| Aims of activity, and main critique being addressed* | Key results from PAI activities |--------------------------|--------------------------| | Key results from comparative evaluation | Central disaster management team (Kathmandu) | | Accuracy and speed of crisis response, and ability to meet the needs of the most vulnerable | • Six out of eight groups of Nepal Red Cross Society staff gave a higher rating to the tool than the current process for accuracy (the ability to match the needs of communities) and vulnerability (the ability to meet the needs of the most vulnerable). | | | • Five out of eight groups thought the tool would make the process of NFRI distribution faster. | | | District-level workshop participants (Mustang) | | | Red Cross staff at the district level gave a higher rating across criteria of speed, accuracy and vulnerability with the tool in comparison to the current process during evaluation activities. But they were only prepared to use the model if it was retrained with data gathered from their own district. | | Phase of crisis response | • Participants identified the highest utility for the tool during the Preparedness phase of crisis management, where they rated the tool more highly than the current process across all criteria. [^15] |

| Overview | Results |--------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ * Ensure there are blank lines before and after lists for proper rendering

• For tables:
  • CRITICAL: Wrap all tables in a container div for horizontal scrolling: <div class="table-container" markdown="1"> before the table and `

    | Header 1 | Header 2 |
    |----------|----------|
    | Data 1   | Data 2   |

    </div>
    ```

• Heading Cleanup and Hierarchy Verification:
  • Delete any titles or headings that contain "(continued)"
  • If a heading is exactly "Financial Reporting Council", convert it to a bold paragraph (**Financial Reporting Council**) instead of a heading.
  • Delete any heading that is identical to the immediately preceding heading at the same level.
  • Before creating an H2 heading, verify: Is this truly a major top-level section, or is it a subsection of the current numbered section? If it's a subsection (like "Assessment Period", "Reporting Requirements", "Summary of Requirements" within a numbered section), use H3 (###) instead.
  • Watch for PDF layout tricks: Sometimes PDFs make subsection titles look prominent (bold, larger font), but they're still subsections. Look at the numbering and logical flow, not just visual appearance.
• Extraneous Elements:
  • Remove any purely decorative horizontal lines (e.g., ---, ***).

1. Content Preservation:

   • Include legal disclaimers, copyright notices, and important notices
   • Preserve all tables, even if they contain legal or administrative information
   • Maintain all substantive text regardless of whether it appears "administrative"
   • Only skip content that is purely decorative or redundant page elements

2. Content Formatting:

   • Emphasis: Preserve bold and italic text using markdown syntax (**bold**, *italic*)
   • Footnotes: Convert to markdown format using ¹ syntax, with footnotes collected at end of document
   • Hyperlinks: Convert ALL hyperlinks to markdown format [text](url). Look carefully for clickable text even if it appears as regular text. Only include URLs that are explicitly present in the PDF - don't invent URLs for plain text.
   • Amendment Documents (e.g. FRED, draft standards): Preserve editing marks:
     • Strikethrough text → ~~deleted text~~
     • Underlined insertions → <ins>inserted text</ins>
     • (Regular underline for emphasis uses standard markdown)
   • HTML Elements with Markdown Content: When using HTML block-level elements that contain markdown-formatted content (bold, italic, lists, etc.), always add markdown="1" attribute to the opening tag. Examples:
     • <blockquote markdown="1">Governments lie; bankers lie; even auditors sometimes lie: gold tells the truth.</blockquote>
     • `
       Chart showing quarterly results:

   Revenue increased by 25% over the previous quarter.

| Aims of activity, and main critique being addressed* | Key results from PAI activities ```

Designing Report and Respond using participatory AI

Throughout the tool development process we designed and delivered participatory AI activities that sought to:

• Increase the usefulness of the tool for frontline users.
• Create high-quality, representative and machine-readable data to reduce the risks of AI model bias.
• Optimise the AI model's performance.
• Create appropriate levels of trust and understanding of the tool by all stakeholders.
• Ensure all stakeholders have a say in evaluating the performance of the tool.

As participation was included at multiple touchpoints along the AI pipeline, and included activities where the technical team interacted with participants, we categorise the level of participation as collaboration.

We held in-person workshops in three different locations (see Figure 16 for an overview) involving two key stakeholder groups:

1. 1Red Cross volunteers from two French speaking regions in Cameroon.
2. 2Red Cross staff working in community engagement and communication.

