Making AI Safe in the Real World: From Governance Principles to Operational Tools

Thomas Schneider argues that governance principles written on paper have no real value unless there are concrete tools and mechanisms to implement them. The central purpose of the session is to examine whether such tools now exist to bridge the gap between aspiration and practice.

He explicitly stated that 'the convention singles out a number of principles, but of course principles without implementation are just aspiration' and framed the discussion as being about 'the distance between the two and whether tools now exist to close the gap between principles written on paper and actually implemented' .

Schneider explains that the Council of Europe's Convention 225 does not create new standards but instead ensures that the human rights and democratic standards developed over the past century are applicable in AI contexts. This is described as a simple but intelligent logic that avoids duplication while filling a specific governance gap.

He noted that the convention 'refers to all national and international standards' and 'tries to make sure that the existing standards no matter where in which region in the world you are that you have on protecting human rights democracy and rule of law also apply to environments where AI comes into play' . He also pointed out that there is a call from governments, civil society, developers, and businesses for clear and reliable guidelines, with even Microsoft's Brad Smith cited as calling for a clear framework .

Schneider emphasises that Convention 225 is designed as a global instrument, not a purely European one. It does not reference anything explicitly European but instead draws on all national and international standards, making it applicable to any country that upholds human rights, democracy, and rule of law.

He stated that the convention 'doesn't refer to anything explicitly European it refers to all national and international standards so it's a global instrument' and that it is 'accessible to all countries in the world that respect human rights, democracy, and rule of law' . The convention had already attracted 21 signatures globally, including from the EU on behalf of 27 member states .

The convention sets out common principles and requires that certain mechanisms with the same functions be in place, but leaves it to each state to determine how to implement these based on their own institutional and legal traditions. This approach creates a form of interoperability between different national systems while respecting diversity.

Schneider explained that the convention 'doesn't tell governments exactly how they need to do it' and that 'it is up to every state to do this based on their institutional and legal history' . By 'defining the same principles and requesting the existence of mechanisms with the same function, it creates a certain level of interoperability between different systems' and 'brings everyone on the same level while allowing different ways of doing this' .

Schneider clarifies that while the convention's principles are directly applied to public actors and those acting on behalf of the public, states also take on a commitment to ensure equivalent protection when private actors use AI. This differentiation in responsibility is described as logical but inclusive.

He stated that 'states also commit to, and even in a more free way, but they commit to make sure that the same level of protection is also applied when it comes to... public, private actors using AI' and that 'it includes also a responsibility for states to make sure that people's rights and democracy and rule of law is also protected when it comes to private actors' .

Schneider describes the Huderia methodology as a well-structured, multi-level tool that caters to users with varying levels of expertise and need. Its three-tier design means it can serve both those seeking a high-level overview and those requiring detailed operational guidance.

He described the three levels: 'One is very general functions. The second level is the 20-page document that explains the different elements of the Huderia model also in a very general way, and then if you want to go deeper, you have like an 80-page document that really has questionnaires and goes into detail about how you could implement this' . He also noted it was developed not just by lawyers but by people from the technical standardisation community, academics, and business people, and tested on concrete cases .

Ayisha Piotti frames the session's discussion around four distinct but interrelated dimensions of the challenge of implementing AI governance principles. She argues that addressing only one or two of these dimensions is insufficient; all four must be considered together for effective implementation.

She outlined the four dimensions explicitly: the normative dimension (the Council of Europe's Convention and Huderia methodology), the country implementation dimension (Egypt's national context), the standards dimension (technical infrastructure for impact assessment and risk management), and the development finance and technical assistance dimension .

Piotti observes that despite its importance, the role of development finance in enabling AI governance implementation is rarely discussed in major governance forums. She calls for more dedicated attention to this topic and identifies it as a key takeaway from the session.

She noted that 'I've been in many many discussions so far in governance and that the whole institutional basically development finance angle is not that much that's covered' , and in her closing remarks she stated that 'development finance... there can be much more that can be done there' and that 'this is a topic that I don't see very much being dealt with, and I think it would be great to have some more sessions on that' .

In her closing remarks, Piotti synthesises the session's key messages by arguing that no single element — normative frameworks, financial resources, or technical assistance — is sufficient on its own. Effective implementation requires all three working in concert, supported by multi-stakeholder collaboration.

She concluded that 'in order for us to do that really, really well, we will need the normative side. We will need, of course, the financial, but we will also need the technical assistance side. And if we do that together in collaboration... try, and even if we fail, still try, take the risk. I think that's the way forward' .

