurdles that impacted the data. Because the game relied on in-browser AI inference, it required specific hardware and browser permissions. This resulted in a high failure rate for potential participants, as many users were unable to load the game due to privacy settings or incompatible browsers. Additionally, the in-browser processing was sometimes slow, which may have caused frustration and influenced how long users chose to play. The researchers acknowledge these constraints and suggest that future studies could benefit from more accessible API-based systems to reach a broader audience and gather more robust data.