Jakub Mareček’s LION 19 keynote presents a framework for reasoning about repeated uses of AI systems and the new Interconnect toolkit for long-run properties therein.

On 18 June 2025, our researcher Jakub Mareček presented a 90-minute keynote at the 19th Learning and Intelligent OptimizatioN Conference entitled “Fairness in Repeated Uses of AI Systems.” He showed how platforms – incl. ride-sharing services, home-rental marketplaces, and virtual power plants – can deliver both strong aggregate performance guarantees and ensure every participant enjoys a fair long-term outcome.

Mareček began with real-world examples: drivers whose earnings diverge despite identical abilities, neighborhoods that receive fewer ride offers, and homeowners shouldering extra wear on local infrastructure. AirBnB researchers in the audience twitched. These cases illustrate that treating everyone the same at each decision isn’t enough. Instead, he argued, we must focus on long-run fairness, where each user’s time-averaged experience converges to the same level.

Figure by PWC used in Mareček's presentation: The Sharing Economy-Sizing the Revenue Opportunity. Source: PricewaterhouseCoopers (London, UK, 2014).

To achieve this, Mareček and his team frame AI platforms as closed-loop systems: an AI system issues prices or incentives, users’ response is modelled using probabilistic models, and the resulting measurements feed back to guide the next round. He demonstrated that common PI-style controllers can break ergodicity, allowing tiny initial differences to amplify into persistent disparities. Then he identified practical conditions that guarantee ergodicity. These guarantees rely on Iterated Function Systems, a tool from applied probability.

All of these insights appear in the full slide deck (PDF) and the 53-page lecture notes (PDF).

Interconnect toolkit

To help researchers and practitioners put these ideas into practice, the AutoFair project has released Interconnect – an open-source toolkit built on top of PyTorch. It provides a simulation engine, formal-verification functionality, modeling templates for both mobility and demand-response use-cases, and Jupyter notebooks that walk through the examples from the keynote.