From scapegoat to spotlight - Regulation is not all you need.

Why we need AI to achieve institutional transparency.

I started writing this text sometime in the first quarter of 2021. I spent countless hours editing it to try to comply with the unclear, inconsistent and changing demands of the leader of an ethics consulting group who had promised to publish it in their newsletter - a promise that was completely empty as, in their own words, "they like regulations".

More recently, as of April 4th, 2023, letters and calls for moratoriums and regulations have become the?loudest voices in the ChatGTP/LLM tsunami that has been flooding social media since November . Nonetheless, the question of practical implementation/enforcement was not explicitly addressed by these voices. Instead, they just implied that this task, incumbent on governments and their institutions, would just magically go well. None seemed concerned then by the obvious regulatory capture threat of Tech Bro Sam - at least within my 'reasonable' friends and professional contacts.

As of October 2023 - it has become clear that this is taking the form of an executive power movement. I feel both worried by this evolution, which had been latent and intangible for a while but is now concrete, and reassured that reasonable and respected voices now publicly share a similar concern. This compelled me to share this text despite its messy incompleteness in the hope that it may contribute some ammunition against what I believe is the biggest threat associated with AI.

Considering the status of regulated technologies and their history, it is reasonable to wonder whether regulations are the best way forward to ensure the development of generative AI benefits humanity or whether it is a trap humanity cannot afford to fall for.?

Technology has been a significant partner in the evolution of our species and our civilization. It has given us the ability to control fire, build shelter, and access new sources of food. It has given us the agricultural revolution and, from there, organized society, division of tasks, and the ability to achieve greatness and destruction at scale.?

Information technologies are tightly linked with our main developmental milestones. And information technologies do not mean only the technology from the digital era. Information technologies started with painting caves, sculpting Venus statues, hollowing stumps or carving bones to make music, and later writing and knotting quipus. These allowed us to expand our communication abilities beyond our biological voices and beyond our instant sensory system and the “here and now” by giving us records. This gave birth to math, sedentarity, governments, taxes and public administration, turning the neolithic revolution into History and founding its institutions.?

The printing press was the first significant example of disruptive information technology. More than any other in history, Gutenber’s invention left its mark by triggering and snowballing the Reform, the Renaissance, the Scientific Revolution, the Enlightenment and the political revolutions that followed shortly. It also spread the adoption of the scientific method and the sped-up circulation of data and scientific knowledge, which in turn supported the Industrial Revolutions. The productivity gains of the industrial revolutions have also supported further the growth of governmental institutions. Growth in terms of scope, by adding more welfare and education to their duties, and in terms of means, by appropriating information technologies for their own activities, developing along the way a proverbial mastery of some technologies like the rubber stamp, carbon-copy paper (millennials and GenZ may not know what "CC" stands for in an email), paper templates and forms, faxes, phone servers and automated answering machines.

The actions of the “Etat Providence ”, ancestor of the Welfare State , needed robust and scalable industrial and business processes. Governmental standards - and the institutions that guide their establishment, emerged in the early 20th century. This scope kept growing, extending the concept of standardization beyond measurements, imposing more and more norms and process regulations throughout many industries and services, such as pharmaceuticals and medical devices (which happen to be my domain of expertise).

In these domains, the constant rise of life expectancy across all developed nations, the ban of several dangerous snake-oil substances or the eradication of deadly diseases like smallpox are often presented indiscriminately like the perfect success stories illustrating the effectiveness of regulations. Yet, there is little objective and quantitative evidence of the specific roles that regulations have played in these collective achievements. Regulations are only a part of a complex system involving public and private policies and stakeholders, researchers, clinicians or logisticians? - and whilst the correlation between the implementation of sophisticated regulations and these achievements is irrefutable, correlation is not causation. The history of modern regulated medical technologies cannot hide many shortcomings, from despicable research practices to obvious failures of the most basic features of regulations (another recent example, still unfolding, is currently affecting millions of sleep apnea sufferers). In Europe, the roll-out of the new Medical Device Regulation? (MDR), voted in 2017, has recently been pushed back to 2028 in an urgent and desperate move necessary to avoid cutting off European patients and doctors from accessing essential technology.

Whilst these whimsical regulatory hurdles are costing billions to patients and the health systems (in case it is not obvious, the industry is just passing these costs to them), evidence shows that regulatory processes contribute more to maintaining market-entry barriers and competitive advantage for existing players than to promoting innovation.

MedTech and pharma are no strangers to the disruptive impact of the Artificial Intelligence Revolution, and AI/ML applications have been front and centre for many forums, professional bodies, investors and innovators in this space. Image classifiers have given existential doubts to radiographers since 2017 at least, whilst the most recent explosion in the field of generative AI, conversational AI and large language models have fueled an uncontrollable afflux of health apps in mental health , clinical decision support and clinical training, only to name a few. The definitions and rules on which traditional MedTech regulations are based are not appropriate to control this afflux .

Independently from technologies traditionally regulated like MedTech and pharma, the recent progress of generative AI applications has raised both public awareness and adoption of the field to new heights. Surfing on the backend of the image generation hype that peaked a few months earlier, two generations of ChatGPT products have beaten new records of scale and reached millions of businesses and end-users within just a few weeks. But besides their massive-scale adoption, the prowess achieved by these applications has also led to a massive interest in their safety and the new risks they raise.

