The Human Influence Observatory

The project builds a public observatory for measuring and monitoring human influence across AI driven societal systems, deliberately tracking human influence and not AI capability. Building on literature by Kulveit et al. (2025), the aim is an indicator framework, a set of measurable proxies for human influence, and a platform where they can be monitored over time.

The observatory will initially focus on domains identified in gradual disempowerment literature where measurement is tractable: economic systems (human influence over labour markets, capital allocation and corporate decision-making), political institutions (human contestability, oversight and participation in governance), and cultural systems (human participation in cultural production and influence over cultural evolution). 

Human influence here means whether people remain able to intervene, contest, and redirect: whether decisions can still be challenged and the objectives behind them shaped, whether increasingly agentic activity remains steerable, and how far institutions still depend on human workers, voters, consumers and participants. Because influence can't be observed directly, the observatory is built from proxies expected to track it, read over time, with attention paid to the interaction effects between domains, the early-warning indicators for feedback loops between economic, political, and cultural influence that drive GD, and the points at which human influence becomes critically compromised. I expect building that to take more than one grant, but this covers its first stage: the economic domain. It is where the measurement task is most tractable, where usable data already exists, and where the influence that propagates outward into culture and politics originates.

The observatory tracks four indicators. AI's share of GDP, considered as a category of its own rather than folded into capital so that output directed by systems can be distinguished from output directed by people. The fraction of consequential corporate decisions made by AI rather than by humans, measured through procurement data and survey data. And the volume of AI spending that no person has approved, as well as the divergence in wealth between industries that still depend on human workers and those that no longer do. The project starts with the US and UK where most data is available, and extends to a few other OECD economies where labour share data permits. Figuring out which ones make up the set depends on what the data supports, and settling that is part of the project.

The grant will help towards four outputs: a method for constructing each indicator (how it is defined, what data it draws on), datasets of the proxy data (documented data behind each indicator), the public observatory as a website where indicators can be read with the underlying data and methods visible alongside them, and a published methodology, including an explanation of which indicators resist measurement and what would need to change for them to be measurable.