Baraka argues that AI governance must be embedded at every stage of the AI system lifecycle, from design through to post-deployment monitoring, rather than being treated as an afterthought or a set of abstract principles. This requires integration into institutions, procurement processes, and the daily decisions of public officials.

She stated that 'AI governance cannot really remain like just a set of high-level principles or a policy paper. It has to become part of how government designs, procures, deploys, monitors, and evaluates' AI systems, and must be 'embedded also in institutions, in our procurement processes, in our sectoral implementation, in capacity building, and in the daily decisions of our public officials' .

Baraka's account of Egypt's governance journey demonstrates that the Council of Europe's Convention and Huderia methodology have practical relevance and utility for countries that are not members of the Council of Europe. Egypt used these instruments as key references when building its own national AI governance framework.

Egypt studied 35 different governance frameworks including 'the Council of Europe's Framework Convention, or AI, the Huderia, the UNESCO recommendation, the OECD AI principles, Hiroshima' , and Huderia was described as 'particularly useful because it helps move the discussion from the very abstract value to structured assessment' . Egypt also incorporated Huderia principles into its AI system guidelines and procurement guidelines .

Baraka highlights that Huderia's key contribution is its ability to translate abstract governance values into concrete, structured assessment questions that governments can actually use. By asking who the stakeholders are, what the context is, and how impact can be evaluated, it becomes a practical operational tool.

She stated that 'Huderia has been particularly useful because it helps move the discussion from the very abstract value to structured assessment. And it asks practical questions. What is the context? Who are the stakeholders? How can you actually evaluate the impact?' .

Recognising that public officials needed more concrete guidance than Huderia alone could provide, Egypt undertook a mapping exercise linking each Huderia process to the relevant ISO standards. This mapping was embedded in Egypt's procurement guidelines to give officials clear, actionable steps.

Baraka explained that Egypt 'did this kind of mapping between every process in the Huderia methodology and what kind of ISO standards well mapped into so that we can provide the public officials with the standards that they have to realize while procurement their AI system' . This was driven by the recognition that 'it's fine that we have the Huderia, this is a framework, but how can we implement it?' .

Baraka outlines Egypt's comprehensive and progressive approach to building AI governance infrastructure, which she describes as a 'full responsible AI governance stack'. This stack has been developed incrementally over several years, with each element building on the previous one.

She described the stack as including the National Council for AI established in 2019 , two editions of the national AI strategy with the second launched in January 2025 , the Egyptian Charter for Responsible AI aligned with OECD and UNESCO principles , and the Egyptian Centre for Responsible AI established the previous year .

Egypt took a comparative and evidence-based approach to developing its national AI governance framework by systematically studying 35 different international frameworks. The goal was not to replicate any single model but to understand the landscape and then adapt it to Egypt's own cultural, institutional, and legal context.

Baraka noted that Egypt studied '35 different governance frameworks' including the Council of Europe's Framework Convention, Huderia, the UNESCO recommendation, OECD AI principles, and Hiroshima , with the objective being 'not to copy, of course, any of these models, but to understand all of this and then make sure that we can localize it and contextualize it to our Egyptian culture and norms and institutions' .

Baraka argues that procurement is not merely an administrative function but a critical governance lever for embedding responsible AI principles into practice. Egypt developed specific procurement guidelines and an AI audit lab to help public officials understand and apply governance requirements in real procurement decisions.

She explained that public officials struggled to translate AI system guidelines into practice, leading Egypt to develop procurement guidelines that map Huderia processes to ISO standards . Egypt is also 'establishing an AI sandbox, an AI audit lab to train public officials on tools of what does it mean governance and what does it mean responsible AI' .

Baraka argues that AI governance and AI systems themselves must be localised to reflect the cultural values, norms, and languages of the communities they serve. Egypt's development of the Arabic LLM Karnak is presented as a concrete example of this principle in action.

She described how Egypt's Applied Innovation Centre developed 'an Arabic LLM, Ozkarnak' to provide a language model aligned with Egyptian and Arab cultural values , and noted that Egypt is 'talking now with a number of African countries on how to take Karnak to the African world' to contextualise foundational models for different regions .

Baraka presents Egypt's geographic and cultural position as a strategic asset in international AI governance cooperation. Egypt sees itself as a bridge between Africa, the Arab world, and the Mediterranean/EU, and is actively working to share its governance experience across these regions.

