The Singularity and Externalities

The unique dangers of artificial intelligence are often used to argue that development should be taken out of the hands of profit-driven private actors. The government is presented as a force for good that can develop AI responsibly for the common good. While this line of reasoning can soothe the minds of well-meaning citizens, it fails to take into account the nature of the externalities involved. Government AI development would force the people to bear the massive risk that frontier labs are taking on. While OpenAI and Anthropic believe that this risk will pay off, the people do not have the information or expertise needed to make that choice for themselves. The labs pouring billions into model training could be wrong, but they will be responsible if they are. The negative externalities of AI innovation would become even more pronounced if the government took control of development. Most significantly, the AI industry is in such an early and undefined stage that eliminating the market process would fundamentally stunt the revelation of information. State AI development would damage the public in the same ways as private development, but with far greater risks. 

Private industry should develop AI because the payoff is by no means guaranteed, and the firms rather than taxpayers should bear that risk. It is almost impossible to tell which leading voices on artificial intelligence actually believe in the singularity, and which are just promoting it for the sake of marketing or fear mongering. Recursive model improvement is expected to overcome power and compute restrictions and get better exponentially. While the singularity seems ever more imminent, it is not certain to happen or even to fundamentally change the world if it does. The causal chain from recursive learning to unbounded benefits for the developers seems strong, but how often have dramatic predictions about the future failed to account for some seemingly innocuous factor? Technology changes, markets adjust, and wealth transfers, but the winners are never certain before the change. The financial burden of AI development would just put one more involuntary and risky bet on the backs of citizens who want nothing more than stable and responsible institutions. The best thing the government can do to accelerate AI development is create a stable and transparent institutional environment. 

The negative externalities of AI would only be increased if it was developed by the government. It is no secret that government research is often less efficient than private research. Reaching the same amount of model development that private companies are creating today would require far more data centers if that model development was being done by the government. The already massive budgets and demand for compute would naturally expand if a government actor was seeking the same outcome as the market. The government’s capability to exercise eminent domain would also lead to catastrophically unpopular data center situations, eliminating even the vestiges of coasean bargaining that have made some data centers embed well in their communities.  The privacy concerns already present in private AI would balloon with state AI. There would be nothing fundamentally stopping the state from using personal information of hundreds of millions of citizens to train models. As powerful as AI companies are, they are in a legal framework that limits their power. State-developed AI would encounter blurrier ethical lines than private AI, and the combination of power and lack of accountability would present more radical risks to the people than private AI. 

AI should be developed by private enterprise rather than the government because the technology is still developing so rapidly that we would lose important developments if it was treated like a field with a clearly defined search cost. The optimal uses for AI and most interesting developments will be best discovered by the price system and multiple actors trying to create something original. The government setting a rigid objective and pouring money into it will not create the forms of discovery necessary in such an early stage technology. While some developed fields can benefit from more externally guided research, AI is changing too dynamically for this to work. The right question to ask changes continuously with new developments, and this does not work well with centrally planned research agendas. There are numerous fundamental questions about the industry with unknown answers that can only be found out through market experimentation. Handing AI development to the government not only slows the approach towards clearly defined ends, but also makes it difficult to know which ends to pursue. While AI innovation certainly has some risks to the public, these risks would only increase if the government took control.