Opening article
The AI Security Institute published a pilot study evaluating how artificial intelligence impacts worker productivity across common workplace tasks. Conducted alongside the UK government's new Future of Work Unit, the research used a randomized controlled trial to measure performance changes. Results indicate varying productivity gains depending on the specific task type.
The rapid growth of AI capabilities and increasing cross-economy adoption , have sparked widespread discussion about the future impacts of AI on global economies and labour markets.
There is great uncertainty about how AI will affect the UK labour market, and researchers are using a range of approaches aiming to answer this question. At the AI Security Institute (AISI), we’ve piloted an independent study targeted at a representative sample of the UK workforce. The research expands on the existing literature by creating a suite of generalisable benchmarks based on the O*NET taxonomy, and testing these in a controlled setting with human participants. This approach underpins a new framework for generating economy-wide impact estimates grounded in robust experimental data.
The UK Government recently announced the establishment of a comprehensive AI and the Future of Work Programme to ensure the UK is prepared to benefit from and adapt to the profound changes AI will bring to jobs, workers, and the labour market. This includes launching a new cross‑government AI and the Future of Work Unit and appointing an independent Expert Panel drawn from industry, academia, civil society and trade unions to guide this work.
We designed experimental tasks to simulate real-world scenarios where these Work Activities would be deployed, without requiring domain-specific knowledge to carry out the task. This ensures that tasks are applicable across multiple professions. We worked with the Institute for the Future of Work (IFOW), their academic network, and experts from Educate Ventures to design and validate the tasks for this experiment.
The study described in this blog post represents an early output from this new Unit’s AISI-based research arm.
In the occupation automation literature jobs are often understood as bundles of tasks, some of which are common across different professions. This is formalised in the Occupational Information Network (O*NET), a comprehensive occupational taxonomy developed by the U.S. Department of Labour.
For the purposes of the pilot study, we use the O*NET Generalised Work Activities. These represent the clearest examples of job behaviours common across occupations, for which it would be feasible to design experimental tasks.
This research provides empirical data on the jagged capabilities of current AI models across different workplace activities, helping policymakers and industry leaders better understand labor market impacts.