Technological development cannot be halted, nor would it be desirable for this to happen. However, an ethical and multistakeholder approach is needed from the start of the AI system project lifecycle to identify, assess, and respond to potential harms while weighing the benefits not just individually but also for the public good before it is released to the public on a large scale. 

In this context, UNESCO published a paper about Foundation Models such as ChatGPT through the prism of the UNESCO Recommendation on the Ethics of Artificial Intelligence. With this paper, UNESCO demonstrates how the lens of the Recommendation can help identify and clarify key ethical concerns related to AI foundation models such as ChatGPT. The report proved the procedural framework to address and mitigate these concerns, including via effective governance models and tools such as ethical impact assessment and complementary approaches such as ethics by design or research ethics committees. 

The report starts with an explanation of what ChatGPT is and why it matters. This part looks into the origin of ChatGPT and various characteristics relevant to the discussion in the paper. It also analyses the model's positive and negative opinions, thereby evaluating its ethical and social implications.  

UNESCO's Ethics of AI 

The UNESCO paper examines ChatGPT in relation to the principles and policy areas of the Recommendation. These encompass concerns related to fairness and nondiscrimination, inaccurate information, responsibility and accountability, safety and security, privacy and data protection, human and environmental flourishing, and education and research. These principles lead to concrete policy actions to ensure fair, inclusive, and sustainable outcomes from AI development. 

• Fairness and nondiscrimination: According to the report, stereotypical and discriminatory outputs are perhaps the most visible and controversial adverse effects of ChatGPT. These occur mainly due to algorithms replicating the biases contained in the data on which they are trained. The principle of Fairness and Nondiscrimination, outlined in paragraphs 28, 29, and 30 of the Recommendation, underlines the importance of safeguarding fairness and nondiscrimination in promoting social justice through AI systems. Specifically, it prioritizes the inclusion of all members of society, emphasizing people with disabilities, women and children, and all marginalized groups, with consideration for their specific needs and language requirements. 
• Economy and labor: Foundation models increase concerns about the impact of AI on labor markets and the speed and depth with which certain jobs will be transformed. Recommendation, Policy Area 10 on Economy and Labour stresses the need to invest in reskilling and upskilling workers, providing them with the tools and education needed to integrate AI effectively.  
• Transparency, explainability and verifiability: ChatGPT is opaque in relation to the data set that has been used to train them and the workings of the system in how it derives its answers. In the UNESCO Recommendation on the Ethics of AI, Policy Area 9 on Communications and Information calls for Member States to improve access to information and knowledge, to respect and promote freedom of expression and diversity of viewpoints, and to promote digital literacy skills. 
• Responsibility and accountability: Accountability is very hard to achieve, especially in the absence of governance frameworks that spell out exact accountability requirements. This delegation of responsibility is blind to the structural effects of ChatGPT and obscures the responsibility of the humans managing OpenAI and the companies developing generative AI products in general. Paragraphs 42 and 43 of the Recommendation state that "the ethical responsibility and liability for the decisions and actions based in any way on an AI system should always ultimately be attributable to AI actors corresponding to their role in the life cycle of the AI system" and demand appropriate oversight and measures to ensure accountability for AI systems and their impacts. 
• Safety and security: Paragraph 27 of the Recommendation states that safety and security risks must be prevented and eliminated from the entire life cycle of AI systems.  
• Privacy and data protection: Paragraphs 32, 33 and 34 of the Recommendation address the issues of privacy and data protection as crucial elements in defending human dignity, autonomy, and agency. Significant importance is given to enforcing national and international law in the collection, use, sharing, storage and deletion of data and the adoption of adequate data protection frameworks and governance mechanisms.  
• Human and environmental flourishing: Policy Area 5 on the environment and ecosystems urges Member States and companies to take responsibility for the direct and indirect environmental impacts of AI.  
• Education and research: In education and research, the debate has focused on the possibility of using ChatGPT to cheat on evaluations. According to the report, setting new guidelines entails the commitment of additional resources and potentially subtracting from time and resources available for teaching. 

Conclusion 

In the report, UNESCO has enumerated some of the concerns related to foundation AI models through the lens of the UNESCO Recommendation. As this analysis demonstrates, tools such as ChatGPT are not currently being designed, developed, and deployed in a manner compliant with the Recommendation. It remarks that the unleashing on the public of the current generation of "experimental" AI tools such as ChatGPT provides a prime example of why Member States must implement the Recommendation to identify, clarify, and mitigate the risks of harm from such models and in so doing govern these models responsibly.  

The analysis demonstrates the urgent need to examine generative models and their applications through the lens of the Recommendation's values, principles, and policy areas and to use tools such as ethical impact assessment to improve future iterations to limit risks and harms and identify and promote the social benefits of such systems.