AI Bias and the Need for Transparency

Artificial intelligence (AI) is central to many major debates today, covering areas from employment and data centers to healthcare and national security. A crucial question remains largely unnoticed in these discussions: will AI models aim for truth, or will they hide undisclosed biases within their responses?

AI tools often present themselves as neutral sources, capable of delivering answers with citations, detailed explanations, and unbiased feedback. However, reality often differs. While apparent biases can be easily identified and dismissed, subtle biases are harder to detect. Most users do not fact-check AI responses, making it easier for the models to influence opinions quietly. Research shows that these models sometimes guide users towards engineered outcomes or offer answers framed in a politically slanted way.

Reports are bringing these hidden biases to light. For instance, The Washington Post tested top AI models on political questions and found a consistent lean towards left-leaning arguments while portraying them as neutral. Similarly, MIT’s Center for Constructive Communication found that reward models exhibit left-leaning biases on topics like climate, energy, and labor unions, even when trained on factual information.

State lawmakers, particularly in New York and California, are taking advantage of the absence of federal regulation to push for laws like Colorado’s Artificial Intelligence Act. This act mandates impact assessments and anti-discrimination measures. These frameworks encourage companies to modify information to avoid legal issues, effectively leading to altered outputs that align with state ideological goals. The FTC views Colorado’s law as coercing alterations in AI outputs to suit state objectives, potentially penalizing accurate responses.

The rapid adoption of AI is unprecedented, with millions of Americans depending on it for information, work, and advice. Many use AI to understand the political landscape. According to a New York Times report, voters increasingly rely on AI chatbots for candidate information, seeing them as objective alternatives to conventional news or voter guides. However, if these tools harbor undisclosed biases, they can influence voter opinions and perceptions.

During his tenure, President Donald Trump addressed these concerns, emphasizing the importance of trustworthy AI. His administration’s AI Action Plan insisted that AI should seek truth, not promote social agendas. Executive orders blocked federal use of biased AI models and established a federal accuracy framework.

The importance of unbiased AI lies in its widespread use. Millions turn to AI for daily guidance. In line with the AI framework, FTC Chairman Andrew Ferguson issued a policy proposal to apply consumer protection laws to undisclosed biases in AI models. According to Section 5 of the FTC Act, a misleading omission or representation is deceptive if likely to affect a reasonable consumer’s decisions. Thus, portraying an AI model as neutral while pushing specific narratives could be a violation.

Ferguson’s proposal also reaffirms federal authority, suggesting AI models should follow a single national standard. Since AI models are designed for nationwide use, the FTC’s strategy resembles existing regulations for products like cars and medicines. If a model contains hidden biases, they must be openly disclosed to prevent violating federal laws. This commonsense approach ensures transparency and aligns with the AI Action Plan’s goals.

Ultimately, American citizens have the right to know if they are being misled. The FTC’s proposed policy is a step toward achieving the objectives set forth in the AI Action Plan and strengthening U.S. leadership in AI development.
— Nicholas Elliot