A Blueprint for the Future: White House and States Issue Guidelines on AI and Generative AI
Since July 2023, eight U.S. states (California, Kansas, New Jersey, Oklahoma, Oregon, Pennsylvania, Virginia, and Wisconsin) and the White House have published executive orders (EOs) to support the responsible and ethical use of artificial intelligence (AI) systems, including generative AI. In response to the evolving AI landscape, these directives signal a growing recognition of the rapid pace of AI development and the need to manage potential risk to individuals’ data and mitigate algorithmic discrimination against marginalized communities.
FPF has released a new comparison chart that summarizes and compares U.S. state and federal EOs and discusses how they fit into the broader context of AI and privacy.
In addition to the state governments, several cities (e.g., Boston, San Jose, and Seattle) have also issued guidelines on generative AI use that seek to recognize the opportunities of AI while mitigating bias, privacy, and cybersecurity risks. In contrast, other jurisdictions, such as Maine, have issued a moratorium on state generative AI use while they perform a holistic risk assessment.
Although each of the state and federal EOs on AI and generative AI has a different scope, at minimum, most charge agencies with the creation of a task force to study AI and offer recommendations.
Here are some overarching takeaways from our analysis of all of the EOs:
1. The White House and California Issued the Most Prescriptive EOs
Of the U.S. state and federal EOs analyzed, the White House requires the heaviest lift. The White House EO mandates dozens of reports and next steps for federal agencies, including the creation of guidance and standards for AI auditing, generative AI authentication, and privacy-enhancing technologies (PETs).
Similarly, of the state EOs, California is the most prescriptive and includes a number of specific mandates and reports tailored to different agencies, such as the creation of procurement guidelines, assessments on the effect of generative AI on infrastructure, and research on the impact of generative AI on marginalized communities.
2. Most State EOs Focus on “Generative AI”
Several state governments, such as California, Kansas, New Jersey, Pennsylvania, and Wisconsin, only focus on generative AI – how the technology should be used by state agencies, the risks it carries, and how it may affect their state industries and workforce. Oklahoma, Oregon, and Virginia take a broader stance and cover generative AI as well as broader types of AI systems in their EOs. Kansas and Pennsylvania are the only two states to explicitly define generative AI.
The White House EO represents an amalgam of the state EOs, as it defines generative AI (similar to Kansas and Pennsylvania) and also broadly covers different types of AI systems (similar to Oklahoma and Virginia).
3. Varying Approaches to Agencies’ Roles
The White House EO charges certain agencies with authority to create binding guidelines and standards for government actors. In contrast, rather than creating new task forces or boards, Kansas and Virginia charge state agencies to study AI technology and provide general recommendations. New Jersey and Wisconsin, two states with less rigorous EOs, emphasize that their task forces serve solely advisory roles. Oklahoma and the White House are the only EOs to require each agency to appoint an individual on their team to become an AI and generative AI expert.
4. Impact to Industry
While these EOs are primarily focused on government use of emerging AI systems, there are major requirements contained in many of them that may have consequential effects on industry.
Procurement Requirements:Companies selling certain AI products and services to government entities will need to satisfy new baseline procurement standards.
Enforcement:Agency-created standards and policies may inform government regulators’ perspectives on AI compliance with data privacy, security, civil rights, and consumer protection laws, particularly given the forthcoming standard setting activity directed by the White House EO.
Influence on Legislation:As mentioned in California’s EO and the White House EO’s accompanying fact sheet, key actors in state and federal executive agencies will work with policymakers to pursue legislative approaches to support the development of responsible AI by the private sector.
These EOs represent a watershed moment for AI system users, developers, and regulators alike. Over the next few years, increased government action in this area will lead to new requirements and opportunities that will have lasting implications for both the public and private sector.
FPF and The Dialogue Release Collaboration on a Catalog of Measures for “Verifiably safe” Processing of Children’s Personal Data under India’s DPDPA 2023
When India’s DPDPA passed in August, it created heightened protections for the processing of personal data of children up to 18. When the law goes into effect, entities who determine the purpose and means of processing data, known as “data fiduciaries,” will need to apply these heightened protections to children’s data. Under the DPDPA, there is no further distinguishing between age groups of children, and all protections, such as obtaining parental consent before processing a child’s data, will apply to all children up to 18. However, the DPDPA stipulates that if the processing of personal data of children is done “in a manner that is verifiably safe,” the Indian government has the competence to lower the age above which data fiduciaries may be exempt from certain obligations.
In partnership with The Dialogue, an emerging research and public-policy think-tank based in New Delhi with a vision to drive a progressive narrative in India’s policy discourse, FPF prepared a Brief compiling a catalog of measures that may be deemed “verifiably safe” when processing children’s personal data. The Brief was informed by best practices and accepted approaches from key jurisdictions with experience in implementing data protection legal obligations geared towards children. Not all of these measures may immediately apply to all industry stakeholders.
While the concept of “verifiably safe” processing of children’s personal data is unique to the DPDPA and not found in other data protection regimes, the Brief’s catalog of measures can aid practitioners and policymakers across the globe.
