Recognize Potential Harms and Risks

Many commenters, especially individual commenters, expressed serious concerns about the impact of AI. AI system potential harms and risks have been well-documented elsewhere.³³ The following are representative examples, which also appeared in comments:

• Inefficacy and inadequate functionality.
  • Inaccuracy, unreliability, ineffectiveness, insufficient robustness.
  • Unfitness for the use case.
• Lowered information integrity.
  • Misleading or false outputs, sometimes coupled with coordinated campaigns.
  • Opacity around use.
  • Opacity around provenance of AI inputs.
  • Opacity around provenance of AI outputs.
• Safety and security concerns.
  • Unsafe decisions or outputs that contribute to harmful outcomes.
  • Capacities falling into the hands of bad actors who intend harm.
  • Adversarial evasion or manipulation of AI.
  • Obstacles to reliable control by humans.
  • Harmful environmental impact.
• Violation of human rights.
  • Discriminatory treatment, impact, or bias.
  • Improper disclosure of personal, sensitive, confidential, or proprietary data.
  • Lack of accessibility.
  • The generation of non-consensual intimate imagery of adults, and child sexual abuse material
  • Labor abuses involved in the training of AI data.
• Impacts on privacy.
  • Exposure of non-public information through AI analytical insights
  • Use of personal information in ways that are contrary to the contexts in which they are collected
  • Overcollection of personal information to create training datasets or to unduly monitor individuals (such as workers and trade unions.
• Potential negative impact to jobs and the economy.
  • Infringement of intellectual property rights.
  • Infringements on the ability to form and join unions. 
  • Job displacement, reduction, and/or degradation of working conditions, such as increased monitoring of workers and the potential mental and physical health impacts.
  • Undue concentration of power and economic benefits.

Individual commenters reflected misgivings in the American public at large about AI.³⁴ Three major themes emerged from many of the individual comments:

• The most significant by the numbers was concern about intellectual property. Nearly half of all individual commenters (approximately 47%) expressed alarm that generative AI³⁵ was ingesting as training material copyrighted works without the copyright holders’ consent, without their compensation, and/or without attribution. They also expressed worries that AI could supplant the jobs of creators and other workers. Some of these commenters supported new forms of regulation for AI that would require copyright holders to opt-in to AI system use of their works.³⁶
• Another significant concern was that malicious actors would exploit AI for destructive purposes and develop their own systems for those ends. A related concern was that AI systems would not be subject to sufficient controls and would be used to harm individuals and communities, including through unlawfully discriminatory impacts, privacy violations, fraud, and a wide array of safety and security breaches.
• A final theme concerned the personnel building and deploying AI systems, and the personnel making AI policy. Individual commenters questioned the credibility of the responsible people and institutions and doubted whether they had sufficiently diverse experiences, backgrounds, and inclusive practices to foster appropriate decision-making.

Potential AI system risks and harms inform NTIA’s consideration of accountability measures. AI system developers and deployers should be responsible for managing the risks of their systems. As AI systems multiply and diffuse into society and the marketplace, customers, workers, consumers, and those affected by AI need assurance that these systems work as claimed and without causing harm. This is especially important for high-risk systems that are rights-impacting or safety-impacting.