In its simplest form, participatory AI means involving a diverse range of stakeholders in building AI systems. This might be members of the public, subject experts or the intended end-users of the AI tool. Participation is particularly important when it comes to using AI for high stakes (and high reward) decisions like climate adaptation, crisis response or health and social care. 

Participatory AI comes in many shapes and sizes. People can input at different stages of the AI development pipeline, from data collection to risk assessments and safety testing. By drawing on the experiences, preferences and grounded expertise of diverse groups we can start to leverage collective intelligence, building more trustworthy AI systems that work in the public interest.

We want AI tools to be better than they currently are – more accurate, fair and trustworthy, as well as focusing on applications that advance the public good. People from different backgrounds should benefit from AI, whether they’re crisis-affected communities from Nepal or doctors from the NHS. To achieve this, we believe it's important to identify and reduce potential harms of technology by working with the groups that are typically furthest away from the design but stand to be most impacted.

AI has the potential to transform the way we work together on problems, helping us to make sense of collective intelligence at scale, while being more creative, collaborative and cooperative. But AI tools can also be biased, spread misinformation or lead to a loss of accountability on important decisions. AI development is also concentrated in the hands of a few powerful companies, mostly based in the US. This means they set the direction of travel, including the problems the technology is trying to solve and the values or preferences the tools are aligned to.

Participatory AI isn’t a silver bullet. It won’t solve all problems associated with AI but it can certainly help. Because this field is still in its infancy, we need to try different approaches and evaluate what works. Our aim is to build more accurate, legitimate and trustworthy tools while also providing meaning and value to the people involved.  All of these factors are critical for public acceptance and uptake across different sectors. 

Collective intelligence methods and participatory design draw on diverse data, preferences and values to solve problems – and we think they can also help when it comes to improving AI. We first looked at the ways that collective intelligence methods could help us build better AI systems in our report on The Future of Minds and Machines.

Since then, we’ve formalised a methodology around participatory AI and identified several points along the AI development pipeline where collective input can lead to more accurate, values-aligned and safer tools. We’ve tried crowdsourcing representative datasets, from crisis-affected communities and value-sensitive design to human-in-the-loop testing of AI models and collective decision-making for the regulation of emerging AI tools. We’ve put these methods into practice, working alongside organisations like the Red Cross and the UK’s Met Office. 

At the moment, we’re pursuing three key strands of work:

• Localising AI - building localised, contextually-relevant AI tools to improve crisis response through participatory design and collective intelligence. 
• AI assurance - developing new methods to involve the public in evaluating risks and assessing the social acceptability of different AI use cases in sectors such as humanitarian aid and public services. 
• Hybrid human-AI technology - applying user-centred design and participatory evaluation methods to develop trustworthy hybrid AI-human tools for decision-making on climate and health.

As we continue to expand our methodological toolbox, we’re always keen to hear from others working in this space, particularly tool developers or public sector organisations starting to work with AI. Get in touch at [email protected] with the subject line Participatory AI.