Policymakers are proposing new transparency requirements for AI developers to disclose information ranging from model design and training data to safety testing procedures and predictions about speculative harms. While transparency is a laudable goal, many of these mandates provide minimal benefit to end users, but threaten to overwhelm Little Tech—our term for startups—with unconstitutional compliance burdens that would make it harder for them to compete with larger platforms.
In this post, we review three basic questions about AI transparency mandates:
- What are mandated disclosures, and why are policymakers proposing them?
- Why do current mandated disclosure proposals put Little Tech at a disadvantage?
- When are mandated disclosures unconstitutional?
We then offer a more targeted, lawful, and useful alternative to current proposals: AI Model Facts, a transparency approach that gives meaningful information to people who use AI products without stifling innovation or undermining American competitiveness.
The what and why of mandated disclosures
State, federal, and international policymakers have proposed a wide range of transparency requirements for AI developers, such as impact assessments, red team performance data, the names of third-party contributors, details of training datasets, and even statements about energy use. Other proposals include disclaimers and watermarks indicating whether AI was used to generate content and notifying users that they are interacting with AI systems. Some proposals require warning labels for AI content, specifying that AI-generated content may be “inaccurate.” Others, like requirements to disclose developers’ use of copyrighted works in model training, seem focused on influencing the outcome of pending litigation about whether training on publicly available data constitutes fair use under current law.
Some model developers already publish some information about their models. These typically take the form of system cards, which describe various features of how the model works and how it was trained. By design, open source models may provide more information about model design, features, and intended use than closed models. In the absence of laws that mandate disclosure, it is possible that over time, the public would understand comparatively less about closed AI systems. Closing this transparency gap could help people make choices about how to use AI products to fit their needs and preferences.
Why do current mandated disclosure proposals put Little Tech at a disadvantage?
Transparency mandates don’t affect all companies equally—and they can hit Little Tech the hardest. Because mandatory disclosure regimes impose penalties (even at times criminal penalties) if companies provide inaccurate or incomplete information, compliance can be extraordinarily expensive to ensure no errors occur. Marshalling the data and information necessary to produce an accurate disclosure takes money and personnel. A large business with a large legal team may be able to bear these costs without diverting energy from running its core business, such as product development, monetization, and partnerships. Larger platforms may have thousands of lawyers on their payroll, a policy team that can work with policymakers and regulators to develop a minimally burdensome compliance regime, and financial resources to hire outside law firms and auditors to share the burden.
In contrast, startups have far fewer employees, and their teams are focused principally on getting a product to market and growing their business. They may not have a compliance team, a policy team, or even a general counsel. They may have fewer financial resources to hire outside law, communications, or accounting firms, and even when they can retain external support, they have more limited internal resources to manage these relationships. Little Tech might also face challenges in complying with transparency mandates that vary across state and international lines. If the pending state proposals are enacted into law, startups would be forced to disclose safety data in some states, copyright-related information in one, and warning labels in another. Compliance with a 50-state patchwork of state transparency obligations would present massive challenges for companies with small engineering, law, and policy teams. As the Treasury Department recently announced with regard to its enforcement of the Corporate Transparency Act, it is important to “appropriately tailor” any disclosure rule to ensure it does not present challenges for small businesses.
The discrepancy in resources between large platforms and Little Tech means that transparency regimes will burden them disproportionately. Those unequal burdens then harm competition: it’s harder for Little Tech to compete with larger platforms with deeper pockets if they’re diverting product and business resources to compliance. The long-term effect will be to increase concentration in AI markets. A mandate should be enacted only if its public benefits clearly outweigh the cost to Little Tech.
When are mandated disclosures unconstitutional?
As recent court cases demonstrate, there are also significant legal concerns with mandated disclosure regimes. The government can compel private entities to disclose commercial information only in limited circumstances. As the Supreme Court has stated, “compulsion to speak may be as violative of the First Amendment as prohibitions on speech,” and “unjustified or unduly burdensome disclosure requirements might offend the First Amendment by chilling protected commercial speech.” The well-established legal test is that disclosures of commercial speech are constitutional primarily when they compel information that is “purely factual and uncontroversial” and if they are not “unduly burdensome.” If they do not meet these standards, if the required disclosure is not considered commercial speech, or if a court finds the disclosures to be “content-based,” then the bar is higher, and the odds are greater that the requirement will be unconstitutional. Compelled speech can be “non-commercial” even if the requirement for making the disclosures in question falls on a corporation.
Courts have applied these principles to recent laws requiring tech companies to make certain disclosures related to their products. In two Ninth Circuit cases, the court found California laws to be unconstitutional because their disclosure requirements violated the First Amendment. In NetChoice v. Bonta, the court found that compelling companies to disclose child safety impact assessments was “highly subjective” and “onerous,” and struck down that component of the law even before the case went to trial. It stated that the “requirement that covered businesses opine on and mitigate the risk that children may be exposed to harmful or potentially harmful materials online … facially violates the First Amendment.” It also held that the First Amendment limits what disclosures the government can require, even if these disclosures are made solely to the government and are not made public.
Similarly, in X Corp. v. Bonta, the court blocked a law that required social media companies to post their terms of service and file reports to the Attorney General twice a year on their content moderation policies and practices. It found that the compelled speech was not commercial because it required a platform to “express a view,” rather than simply “communicat[ing] the terms of an actual or proposed transaction.” The requirement would have obligated the company to “convey the company’s policy views on intensely debated and politically fraught topics.” Ultimately, the law was unconstitutional because it was “more extensive than necessary” to serve the state’s goal of helping consumers to make informed decisions.
