AI, Digital Platforms, and the New Systemic Risk
This paper critiques the narrow definitions of systemic risk in current EU legislation like the AI Act and DSA, proposing a comprehensive, multi-level framework drawn from finance and complex systems theory to better address emerging threats such as multi-agent failures, large-scale discrimination, and systematic hallucinations in AI-platform hybrid infrastructures.
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 building a city. In the past, if a single house caught fire, it was a tragedy for that family, but the rest of the city kept running. That's what we used to call "individual risk." But today, our city is built on a new kind of foundation: a massive, interconnected web of digital platforms and Artificial Intelligence (AI).
This paper argues that we are facing a new kind of danger called Systemic Risk. Think of this not as a single house fire, but as a situation where a small spark in one part of the city could trigger a chain reaction, causing the entire power grid to fail, the water supply to stop, and the whole city to collapse.
Here is a breakdown of the paper's main ideas using simple analogies:
1. The Old Rules Don't Fit the New City
The authors look at how we used to handle big disasters, specifically in finance (like the 2008 banking crash). In finance, regulators learned that if one big bank fails, it can drag down the whole economy. They built rules to stop that.
Now, the EU has tried to apply similar rules to AI and digital platforms with two new laws:
- The Digital Services Act (DSA): This is like a rulebook for the "city squares" (social media, search engines). It says, "If your square is huge (45 million people), you must watch out for big problems like illegal content, election interference, or harm to kids."
- The AI Act: This is a rulebook for the "engines" (the AI models themselves). It says, "If your engine is the most powerful one ever built, you are dangerous and must be regulated."
The Problem: The authors argue the AI Act is looking at the wrong thing. It assumes only the "most advanced" engines (the biggest, most expensive ones) can cause a city-wide collapse. They say this is like thinking only a massive nuclear power plant can cause a blackout, ignoring the fact that a million small, poorly maintained generators could also cause a grid failure if they all fail at once.
2. The Four Levels of Danger
The paper creates a new map to understand how these digital failures spread. They identify four levels where things can go wrong:
- Level 1: The Single Engine (Single-Model Risk). Imagine one specific AI model is used by almost everyone for hiring, lending money, and writing news. If that one model has a bug or a bias, everyone gets the same bad advice at the same time. It's a "monoculture" disaster.
- Level 2: The Herd Mentality (Multi-Model Risk). Imagine thousands of different companies all using AI models trained on the exact same data. If that data is flawed, all those different companies will make the exact same mistake simultaneously. It's like a herd of cars all swerving left at the same time because they saw the same fake road sign.
- Level 3: The Engine in the Square (Model-Platform Integration). This is when an AI engine is plugged directly into a social media platform. The AI writes a post, and the platform's algorithm pushes it to millions of people instantly. The paper uses the example of an AI generating harmful images that then go viral. The AI creates the poison, and the platform is the delivery truck that spreads it everywhere.
- Level 4: The Engine in the Government (Model-Institution Integration). This is when AI is used to run the city's critical systems, like police predictions or court decisions. If the AI is biased, it doesn't just hurt one person; it changes the entire justice system, locking up innocent people or denying loans to entire groups of people systematically.
3. Testing the Rules with Real Examples
The authors test their new map against three specific problems to see if the current laws catch them:
Discrimination (Bias):
- Current Law: The AI Act says only the "most advanced" models are risky. But the authors say even older, smaller models can discriminate against millions of people if they are used everywhere.
- Verdict: The current laws might miss this because they are too focused on how "smart" the model is, rather than how widely it is used.
Hallucinations (Lying):
- Current Law: The AI Act focuses on "high-impact" risks. But the authors point out that AI lies (hallucinations) are actually less common in the newest, most advanced models, but they are still very common in the slightly older ones.
- Verdict: If the law only regulates the "smartest" models, it might ignore the models that are actually lying the most. If an AI confidently tells a lie to a million people, that's a systemic risk, even if the AI isn't the "smartest" one.
Climate Change (Environment):
- Current Law: The AI Act defines risk as something that hurts the "Union market" (the economy).
- Verdict: The authors argue this is a mistake. Burning massive amounts of energy to train AI hurts the planet, which hurts everyone, not just the economy. You can't put a price tag on a habitable planet. The current law is too focused on money and misses the environmental disaster.
4. The Solution: A Better Map
The paper concludes that we need to change how we think about risk.
- Stop looking only at the "Smartest" AI: Risk isn't about how powerful a model is; it's about how connected it is. A small AI used by a bank, a hospital, and a school is more dangerous than a super-AI used only for writing jokes.
- Look at the "City," not just the "Engine": We need to regulate how AI and platforms work together. A biased AI output is bad, but a biased AI output that is amplified by a social media algorithm to reach millions is a systemic disaster.
- Care about People, not just Markets: The laws need to protect fundamental rights (like fairness and truth) and the environment, not just the stability of the stock market.
In short: The paper argues that our current rules are like trying to stop a flood by only checking the biggest dams, while ignoring the thousands of small cracks in the levee that are actually causing the water to rise. We need a new framework that looks at the whole interconnected system to prevent the digital city from collapsing.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.