Collaborative problem framing and initial tool design took place in August and September 2021. All other tool development and participatory AI activities happened between January-April 2022.

We identified suitable intervention points during model development and the tool design using an updated version of Nesta's Participatory AI framework. See Figure 17 for an overview of activities we carried out.

Firstly, we worked with Cameroon Red Cross (CRC) staff to curate our dataset, verify labels for the dataset, create new rumours to test the classification algorithm, and create responses for existing rumour categories. With volunteers, we focussed on testing the usability of the interfaces and the reporting process, testing the clustering algorithm, and discussing potential impacts of the tool, including issues related to trust and uptake.

System Integration and Deployment and Oversight appear lighter as they were out of scope for the project. CRC = Cameroon Red Cross. Adapted from the framework published in Participatory AI for humanitarian innovation: a briefing paper.

Left: Individual testing of user-interface prototypes for the Report and Respond tool. Right: Group work.

Report and Respond Results

HOW WELL DOES THE TOOL WORK ON A TECHNICAL EVALUATION?

In this section, we provide an overview of how well the Report and Respond tool performed in both technical and participatory evaluations, as well as describing to what extent we were able to address the critiques of AI through participatory activities. We also cover the perceived trustworthiness and usefulness of the tool for frontline stakeholders.

Our technical evaluation consisted of measuring the performance of our models using multiple test datasets. We also tested for model bias to ensure that any limitations of the model were well documented for future users.

Complete results from evaluation activities can be found in the detailed Technical Report.

Photo: Edouard Tamba, Unsplash

Although the overall accuracy of the model on the test dataset of rumours from Cameroon appears high, we observed variance across different rumour categories. For example, the model has an accuracy of 91.7% for the rumour category 'Croyance que certaines personnes/institutions gagnent de l'argent à cause de la maladie' ('Belief that some people or institutions are making money from the disease') and only 58% for the category 'Croyance que la maladie existe ou est réelle' ('Belief that the disease exists or is real').

We also tested the model's performance using two additional datasets:

1. 1A comparable community feedback dataset from the Democratic Republic of Congo (DRC), from the IFRC Go platform.¹²
2. 2A new rumour dataset created by the Cameroon Red Cross staff.

Although the model shows a similar decrease in accuracy for both datasets, the performance is still relatively high. This suggests that the model can be generalised to other French language countries and contexts in the future. The reduction in accuracy we observed for the test dataset generated by the Red Cross staff¹³ highlighted a limitation of our model – that it struggled to assign multiple labels when rumours spanned more than one category.¹⁴ Through discussions of this result with Cameroon Red Cross staff, we identified that the ability to assign multiple labels would be important for future iterations of the model.

HOW WELL DOES THE TOOL PERFORM IN A PARTICIPATORY STAKEHOLDER EVALUATION?

As part of our participatory AI activities, we involved Red Cross volunteers in a comparative evaluation. For this, we asked them to rate a hypothetical scenario of using the tool during their fieldwork in comparison to their usual process for collecting community feedback. The evaluation criteria were inspired in part by the recommendations of the OECD DAC Network,¹⁵ adapted to make them more accessible for the Red Cross volunteers.

The results show that participants thought that the largest improvements would be for the timeliness of the response and the ability to respond to community needs. They anticipated that the tool would give more reliable responses to rumours, and that the benefits of this would be felt by both community members and volunteers. Table 6 provides a summary of the key results. Figure 19 shows the average result for six groups (17 participants in groups of two to four people) made up of Cameroon Red Cross volunteers. Volunteers thought the Report and Respond tool has the potential to increase the speed of localised humanitarian response, as well as improving the ability to meet the needs of communities.

Radar chart shows average scores of six groups (17 participants). Participants were asked to assign a rating to statements about the criteria (e.g. how well are you able to meet the needs of the community? 1 = not well, 5 = very well).

“The community benefits from the tool. It's important because we'll respond rapidly to the needs of the community."

Red Cross Volunteer (Kribi)

HOW DID THE PARTICIPATORY AI ACTIVITIES INFLUENCE THE TOOL AND HELP US ADDRESS SOME COMMON CRITIQUES OF AI?

We tried multiple participatory AI interventions despite the challenges of designing, delivering and interpreting engagement activities with the project team split between multiple time zones and languages. Table 7 provides an overview. These activities showed that participation can start to address some of the critiques of humanitarian AI – and have an impact on AI performance as well as trust or acceptance of the tool by different stakeholders.

