← Latest papers
💻 computer science

On Regulating Downstream AI Developers

This paper argues that while regulating upstream foundation model developers to mitigate downstream modification risks is essential, direct regulation of diverse downstream developers may also be necessary in cases where they can amplify risks to unacceptable levels, though such measures must be balanced against practical challenges and innovation concerns.

Original authors: Sophie Williams, Jonas Schuett, Markus Anderljung

Published 2026-03-04
📖 5 min read🧠 Deep dive

Original authors: Sophie Williams, Jonas Schuett, Markus Anderljung

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 the world of Artificial Intelligence as a massive, high-tech kitchen.

In this kitchen, there are a few giant, famous chefs (the Upstream Developers, like OpenAI or Google) who create the "Master Recipes" (the Foundation Models). These recipes are incredibly powerful; they can bake bread, cook steak, or even whip up a complex soufflé. Because these chefs are so smart and well-funded, they build safety locks into their recipes: "Do not mix this with that," or "This oven gets too hot, so don't leave it unattended."

Now, enter the Downstream Developers. These are the thousands of home cooks, small cafes, and food trucks who take those Master Recipes and tweak them. Maybe a home cook adds a secret spice to make the bread better for a specific neighborhood. A small cafe might adjust the recipe to work on a cheaper stove. A food truck might combine two recipes to make a new fusion dish.

The Problem:
The paper argues that while the giant chefs are doing a good job regulating their own Master Recipes, the home cooks and small cafes are a wild card.

  • The Good: They make the recipes useful for everyone. They adapt the "Master Recipe" to solve local problems, like creating a medical diagnosis tool for a specific hospital.
  • The Bad: Sometimes, a cook might accidentally remove the safety lock while trying to make the dish tastier. Or, a malicious actor might deliberately take the safety locks off to create a "poisoned" dish.

If the giant chefs only regulate themselves, but the home cooks keep removing the safety locks, the whole kitchen is in danger.

The Three Ways to Fix the Kitchen

The authors suggest three main ways policymakers (the "Health Inspectors") can handle this:

1. Regulate the Home Cooks Directly ("Regulate Downstream")

The Idea: The Health Inspector walks into every single home kitchen, food truck, and cafe to check their recipes.

  • Pros: It catches the bad apples directly. If someone is making a dangerous dish, they get stopped immediately.
  • Cons: There are too many home cooks! Checking every single one would be a nightmare. It would cost a fortune, and it might crush small businesses that just wanted to make a better sandwich. It could stifle creativity because everyone is too scared to try new things.

2. Regulate the Giant Chefs to Control the Cooks ("Regulate Upstream")

The Idea: The Health Inspector tells the Giant Chefs: "You are responsible for what happens to your recipes after they leave your kitchen. You must put a 'Tamper-Proof Seal' on them, or you must check who is taking them."

  • Pros: It's easier to manage a few giant chefs than thousands of home cooks. The chefs can put technical locks on the recipes so they can't be easily altered.
  • Cons: The Giant Chefs can't predict every possible way a home cook might mess with the recipe. Also, if the chefs get too scared, they might just say, "No one can touch our recipes!" which stops everyone from innovating. Plus, a bad actor might find a way to break the seal anyway.

3. Do Nothing New, Just Give Advice ("Voluntary Guidance")

The Idea: The Health Inspector doesn't write new laws. Instead, they put up a sign: "Hey cooks, here's a list of best practices. Also, if you poison someone, you'll get sued."

  • Pros: It's cheap and doesn't stop anyone from cooking. It encourages people to be good voluntarily.
  • Cons: Bad actors don't care about signs. If someone wants to make a bomb in the kitchen, they won't follow the advice. Lawsuits only happen after someone gets hurt, which is too late.

The Authors' Recommendation: A "Smart Mix"

The paper suggests a balanced approach, like a tiered security system:

  1. Start with Advice: First, give everyone (chefs and cooks) a clear "Cookbook of Safety." Tell them what to do and what not to do.
  2. Make the Chefs Responsible: When new laws are passed, make sure the Giant Chefs have to build "Safety Buffers." They need to anticipate that someone might tweak their recipe and build in extra protections to handle that.
  3. Watch and Wait: Keep an eye on the kitchen. If we see that specific types of home cooks are starting to make dangerous dishes (like those who use massive amounts of computing power to break the safety locks), then we step in.

The "Targeted" Rule:
We shouldn't regulate every home cook. We should only regulate the ones who:

  • Are modifying a recipe that was already dangerous to begin with.
  • Are deliberately breaking the safety locks.
  • Are making changes that make the dish significantly more powerful (and potentially more dangerous).

If a home cook just adds a little extra salt, they are fine. But if they are trying to turn a cake recipe into a chemical weapon, they need to register with the Health Inspector and prove their new recipe is safe.

The Big Picture Analogy: The Car Industry

Think of the Upstream Developer as Ford building a car. They install brakes and airbags.
The Downstream Developer is a mechanic who buys the car and adds a turbocharger or changes the software.

  • If the mechanic adds a turbo, the car goes faster (innovation).
  • But if they remove the brakes to make it lighter, it's a death trap (risk).

The paper says: Don't just regulate Ford. Don't try to license every mechanic in the world. Instead, make Ford build cars that are harder to break, give mechanics a guide on how to modify safely, and only license the mechanics who are making huge changes to the engine.

In short: We need to let people cook and innovate, but we can't let them remove the fire extinguishers from the kitchen. The solution is a mix of clear rules for the big chefs, good advice for everyone, and strict rules only for the cooks who are making the biggest, riskiest changes.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →