IGF 2023 WS #313 Generative AI systems facing UNESCO AI Ethics Recommendation

As moderator of the panel, Yves POULLET explained what we call “Generative AI”, why these systems and their applications might be fruitful for the society but also encompass a lot of risks not only for our individual liberties but also for societal concerns. He raised the question to the audience “who has already used generative AI systems?”, to which a majority of persons answered positively.

He explained that initially, OpenAI proposed to restrict the use of ChatGPT only to professional users because it could be dangerous if deployed to the general public. However, OpenAI launched a general public application in 2022 and since then, many companies have developed foundation models as well. The foundation models are general purpose models and are not per se developed for specific purposes. We spoke about ‘transformers’, Large language machines, multimodal generative systems, …. Dall E, Chat GPT, Midjourney, Bard, ERNIE 3.0, KoGPT. A lot of generative AI applications derived from foundation models are possible.

Gabriela RAMOS, Assistant Director-General for the SHS UNESCO department in charge notably of the AI Ethics Recommendation, delivered initial remarks, introducing the UNESCO Recommendation on the Ethics of AI, and the recent UNESCO report on ChatGPT comparing the latter with the provisions of the recommendation. She enumerated the various issues generated by generative AI systems and claimed for more reflexion, action and initiatives in this field. She pleaded for governance of these technologies. On that point, she introduced the impact assessment tool developed by UNESCO.

Dawit BEKELE introduced the technical peculiarities of Gen AI: they are designed to generate human like content. They generate coherent and context-related inputs based on the input they receive. He pointed out that they are used on large scale platforms. He then stressed the numerous benefits of generative AI systems as they can be used directly for filtering, rewriting, formatting, etc. They are a major factor of risks for our societies because of potential misuses, creation of harmful content, the fact that many people trust what they see online, risks to information and education, risks for jobs (writers, etc.). Moreover, because of the bias challenge (biased data sets), some countries banned their use. The applications of language models are diverse and include for instance text completion, text-to-speech conversion, language translation, chatbots, virtual assistants, and speech recognition. They are working with big data from a lot of different sources: public ones (as Wikipedia, administration data bases, …) but also from private ones. However, these resources are not representative of the whole world.

The first roundtable focused on the Generative AI governance and faced the following questions: Do we need a regulation? What do you think about soft law based just on ethical recommendations or voluntary code of conduct? What do you think about the “moratorium” requested by certain companies and academics?  What about a global regulatory model as UN is thinking?

Changfeng CHEN first mentioned a concept: cultural lag. Cultural lag is a term coined by sociologist William F. Ogburn to describe the delayed adjustment of non-material culture to changes in material culture in the 1920s. It refers to the phenomenon where changes in material culture (such as technology, tools, infrastructure) occur more rapidly than changes in non-material culture (such as beliefs, values, norms, including regulations). She applied this concept to generative AI. Under her opinion, first, we need regulation for generative AI because it is a powerful technology with the potential to be used for good or for harm. But generative AI is still developing, and the scientists and engineers who create it cannot fully explain and predict its future. Therefore, we need to regulate it prudently rather than nip it in the cradle through regulation. Furthermore, we need to be more inclusive and have the wisdom to calmly deal with the mistakes it causes. That only shows human civilization and self-confidence. Second, a moratorium on generative AI being a temporary ban on the development and use of this technology, it would be a drastic measure. It is unlikely to be effective in the long term. Generative AI is a powerful technology with the potential to be used for good, and it would be unwise to stifle its development entirely. Third, a global regulatory model for generative AI would be ideal but it will take time to develop and implement. AI, including generative AI is developing very rapidly in China and has been widely used. Fourth, she explained that China has been at the forefront of developing and regulating generative AI. In 2022, China released the Interim Administrative Measures for Generative Artificial Intelligence (AI) Services. It was published in July 2023. These measures require providers of generative AI services to: Source data and foundation models from legitimate sources; Respect the intellectual property rights of others; Process personal information with appropriate consent or legal basis; Establish and implement risk management systems and internal control procedures; Take measures to prevent the misuse of generative AI services, such as the creation of harmful content.