*Relevance of activities to specific AI critiques are the result of post-hoc analysis.

The vignettes that follow provide a detailed description of two participatory AI activities we tried during the project. The first is an example of the importance of user-friendly task design and tools, as well as careful facilitation, for the success of participatory AI. Both vignettes describe how Red Cross staff influenced the development of the classification model.

Figure 21: Example activity for labelling with CRC volunteers using KoBo Toolbox

HOW TRUSTWORTHY IS THE TOOL AND SHOULD IT BE OPERATIONALISED?

The trustworthiness of AI systems is a key implementation barrier, especially when tools replace well established processes. We used focus group discussions to understand how biases might impact the trustworthiness of the tool. We also measured participants' general attitudes towards the idea of computers making decisions about misinformation using post-workshop questionnaires, as well as specific positive or negative associations towards the tool and its future development. Table 8 provides an overview of the results.

Photo: Edouard Tamba, Unsplash

What next?

Further work is necessary to make the tool ready for testing in operational settings and full deployment. As a priority: a) the coding system of the Red Cross should be refined on a regular basis to eliminate redundant labels and enable the distinction between closely related topics and; b) the system should allow for multiple labels per rumour so the model can learn to respond appropriately to community feedback that spans multiple categories.

Further work could focus on training the model on the full training dataset from IFRC Go, incorporating all of the additional rumour categories¹⁸ and using rumours from other sources, such as social media platforms. Training models on a variety of data sources would enable triangulation of their performance for detecting rumours. Using additional, timely sources of data would also help to more reliably determine which rumours are emerging at any point in time. The model we have developed is only appropriate for a French language context, however the approach is generalisable. Extending the functionality of the tool to cover other languages would rely on additional data and modelling.¹⁹

The short timeframe of the project only allowed us to test discrete parts of the workflow in an isolated fashion, outlined in the description of 'How the tool works'. We weren't able to fully develop and test the new internal workflows that would be necessary for the Cameroon Red Cross Society to make the most of this tool. For this, the Community Engagement and Accountability team needs to develop regular processes for: a) reviewing the outputs of the clustering algorithm to keep track of emerging rumour categories; b) labelling and verifying rumours as they come in; c) assigning appropriate responses for new rumour categories and; d) updating responses for existing rumour categories if new guidance emerges.

At present, the Cameroon Red Cross does not have permanent staff with the technical capabilities required to develop AI models, or the digital systems infrastructure to process and maintain datasets. This is a major barrier to the application of CCI in local humanitarianism.

WIDER RELEVANCE FOR THE HUMANITARIAN SECTOR

The scale and spread of misinformation and disinformation is a growing challenge for humanitarian organisations. The COVID-19 pandemic has highlighted the problem, with misinformation helping to exacerbate the crisis and hindering humanitarian response. The Report and Respond tool shows the potential of involving frontline staff and volunteers in the efforts to address this problem. Putting CCI tools in the hands of those in regular contact with communities can help humanitarian organisations address misinformation in a more timely manner, provide locally appropriate responses, and adapt their interventions accordingly through better real-time monitoring.

We hope other initiatives to address misinformation, such as Médecins Sans Frontières' 'MSF-Listen' project²⁰ and the Katikati tool being developed by the team behind the Africa's Voices project,²¹ may be able to adapt and incorporate elements of our tool into their own operations.

Photo: Aalok Atreya, Unsplash

What we learned about the benefits of collective crisis intelligence and participatory AI

The results presented in earlier sections confirm that collective crisis intelligence approaches can lead to better solutions, and that participatory activities can have a substantive impact on AI performance, trust, and acceptance of the tool by different stakeholders. We summarise the key findings below.

Although promising, they offer only an initial view of the potential of these emerging innovation approaches. Further experiments are needed to demonstrate how they can be applied most effectively to confer maximum benefits to local humanitarian responders and the communities they serve.

Collective crisis intelligence has the potential to make humanitarian action more timely and appropriate to local needs

Collective crisis intelligence tools combine locally-generated data and insights from the people closest to a crisis with the processing power of AI. Our comparative evaluation suggests that these tools have the potential to significantly improve local humanitarian action.

• In Nepal, Red Cross Society staff rated the NFRI-Predict tool more highly for accuracy (the ability to match the needs of communities) and vulnerability (the ability to meet the needs of the most vulnerable) than the current process for determining NFRI distribution. A majority also thought the tool would make the process of NFRI distribution faster.
• In Cameroon the results of the comparative evaluation showed that Red Cross staff and volunteers thought that the collective crisis intelligence Report and Respond tool would lead to significant improvements in the speed of the response and the ability to respond to community needs.

