Democratising AI: the promise and pitfalls of open-source LLMs

The session focused on the potential of open-source large language models (LLMs) to democratise access to AI, particularly in fostering innovation and empowering smaller economies and the Global South.

Adult, Female, Person, Woman, Male, Man, Chair, Furniture, Tripod, Indoors, Couch, Accessories, Jewelry, Necklace, Bag, Handbag, Head

At the Internet Governance Forum 2024 in Riyadh, the session Democratising Access to AI with Open-Source LLMs explored a transformative vision: a world where open-source large language models (LLMs) democratise AI, making it accessible, equitable, and responsive to local needs. However, this vision remains a double-edged sword, revealing immense promise and critical challenges.

Panelists, including global experts from India, Brazil, Africa, and the Dominican Republic, championed open-source AI to prevent monopolisation by large tech companies. Melissa Muñoz Suro, Director of Innovation in the Dominican Republic, showcased Taina, an AI project designed to reflect the nation’s culture and language. ‘Open-source means breaking the domino effect of big tech reliance,’ she noted, emphasising that smaller economies could customise AI to serve their unique priorities and populations.

Yet, as Muñoz Suro underscored, resource constraints are a significant obstacle. Training open-source models require computational power, infrastructure, and expertise, which are luxuries many Global South nations lack. A Global South AI expert, Abraham Fifi Selby echoed this, calling for ‘public-private partnerships and investment in localised data infrastructure’ to bridge the gap. He highlighted the significance of African linguistic representation, emphasising that AI trained in local dialects is essential to addressing regional challenges.

The debate also brought ethical and governance concerns into sharp focus. Bianca Kremer, a researcher and activist from Brazil, argued that regulation is indispensable to combat monopolies and ensure AI fairness. She cited Brazil’s experience with algorithmic bias, pointing to an incident where generative AI stereotypically portrayed a Brazilian woman from a favela (urban slum) as holding a gun. ‘Open-source offers the power to fix these biases,’ Kremer explained but insisted that burdensome regulation must accompany technological optimism.

Despite its potential, open-source AI risks misuse and dwindling incentives for large-scale investments. Daniele Turra from ISA Digital Consulting proposed redistributing computational resources—suggesting mechanisms like a ‘computing tax’ or infrastructure sharing by cloud giants to ensure equitable access. The session’s audience also pushed for practical solutions, including open datasets and global collaboration to make AI development truly inclusive.

While challenges persist, trust, collaboration, and local capacity-building remain critical to open-source AI’s success. As Muñoz Suro stated, ‘Technology should make life simpler, happier, and inclusive, and open-source AI if done right, is the key to unlocking this vision.’

All transcripts from the Internet Governance Forum sessions can be found on dig.watch.