Stefaan VERHULST stressed the importance of a responsible approach and raised the question of the extent to which the development of AI should be open or closed. He praised for open data and open science to avoid digital asymmetries. He underlined the fact that the US is again member of UNESCO and endorsed its Recommendation on AI Ethics. He pointed out that the principles underpinning the ethical values are aligned in multiple documents: the US Blueprint for an AI Bill of Rights, the UNESCO Recommendation on AI Ethics, the EU documents, etc. The US approach is based on a co-regulation approach. He stressed the need for notice and explainability rather than a complete regulatory system. On that point, he underlined the fact that States in the US are much more active in working on legislative regulations than the federal authority. Cities are particularly active as well in this field and Stefaan Verhulst underlined the interest of this bottom-up approach more respectful of local disparities and permitting a real participation of the citizens.

During the Q&A, Omor Faruque, a 17-year-old boy from Bangladesh and the founder and president of Project OMNA suggested for the policy questions as a global child representative to: 1. Establish clear ethical guidelines for the development and use of foundation models, with input from all stakeholders, including children and young people; 2. Create a public registry of foundation models, including information about their ownership, purpose, and potential risks; 3. Develop mechanisms for public oversight and accountability of foundation models; 4. Convene a global forum on generative AI to discuss the ethical, legal, and social implications of this technology; 5. Support research on the impacts of generative AI on everyone including children and young people; 5. Promote digital literacy and critical thinking skills among children and young people so that they can be informed users of generative AI; 6. Consider a moratorium on the development and use of generative AI systems until appropriate safeguards are in place to protect children and young people from potential harms.

Steven VOSLOO (UNICEF) stressed that UNICEF is also concerned that we don't yet know the impacts of generative AI (positive and negative) on children's social, emotional and cognitive development. Research is critical, but will take time. He asked how we can best navigate the reality that the tools are out in the public, and we need to protect and empower children today when we'll only fully know how later.

Torsten Krause said that responsibility is a question for all and not only for children and asked if we would need official institutions’ certificates or permission before distribution of technologies likes generative AI systems.

On this matter, Stefaan Verhulst and Changfeng Chen agreed that young people must be involved.

A Finnish attendee stressed that it would be complicate to regulate a technology that is used by the general public.

Doaa Abu Elyounes, Programme Specialist in the Bioethics and Ethics of Science and Technology section, said that it is of course tempting to use these systems because they allow to be faster in writing for example, and therefore that we should be more aware of the risks involved.

The second roundtable was dedicated to specific socio-economic topics linked with Gen AI. First, the problem of non-representativeness of certain languages in big data excluding certain populations and creating a cultural dominance. Second, the fact that the use of most of these generative AI applications are based on a business model which requires payment for the proposed services. 

Siva PRASAD stressed that people developing the technologies look for profit and accentuate the digital divide because they are not interested in populations which are not a source of profit. The use of technology is affecting social innovations, and it is the role of public authorities to pay attention to the digital divide, especially as regards marginalised communities. He evoked the specific problem of the use of generative AI in the education system and underlined the right for young people to use the technology for building up their personality and the obligation of the teachers to help to be aware of the risks. Referring to the opinion of Stefaan Verhulst, he said that the local approach is the only way to develop sustainable and equal societies. The international and national authorities have to finance this local approach. On that point, he asserted that while there is a universal bill of rights, local answers are needed.

Fabio SENNE focused on aspects of measurement of the socioeconomic impacts of AI and data governance. He called for attention to the current scenario of digital inequalities and how it shapes the opportunities and risks related do generative AI. Disparities among countries and regions can impact diversity in data used to train AI based applications. In the case of Brazil there is not only the availability of content in Portuguese, but also in more than 200 indigenous languages. Digital inequalities affect diversity in data used to train models. In terms of access to connectivity and devices, we see persistent patterns of inequalities (ethnicity, traditional population, rural/urban, income, level of education, age). Diversity and inclusiveness have to be principles. The use of Generative AI tools also can be affected by poverty and vulnerability [e.g. income]. Early-adopter tend to benefit more when a new application is available. The impacts tend to be more disruptive in the early phases of dissemination of those tools. Fairness and non-discrimination/ inclusive access for all are also principles.

As concluding remarks, the panellists indicated what are the more crucial issues raised by generative AI in their opinion. We summarised these in the key takeaways and call for actions (see above).

As final remarks, Marielza OLIVEIRA thanked the Working Group on Information Ethics of IFAP-UNESCO for this panel and developed recommendations as regards the IFAP and UNESCO future works. She addressed to the IGF the need to go forth with that topic as a major challenge for all people and our society. She pleaded for continuing the discussion and trying to solve the delicate issues raised by these technologies.