Collective crisis intelligence can enhance the utility of locally-generated data and drive new forms of action

Citizen-generated data and frontline insights are at the heart of collective crisis intelligence tools. Although this often means creating new datasets, tools can also draw on existing data held by humanitarian organisations. AI models can help transform these community datasets into insights that drive locally appropriate decisions and more anticipatory action.

• In Cameroon, we showed that collective crisis intelligence could elevate the value of underutilised community feedback data. We repurposed data already held by the IFRC, and used it to prototype a faster way to report, monitor and respond to COVID-19 misinformation, with the goal of containing it more effectively.
• In Nepal we demonstrated that applying AI to a new dataset of community NFRI preferences from geographically distinct regions and 3000 diverse households could give the Nepal Red Cross important insights into how they can predict which aid items will be needed in advance of a crisis. This has the potential to enable more anticipatory action, including through the stockpiling of the most essential goods for people in different regions.
• The Nepal Red Cross are already considering other ways they could use CCI approaches to develop more anticipatory operating models – predicting volunteer deployment needs depending on crisis type and community needs, predicting blood donation requirements, or more accurately estimating the value of cash transfers.

Collective crisis intelligence and participatory AI can help increase trust in AI tools

The perceived trustworthiness of AI systems is important for their uptake and, ultimately, their social licence to operate. In our project, stakeholders had high levels of trust in the collective crisis intelligence tools we developed, but more work is needed to untangle the factors which had the most influence on building that trust.

• In both countries, the majority of frontline staff and volunteers rated the CCI tools we developed more highly on 'trust' than their current operational processes for tackling those problems.
• In Cameroon we saw an increase in the proportion of volunteers who felt positive (rather than neutral) about the CCI tool in successive workshops.
• In Nepal, knowing that technology had been developed with the involvement of local stakeholders and using local data increased trust in the technology and the likelihood of using it amongst frontline staff, but only if it isn't the only input into decision making.

Participatory AI can overcome several critiques and limitations of AI to improve model performance

The curation of datasets, prioritisation of input features, and other specifications of model design are typically chosen by data scientists in combination with automated methods. We showed that, when activities are well designed, participatory approaches can be a useful input into any of these stages of technical development and can lead to better performing models.

• In Nepal, this meant validating and refining the data inputs used to train the AI model – specifically removing Ethnicity as an input – due to local preferences. This resulted in an algorithm of equivalent accuracy that also took into account the concerns of stakeholders and communities on the frontline of crises.
• In Cameroon, involving Red Cross staff in generating a dataset of new rumours to test the model, helped identify a 'blindspot' in the functionality of the AI. As a result, this issue can be addressed to improve the model before deployment.
• In Cameroon, we demonstrated that it is possible to help improve the quality of community data by designing collective crisis intelligence tools that build validation into the workflow. We also showed that Red Cross staff could successfully complete labelling activities to generate datasets of higher quality data for training an AI model. Ultimately, this should help increase trust in citizen-generated data, helping it to be taken more seriously as a source of intelligence.²²

Participatory AI helps to surface tensions between the assumptions and standards set by AI gatekeepers versus the pragmatic reality of implementation

Local stakeholders are rarely involved in setting the benchmarks that models are measured against. Likewise, recommendations for how AI tools should and shouldn't be used are often set by technologists based in the Global North who are far from understanding the local context.

• During evaluation activities in both countries, we learned that frontline stakeholders were more pragmatic about using AI across a broader range of circumstances. They preferred to use a tool that was 'good enough' rather than 'perfect', as long as it improved on the current process.
• In Nepal, we identified key areas of consensus and differently held values between stakeholder groups in order to set more appropriate evaluation criteria for the model. Alongside the technical evaluation, this approach can help give a more nuanced understanding of AI performance.

Participatory AI creates opportunities for building and sharing new capabilities among frontline staff and data scientists

Our previous research highlighted that frontline staff, including volunteers, are often underutilised as key stakeholders in CCI innovations. The interdisciplinary environment of collective crisis intelligence projects creates multiple opportunities for upskilling among project team members and participants alike.