Many of the AI-related disclosures that have been proposed by state and federal policymakers would be unconstitutional for the same reasons that these two California laws were struck down. Requiring AI developers to produce speculative information—such as hypothesizing about future safety risks or potential bias—obligates them to “express a view,” involves a “highly subjective” policy assessment on “intensely debated and politically fraught topics,” and is likely to be impermissibly “onerous” for AI developers.
What types of mandated AI transparency might work for Little Tech?
It is possible that public understanding of and trust in AI could be improved through targeted requirements for model developers to disclose useful, factual, noncontroversial information about their models. The central challenge for policymakers is to develop a regulatory framework that is constitutional, helpful for users, and not unduly burdensome for startups. Finding policy options that meet all three of these criteria is difficult but not impossible.
One option is to require base model developers to publish “AI Model Facts:” a brief fact sheet that helps people make informed choices about which models to use and how to use them. If a developer could easily link to these data points in its user interface or in its terms of service, then AI Model Facts would likely not be too burdensome for companies of any size to produce. For example, AI Model Facts could include the following:
- Knowledge cutoff date: it could be helpful for users to know the ”knowledge cutoff date” for a model’s training data. If the cutoff is April 2025, for instance, you would not expect a correct response to a question about the winner of the NBA Finals in June 2025.
- Release date: given the pace of innovation and the proliferation of competing models, it is easy for consumers to be confused about whether they are using a developer’s most recent model, or whether the model was released before or after a competing model from another developer.
- Language support: many AI model developers already provide information about supported languages. This information may benefit users trying to figure out what prompts are likely to generate the most accurate responses.
- Location of a company’s headquarters: some people might prefer to use a model that is based in their country of origin or that is more likely to reflect their cultural preferences. For instance, some users might prefer an American model to a Chinese one, or vice versa. Other users might choose a European model over an American one, or vice versa.
- Links to the terms of service: providing easy-to-access links to a company’s terms of service, and, where relevant, content standards or privacy policies, might help users to better understand how a product functions.
- Performance against objective, standardized benchmarks: it is currently standard practice for model developers to publish a model’s performance against objective, standardized benchmarks in their system cards. When these benchmarks are so well-established and so accessible that they are neither “controversial” nor “unduly burdensome” for Little Tech to produce, then it is conceivable they could be a required component of AI Model Facts.
In developing the components of AI Model Facts, policymakers might consider the website domain name registration (DNS) system as a model. Registering a website requires an operator to provide some basic information, such as the name and contact information of the registrant, the date the domain was first registered, and the name of the company through which the domain was registered. This data is factual, useful, uncontroversial, and not unduly burdensome for startups, and it provides interested users with some basic information they can use to understand a website they visit or to contact a website administrator, if needed.
Another option is Twilio’s AI nutrition label concept, which includes a description of the model, the type of model, and then several additional “ingredient” categories. Of course, to ensure that a nutrition label requirement is constitutional, all the required categories of information would need to be factual, uncontroversial, and not unduly burdensome for the developer to collect and provide.
A base model developer should be required to publish AI Model Facts each time it releases a new model, and it would be required to update the fact sheet if it has knowledge that a material component has changed or is incorrect. To ensure AI Model Facts are both comparable and practical, developers should present the required information in a consumer-friendly format. Policymakers might prescribe some basic guidelines in how the information should be presented so as to promote a baseline level of consistency that could be helpful to consumers, but they should also allow some flexibility in format. This flexibility is key to ensuring the information is genuinely useful to users within different product environments and not unduly burdensome for developers to implement.
A requirement to publish AI Model Facts should apply only to developers of base models: downstream model developers, such as startups that fine-tune an open source base model, should not be obligated to provide the same type of transparency. Downstream developers are not well-positioned to provide facts about how a base model was trained, so requiring them to do so may lead to inaccurate or outdated information, undermining the very purpose of transparency: to provide people with accurate information that they can use to inform which products they use and how they use them. If policymakers want disclosures from downstream developers, those requirements should be limited to facts they actually know, like their own location or performance benchmarks of their own model.
Of course, AI developers would be free to voluntarily publish additional information beyond any disclosures required by law. Many model developers now publish system cards that describe the training process in detail, and that practice could continue. Open source providers often go beyond proprietary model developers to detail the attributes of their models and how they might be used by downstream developers. That practice could and should continue on a voluntary basis.
Whatever the baseline level of transparency required by law, companies might voluntarily choose to compete to offer the most helpful transparency to consumers. For instance, Anthropic has chosen to develop and publish a Responsible Scaling Policy that describes its safety and security practices. Other companies might seek to mimic Anthropic’s disclosures, or might attempt to develop an alternative that they believe provides more useful information in a better format. This type of voluntary competition has occurred in other areas of the tech ecosystem, such as the trend for search and social media companies to publish transparency reports that include qualitative and quantitative data on how they moderate content and provide data to governments. This type of competition between developers to offer best-in-class transparency could be a positive force for the AI industry if it incentivizes them to build better products and provide more useful information to consumers. But to ensure that Little Tech can compete on a level playing field with larger platforms with deeper pockets, the choice to go beyond the information the government requires to be disclosed in AI Model Facts should be entirely voluntary.
Transparency in AI model development is a worthwhile goal, but the path toward achieving it must be carefully designed to provide useful information for people who use AI products, avoid unconstitutional speech mandates, and steer clear of disproportionately burdening Little Tech. Mandated disclosure requirements that are speculative, subjective, or controversial are not only likely to be struck down in court—they also risk increasing concentration in AI markets and benefiting larger AI platforms at the expense of Little Tech.
Instead, we propose AI Model Facts, a fact sheet that provides factual, useful, and low-burden disclosures by base model developers. This approach would give users meaningful transparency without stifling innovation, violating the First Amendment, or favoring larger platforms. A measured, evidence-based standard like this can foster a more competitive AI ecosystem—one where entrepreneurs and startups have a real shot.