The Brief outlines the following measures that can amount to “verifiably safe” processing of personal data of children, proposing additional context and actionable criteria for each item:
1. Ensure enhanced transparency and digital literacy for children.
2. Ensure enhanced transparency and digital literacy for parents and lawful guardians of very young users.
3. Opt for informative push notifications and provide tools for children concerning privacy settings and reporting mechanisms.
4. Provide parents or lawful guardians with tools to view, and in some cases set, children’s privacy settings and exercise privacy rights.
5. Set account settings as “privacy friendly” by default.
6. Limit advertising to children.
7. Maintain the functionality of a service at all times, considering the best interests of children.
8. Adopt policies to limit the collection and sharing of children’s data.
9. Consider all risks of processing their personal data for children and their best interests via thorough assessments.
10. Ensure the accuracy of the personal data of children held.
11. Use and retain personal data of children considering their best interests.
12. Adopt policies regarding how children’s data may be safely shared.
13. Give children options in an objective and neutral way, avoiding deceptive language or design.
14. Put in place robust internal policies and procedures for processing personal data of children and prioritize staff training.
15. Enhance accountability for data breaches through notifying the parents or lawful guardians and adopting internal policies such as Voluntary Undertaking if a data breach occurs.
16. Conduct specific due diligence with regard to children’s personal data when engaging processors.
We encourage further conversation between government, industry, privacy experts, and representatives of children, parents, and lawful guardians to identify which practices and measures may suit specific types of services and industries, or specific categories of data fiduciaries.
ICYMI: FPF Webinar Discussed The Current State of Kids’ and Teens’ Privacy
Privacy by design for kids and teens has expanded across the globe. As policymakers, advocates, and companies grapple with the ever-changing landscape of youth privacy regulation, the Future of Privacy Forum recently hosted a webinar discussing the current state of kids’ and teens’ privacy policy. The webinar explored the current frameworks that are influential worldwide, the variations in youth privacy approaches, and the nuances of several emerging trends.
The virtual conversation, moderated by FPF’s Chloe Altieri, included a discussion about the industry’s work on compliance with a variety of regulations across jurisdictions. The webinar began with presentations from the panelists, setting the stage with their current work on youth privacy issues. The panelists were Phyllis H. Marcus, Partner at Hunton Andrews Kurth LLP, Pascale Raulin-Serrier, Senior Advisor in Digital Education and Coordinator of the DEWG at the French CNIL, Michael Murray, Head of Regulatory Policy at the U.K. ICO, and Shanna Pearce, Managing Counsel, Family Experience at Epic Games.
Phyllis led the audience through the current U.S. legislative and regulatory landscape. The U.S. States have been incredibly active in children’s and teens’ privacy legislation, with 11 states having enacted bills and an additional 35 states have considered legislation for youth online safety. The trends emerging from the online safety legislation being considered show four primary categories of laws being considered by state regulators: Platform Accountability Laws, Age Verification Laws, Social Media Metering Laws, and the California Age-Appropriate Design Code (CA AADC). In recent actions, the Federal Trade Commission has stepped up enforcement of the Children’s Online Privacy Protection Act (COPPA). Finally, the U.S. Congress has taken an interest in enhancing youth online safety measures through the introduction of COPPA 2.0 by Sen. Markey and Sen. Cassidy and the Kids Online Safety Act (KOSA) by Sen. Blumenthal and Sen. Blackburn.
Pascale discussed the work of the French CNIL, both nationally and internationally, in making privacy and youth online safety an effective initiative. Key international initiatives include the International Resolution on Children’s Digital Rights adopted in 2021, which harmonized a regulatory vision and set of core principles among Data Protection Authorities Worldwide around youth online safety. Additionally, the CNIL is working to improve digital education to combat the issues of online safety. The CNIL published eight recommendations to enhance the protection of children online in 2021 with the aim of providing practical advice to a range of stakeholders. Such recommendations include strengthening youth awareness of risks online and privacy rights as well as encouraging youth to exercise their privacy rights in order to stay safe. Youth education efforts are being undertaken in conjunction with digital literacy efforts that empower parents and caretakers to have meaningful conversations about online safety with their children.
Michael discussed the United Kingdom’s Age Appropriate Design Code (U.K. Children’s Code), which is a statutory code of practice under the UK’s General Data Protection Regulation(GDPR). The Children’s Code is grounded in the principles established by the United Nations Convention on the Rights of the Child (UNCRC). Michael explained that the Code sets out 15 interlinked standards of age-appropriate design for online services that are likely to be accessed by children. The U.K. Information Commissioner’s Office (ICO) has undertaken a wide-ranging effort to supervise the code by exploring the possible ways the ICO could provide guidance for online services to meet the Code’s objectives. The ICO not only looks to submitted complaints but also to industry engagement efforts and engagement with other regulators or government offices to inform its guidance. According to recent industry surveys, this new method of supervising code implementation has been effective.
Shanna discussed the safety and privacy considerations that are necessary when building online experiences for kids and teens, specifically on gaming platforms. Shanna spoke about the balance and thoughtful product design required to create a positive experience for younger players and their guardians in compliance with global regulations. One product of this balancing act for Epic Games was the deployment of a suite of protections across its ecosystem of games, including age-appropriate default settings, parental controls, and Cabined Accounts, an Epic account designed to create a safe and inclusive space for younger players. Players with Cabined Accounts can still play Fortnite, Rocket League, or Fall Guys but won’t be able to access certain features, such as voice chat, until their parent or guardian provides verifiable parental consent (“VPC”).