• Participatory activities helped to convince data scientists of the importance of involving a wider community of stakeholders in model development, particularly for problem definition, data wrangling and surfacing post-deployment challenges.²³

In Cameroon, both staff and volunteers benefited from learning new skills through labelling and design workshops. The activities also motivated the volunteers to play a more active role in the battle against misinformation.

The project offered multiple opportunities for skills exchange between individuals and partners. For example, the Red Cross valued learning about AI and trying out digital tools for prioritisation and prototyping during community engagement activities. At the same time, data science fellows developed new capabilities in designing and delivering participatory activities.

05. Barriers, challenges and how to overcome them

Innovators who want to make use of collective crisis intelligence may encounter several potential technical, organisational and methodological challenges as they develop new tools. Our landscape analysis, 'Collective crisis intelligence for frontline humanitarian action', set out a comprehensive overview of barriers to the application of CCI by humanitarian organisations. Rather than repeating these, we outline the five main barriers we faced during project implementation. Addressing these barriers should be a priority for the sector in order to reap the full benefits of collective crisis intelligence.

Local humanitarian organisations have data, but it's not AI-ready and open datasets can't fill the gaps

Our experience shows the majority of community datasets held by humanitarian organisations are small scale. We found that even larger, labelled datasets can suffer from poor quality labels, particularly when categories are subjective and coding taxonomies are not reviewed regularly. Although we hoped to enrich our models with open data, we faced similar issues of gaps or insufficient granularity to be easily repurposed for AI development.

Unless we invest in levelling the data playing field and develop AI techniques that can function under these constraints, AI will remain inaccessible and inappropriate for the countries most affected by humanitarian crises and local humanitarian responders.

Recommendation

Technologists should develop AI methods that can function in low resource settings, data sparse environments, or with poorly labelled datasets, rather than data hungry supervised methods and increasingly larger scale language models.

Recommendation

The humanitarian sector should also take a more coordinated approach towards filling data gaps. It must invest in building up open datasets that are relevant and useful to local frontline humanitarians, and help ensure they are up-to-date and high quality.

Local organisations are currently ill-equipped to participate in AI development

During our partnership, we were told of multiple organisational datasets and existing models that could be relevant to our work together, but local Red Cross staff could not provide a description of what the data contained or how the models worked due to poor historic documentation. Negotiating access to existing data took many months, while the capabilities and technology infrastructure to collect, analyse and securely manage new data in the long term was limited.

Locally-developed AI that can address local needs will only be possible if international funders and humanitarian organisations invest in locally-grown data science talent and digital infrastructures, and build data capabilities across the board, not just in information management teams.

Recommendation

The humanitarian sector should insist on properly documenting existing AI models and support local organisations to develop and maintain data inventories they can share with collaborators.

Recommendation

Make the shift towards widespread data capability across all departments, including AI literacy, to understand when it can or can't be used. Create local embedded data fellowships, building on the UKHIH model, to bring technical talent into local organisations.

Shifting from 'implementation' partnerships to multi-disciplinary 'co-design' partnerships

It took time for our core project team to transition away from classic (imbalanced) roles in programme delivery towards multi-disciplinary 'co-design', with equal contributions to defining agendas and success. During early stages of the project, misalignment about these different operating modes and lengthy contract negotiations caused delays and frustration for all partners. These challenges were further exacerbated by our inability to meet in person due to COVID-19 travel restrictions.

It's unfair to expect local partners to be able to make the radical shift required without dedicated support to transition to a different way of collaborating. Supporting this transformation should be a sector-wide priority as it moves towards local development and ownership of new AI tools.

Recommendation

Allocate sufficient time dedicated to developing enabling environments for co-design, as well as prioritising alignment on roles and expectations.

Participatory AI requires new community engagement skills, and data scientists who engage with participation

AI can be difficult to explain, especially to stakeholders from very different backgrounds. A key challenge for the field is creating activities that help participants make the conceptual leap so they can contribute meaningfully to participatory AI activities. We had to design most of our activities for explaining AI and engaging affected communities and frontline volunteers from scratch. Delivering the activities required a basic level of AI literacy, as well as flexibility from facilitators. Our experience also highlighted the challenge of translating outputs from participatory AI activities into actionable changes to the model in the absence of a shared vocabulary between community engagement and technical practitioners.

There is still a long way to go before we'll be able to say which participatory AI methods are the most effective. But as the evidence builds, activities need to be codified and made simple for others to use.

Recommendation

Create easy-to-use guides for running participatory AI activities, with a clear role for technical teams and provide specialised training for community engagement practitioners.

