Through its Epistemic Agency initiative, aihorizon R&D is pioneering a reproducible framework for detecting and measuring censorship-like behaviors in AI language systems. Currently in its mid-phase, the project builds on our own research design and is strengthened by Quantpi’s advanced testing infrastructure, allowing us to pair rigorous quantitative analysis with transparent protocols. The goal: to shed light on one of the most urgent issues in AI governance — the invisible shaping of information access.

​

Our framework examines AI responses using three complementary metrics:

• Sensitivity Score (SSS) — gauges the political, cultural, or historical sensitivity of a topic.

• Linguistic Elaboration Score (LES) — measures expressive richness and structural complexity.

• Truth Score (TS) — benchmarks factual completeness against a vetted “Oracle” reference.

​

Together, these measures reveal whether suppression occurs through omission, distortion, or subtle rhetorical shifts — and help distinguish between sounding informative and being informative.

​

We run large-scale tests across politically, historically, culturally, and technologically sensitive topics, archiving every prompt, output, and score for public audit. This enables transparent, side-by-side comparisons between different AI systems and between open and restricted configurations.

​

Why does this matter? AI models are becoming the primary gateways to knowledge worldwide. Like traditional media, they can be shaped by policy, bias, and regulation. By making censorship effects visible and measurable, we aim to strengthen the public’s epistemic agency — the right and ability to access, assess, and act on truthful information.

​

Once complete, the full methodology and results will be published in a peer-reviewed study, along with an open artifact archive, enabling researchers, NGOs, and policymakers to replicate, compare, and extend our work.

​

​