After the panelists’ presentations, we launched into a discussion on several of the most salient issues unsettled in youth privacy policy, such as online safety, age assurance, parental consent, new regulatory requirements, and guidance for the industry. We have summarized the discussion and included key takeaways.
How are policymakers and industry members working to resolve points of tension between privacy and safety? How is this tension and its resolution approached differently across the globe?
Michael Murray gave a three-fold answer on the difference between U.S. and U.K. child privacy & safety policy. The first key difference is that in the U.K. context, the ICO and OFCOM lead a dual effort where the ICO focuses on privacy, and the U.K.’s Office of Communications (Ofcom) focuses on safety and content regulation. In the U.S., there is no defined, systemic privacy and safety regulatory effort. The second is that the U.K. is a signatory to the UNCRC, which defines children as anyone under the age of 18, and that is reflected in U.K. law. In contrast, the U.S. is not a UNCRC signatory, and the current U.S. federal protections define children as individuals under the age of 13 years old. Finally, third, the U.S. operates largely under an actual knowledge standard that an online site or service is directed to children. Whereas the U.K. Code and, recently, the California AADC operate under a “likely to be accessed” by children standard.
Phyllis H. Marcus elaborated on some of the points Michael brought up, describing the knowledge standard and tensions between privacy and safety we see in the U.S. COPPA is a privacy rather than a safety regime, but it does have safety components as one of its statutory underpinnings. It is important to know that this current regime is being somewhat upended by the patchwork of state laws being passed. Additionally, this regime may change if COPPA 2.0 and KOSA make their way through Congress and especially if they go into effect, where, combined, they will regulate privacy and safety regimes in tandem.
Pascale provided insight into the approach by France and the European Union (EU), where safety and privacy are not opposing or differentiated regulatory efforts. Safety is an element of privacy regulation and is considered within Data Protection legislation.
Age assurance is a large part of the policy conversation on youth privacy and safety online, especially as privacy protections for teens are expanding. What are your thoughts on the current issues around age assurance?
According to Michael, the area of age assurance and age verification is rapidly evolving, but there is “no silver bullet” for establishing or verifying user age on digital platforms and services. The U.K. Code takes a risk-based and proportional response to age assurance. The lower the risk of processing a child’s data, the less intensive age assurance or verification mechanisms need to be. For example, self-declaration might be a suitable mechanism for a lower-risk service. However, where risk is higher, more assurance and verification are needed to protect against processing children’s data, which is a GDPR violation. For higher-risk services, age estimation software with a buffer, a form of cabined accounts, or age verification through mobile contracts and digital IDs could be employed. These methods are all still immature, and there is no clear one-size-fits-all solution.
Pascale expressed that there are no clear solutions for age assurance and verification at this time. The CNIL, along with working groups such as DEWG, are still experimenting with age assurance methods. The CNIL is working towards developing a technical approach among different stakeholders in both the government and private sectors to harmonize methods across the chain of actors concerned. However, it is clear that the solution developed will not rely on biometric data collection for age verification efforts, though scans and estimations of face shape may be permitted. The end goal will be to find a strong mechanism for verification while balancing privacy concerns.
What are the important considerations when trying to strike a balance between data minimization, collecting information for age assurance, the level of accuracy that is appropriate, and the evolving landscape of age assurance technologies?
Shanna discussed the challenges the industry faces with respect to the state of age assurance technology in certain scenarios. Those challenges include imprecision with some age assurance technology and methods that cannot be used across all types of devices where users access online services, such as gaming consoles. These challenges are reduced when several methods are offered for verification of adults providing parental consent but may be significant where a single method is used as a gate for users to access services. While alternative age assurance technologies continue to develop, Shanna observed that the industry can find creative ways to improve the reliability of existing methods like self-reported age gates–such as providing child experiences that reduce the incentive to misstate age (Epic’s Cabined Accounts are one example) and using trusted data intermediaries to reduce friction and privacy risk to parents providing parent verification.
Michael echoed concerns about data minimization complicating age verification techniques. He added that when given a choice, a lot of parents prefer age estimation mechanisms as opposed to giving hard identifiers or personal information such as ID numbers or credit cards for age verification purposes.
These questions around age assurance are sometimes linked to discourse about parental consent. Can you speak to these two topics and share a bit about the emerging methods for each?
Phyllis provided more insight into age assurance and parental consent practices in the U.S., noting that, at least in the U.S., age assurance and parental consent “are really two different things.” When it comes to children’s use of online services under federal law, there is no requirement to verify the age of users. Rather, the U.S. requirement is to obtain consent from someone whom the company has reasonable assurance to be the parent of a child requesting access to a service. There are some parental consent mechanisms that have been whitelisted by the FTC and have been in place for decades, while others are still being reviewed by the FTC. The idea of age assurance, on the other hand, is relatively new in the U.S., and there are a number of actors considering the possibility of deploying age assurance methods in the states. Key considerations for exploring the use of age assurance technology in the U.S. include looking at less-intrusive, less risky verification methods and making data minimization a priority. Finally, Phyllis made clear that when using age estimation systems for age assurance, the over/under age estimations could be risky if not adequately tested. When estimating a user’s age with just a few years for margin of error, that would be the difference in compliance and non-compliance.