Too few digital tools work in low-resource settings

There are relatively few digital tools to support participatory AI activities that function well in low-resource settings. KoBo Toolbox is a rare example, which we had to repurpose for multiple tasks in the absence of viable alternatives. In both countries, we saw the value of demonstrations and visualisations in making AI more tangible. In Cameroon, high-fidelity prototypes helped participants understand how the AI model could support operational tasks in the field. However, we found that existing digital prototyping tools hadn't been developed with low-resource settings in mind. Although interacting with prototypes helped participants connect the concept of AI with their operational reality, their frustration with technical issues sometimes outweighed the benefits.

Our unsuccessful attempts to involve volunteers in data labelling also highlighted the need for new open source labelling tools with user-friendly workflows, high data protection standards, and flexible control over task assignment. In general, participatory AI needs more creative digital tools, especially those that allow communities to enter data in their own words.

Recommendation

Develop adaptable, open-source, and user-friendly digital tools that function in low resource settings to support participatory AI.

06. A call to action

NO AI ABOUT US WITHOUT US ²⁴

It is time to upend and localise the development of humanitarian AI. We must move away from top-down initiatives and proprietary systems that risk reinforcing colonial power dynamics, undermining the localisation agenda, and potentially cause more harm to crisis affected communities.

The sector must stop seeing crisis affected communities as mere data points, and start to create opportunities for them to actively shape and own the tools that can help them address the crisis they face. This needs a coordinated commitment to a different way of doing things from humanitarian funders, innovators, and researchers across the board. At a minimum, humanitarian technology projects should be required to demonstrate how affected communities have been involved in the development and oversight of new tools.

Our project has shown that it is possible to build AI with local infrastructure, local data, and local talent, and that it is possible to build AI that responds to local values and priorities. But much more investment is needed to realise a future where locally-developed and owned AI becomes 'business as usual'.

Through our Accelerated Innovation Collaboration, we have started to demonstrate the potential for collective crisis intelligence — the combination of intelligence from crisis-affected communities with the predictive power of AI — to create new operating models and more anticipatory and appropriate humanitarian action. But there is much more to discover, and a need for more rigorous experimentation.

In our previous report, 'Collective Crisis Intelligence for Frontline Humanitarian Innovation',²⁵ we set out 10 key research and development opportunities in this space (Table 9). We are grateful for the support of the UK Humanitarian Innovation Hub, which has allowed us to start to explore some of these opportunities. We call on more humanitarian innovation funders to identify the role they can play in nurturing an ecosystem of actors committed to advancing this emerging field.

Table 9: Ten key research and development opportunities for collective crisis intelligence

Appendix

The Project Outputs

Over the course of the project we developed and tested discrete parts of the workflows described above. In summary, we created:

Nepal

• A new dataset about NFRI preferences at the household level.
• A new dataset about additional items that communities felt should be included in Red Cross NFRI distribution.
• A classification algorithm to predict the 'essentialness' of Shelter (housing and clothing) items.
• A classification algorithm to predict the 'essentialness' of Wash (health and sanitation) items.
• A prototype of a new desktop-based NFRI-Predict tool interface.

Full model documentation: CCI Nepal. Detailed methodology and results: Technical Report Nepal.

Cameroon

• A classification algorithm for matching rumours to existing COVID-19 rumour categories.
• A clustering algorithm for grouping new rumours into emerging categories.
• A prototype of the SMS-based Report and Respond tool interface.
• A prototype of the mobile-based Labelling tool interface.

Full model documentation: CCI Cameroon. Detailed methodology and results: Technical Report Cameroon.

Acknowledgements

We would like to thank all workshop participants in Sindhupalchok, Mustang, Mahottari and Kathmandu (Nepal), and Tibati and Kribi (Cameroon) for their generosity in sharing their time and expertise with us. Special thanks to the District Heads of the Red Cross teams, who helped to organise regional workshops, to the volunteers who helped with data collection in Mahottari and Sindhupalchok, and to all community members who contributed to the generation of a new dataset.

Project team

Nepal Red Cross Society: Rudra Adhikari, Dharma Datta Bidari, Deepak Dawadi, Sarita Dhungana, Tara Gurung, Sakun Joshi, Anupa Koirala, Uddhav Nepal, Janardan Pokharel, Sachin Raut.

