On June 16, 2022, the Federal Trade Commission (FTC) issued a strongly worded report to Congress, “Combatting Online Harms Through Innovation,” warning that policymakers must use “great caution” when mandating the use of artificial intelligence (AI) as a policy solution to combat online harms. Over-reliance on AI could introduce new harms potentially endemic to AI systems, including inaccuracies, bias, discrimination and commercial surveillance creep, the report claims. It notes that Congress and regulators should focus on developing legal frameworks to ensure that AI tools are transparent and accountable, and do no harm.

The report is the first statement on AI made by the full Commission now that the Democrats hold a majority on the FTC, suggesting that the FTC may be gearing up for anticipated AI rulemaking.¹

Background

In the 2021 Appropriations Act, Congress directed the FTC to examine and report on ways that AI may be used to combat a variety of “online harms” such as online fraud, impersonation scams, bots, deepfakes, hate crimes, cyberstalking and misinformation campaigns aimed at influencing elections, among others. Congress also instructed the FTC to recommend laws that could advance the use of AI to address online harms.

The FTC’s concerns regarding the use of AI

While the FTC report addresses the use of AI to combat online harms, the FTC takes the opportunity to discuss its position on the use of AI more broadly.² The report explains how AI tools can be inaccurate, biased, and discriminatory by design, and may incentivize reliance on increasingly invasive forms of commercial surveillance.

• Inaccuracies: AI tools used to detect online harms are blunt instruments with built-in imprecision and inaccuracy. Because the datasets supporting the AI tools are trained on previously identified problems, the AI may have difficulty identifying rapidly emerging new phenomena and avoiding false positives or false negatives. Further, the algorithms may have difficulty processing content that is too complex and dynamic for them to capture, meaning that AI tools are necessarily reactive and “need constant adjustment even when they are built to make their own adjustments.”³
• Increased surveillance: Ironically, the report finds that increasing the accuracy of AI tools has its own downsides and could lead to increased surveillance, requiring more extensive data extraction practices and more invasive forms of shadowing in order to accurately train the AI tools.
• Bias and discrimination: The report analyzes how AI tools can reflect the biases of their developers, datasets, and algorithms, leading to illegal discrimination, unfair results and censorship, depending on how the AI tools are used. 
• May not be fit for purpose: Online harms often result from adversaries with harmful agendas seeking to actively evade or manipulate AI detection tools. While this state of affairs is not going away, the main struggle is to ensure adversaries are not in the lead. The question is whether AI tools can be made sufficiently robust and flexible to meet this challenge, since the report finds that AI tools currently are brittle and can fail even with small modifications to inputs.⁴

 The report concludes that legal frameworks should be developed to prevent these harms from occurring and urges Congress, regulators, scientists, developers and users to focus on several related considerations, including:

• Human intervention: Trained humans are needed to monitor the use and decisions of AI tools, but even extensive human oversight will not solve for underlying algorithmic design flaws in AI tools used to combat online harms;
• Meaningful transparency, explainability and contestability: AI use should be transparent, explainable and contestable, especially when people’s rights are involved or when personal data is being collected or used;
• Accountability: Platforms and other companies that rely on AI tools must be accountable both for their data practices and for their results, including implementing meaningful consumer appeal and redress mechanisms. The report also recommends the use of independent audits and algorithmic impact assessments.

The report focuses on the importance of the transparency and accountability when relying on AI tools.⁵ It also makes a series of recommendations: require appropriate documentation of datasets and models, keep privacy and security in mind, take responsibility for the both inputs and outputs of AI tools, strive to hire diverse teams and avoid using training data and classifications that reflect existing societal and historical inequities.

These principles and recommendations are similar in many ways to those of other regulators and organizations that have addressed AI, including the Organization for Economic Cooperation and Development (OECD),⁶ the US National Institute of Standards and Technology (NIST),⁷ the National Association of Insurance Commissioners (NAIC),⁸ the European Commission,⁹ United Kingdom government,¹⁰ and the Alan Turing Institute.¹¹

Conclusion

The FTC’s discussion of the responsible use of AI applies across sectors. We encourage companies that use, or anticipate using, AI tools to reflect on whether their AI practices align with the FTC’s approach and to consider what governance and technical measures might be needed to mitigate legal risk going forward. 

---------------------------------------------------

[View source.]