There is a lot of work being done globally by lawmakers, regulators, advocates, industry leaders, and researchers to answer these policy questions we have discussed today. How are recommendations created, and how is this guidance impactful for remedying noncompliance or figuring out solutions to protect youth privacy online?
Pascale noted that in addition to the CNIL’s eight recommendations previously mentioned, global IT experts are exploring age verification technologies to be able to create recommendations for compliance and enforcement. Work is also being done on digital education for parents as a way to increase awareness and understanding of child privacy and safety online. There is a balance between allowing parents to be involved and requiring online services to add protections. There are also important nuances around teen autonomy, developmental stages, and parents sharing too much of their child’s information. There is more to come on recommendations providing topics of discussion with parents as well as developing cooperation on a voluntary basis with the industry.
Michael agreed that digital education is a vitally important part of the solution, and research shows that parents want to have a say in the online services their kids use. Still, it cannot be the entire solution, and parents will not always be able to make informed decisions. Children’s design codes are placing an emphasis on the design of online services to avoid placing an overwhelming burden on the shoulders of parents. This emphasis works productively in tandem with developing resources for parents to have productive conversations with their children.
Recent youth privacy legislation has included a variety of standards for the level of knowledge of an online service’s audience’s age. These variations in legislation have led to companies needing to consider youth privacy issues, like age assurance, that previously did not. How is this impacting emerging technology and the practical implementation of new products?
Phyllis responded that the development of new standards for determining what services do or do not fall under the scope of regulatory scrutiny and age assurance requirements is one of the most hotly contested and highly discussed issues in the evolving U.S. landscape. Under COPPA, the requirements are defined clearly into buckets, which then clearly define the scope of the law. The current federal standard in the U.S. is actual knowledge that a service is directed to children. Phyllis cautions that it’s important to note that most new initiatives change this paradigm, and the jury is still out on what the new standard will ultimately entail with COPPA 2.0 and KOSA.
Shanna noted that while many services were developed and deployed prior to legislation going into effect, those services may be brought into scope later. Retrofitting an existing service to address things like parental consent and default settings may require significant design and technical effort, and the process is complicated further as the age of digital consent differs across regions. Shanna stated that engagement by regulatory bodies and issuing of guidance is invaluable to companies trying to comply with these evolving requirements in the tech space.
Pascale added that this is not the first time that the industry has faced technical difficulties like the ones we see today in age assurance and verification. According to Pascale, innovation is a key element to prioritize in each company’s approach because big innovations can guide smaller ones.
We asked a final question to all of our panelists: What do you foresee as the near future of youth privacy policy? What issue should companies or policymakers have top of mind right now?
Phyllis observed that there is a lot on the horizon and that it will be easy for actors to fall behind if they are not intentionally keeping up with youth privacy. It is clear that developments in the U.K. have had an effect on U.S. policy at both the state and federal levels. These initiatives will continue to be momentum to keep an eye on.
Pascale opined that privacy by design is one of the best policy options. While digital education is important to aid in solving these issues, integrating privacy by design at the conception of tech innovation will help to distribute the pressure of protecting youth online.
Michael noted that age assurance is an obvious answer. Additionally, the resolution of First Amendment questions presented in the litigation of the California Age-Appropriate Design Code will be critical. The suit brings up fundamental issues around how to protect data without impacting U.S. constitutional rights that will be an important debate.
Shanna is interested in seeing how companies balance privacy with uses of emerging technologies that improve online safety. She also observed that a variety of laws are currently taking shape around the globe, and there’s an opportunity to improve consistency and clarity of forthcoming guidance so companies can comply effectively.
Each of the panelists shared helpful resources, which we have listed and linked below, along with a few of our own. You can also find the panelist’s presentation slides and additional resources here.
Coming up soon! You won’t want to miss FPF’s final session in our virtual Immersive Tech Panel Series on December 6 at 11 am ET. The December session will dive into designing immersive spaces with kids and teens in mind. You can register for this event here.
For more information or to learn how to become involved with FPF’s youth privacy analysis and initiatives, please contact Chloe at [email protected]. Subscribe here to receive monthly newsletters from the Youth and Education Team.
FPF Offers Input on Massachusetts Student Data Privacy Proposal
On October 30, FPF provided testimony before a hearing of the Massachusetts Joint Committee on Education regarding H.532/S.280, an Act Relative to Student and Educator Data Privacy.
FPF’s Director of Youth & Education Privacy, David Sallay, discussed his previous experience as chief privacy officer for the Utah State Board of Education and applauded policymakers for calling for the creation of a similar role in H.523/S.280. During the hearing, he also highlighted how the role could help address several of the other bills that were discussed, including providing support to rural schools and to Massachusetts’ educator-to-career data center. By designating privacy-focused personnel and requiring training, Massachusetts has an opportunity to improve structure, transparency, and consistency for schools, districts, and parents.
FPF’s testimony also included several recommendations and improvements, including expanding the bill’s student data privacy training requirements beyond educators, as procurement, IT, and other administrative staff often also have access to covered data. One component of the bill that will likely prove to be controversial is the broad private right of action included in the enforcement provisions. We expect this to be the subject of continued discussion and debate in the legislature. Citing his experience in Utah, David noted that granting the chief privacy officer role the authority to investigate alleged violations of student privacy laws, could help streamline and simplify enforcement.