Cameroon Red Cross Society: Henshaw Arrey, Guy-Stephane Djob, Hyacinthe Olinga Eloundou, Cathy Essouma, Rodrigue Etono, Fabrice Ewane, Princewill Nkongho, Santana Nnang.

IFRC: Thierry Balloy, Sophie Louise Everest, Laurent Fernandez, Viviane Fluck, Manorama Gautum, Herve Gazeu, Ramya Gopalan, Heather Leson, Patricia Mugenzi, Yves Stephane Ngaleu, Sushama Pandey, Annemarie Poorterman.

Nesta: Aleks Berditchevskaia, Jyldyz Djumalieva, Rob Fuller, Isabel Gill, Nasra Hussein, Pius Ngwa, Lusine Mehrabyan, Kathy Peach, Saurav Poudel, George Richardson, Joana Murta Rosa, Meera Shah, Isabel Stewart, Oli Whittington.

Open Lab, Newcastle University: Simon Bowen, Tom Feltwell, Ian Johnson.

Advisory Group: João Porto de Albuquerque, Maesy Angelina, Nell Gray, Weisi Guo, George Hodge, Ewan Oglethorpe, Mona Sloane, John Twigg.

UKHIH Team: Mark Beagan, Ben Ramalingam, Tonia Thomas.

A special thanks to Sharath Srinivasan and Luke Church from the Katikati platform who helped us to refine the design of the Report and Respond tool. We're also grateful to Shreyasha Poudel, Robert Soden and Manveer Kalirai for the two workshops that helped us explore the potential unanticipated impacts of the NFRI-Predict tool.