Further triggered by apocalyptic scenarios involving Artificial General Intelligence takeover and gruesome stories like the recent suicide of a chatbot user, a consensus seems to be emerging within the industry, opinion leaders, advocacy groups, all levels of government and intergovernmental agencies to call for more regulations of AI, more guidelines, more frameworks and more declarations of principles. Nonetheless, whilst regulations are only starting to be signed into law, frameworks, guidelines, and other standards have been plenty already.

Are regulations and moratoriums the best solution we - creative humans - can come up with to face this challenge? As we have seen above, the effectiveness of the former is yet unproven in regulated industries . And, in our global economy, where tech and social media companies are knowingly harvesting their profits from the mental health capital of millions of teenagers, I think it is delusional to believe that either can be effective. Evidently, the development of larger models is not going to stop, regardless of how many signatories agree on a moratorium, whether its scope is research or deployment, and whether governmental institutions attempt to circumvent these developments. Many sizable countries and unscrupulous individuals and corporations everywhere are completely unaffected by whatever agreement other actors may reach. Does this mean giving up and just bracing for what’s to come? On the contrary, I believe this is an opportunity for our civilization to change our perspective entirely and adopt a new social and institutional paradigm.?

***********************************************************

In the current status quo, artificial intelligence is cast in the role of “Scapegoat AI”.

Like its biblical counterpart, it is the perfect recipient of our sins. And, as mentioned earlier, there are many calling for it to be sent, with our sins, into a developmental desert far from our view. “Bias? Inequities? Injustice? Let’s blame all that on these mad scientists; let’s close the lid now and pat ourselves on the back for our efficient institutions.”?

Instead, AI needs to be given a new role as “Spotlight AI”. A spotlight is a tool, first and foremost. An exploration tool. In the case of AI - Large Language Models and other AI/ML applications in particular, the most interesting and fascinating characteristic of these tools is its ability to explore us - humans. Whilst they lack many elements of human (or other biological) intelligence, their capability for churning gigantic amounts of human-generated data can detect and isolate patterns intrinsic to our species’ biological and social characteristics beyond our explicit understanding.?

An algorithm producing biased results is like a microscope showing pathogens in our blood.

Consider bias, inequities, injustice, and the temptation to invent an answer when not knowing the result. These fundamental sins of our societies have existed before advanced technology. And more often than not, they are deeply embedded in our institutional culture, which in turn maintains and even reinforces them.

When the results of the research conducted on algorithmic bias reached the public mainstream , what shocked me the most were not the research results themselves. What shocked me was that people were acting like they were surprised by these results. Like the widespread adoption of camera cell phones exposing to the masses the exactions of many American police forces, AI was a revelator, not a cause. When an algorithm produces a harmful result, there is a very long and complicated chain of rationale that explains this result and the harm behind it.?

Mandating regulations that force developers to use training datasets that have been ‘unbiased’ may sound reasonable at first sight. But, when taking into consideration the historical context, this appears as an ineffective corrective measure. "Corrective", in that it only addresses the symptoms, not the root cause. And ineffective because as seen above, the institutional processes that create these regulations are working on timescales that cannot cope with the pace of AI development. Moreover, it is paradoxical to expect governments to promote, implement and enforce such regulations when their track records show that they are the worst offenders .?

Humanity is facing significant existential challenges. Besides the “AGI takeover” doomsday scenario, other issues such as global warming, mental health epidemics , ethnic cleansing and the threat of a major nuclear conflict are challenges that our institutions have not been able to address through regulations.?

And while civil nuclear safety and other large regulated industries are often used as an argument FOR regulations, I disagree. Regulations did not prevent Chernobyl, Fukushima, Challenger and much too many deaths on the roads. Rather than wasting resources on the desperate bet that such regulations could impact AI development beneficaly, I believe that AI shall be the tool that allows us, social humans, to shine a light on our institutional and regulatory shortcomings and, from there, create another inflexion point on our the social progress, like Gutenber’s printing press, on the path to a more transparent future.?

And this is an opportunity. It is an opportunity for us humans to codesign our institutions. Bennett and Roy demonstrated eloquently in Our Transparent Future how our institutions have reached their current development through an evolutive process - not through design. This is evident when considering the lack of theoretical premises on which regulations are built.? Our institutions need to rely on solid technological foundations to remain relevant in a world where the pace is set by technological development, and this is a role more suited to engineering design than to evolution.

Now that we have reached the third level in our own evolution (I paradoxically love this concept from Max Tegmark ), it is time to engage in this collaborative design exercise as a species. "Design" aims at addressing unmet needs. Human needs, in priority to institutional needs. The threat of technological singularity is a fantastic opportunity for a social and institutional singularity. Honest introspection is a necessary tool to understand and, therefore, treat the root causes of mental illnesses. Similarly, the adoption of transparent, spotlight AI is an essential exercise for our civilization to understand its needs and drive the design of - rather than ‘evolve into’ - the next stage of its existence.