FPF and OneTrust Release Collaboration on Conformity Assessments under the proposed EU AI Act: A Step-by-Step Guide & Infographic
Today, the Future of Privacy Forum (FPF) and OneTrust released a collaboration on Conformity Assessments under the proposed EU AI Act: A Step-by-Step Guide and accompanying Infographic. Conformity Assessments are a key and overarching accountability tool introduced in the proposed EU Artificial Intelligence Act (EU AIA or AIA) for high-risk AI systems.
Conformity Assessments are expected to play a significant role in the governance of AI in the EU, and the Guide and Infographic provide a step-by-step explanation of what a Conformity Assessment is–designed for individuals at organizations responsible for the legal obligation to perform one–along with a roadmap outlining the series of steps for conducting a Conformity Assessment.
The Guide and Infographic can serve as an essential resource for organizations who want to prepare for compliance with the EU AIA’s final text, which is expected to be adopted by the end of 2023 and become applicable in late 2025.
Information and background about the proposed EU AI Act & Conformity Assessments. The proposed EU AIA is a risk-based regulation with enhanced obligations for high-risk AI systems, including the obligation to conduct Conformity Assessments. In the EU context, the Conformity Assessment obligation is not new: the EU AIA aims to align with the processes and requirements found in laws that fall under the New Legislative Framework (NLF), and Conformity Assessments are also part of several EU laws on product safety, such as the General Product Safety Regulation, the Machinery Regulation, or the in vitro diagnostic Medical Devices Regulation.
The Conformity Assessment applicability for AI systems. A Conformity Assessment is the process of verifying and/or demonstrating that a high-risk AI system complies with the requirements enumerated under Title III, Chapter 2 of the EU AIA. The first step in the Conformity Assessment journey is determining whether an organization’s AI system falls under the Conformity Assessment legal obligation, and the Guide and Infographic include a flowchart of questions for an organization to answer in order to determine whether they need to comply with the Conformity Assessment obligation.
Conformity Assessment requirements for high-risk AI systems. The Guide describes each Conformity Assessment requirement, its meaning, and at what phase of the AI system’s life cycle each requirement should be met. These requirements include Risk Management System; Data and Data Governance; Technical Documentation; Record Keeping; Transparency Obligations; Human Oversight; Accuracy, Robustness and Cybersecurity.
Overview of EU Plans for Standards & Presumption of Conformity. The European Commission is looking to obtain standards that provide “procedures and processes for conformity assessment activities related to AI systems and quality management systems of AI providers.” Such standards will be crucial to developing operational guidance for the implementation of Conformity Assessments and are expected to facilitate compliance with the technical obligations prescribed by the EU AIA. Given that the EU AIA is still under negotiation, the draft standardization request that was issued by the European Commission in December 2022 may be amended when the AIA is finally adopted.
For more information about the EU AIA, Conformity Assessments, and the Guide and Infographic, please contact Katerina Demetzou at [email protected].
Click here to view the Updated Guide on Conformity Assessments under the EU AI Act.
FPF Submits Comments with the National Telecommunications and Information Administration (NTIA) on Kids Online Health and Safety
On November 15, the Future of Privacy Forum filed comments with the National Telecommunications and Information Administration (NTIA) in response to their request for comment on Kids Online Health and Safety as part of the Biden-Harris Administration’s Interagency Task Force on Kids Online Health & Safety.
Young people increasingly engage with their peers online and lawmakers continue to introduce legislation to expand protections for the privacy and safety of minors beyond the existing COPPA framework. However, adopting a one-size-fits-all approach to developing policies for minors online presents challenges, as protections that are appropriate for very young children may not be suitable for older teenagers with greater agency and autonomy. While addressing online experiences for minors is a multi-faceted issue, as evidenced by the interagency task force, FPF has identified four of the most impactful areas for privacy that the Task Force should consider as they develop voluntary guidance, policy recommendations, and a toolkit on safety-, health-, and privacy-by-design for industry to apply in developing digital goods and services.
1. Children and teens have varying privacy needs across developmental stages, and overgeneralized restrictions may exacerbate health risks and undermine the developmental benefits of social online experiences. In particular, limitations on access to content and connecting with peers may have negative consequences on the ability of adolescents to explore and develop independence and identity.
2. While many stakeholders agree on high-level policy goals, such as extending heightened protections to both children and teens or minimizing unnecessary data collection, there is little consensus on how best to implement broadly agreed-upon policy goals. In some areas, such as age assurance, there is significant disagreement on how best to grapple with conflicting equities on privacy and safety.
3. Companies building new features to protect the privacy and safety of minors online currently take into account the varying developmental stages of minors and the interaction between minors’ autonomy and parental involvement. These two considerations inform how companies balance privacy and safety before introducing new features and reviewing existing tools as research and societal norms evolve.
4. FPF recommends additional research investigating minors using online services for educational purposes versus recreation, shifts in privacy risks at different ages and stages of development, and the relationship between privacy and safety in applying heightened protections to teens. This research is necessary to identify appropriate safeguards for minors online in both policy and practice.
