Conference Agenda

Transparency by Inquiry:

Unveiling Speech Norms in AI Content Moderation

Transparency has become a central governance tool in Artificial Intelligence (AI) regulatory frameworks[1] and it is widely presented as the cure for the opacity of AI systems. Existing legal initiatives, including the recently enacted AI Act and the Digital Services Act (DSA), seek to enhance AI transparency primarily by imposing various disclosure duties. Those include transparency reports of aggregate statistics, statement-of-reasons microdata, and periodical bias audits. These regulatory duties reflect the assumption that detailed data sharing obligations should yield procedural fairness and enable meaningful oversight by impacted stakeholders.

We term transparency that is operationalized primarily through self-reported disclosures of summary statistics and data: Transparency-by-Disclosure (TbD). We contend that TbD is insufficient because it fails to facilitate the full range of systemic investigations needed to meaningfully hold AI systems accountable. TbD risks reproducing the very opacity it seeks to undo for several reasons. First, issued by the same corporations that design or deploy AI systems, disclosures are inherently bounded by regulatory requirements and their discretional implementation by the platforms. Platforms often assert Intellectual Property protection or use privacy as pretext,[2] practically disclosing what they are willing to reveal, rather than what stakeholders seek to know.[3] Consequently, disclosures often contain partial data. Second, TbD does little to surface the underlying value choices encoded in AI systems. Their models learn patterns, make predictions, generate content, and operationalize norms that are not directly observable from disclosure outputs.[4] Making these latent normative choices legible to those humans impacted by the system adds another layer of complexity. Consequently, TbD creates and perpetuates systemic gaps in public knowledge: about how AI systems actually function, whom they affect, and which values they encode, yielding a governance landscape shaped as much by what remains unknown as by what is disclosed.

Accordingly, we propose Transparency-by-Inquiry (TbI) as a complementary approach to TbD. TbI positions transparency not as a data-availability exercise, but as a distributed practice of multiple stakeholders developing testable inquiries about AI systems, guided by their own interests and values and operationalized by their own tools. In TbI, the power to define what matters is hence redistributed outwards from the platform to society: The questions come first; the required data, methodology and access mechanisms follow from those inquiries.

We use AI content moderation systems as a case study to explore TbI. These systems encode speech norms that shape digital public discourse. Yet, despite their profound implications for free expression and democratic life, the underlying values and tradeoffs these specific speech norms operationalize remain concealed within layers of design choices. To illustrate the limits of TbD and the promises of TbI for extracting meaningful transparency, we examine how platforms treat antisemitic content. As a value-laden norm, antisemitism has various acceptable definitions which have become the subject of academic and political controversies. Nevertheless, even when platforms disclose data concerning hate speech moderation practices, they may not clarify which specific definition their model is applying in practice. This lack of transparency limits public understanding and oversight of how critical decisions about online speech are being made.

The paper critically examines the limitations of TbD in the context of AI content moderation and makes five key contributions: First, it introduces TbI as a complementary approach for achieving meaningful transparency and specifies its building blocks. Second, responding to growing calls for alternative transparency paradigms, it reframes AI transparency based on the ability of diverse stakeholders to articulate and answer targeted inquiries, rather than relying primarily on data disclosure. Third, it develops a typology that structures transparency inquiries for AI content moderation systems based on the ability to specify questions about components of the AI systems (and as a whole), data flows, speech norms, and the type of information sought. This deconstruction can support different stakeholders to bridge the gap between abstract accountability concerns and concrete, testable inquiries. Fourth, it demonstrates the potential of TbI through a case study exploring how antisemitism is defined and operationalized in content moderation AI systems. Fifth, it analyzes the broader policy implications of TbI.

[1] Hannah Bloch-Wehba. 2019. Access to algorithms. Fordham L. Rev. 88 (2019), 1265.

[2] Niva Elkin-Koren, Maayan Perel, and Ohad Somech. 2025. Unlocking Platform Data for Research. Ind. L.J. 100, 4 (2025); Rory Van Loo. 2022. Privacy Pretexts. Cornell L. Rev. 108 (2022), 1.

[3] Evelyn Douek. 2022. Content moderation as systems thinking. Harv. L. Rev. 136 (2022), 526.

[4] Bernhard Rieder and Jeanette Hofmann. 2020. Towards platform observability. Internet policy review 9, 4 (2020), 1–28.