A University of Glasgow-led initiative has unveiled an open-source auditing workbench designed to help organisations assess AI risks, strengthen governance and involve affected communities in evaluating the real-world impact of Artificial Intelligence systems.
A University of Glasgow-led research project is releasing a free tool to help organisations, policymakers and the public maximise the benefits of Artificial Intelligence applications while identifying potential harms.
The tool, developed as part of the Participatory Harm Auditing Workbenches and Methodologies (PHAWM) project, aims to address the urgent need for rigorous assessments of AI risks caused by the rapid expansion and adoption of the technology across multiple sectors.
It is designed to support the aims of regulations such as the European Union’s AI Act, introduced in 2024, which seek to balance AI innovation with protections against unintended negative consequences.
PHAWM’s new open-source workbench tool empowers users without extensive AI backgrounds to conduct in-depth audits of the strengths and weaknesses of any AI-driven application. It also actively involves audiences who are typically excluded from audit processes, including individuals affected by AI-driven decisions, to improve outcomes for end-users.
The tool is the first public outcome from PHAWM, launched in May 2024 with £3.5 million in funding from Responsible AI UK (RAi UK). The project brings together more than 30 researchers from seven leading UK universities alongside 28 partner organisations to address the challenge of developing trustworthy and safe AI systems.
The tool and its accompanying framework are publicly available and free to download from the project’s website.
Professor Simone Stumpf, of the University of Glasgow’s School of Computing Science, leads the PHAWM project. She said:
“Generative and predictive AI applications have the potential to give organisations valuable new ways to deliver improved services for end users. They are already influencing decisions in areas including housing, employment, finance, policing, education, and healthcare.
“However, they can be afflicted by flaws like bias and inaccuracies. In order to avoid building AI applications which enforce unfair outcomes in critical services, they must be carefully monitored and regularly audited by humans.
“Until now, those audits are usually conducted by people with a deep understanding of the processes which drive AI, but who may lack insight into the social or cultural impacts those systems may create. There is rarely an opportunity for people who will regularly use or will be affected by AI decision-making to help guide their development.
“Our new workbench tool is designed to help organisations create better, fairer, more transparent AI systems by providing diverse perspectives on AI applications which might otherwise go unexamined.”
The tool and framework were developed through co-design workshops with partners and stakeholders in the health and cultural heritage sectors. These are two of the four areas PHAWM was established to investigate, alongside media content and collaborative content generation.
The PHAWM tool works by systematically gathering diverse perspectives through a four-stage auditing process.
First, the audit instigator provides information about the AI system in accessible, non-technical language.
Second, relevant stakeholders, including system users and those affected by AI decisions such as members of the public or patients in health applications, are invited to participate.
Third, participants align the audit with their concerns and lived experience. They identify potential positive and negative impacts, create metrics to measure them and assess whether the application meets their criteria. The system receives a pass or fail grade based on participant-defined standards.
Finally, the instigator gathers insights raised during the audit and develops action plans to inform decisions about deployment or integration.
Professor Stumpf added: “The tools and processes we’ve developed offer a practical, community centred approach to evaluating the real world impacts of Artificial Intelligence. The workbench is a flexible tool which can be used to run in-depth audits of AI applications an organisation has developed in-house, as well as being used to investigate whether off-the-shelf AI applications will meet organisations’ needs before they are purchased.
“Being able to look in such depth and from so many different angles will help organisations make properly informed decisions which assess the balance of risk and reward which comes from adopting new technologies. Our hope is that organisations will be encouraged to use the tool and framework we’ve developed with our partners and stakeholders will enable them to reap the benefits of AI while avoiding any potential for harms.”
The PHAWM team continues to refine the tool in collaboration with representatives from its four focus areas. Public Health Scotland and NHS National Services Scotland contributed to the health use case, while Istella contributed to media content. The National Library of Scotland, Museum Data Service and David Livingstone Birthplace Trust participated in the cultural heritage use case, and Wikimedia contributed to collaborative content generation.
The team is also developing comprehensive training and certification support to help organisations adopt the auditing tools effectively.