FPF Statement on Biden-Harris AI Executive Order
The Biden-Harris AI plan is incredibly comprehensive, with a whole of government approach and with an impact beyond government agencies. Although the executive order focuses on the government’s use of AI, the influence on the private sector will be profound due to the extensive requirements for government vendors, worker surveillance, education and housing priorities, the development of standards to conduct risk assessments and mitigate bias, the investments in privacy enhancing technologies, and more. Also important is the call for bipartisan privacy legislation, the most important precursor for protections for AI that impact vulnerable populations.
FPF Submits Comments to the FEC on the Use of Artificial Intelligence in Campaign Ads
On October 16, 2023, the Future of Privacy Forum submitted comments to the Federal Election Commission (FEC) on the use of artificial intelligence in campaign ads. The FEC is seeking comments in response to a petition that asked the Agency to initiate a rulemaking to clarify that its regulation on “fraudulent misrepresentation” applies to deliberately deceptive AI-generated campaign ads.
FPF’s comments follow an op-ed FPF’s Vice President of U.S. Policy Amie Stepanovich and AI Policy Counsel Amber Ezzell published in The Hill on how generative AI can be used to manipulate voters and election outcomes, and the benefits to voters and candidates when generative AI tools are deployed ethically and responsibly.
With contributions from Aaron Massey, FPF Senior Policy Analyst and Technologist, Keir Lamont, Director for U.S. Legislation, and Tariq Yusuf, FPF Policy Intern
Several technologies can help individuals configure their devices to automatically opt out of web services’ requests to sell or share personal information for targeted advertising. Seven state privacy laws require that organizations honor opt-out requests. This blog post discusses the legal landscape governing Universal Opt-Out Mechanisms (UOOMs), as well as the key differences between the leading UOOMs in terms of setup, default settings, and whether those settings can be configured. We then offer guidance to policymakers to consider clarity and consistency in establishing, interpreting, and enforcing UOOM mandates.
The legal environment behind Universal Opt-Out Mechanisms
Online advertising continues to evolve, specifically in reaction to new regulatory requirements as an increasing number of international jurisdictions and U.S. states have enacted comprehensive privacy laws. As of October 2024, twelve states grant individuals the right to opt out of businesses selling their personal information or processing that data for targeted advertising. Of these twelve state privacy laws, seven include provisions that make it easier for individuals to opt out of certain uses of personal data. This includes the kind of personal and pseudonymized information that is routinely shared with websites, such as browser information or information sent via cookies.
Historically, a significant practical hurdle existed in the implementation of opt-out rights: users wishing to exercise the right to opt out of the use of this information for targeted advertising must locate and manually click opt-out links that businesses provide on their web pages, and they generally must do so for every site they visit. To make opting out easier, seven state’s privacy laws (California, Colorado, Connecticut, Delaware, Montana, Oregon, and Texas) require businesses to honor individuals’ opt-out preferences transmitted through Universal Opt-Out Mechanisms (UOOMs) as valid means to opt out of targeted advertising and data sales. UOOMs refer to a range of desktop and mobile tools designed to provide consumers with the ability to configure their devices to automatically opt out of the sale or sharing of their personal information with internet-based entities with whom they interact. These tools transmit consumers’ opt out preferences by using technical specifications, chief among these the Global Privacy Control (GPC).
California became the first state to establish the force of law for opt-out signals as valid opt-outs through an Attorney General rulemaking process in August, 2020. Specifically, businesses who do not honor the Global Privacy Control on their websites may risk being found in noncompliance with the California Consumer Privacy Act (CCPA), which was the central topic in the recent enforcement action against Sephora, an online retailer. In the complaint, state authorities alleged that Sephora’s website was not configured to detect or process any GPC signals and, as a result, failed to honor users’ opt-out preferences by not opting them out of sales of their data.
Although other UOOMs exist (and more are likely to emerge), we focus exclusively on the tools endorsed by the creators of the Global Privacy Control specification. In 2023, the FPF team downloaded and installed each tool and evaluated each tool’s installation process, whether GPC signals were sent without additional configuration, and whether those settings could be adjusted (see Figure 1 below).
Installation
GPC Signals Sent without Additional Configuration
Can the Configuration Be Adjusted?
IronVest
Requires account sign-up
❌ No
Yes; GPC can be enabled only on a per-site basis, not globally.
Brave Browser
No steps required after installation
✅ Yes
No; GPC cannot be disabled, either globally or per-site, even when other protections in the “Shields” feature are turned off.
Disconnect
No steps required after installation
❌ No
Yes; GPC can be enabled globally but not on a per-site basis using a checkbox in the main browser plugin window.
DuckDuckGo Privacy Browser
No steps required after installation
✅ Yes
Yes; GPC can be disabled globally but not on a per-site basis.
DuckDuckGo Privacy Essentials
No steps required after installation
✅ Yes
Yes; GPC can be disabled both globally or on a per-site basis by disabling “Site Privacy Protection.”
Firefox
Requires technical configuration
❌ No
Yes, GPC can be disabled globally in the browser’s technical configuration but not on a per-site basis.
OptMeowt
No steps required after installation
✅ Yes
Yes; GPC can be disabled both globally or on a per-site basis by disabling the “Do Not Sell” feature.
