Prompts for Public-Sector LLMs Should Be Governed as Commons
This paper argues that prompts for public-sector large language models should be treated as governed commons rather than private inputs, proposing a "Prompt Commons" framework with versioned repositories and stakeholder negotiation mechanisms to enhance transparency, contestability, and auditability in AI deployment.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are running a city council meeting where a very smart, but slightly opinionated, robot assistant is helping you draft decisions. You might think the robot's personality is fixed because it was built by a big tech company. But this paper argues that how you talk to the robot (the "prompt") is actually the most important part of the conversation.
Here is a simple breakdown of the paper's main ideas, using everyday analogies.
1. The Problem: The "Hidden Recipe"
Think of the Large Language Model (LLM) as a giant, high-powered kitchen. The ingredients (data) and the chef's training (the model's code) are already set.
However, the paper argues that the recipe you give the chef is what actually determines the meal.
- If you tell the chef, "Make a spicy dish for a budget party," you get one result.
- If you tell them, "Make a healthy, expensive dish for a VIP," you get a totally different result.
In public government jobs, these "recipes" (prompts) are often written by one person, passed around informally, or kept secret. If a recipe has a hidden bias (like assuming everyone drives a car), the robot will make decisions that hurt people who take the bus. Because the recipe is hidden, no one can see the bias, argue about it, or fix it.
2. The Solution: The "Community Cookbook" (Prompt Commons)
The author proposes a new system called Prompt Commons. Imagine a public, community-run library of recipes instead of a secret notebook in a single chef's pocket.
This library has three special rules:
- Version Control: Every recipe has a version number and a history log. If someone changes an ingredient, everyone knows exactly when and why.
- Provenance (The "Who"): Every recipe must say who wrote it. Was it written by a senior citizen? A disability advocate? A neighborhood group? This ensures we know whose voice is in the recipe.
- The "Veto" Button: If a specific group (like a local disability council) sees a recipe that harms their community, they can hit a "pause" button. The recipe is locked in a quarantine zone until the issue is discussed and fixed.
3. How It Works in Practice: The "City Pilot"
The researchers tested this idea in a large North American city. They gathered 443 different people from various community groups (seniors, immigrants, LGBTQ+ groups, etc.) to write their own "recipes" for how the city robot should talk about streets and parks.
They turned these into a dataset of over 3,000 variations and tested three different ways of managing the library:
- Open: Anyone can add a recipe (like a public bulletin board).
- Curated: A team checks that every recipe has the right tags and covers different groups (like a strict editor).
- Veto-Enabled: The curated system, but with the added power for specific groups to block bad recipes.
The Results:
- When they used a single person's recipe, the robot was very "committed" to one side of an argument (e.g., always prioritizing cars).
- When they used the Community Cookbook (especially the Veto version), the robot's answers became more balanced. It was more likely to suggest "compromises" (e.g., "let's have a bike lane and a parking spot") rather than picking a winner.
- Most importantly, when a "bad" answer came out, the team with the Veto system fixed it much faster (in about 5 hours) compared to the open system (which took over 30 hours).
4. Why This Matters for Government
The paper suggests that governments should stop treating these prompts as temporary notes. Instead, they should treat them as official policy documents.
- Transparency: Just as a city must publish its budget, it should publish the "recipes" it uses to make decisions.
- Accountability: If a decision goes wrong, you can trace it back to the specific version of the recipe that caused it.
- Negotiation: Instead of one person deciding the robot's personality, different groups can submit their own versions, and the system can be designed to find a middle ground.
5. What the Paper Does Not Say
It is important to stick to what the paper actually claims:
- It does not say this system is perfect or that it solves all AI problems.
- It does not claim that the "compromise" answers are always the best moral choice; it just says they are more visible and easier to debate.
- It does not suggest that the robot's brain (the model) needs to be changed; the change is entirely in how we talk to it.
The Bottom Line
The paper argues that how we ask the question is just as important as the answer. By treating the "question" (the prompt) as a public, shared, and regulated resource, we can make AI in government more fair, transparent, and responsive to the people it serves. It turns a secret, one-person instruction into a public conversation.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.