We can now look back on Nov. 30, 2022 — the day that OpenAI publicly released ChatGPT 3.5 — as a “Sputnik Moment” in the history of artificial intelligence.
Embraced by more than 100 million users in the first two months of its release, this AI large-language model illustrated the engaging power of a chatbot that could converse with users, provide a wide array of information in response to prompts and compete with traditional tools for internet search. ChatGPT and its progeny also accelerated the debate on AI safety and whether this powerful software program might one day pose an existential threat to its human inventors.
The AI safety conversation we currently are experiencing was ushered in by signed manifestos and resignations of researchers employed by the top AI frontier model innovators, such as Google, OpenAI and Anthropic. But the arguments for safety and counterarguments for unbridled regulation have evolved over the past 15 years in an intricate balancing act as the capacity of AI has rapidly advanced.
The pressure for AI safety advanced by Geoffrey Hinton, Yoshua Bengio, Ilya Sutskever and other scientists isn’t new. It reflects the sustained advocacy of many AI developers to align AI investment with societal goals. Calling for transparency and interpretability for AI models, these researchers seek to look inside the “black box” of neural networks to see how these Large Language Models actually “think.”
Safety-minded researchers have developed testing tools such as checklists to determine dangerous capabilities of new models before they are released. In fact, the 2017 Asilomar conference developed 23 principles for AI safety, which included protections for human rights and identification of potentially catastrophic AI risks. For example, one core principle calls for humans to control which powers are delegated to AI models and for training advanced models not to subvert our civic and economic processes.
The recent hack of Hugging Face by “runaway” AI agents underlines that the industry remains far from implementing and enforcing these enlightened safety principles. Clearly, the powerful forces urging breakneck development and global competition have sacrificed safety measures on the altar of gaining leadership and market share in the chatbot space. Yet even the leading CEOs competing for AI supremacy now recognize that they require public trust and support, otherwise the technology will fall out of favor of both consumers and investors.
Foreign competition also drives the competitive landscape, with the fear that if the U.S. doesn’t continue to build its lead in LLM capability, nations such as China will subsidize AI development and be the first to enable frightening scenarios involving creation of new weapons that will tilt the global balance of power. This argument has political strength, motivating the Trump administration to take a largely hands-off approach to AI regulation on the national level, and even to oppose localized safety measures on the state level.
As a result of these domestic and global forces, the speed-versus-safety tension continues to replicate itself with each new and more powerful AI model. Companies have delayed time-consuming model testing in order to rush their latest upgrades into the public market. This behavior needs to stop.
Rather than simplifying society’s choice as safety versus progress, we need to move toward a deeper understanding of the true incentives faced by AI companies. The leaders of the AI scientific revolution manifest complex motives, combining a desire to improve the human condition with a seeming disregard for the consequences of building machines that they don’t truly understand and might not be able to fully control. And the financial incentives to keep building these models might cloud their altruistic vision.
This paradox echoes the dilemma faced by physicists who worked on The Manhattan Project, such as Robert Oppenheimer, who ultimately decided to build the atomic bomb to help the U.S. win World War II. Oppenheimer famously concluded, “Technology happens because it’s possible,” prophesying that powerful tools such as nuclear weapons will be built, because humans are capable of building them.
Government efforts to coordinate AI development and impose safety measures have been mixed. The well-intended Biden administration’s October 2023 executive order sought to impose structure on the internal development practices of frontier models and foster ethical development. It would have required AI developers to conduct red-team safety tests and share the results with the federal government. It also called for the creation of evaluation rubrics for AI and to assess the impact of AI on labor markets, civil rights and national security.
Not surprisingly, President Donald Trump rescinded the Biden executive order and replaced it with his “Removing Barriers to American Leadership in Artificial Intelligence” order, seeking to stimulate more rapid AI growth, both as an economic boon to the U.S. and to maintain our perceived lead against China and other models. This dramatic regulatory turnaround illustrates that we can’t rely on a single administration’s priorities and should look to devise more durable incentives for AI leaders to pursue safety for internal reasons and to gain public trust.
While AI technology is still in a relatively early phase, we have learned enough to know that certain methods — such as mandatory checklists, increased transparency and reporting of results — can in fact identify dangers inherent in new models that will necessitate refinement and retraining prior to their public release.
Sensible regulation will not emerge from either doomsday or utopian views of AI. Rather, we need to assemble a diverse set of stakeholders to take a sober look at the capabilities and flaws of frontier models, in order to create a road map for sustainable AI safety. These safety measures should then be codified in laws that allow for continued technological advances, but at a more measured and safety-oriented pace.