Privacy Badger
No steps required after installation
✅ Yes
Yes; GPC can be disabled both globally or on a per-site basis by disabling the “Do Not Sell” feature.
Figure 1: Observations of eight leading UOOM toolsas of October 12, 2023
Our survey allows us to make four key observations about the state of these UOOMs.
Current GPC implementations are largely limited to browser plugins for desktop environments. Google Chrome, Microsoft Edge, and Safari do not natively support the GPC signal. Mozilla Firefox supports sending the GPC signal, but configuring was the most challenging setup of all the tools we tested. Brave and DuckDuckGo are the only browsers that natively support the GPC. In addition, Brave and DuckDuckGo are the only desktop and mobile browsers with GPC enabled by default.
GPC tools significantly differ from one another in user experiences for both installation and use. The installation process for six of the tools was direct and, therefore suitable to a broad range of consumer knowledge. Two of the tools, IronVest and Firefox, require additional steps to enable GPC. Ironvest requires the creation of an account upon downloading the tool, and through that account offers not only GPC but also a subscription-based suite of further online security services like password managers and email maskers. By contrast, Firefox does not require an account, but it requires users complete more steps to enable the GPC that require technical knowledge or experience. Specifically, users must access the about:config settings page in Firefox, which warns the user to “Proceed with Caution” and requires users to know how to find the GPC configuration options. Users with limited experience configuring about:config settings on this browser may struggle to enable the GPC signal on Firefox. Following FPF’s study on September 25, 2023, Mozilla enabled a graphical UI setting for GPC in Firefox Nightly. Firefox Nightly provides tech savvy users with more experimental builds of Firefox. Features typically migrate from Nightly to the more broadly available Firefox browser over time.
GPC tools differ significantly in their default settings after installation, potentially creating consumer confusion in switching from one service to another. Three of the tools leave the GPC off by default following final installation; four of them enable the GPC by default. Firefox, for example, does not enable GPC by default, and it requires the most work to enable, whereas Brave enables GPC by default without notifying users or allowing them to disable it. Many tools include other privacy features in addition to GPC, such as Privacy Badger’s ability to block surreptitious tracking mechanisms like supercookies. These tools were not examined in this report, though they may create divergent user experiences that can cause consumers to draw different conclusions as to each tool’s utility and effectiveness. Users installing a privacy-focused browser extension or using a privacy-focused browser may be unaware that in certain cases privacy features are disabled by default and require additional configuration after installation.
Finally, we observe that these tools significantly differ in configuration options for when and where to send the GPC signal. The tools collectively deploy two types of configuration: globally sending the GPC to every site and/or selectively sending the GPC on a per-site basis. None of the tools have pre-configured profiles or “allow / deny” lists for when to send the GPC, and about half of the tools allow users to set the GPC both as a global setting and on a per-site basis. IronVest only allows sending the GPC on a per-site basis, while Brave only enables the GPC on a global basis. However, given that most state laws that require compliance with a UOOM also require affirmative consent to opt back in following an opt-out, it is unclear whether disabling the GPC signal for a site after visiting it will have legal effect.
Next Steps & Policy Considerations
In 2023 alone, six states passed comprehensive privacy laws. In the years ahead, we expect that more states will be added to this list, and many are likely to include provisions regarding UOOMs. Policymakers must ensure that all UOOM requirements offer adequate clarity and consistency.
One place where greater detail from policymakers would provide benefit to organizations seeking to comply with legal requirements is in guidance not only for covered businesses, but also for vendors of consumer-facing privacy tools. Specifically, guidance would be useful regarding how a UOOM must be configured or implemented to give assurance that the GPC signals being sent are a legally valid expression of individual intent. For example, a minor detail such as whether a tool contains a “per-site” toggle for the GPC may be significant in one state, but not another.
Similarly, the question of “default settings” and their legal significance requires greater clarity in many jurisdictions. For example, to be considered a valid exercise of individuals’ opt-out rights under Colorado law, a valid GPC signal occurs when individuals provide “affirmative, freely given, and unambiguous choice.” This requirement creates an engineering ambiguity for publishers and websites over the validity of GPC signals they receive. For example, users installing a browser extension that requires a separate, affirmative user configuration prior to sending the GPC signal will unambiguously be a valid expression of individual choice. On the other hand, an individual using a browser marketed with a variety of privacy preserving features, including the GPC, may be sending a GPC signal that does not meet the law’s standards for defaults if those features are enabled by default and they do not provide notice to users. The user may have wanted a privacy feature other than GPC and not been aware that the GPC signal would be sent. On the other hand, another user may both be seeking and appreciate a default-on GPC and not want it to be legally ignored because they didn’t affirmatively enable it. Publishers and websites do not have an engineering mechanism to differentiate between these scenarios, incentivizing them to use nonstandard techniques, like fingerprinting, for the purposes of discerning which GPC signals are valid.
New states implementing comprehensive privacy laws also increase the odds that specific privacy rights may fracture across jurisdictions in ways that are either cohesive or irreconcilable. The current GPC specification does not support conveying users’ jurisdictions, so it is unclear how organizations must differentiate between signals originating from one jurisdiction or another. The result could be that entities must choose which state to risk running afoul of the law in such that they may follow the requirements of a conflicting jurisdiction.
