Aiden Gonzalez

The Governance Gap: The widening distance between the development of AI and regulations that govern it

There exists a disparity between the rate at which artificial intelligence (AI) is developed and deployed and the rate at which state and federal governments are able to regulate it. This is referred to as the “governance gap”. This brief will argue that this gap is both dangerous and preventable. It will further improve on extant AI governance frameworks, which tend to be vague and lack a deep contextual understanding of AI, by providing a concise, actionable set of protective principles for the US Congress to enact in order to close the gap, thereby protecting minors, workers, and human decency.

The United States is no stranger to technological innovation exceeding the capacity of Congress to implement safeguards. From 19th century factories to 21st century data applications, the enterprising American has always developed new and productive technology before legislators were prepared to deal with it. Indeed, the AI governance gap has created an environment reminiscent of the late 20th century’s pre-COPPA days, where online games targeted at young children existed in a sort of Wild West while Congress lagged behind.1 All agreed the potential for danger was present, but those vested with the duty to regulate it seemed paralyzed by this brave new world, unable to articulate the regulation need.

AI, however, has the capacity to be even more harmful than online games. From breaching system vulnerabilities to spoofing world leaders and market manipulation, both major and minor harms can result from intelligent systems and will do so, if AI development proceeds without safeguards. Unfortunately, the stage for our present danger was set long ago, and our present governance gap is decades in the making. Beginning in the 1990s, Congressional leadership slashed staff and funding for the research bodies available to legislators, including disbanding the Office of Technology Assessment. According to the Government Accountability Office,

The Office of Technology Assessment (OTA) was established to examine issues involving new or expanding technologies, to assess their impacts, to analyze alternative policies to avert crises, and for scientific expertise to match that of the executive branch.

Source: Government Accountability Office

The Congressional Research Service (CRS) too has had its funding held largely constant in real terms since the year 2000, despite the increase over the last two decades in the myriad of emergent issues and complex topics confronting our nation today. The net effect of the systematic dismantlement of the expert bodies on which Congress relies is that it is incapable of understanding the newest technological innovation and thereby ill-equipped to and legislate safeguards for it. This carries with it risks to both safety and progress. Additionally, given almost unfettered lobbying in the halls of the Capitol, regulatory and Congressional capture by well funded AI companies also contributes to the relative dearth of regulation imposed on those companies.2

But not all of the reasons for this governance gap are governmental failings. AI researchers themselves are frequently surprised by the progress they make in one version of a model over the last. Because these models are essentially ubiquitous, AI companies are heavily incentivized to have the best LLM or the most efficient ML-powered drug research, because the competition is only ever one Google search away. Although the major AI companies, most notably OpenAI and Anthropic, have their own ethical development frameworks,3 the United States has never relied on self-policing for industry, and with something as paradigmatic as AI, we should not start today.

As a result, this sort of “AI arms race” combines with an inability of lawmakers to fully grasp the complexities of the landscape to exacerbate the governance gap. The current state of AI governance frameworks lack specifics and a detailed knowledge of AI system development and use, but although Congress lacks a robust research apparatus this need not be the case. To combat these systemic issues, Congress should 1) dramatically increase the funding to the non-partisan CRS so that legislation in all areas and especially in technology, data, and AI is well researched and effective; 2) create a new regulatory body - a sort of Digital Technology Commission - to regulate systems and AI models and to protect consumers, businesses, and bystander; and 3) enact federal legislation which implements safeguards for AI models.

Legislation of AI governance must be both practical and specific. Requiring models to be “fair” lacks specificity and allows multibillion dollar companies to exploit that ambiguity. Similarly, even highly specific requirements of “reproducibility” of answers or content are impractical and demonstrate a lack of understanding of how LLMs function. Further, although it is likely that most people will agree with the author’s opinion that humanity should own creative output, like literature, art, and movies, and that AI should simply reduce or eliminate menial, repetitive, or trivial labor, that is a societal orientation question rather than one of protecting human rights or economic integrity, and legislation on such matters is antithetical to the American mode of governance. This brief, therefore, attempts to strike the balance between concision and completeness to produce a robust list of actionable and practical principles, viz.:

  1. Personally identifiable information (PII) from minors, including photographs, facial scans, and voice recognition, must never be used in production applications or training data.
  2. PII from adults must only ever be used in production applications or in training data with informed consent of the individual, including where it will be and its purpose.
  3. PII should not be sold between corporate entities.
  4. Images generated with AI must include identifiable metadata, including which model produced the image and the date on which it was produced.
  5. Because AI cannot be held accountable, it can never be allowed to make the final decision on 1) killing an individual, including in autonomous weapons systems, 2) medical care involving significant probability of death, or 3) terminating employment.
  6. AI must be prevented from creating pornography featuring photographs of real individuals and featuring minors whether real or fictional.
  7. AI must be prevented from exploiting vulnerabilities in software and infrastructure. It should further prevent identification of vulnerabilities unless explicitly given authenticated permission from the owner of that software.4

It is owned that some details of implementation are not included here; nevertheless this brief is published in the hope that the principles are sufficiently specific to allow a point of origin in discussing and debating legislation. Finally, it should go without saying that willful violation of these principles once enacted through legislation should carry more than a nominal penalty.

Closing the governance gap must be done quickly in order to avoid catastrophic harm to individuals and to society. Congress must act to restore its consultative and informational appendages so that legislation can again be well researched and well considered. Moreover, a new regulatory commission with the mandate and tools to prevent and remediate exploitation through digital applications, including AI, should be instantiated to keep pace with and proactively address new developments before they cause harm. Finally, Congress should again assert itself, a near first in the 21st century, to legislate safeguards and protections for individuals and society.


Footnotes

  1. Although COPPA was passed and subsequently amended by Congress, the impetus for its enactment came from the Federal Trade Commission. This brief will shortly argue that Congress would be best served in creating a new advisory and enforcement body. ↩

  2. For a particularly alarming instance of the efficacy of AI lobbying, see NBC New York’s investigation into AI lobbying in US statehouses. ↩

  3. In many cases, these really are quite robust. OpenAI, for instance, disallows its marketed models from spoofing voices or collecting identifiable private information. ↩

  4. There are legitimate instances in which we might want LLMs to identify vulnerabilities or inefficiencies in software in order to patch or refactor them. Instead of copying and pasting a code block into a model and asking for vulnerabilities, however, authenticated permission access to a repository must be obtained. ↩