The Trump administration has just released the National AI Legislative Framework. As I understand what a legislative framework means, the goal is to establish goals for legislators to guide related laws on identified topics in constrast say to temporary executive orders.
The topics selected emphasize a) protecting minors and empowering parents, b) limiting unnecessary regulations that would slow the progress of AI advancements, c) protecting taxpayers from an increased burden resulting from AI power consumption, and d) creating systems that are free of “ideological biases”. The goals and supporting materials have only been available for a few days, but there are plenty of online reactions for those (e.g., ABC) and a search will reveal many more.
I have opinions on many points raised in the various analyses, but my major focus is on the section dealing with minors as it relates most strongly to classroom issues. A separate section focused on developing an
AI-ready work force also relevant, but I focus here on the section on protecting minors. I have found one existing in-depth reaction to this section.
Major expectations include:
- Protect children and adults from deep fakes. This goal specifically references Melania Trump’s personal interests and the Take it down act.
- Parents/guardians should have ways to control over minors’ privacy settings, screen time, and content selection.
- Reasonable commercial means should be available to address the issues identified above and to ensure that users meet age requirements for AI and other online tools.
- Minors should be protected from data harvesting for use in AI model development and advertising.
- Federal laws should not unnecessarily preempt state-level requirements. I hope I have expressed this accurately in my effort to be brief, as the document seems to imply both federal and state-level expectations have their place.
- When federal standards are required, they should be unambiguous limiting the likelihood of constant court battles.
The devil and the details.
Yes, I did alter the more common expression. Details and circumstances matter, and a framework avoids details. Expectations may sound good but be unrealistic, and expectations of who must execute the expectation unfair. Expectations, if implemented, also seem to me, based on what I would expect would be required, to be found inconsistent. If you also generalize from state differences in what is labeled “woke” (e.g., historically accurate accounts of our past) and “inappropriate” (e.g., book bans), I can foresee some AI experiences being deemed inappropriate.
The unpredictable nature of AI and user interactions, personal privacy protections, and age/parental controls, when combined, can create expectations that are frequently interpreted as inconsistent.
Here are some of the challenges I anticipate. To allow parental controls, a system (meaning a combination of the online service and local tool) must identify that the active user is a minor, the minor is associated with specific parents/guardians, and there are preferences for specific topics/content sources. These data points must be stored or calculated/determined on the fly. AI tools would have to be moderated at the level of the input (e.g., are questions asked age-appropriate and parent-approved) and/or at the level of the output. Even social media offers a version of this challenge as even the content appearing on the most common social media platforms (e.g., Facebook, Instagram, X) cannot be completely anticipated. This reality would mean that perfect safety would require blocking all use of these platforms. However, expectations of these platforms are to some extent protected by Section 230, implying that the platforms cannot be responsible for what users contribute.
The ambiguity of AI input and output, combined with the guideline to create expectations that avoid legal ambiguity, creates what seems like a nearly impossible challenge. Judgment rather than strict standards would seem to be required, but judgment admits to ambiguity.
The data collection and storage requirements for decision-making are, to some extent, at odds with personal privacy requirements. AI companies are somehow to solve this riddle. How do you ensure a user is more than 13 or 18 without collecting and retaining personal information?
The AI Framework and COSA
An attempt to imagine how the legislative framework would be implemented in practice might use the Kids Online Safety Act (KOSA) as an example. Both the content and the processes to create this legislation are illustrative. KOSA is, in some ways, a broadening of the issues from COPPA (Children’s Online Privacy Protection Act), which gave us the 13-year-old minimum we now know well.
In recognizing the process issues associated with KOSA, I am referring to the history of this legislation and the challenge in operationalizing specifics. KOSA was first introduced in 2022 by the bipartisan pair of Richard Blumenthal and Marsha Blackburn, and although online safety for children is generally supported by both parties, a final version has yet to be approved by both houses.
A Wikipedia article describes the history of this act and the issues that have limited the creation of a final version review, similar to those I have already identified. Verification of age restrictions has been an issue with teens who often find it a simple matter to get around parental consent requirements and additionally, with many parents lying about their children’s ages so they could use some media platforms. Parents and the government sometimes obviously differ in what access is appropriate. The collection of personal information, what information constitutes a “persistent identifier” over time and across online locations, and how a system can avoid identifiers while simultaneously determining age and approval to view parent-designated categories of information, struggle to achieve group support. A final decision proves elusive. Finally, the topic of free speech and children’s autonomy presents challenges. When do children have the right to investigate personal issues that might reveal topics on which they and their parents disagree?
AI was not an issue for KOSA, but makes the same challenges even more complicated as inputs and outputs are less predictable.
Summary
I have read “what about the kids” is always a go-to issue when promoting federal legislation that addresses restricting behavior in one way or another. This is not intended to downplay the issue, but to recognize that it is often a point of reference in regulating personal decision-making. Younger users have again been singled out in the use of AI. Recent political promotion of a legislative framework for AI guardrails and priorities contains this same concern. My reaction to this topic is often shaped by the complexity of possible solutions to multiple challenges posed by authorities who seek the development of powerful tools and offer little guidance when different issues seemingly require conflicting fixes. When each of us encounters imperfect situations in our lives, we wrestle with priorities and trade-offs because decisions must be made. Sometimes I wonder if politicians recognize this same reality, but the pressures of governance do not allow for imperfect recommendations and requirements.