As user-facing privacy tools are developed and updated, responsible businesses will likely err on the side of over-inclusion by treating all GPC signals as valid UOOMs. However, increased user adoption and the expansion of the GPC into new sectors (such as connected TVs or vehicles) could change expectations and put more pressure on different kinds of advertising activities. In the absence of uniform federal standards that would create guidance for such mechanisms, most businesses will aim to streamline compliance across states, providing a significant opportunity for policymakers to shape the direction of consumer privacy in the coming years. Policymakers must be aware of these developments and strive for clarity and consistency in order to best inform organizations, empower individuals, and set societal expectations and standards that can be applied in future cases.
FPF Weighs In on the Responsible Use and Adoption of Artificial Intelligence Technologies in New York City Classrooms
Last week, Future of Privacy Forum provided testimony at a joint public oversight hearing before the New York City Council Committees on Technology and Education on “The Role of Artificial Intelligence, Emerging Technology, and Computer Instruction in New York City Public Schools.”
Specifically, FPF urged the Council to consider the following recommendations for the responsible adoption of artificial intelligence technologies in the classroom:
Establish a common set of principles and definitions for AI, tailored specifically to educational use cases;
Identify AI uses that pose major risks – especially tools that make decisions about students and teachers;
Create rules that combat harmful uses of AI while preserving beneficial use;
Build more transparency within the procurement process with regard to how vendors use AI; and
Take a student-driven approach that enhances the ultimate goal of serving students and improving their educational experience.
During this back to school season, we are observing school districts across the country wrestle with questions about how to manage the proliferation of artificial intelligence technologies in tools and products used in K-12 classrooms. In the 2022-2023 school year, districts used an average of 2,591 different edtech tools. While there is no standard convention for indicating that a product or service uses AI, we know that the technology is embedded in many different types of edtech products and has been for a while now. We encourage districts to be transparent with their school community regarding how AI is utilized within the products it is using.
But first, it is critical to ensure uniformity in how AI is defined so that it is clear what technology is covered and to avoid creating overly broad rules that may have unintended consequences. A February 2023 audit by the New York City Office of Technology and Innovation on “Artificial Intelligence Governance” found that the New York City Department of Education has not established a governance framework for the use of AI, which creates risk in this space. FPF recommends starting with a common set of principles and definitions, tailored specifically to educational use cases.
While generative AI tools such as ChatGPT have gained public attention recently, there are many other tools already used in schools that fall under the umbrella of AI. Uses may be as commonplace as autocompleting a sentence in an email or speech-to-text tools to provide accommodations to special education students, or more complicated algorithms used to identify students at higher risk of dropping out. Effective policies governing the use of AI in schools should follow a targeted and risk-based approach to solve a particular problem or issue.
We can look to the moratorium on adopting biometric identification technology in New York schools following the 2020 passage of State Assembly Bill A6787D as an example of how an overly broad law can have unintended consequences. Although it appeared that lawmakers were seeking to address legitimate concerns stemming from facial recognition software used for school security, a form of algorithmic decision making, the moratorium had broader implications. Arguably, it could be viewed to ban the use or purchase of many of the computing devices used by schools. This summer, the NY Office of Information Technology Services released its report on the Use of Biometric Identifying Technology in School, following which it is likely that the Commission will reverse or significantly modify the moratorium on biometric identification technology in schools. This will present an opportunity for the city to consider what additional steps should be taken if it resumes use of biometric technology and will also likely open a floodgate for new procurement.
Accordingly, this is an important moment for pausing to think through the specific use cases of AI and technology in the classroom more broadly, identify the highest risks to students, and prioritize developing policies that address those higher risks. When vetting products, we urge schools to consider whether that product will actually enhance the ultimate goal of serving students and improving their educational experience and whether the technology is indeed necessary to facilitate that experience.
We urge careful consideration about the privacy and equity concerns associated with adopting AI technologies as AI systems may have a discriminatory impact on historically marginalized or otherwise vulnerable communities. We have already seen an example of how this can manifest in classrooms. Commonly deployed in schools, self-harm monitoring technology works by employing algorithms that rely on scanning and detecting key words or phrases across different student platforms. FPF research found that “using self-harm monitoring systems without strong guardrails and privacy-protective policies is likely to disproportionately harm already vulnerable student groups.” It can lead to students being needlessly put in contact with law enforcement and social services or facing school disciplinary consequences as a result of being flagged. We recommend engaging the school community in conversation prior to adopting this type of technology.
It is also critical to note that using any new classroom technology typically comes with increased collection, storage, and sharing of student data. There are already requirements under laws like FERPA and New York Ed Law 2-D. Districts should have a process in place to vet any new technology brought into classrooms and we urge an emphasis on proper storage and security of data used in AI systems to protect against breaches and privacy harms for students. School districts are already vulnerable as targets for cyber attacks, and it is important to minimize risk.
Finally, we flag that there are disparities in the accuracy of decisions made by AI systems and caution that there are risks when low accuracy systems are treated as gospel, especially within the context of high impact decision making in schools. Decisions made based on AI have the potential to shape a student’s education in really tangible ways.
We encourage you to consider these recommendations and thank you for allowing us to participate in this important discussion.