Endnotes

1NFRI = non-food relief items 2Responsible AI for Disaster Risk Management. The Open Data for Resilience Initiative and the World Bank. https://www.gfdrr.org/en/publication/responsible-artificial-intelligence-disaster-risk-management Accessed on 28 June 2022 3Responsible AI for Disaster Risk Management. The Open Data for Resilience Initiative and the World Bank. https://www.gfdrr.org/en/publication/responsible-artificial-intelligence-disaster-risk-management Accessed on 28 June 2022. 4Berditchevskaia, A., Peach, K., and Malliaraki, E. (2021). Participatory AI for humanitarian innovation: a briefing paper. London: Nesta. 5Berditchevskaia, A., Peach, K., Gill, I., Whittington, O., Malliaraki, E., and Hussein, N. (2021). Collective crisis intelligence for frontline humanitarian response. London: Nesta 6Berditchevskaia, A., Peach, K., and Malliaraki, E. (2021). Participatory AI for humanitarian innovation: a briefing paper. London: Nesta. 7This tool was developed specifically to meet the needs of the Nepal Red Cross, but we believe that it could also be useful for other organisations involved in crisis response. 8We selected a logistic regression model to optimise for interpretability. 922 NFRI items were included in the survey. These items are typically distributed by the Red Cross within two standard packages: Shelter (covering housing and clothing items); and Wash (covering health and sanitary items). 10Our sample was balanced between male and female respondents (~50:50) and featured households from both rural and urban locations. 11Stakeholder mapping was carried out by Nesta and Nepal Red Cross Society/IFRC community engagement staff 12Measured using the global micro F1-score across all items. The F1 score measures performance through a weighted average of Precision and Recall. 13We evaluated the performance of our model on subsets of the test datasets based on the following protected characteristics: Ethnicity, Respondent Gender, Location. 14The error was largely due to an imbalance in our training dataset: many respondents identified the majority of NFRI items as 'essential'. As the current process followed by the Red Cross deems all items essential, we consider this type of error as causing less harm than other types of errors. 15We asked groups of participants in Kathmandu to imagine using the tool during Preparedness, Early Warning and Response. All groups identified potential benefits of using the tool in comparison to the current process across all three stages of crisis response. See Technical Report for a detailed breakdown of results. 16Measured using the F1 micro score. 17Here we report the combined % of participants who selected 'a little positive' or 'very positive' in the post-workshop survey. 18We posed the question in this way to capture broader attitudes towards AI throughout the project. Using this language also helped to ensure that the question was more accessible to all stakeholders. 19Here we report the combined % of participants from Kathmandu and Mustang workshops. 20A balanced dataset has the same number of input samples for each output category of interest. 21An important priority for new data collection is including a district from the mountain region, e.g. Mustang. Sindhupalchok and Mahottari were chosen for their representativeness of other districts in the Hill and Terai regions, respectively. These two regions have historically seen the most flooding. 22The technical specification published on GitHub alongside this report provides an overview of what is needed. 23Berditchevskaia, A., Peach, K., Gill, I., Whittington, O., Malliaraki, E., and Hussein, N. (2021). Collective crisis intelligence for frontline humanitarian response. London: Nesta 24See detailed analysis in Technical Report. 25We use the terms label, code and category interchangeably. 26Erina Mahmud. Community engagement to counter misinformation in Rohingya refugee camps https://www.oecd-ilibrary.org/sites/e8f2389a-en/index.html?itemId=/content/component/e8f2389a-en%20%20 Accessed 26 June 2022. 27Islam, M. S., et. al. (2021). COVID-19 vaccine rumors and conspiracy theories: The need for cognitive inoculation against misinformation to improve vaccine adherence. PLOS ONE, 16(5), e0251605. https://doi.org/10.1371/journal.pone.0251605 28We report accuracy using the global micro F1 score, which gives an average across all labels. The F1 score measures model performance through a weighted average of Precision and Recall. 29Clustering accuracy is reported using the average homogeneity score. Homogeneity score measures the proportion of comments in a cluster that convey similar meaning. It is assessed using manual inspection of clusters. This was carried out by the local data science fellow and the Red Cross volunteers during the evaluation workshops. See Technical Report for details. 30The IFRC GO dataset contained incomplete information about the age and location of respondents so we limited our bias audit to gender. 31https://go.ifrc.org/ 32For three of eight rumour categories, the model showed a drop in accuracy of more than 10% for the crowdsourced dataset. 33Our classification model is a multi-class model, which should be able to assign multiple labels. In most cases, the model struggles to assign more than one code per data point. This is a limitation of the underlying training dataset, which didn't contain many rumours with multiple codes. 34https://www.oecd.org/dac/evaluation/daccriteriaforevaluatingdevelopmentassistance.htm 35We interpret this as referring to the multiple steps involved in the current system which involve switching between technologies. Volunteers start by photocopying paper to collect community feedback in the field. After collecting feedback in the field they collate all of the pieces of paper and share them with the Red Cross who transfer the feedback into a digital format so it can be sent to the IFRC GO platform. Volunteers referred to the current process as cumbersome, inefficient and costly. 36Here we report the combined % of participants who selected 'a little positive' or 'very positive' in the post-workshop survey. The remaining participants (17%) reported neutral feelings about the tool (neither positive or negative). 37We posed the question in this way to capture broader attitudes towards AI throughout the project. Using this language also helped to ensure that the question was more accessible to stakeholders. 38The model was trained using eight rumour categories that were identified as priorities by the Cameroon Red Cross Society. 39As a starting point, the model should be retrained using different text pre-processing and transformer models. 40https://msf-listen.org/ 41Education in Emergencies in Protracted Crisis 2019-2023. https://www.africasvoices.org/wp-content/uploads/2021/04/AVF_FCDO-EiE-Final-Report-26th-February-2021.pdf Accessed 28 June 2022. 42The Premise tool is an example of an existing tool that makes use of verification mechanisms to ensure data quality. Fiorella Riccobono. Getting Started with Hyperlocal Data Collection. Premise. (https://www.premise.com/blog/getting-started-with-hyperlocal-data-collection/). Accessed 26 June 2022. 43In particular, problem definition, data prioritisation and helping them to think more robustly about deployment. See: Saurav Poudel. My Nesta Experience. Nepali Wanderer, 12 May 2022. https://nepaliwanderer.com/2022/05/12/my-nesta-experience-%ef%bf%bc/. Accessed: 26 June 2022. 44Inspired by a participant quote during evaluation activities in Nepal and the widely used slogan Nothing About Us Without Us. See: https://en.wikipedia.org/wiki/Nothing_About_Us_Without_Us. Accessed 26 June 2022. 45Berditchevskaia, A., Peach, K., Gill, I., Whittington, O., Malliaraki, E., and Hussein, N. (2021). Collective crisis intelligence for frontline humanitarian response. London: Nesta

58 Victoria Embankment London EC4Y 0DS +44 (0)20 7438 2500 [email protected] @nesta_uk www.facebook.com/nesta.uk www.nesta.org.uk

ISBN: 978-1-913095-65-9

Nesta is a registered charity in England and Wales with company number 7706036 and charity number 1144091. Registered as a charity in Scotland number SC042833. Registered office: 58 Victoria Embankment, London, EC4Y 0DS.

Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.