She stated that 'Egypt is always ready to act as a bridge between Africa, between the Arab region, between also the Mediterranean, the EU. I think our location is really in the middle of everything. So we are happy to help and to provide, to exchange our experience in this important emerging technology' .

Baraka acknowledges that while Egypt has been able to make significant progress on policy, guidelines, and capacity building using national funds, compute infrastructure represents a major financial barrier. The inability to fund a full cloud-based AI sandbox illustrates the broader challenge facing Global South countries.

She explained that 'compute infrastructure for us is a big challenge' and that while policy and capacity building work 'can be really done by our local national funds', hardware and compute infrastructure require different financing . Egypt was only able to establish 'a lab with a number of desktops with GPUs' rather than 'the full-fledged cloud-based AI sandbox because, of course, of funding' .

Baraka draws a clear distinction between what multilateral technical assistance can and cannot fund. While organisations like the World Bank and GIZ are useful for policy and capacity work, they do not address the hardware and compute infrastructure needs that are critical for full AI governance implementation.

She noted that 'with technical assistance, you can do some work related to capacity building, developing the curricula, having the national standards, having the policies' and that organisations like the World Bank and GIZ can help with 'reviewing, the revision' . However, for hardware, Egypt is 'exploring different ways of funding and financing mechanisms' to enable full implementation of the strategy .

Peter Deussen argues that one of the most important functions of international standards is that they speak the language of industry. This makes them an essential bridge between high-level governance principles and the practical realities of organisations developing and deploying AI systems.

He stated that ISO documents 'provide you with language that industry understands. It provides you with industry best practices. So, if you want to talk about companies who, you know, developing or, you know, bringing AI in the market, that's the language they understand' .

Deussen argues that Huderia and ISO standards are not competing frameworks but complementary ones that address different aspects of the implementation challenge. Huderia focuses on the substance of human rights and democratic impacts, while ISO standards provide the organisational process infrastructure needed to act on those assessments.

He described the two as 'like two sides of the same things. ISO gives you more like the connection to what to do actually within your organization... while Hedaria focuses more on the matter as such, on impacts or risks on human rights, democracy and rule of law' . He concluded that 'bringing these two things together... might be a very good approach... to make it more easy for organizations to implement Huderia' .

Deussen outlines the suite of AI-related standards developed by the ISO/IEC Joint Committee SC42, explaining how each addresses a different aspect of responsible AI governance. Together, these standards form a comprehensive framework for organisations seeking to manage AI responsibly.

He described the development of a risk management document for AI , the AI management system standard ISO 42001 providing 'general requirements on how to set up processes within an organization on how to deal responsibly with artificial intelligence' , and a standard on AI system impact assessment 'which explicitly focused on impact of AI systems to persons, to individuals, to group of individuals, and to society in general' .

Deussen identifies a specific gap that ISO standards fill which Huderia does not address: the internal organisational processes needed to act on impact assessments. ISO standards specify what review cycles, approval processes, and documentation are needed to embed AI governance within an organisation.

He explained that ISO's impact assessment standard 'talks about what kind of review cycle do you need? What kind of approval processes do you need? What kind of documentation would be appropriate to understand this?' , providing 'the connection to what to do actually within your organization' that Huderia does not offer .

Deussen identifies a significant remaining gap: while ISO has developed comprehensive standards for industry, these have not yet been translated into guidance that policymakers and regulators can readily use. A bridging document or mapping between the two worlds is needed.

He stated that ISO 'have not done yet... to describe how these documents can be used in a regulatory context, how these documents can be used by policymakers' , and suggested that 'having some kind of guidance document, some kind of mapping that, you know, bridging between those documents and the regulatory world... might be a very useful way' .

In response to an audience question about the lack of structured information on AI deployment, Deussen argues that standardisation bodies can play a useful role in defining the structure and content requirements for an AI information database, even if they cannot populate such a database themselves.

He stated that 'what ISO can provide is the structure of such a database, explaining what kind of information put into this database and how to maintain it, how to organize it' , while acknowledging that 'standardization cannot provide you with a catalog about all industry applications of artificial intelligence' .

An audience member argues that the fundamental barrier to effective AI governance is not a lack of guidelines but a lack of standardised, accessible information about where and how AI is actually being deployed. Without this foundational information, governance frameworks cannot be effectively applied.

The audience member stated that 'we don't have information where exactly is used AI in any different industries and companies, how particularly they are using, and what is the interaction between the human oversight and AI as an autonomy level' , and asked whether 'this kind of information in a standardized way... can build the gap of foundation for governance and for action taking' .