Inside Germany's drive to democratize AI

SAN FRANCISCO — Hundreds of volunteers across Rwanda have spent hours recording their voices during the COVID-19 lockdowns as part of an effort by the German development agency and other partners to improve speech and voice recognition for Kinyarwanda, the national language of Rwanda. Last year, GIZ launched an artificial intelligence initiative called FAIR Forward - Artificial Intelligence for All, as part of its Sector Programme Digital Development. GIZ, the implementation organization of Germany’s Federal Ministry for Economic Cooperation and Development, or BMZ, is working around the world to strengthen local expertise in AI, remove barriers to entry with a key focus on language data, and support policy frameworks that are ready for AI. In May 2019, GIZ, in partnership with the government of Rwanda, launched the Digital Transformation Center in Kigali, Rwanda, with the goal of supporting digital solutions, strengthening the local ecosystem for innovation, and scaling up digital solutions across the region. “Language is the future of human-machine interaction.” --— Lea Gimpel, co-lead, FAIR Forward At the same time, Germany was beginning the implementation of its own AI strategy, which articulated the importance of guaranteeing “a responsible development and deployment of AI which serves the good of society.” Leaders within BMZ saw the synergy between these efforts and launched FAIR Forward, along with partner countries Rwanda, Uganda, Ghana, South Africa, and India, in 2019. “The ministry decided ‘Let’s connect those two stories.’ That’s how FAIR Forward was born,” Balthas Seibold, co-lead of the initiative, told Devex. Several things set GIZ apart from other bilateral donors supporting AI work, said Lea Gimpel, the other co-lead of FAIR Forward. “We focus on building markets and structures for local AI development and emphasize long-term sustainability,” she said in an email to Devex. In its work on AI, GIZ focuses on skills, open AI training data, and policy frameworks, rather than building applications for use in an international development context or supporting the collection of open data sets that go unused, Gimpel continued. “We aim to enable local developers to build solutions to address local challenges,” she said. GIZ has 500 projects implementing digital technologies, Björn Richter, who leads GIZ’s digital development work, said in a panel during the virtual Concordia summit. While GIZ is advising 15 partner countries on digital transformation, the agency has discovered that in many cases it has as much to learn from its partner countries as they do from Germany, Richter said. “We are all learners in the area of digital for development,” he explained during the panel. “We are here for a joint walk.” GIZ believes in taking an ecosystem approach, working not just with digital ministries, but also with startups, tech companies, and civil society organizations in its digital transformation efforts, he said. Germany is prioritizing Digital4Development, or D4D, as a key priority of its presidency of the Council of the European Union, and the country is partnering with Belgium, Estonia, France, Luxembourg, and their development agencies to launch a pilot Africa-Europe D4D Hub next year. Germany seeks to promote what Richter describes as a “human-centric approach” to digital transformation that prioritizes “the digital sovereignty of the partner countries,” which he describes as “benefiting locally” even if it is “steered globally.” “We need to democratize AI, otherwise we will be in deep trouble.” --— Balthas Seibold, co-lead, FAIR Forward Opening up voice data In June, FAIR Forward joined forces with the Rockefeller Foundation, Google.org, and Canada’s International Development Research Center to launch the Lacuna Fund to expand AI training data in low- and middle-income countries. The fund aims to unlock the power of machine learning globally by providing aspiring machine learning users with resources to produce labeled datasets in areas including health and agriculture, Daphne Luong, director of Google AI, told Devex. GIZ directed its support of the Lacuna Fund specifically to African language data. “Language is the future of human-machine interaction,” FAIR Forward co-lead Gimpel said. FAIR Forward focuses on languages for two reasons, she said. First, local language interfaces can enable underserved access to information across multiple issues related to the Sustainable Development Goals including health, education, agriculture, or financial inclusion, Gimpel said. “We see language technologies as an enabling technology for many applications,” she said. Second, there is a lower risk profile associated with language data and language technologies than other applications of AI, Gimpel said. Speech recognition has revolutionized the way people interact with devices, but the benefits of this software are not distributed equally. A few technology companies handle the majority of speech recognition interactions, often using proprietary data. Their speech recognition services are only available in major languages, and they face a number of challenges, such as understanding people with different accents, due to “the biases within the data on which they were trained,” said Remy Muhire, community lead for voice technology at Mozilla, which is partnering with GIZ and the Rwandan startup Digital Umuganda to upload Kinyarwanda to Common Voice, an open-source voice database. He pointed out that none of the dominant voice assistants support a single native African language. “GIZ has helped to create an ecosystem around voice tech in Rwanda by hosting meetups, workshops, and hackathons,” he said in an email to Devex. “They've also provided mentorship and spaces to companies.” The collaboration between Mozilla and BMZ on Kinyarwanda is the start of a larger partnership to open up voice data for other African languages. Democratizing AI The FAIR Forward team sees a role for donors to play in ensuring the benefits of AI extend to the most marginalized in society: “We need to democratize AI, otherwise we will be in deep trouble,” Seibold said. He said FAIR Forward is part of an effort within GIZ to go beyond using digital tools to work on building digital programs. There are a large number of pilots demonstrating the many ways AI can be used to advance progress on the SDGs, but a lack of efforts to build ecosystems for these AI use cases, he noted. “Where can we create use cases and models in this field of democratizing AI that can then be scaled?” he asked. For example, building on the Kinyarwanda voice data work, the Rwandan government might build a COVID-19 chatbot, a startup might build crop disease hotline, and a Rwanda-based NGO might do an advocacy campaign that could reach illiterate people, Seibold explained. “The idea is not to just have one chatbot that understands Kinyarwanda,” he said.