When artificial intelligence executives warned of existential risks in September, followed days later by President Trump's signing of a voluntary industry framework for AI governance, the disconnect exposed a deeper problem: the public and researchers lack the visibility needed to evaluate whether AI poses genuine dangers or whether oversight measures are working. Industry secrecy around AI systems mirrors the long struggle researchers have faced accessing social media data, but the stakes are considerably higher.

Restricted access to data and models hands significant leverage to AI companies while constraining the work of independent researchers and public scrutiny. Self-regulation only functions when corporate interests align with public ones or when external pressure forces accountability. Yet external pressure requires that outside researchers and the broader public possess sufficient insight to judge impact. Without transparency requirements, neither policymakers nor citizens can determine whether any slowdown or oversight is warranted.

The social media precedent

Digital trace data—records of online activity created for purposes like advertising or audience measurement rather than research—has proven valuable for studying phenomena ranging from human trafficking networks to propaganda campaigns. Yet after more than two decades of social media research, fundamental questions about platform harms remain unsettled, largely because researchers still lack reliable, durable access to the data they need.

Analysis of 113 researcher requests at the Princeton Accelerator, which develops tools for studying the information environment, reveals consistent demand for straightforward capabilities: the ability to independently analyze digital trace data, preserve it, and share it with collaborators. These sound like basic scientific requirements, yet the social media research space has never solved them adequately.

Seven categories of resources currently exist for researchers studying the information environment, each with significant limitations:

  • Company platforms and systems, accessible only through restricted APIs or controlled research environments. Four years ago, most peer-reviewed research on the information environment focused on X/Twitter; after the platform raised academic access fees from free to $42,000 per month, that changed. Even when companies offer researcher data programs, access is limited, discretionary, and typically non-permanent, making replication and validation difficult.
  • Survey panels, which are costly and capture only specific questions at a single moment in time.
  • For-profit data vendors, often prohibitively expensive and lacking scientifically representative datasets, creating risk that researchers draw conclusions about available data rather than the actual phenomenon.
  • Open-source and free tools, more accessible but suffering from the same representativeness problem.
  • Archives and preservation initiatives, helpful but with uneven coverage across populations, platforms, and data types.
  • Researcher support services including legal guidance, ethics review, and protection against harassment and intimidation.
  • Legal pathways, exemplified by Article 40 of the Digital Services Act, which granted European researchers a formal right to access social media data.

The DSA route, however, has proven slow and unreliable in practice. Approvals take months. X has denied requests without justification. TikTok has returned incomplete datasets. Meta excluded European politicians with fewer than 25,000 followers. These failures prompted the European Commission to issue preliminary findings against Meta and TikTok and impose a €120 million fine on X. If companies ignore legal penalties, voluntary compliance frameworks appear unlikely to succeed.

AI presents a harder problem

Studying AI systems poses greater obstacles than social media research across multiple fronts. Privacy risks around chatbots are particularly acute: because interactions feel like private conversations with something resembling a person, users often share sensitive information. Researchers could study such interactions with consent, but the training data, model architectures, developer decisions—including safety rules and ongoing A/B testing—remain hidden from view.

Early efforts to address this gap exist. The European Union recently designated ChatGPT a "Very Large Online Search Engine," triggering transparency and researcher access obligations. Yet conversations where the chatbot contributes its own responses may fall outside these requirements.

AI may generate substantial benefits or prove a national security necessity. Neither justifies the absence of systematic measurement and study, even if the most sensitive capabilities remain restricted for safety and security. Even if DSA access improved, gaps would persist—particularly for researchers outside Europe. A global-scale experiment is unfolding, yet the companies designing the systems are also writing the rules, and independent researchers cannot assess the effects.

No formal pathway currently exists for independent researchers to study AI systems or tools at scale, let alone their broader effects on people, economies, and societies. Only after such a pathway is established can decisions about guardrails and development pace be properly informed. The United States never created a data access framework for social media research. The question now is whether it will do so for AI, as the technology becomes an increasingly dominant source of information.