From Hugging Face to Model Citizens: Why AI Governance Is Having Its 2018 GDPR Moment

I’ve been speaking recently about AI ethics and risk to different audiences; from compliance and legal professionals, through online dating executives to European marketers. I used this privileged position to ask each group a couple of questions: their answers have me thinking we are both still in the early stages of AI’s evolution to a truly transformational,  general purpose technology, and how urgent it is to get some governance structure around to support this now, before the storm. 

My questions stemmed from a gut feeling that we are having a ‘governance moment’ where previously specialised topics suddenly break out into mainstream consciousness. My surprised observation that people ‘In Real Life’ know what is meant by the words “hugging face” is supported by what they search for on Google - “Spain AI data breach”, “Microsoft Anthropic AI Safety” and most tellingly, “What does AI know about me?”

I tested this with live audiences, asking my first question - “what links King Charles III, Christine Lagarde and Pope Leo XIV?”. The answer is, they have each issued urgent, public warnings on AI, from moral and social dangers to data sovereignty, economic threats and climate change. 

Every audience responded with a good number of raised hands from folks who confidently made this connection, while pointing out I should add Bill Gates to the list. Such themes around AI risk and governance are clearly not just the musings of Silicon Valley insiders, but of concern to leading figures across society and most importantly, the public. It reminds me of the days just before GDPR where I overheard two old ladies on the London bus discussing privacy and data rights, and realised we had reached a tipping point.

It does seem that public and AI leaders are on the same page believing it is advancing too quickly, especially with the arrival of RSI, Recursive Self Improvement which sets up a compound loop inside AI models. Media debate has focused on a 4,000-word essay by Dario Amodei, CEO of Anthropic, entitled “We must pace the frontier”. 

This led to my second question: “who here has read the full essay?” The result was very different; across 3 events and over 250 professionals, only one hand was raised (a global DPO, as it turned out). 

Before you think I'm putting myself on a moral platform here, I must share my first thought was to take the essay and ask Germini to condense it down to 250 words, using what it already knows about me. But doesn’t that make me part of the problem? Asking AI to tell me what I need to know about the unintended consequences of AI seemed a bit fishy. So I took the advice of my teenage daughter to ‘touch grass’, read the essay and considered it in the old-fashioned way. 

So what did I learn? Three things struck me as both familiar and an evolution from the GDPR moment, starting with the concept of Evaluators. This role clearly embodies the principles of Privacy by Design, where privacy and safety measures are embedded in the build of systems and processes from day one and continue to apply by default. 

In Amodei’s approach, Evaluators are independent, but enjoy the same physical and systems access as an internal staff member. Complementing GDPR, they can speak truth to power with an unfiltered right to publish their findings, and be protected as whistleblowers when they expose risky or unfair practices. This is a crucial bridge between ethical statements and practice, and an exciting evolution of the DPO role.

The second pillar is a focus on operational excellence, processes and tools to manage risk. Many process flows and checkpoints Amodei calls out are well-known features of the Software Development Life Cycle, and particularly cyber security. My wife is a cyber security professional and made the observation the other day, that cyber folks understand a lot more about AI, than many AI folks do about cyber. 

Like cyber, this approach requires organisations to cross silos with the goal of getting assurance, not reassurance that risks are being identified, managed and monitored and there is a plan when, not if something goes wrong. As privacy professionals, we can work with cyber allowing data and cases to flow through new AI systems and analyse how these processes are working. 

This comes down to auditing the codebase in Amodei’s approach, which means both understanding of technical requirements by Evaluators and DPO’s, and dedicated tools for diagnosis and oversight. As someone involved at the birth of what we now call PrivTech, I’ve been frustrated by providers stuck in narrow regulatory compliance lanes instead of serving both consumers and institutions. AI presents huge challenges in scale and speed, but can give us new solutions: moving beyond cookie banners to automated, agentic systems that integrate seamlessly with gatekeepers. 

The final, most important  part of the Pacing Essay is the need for transparency – again something familiar from GDPR, but only partly realised since 2018. Similar to the “what goes on Apple stays on Apple” slogan, privacy and transparency are seen by AI leaders as offering a long-term competitive advantage. This ‘race to the top’ gives us the opportunity to revisit concepts which struggled to gain traction around GDPR, such as the model card, or nutrition label for AI algorithms. 

In an industry already famous for opacity and impenetrable jargon, model cards translate multi-page technical documents into digestible metrics which can, crucially be compared and audited across sectors and demographics, revealing biases and risk blind spots. In the hands of an experienced evaluator or marketer, model cards generated directly from development and risk platforms become powerful artefacts which can make the jump into public awareness like food nutrition labels have done. This gives consumers the transparency, confidence and trust they need to return on businesses’ investment in AI.

If it seems like a full-time job to keep up with just the latest developments in AI, let alone the opining of Amodei and his peers, that’s because it is one. We must focus first on just the key facts needed to leverage our own experience, and work with others who have sector or vertical expertise to mitigate AI risks. I believe model cards could help us in this task and crucially, get AI deployers and not just builders on the record, while bringing the public along. Do reach out to me if you have thoughts on this topic or would be interested to collaborate…