Abstract
The public release of ChatGPT in late 2022 rapidly turned generative AI from a technological novelty into a central governance concern, prompting legislation, international summitry, and sustained public debate about regulation, safety, and accountability. Online news plays a central role in shaping how these issues are publicly understood, making web-based news infrastructure a critical object of Web Science inquiry.
This paper presents one of the first large-scale longitudinal framing analyses of generative AI governance in online news. Drawing on 1,116,091 articles from the GDELT 2.0 Global Knowledge Graph, spanning November 2022 to June 2026, we apply a six-category governance frame taxonomy (Innovation & Opportunity, Risk & Safety, Regulation & Governance, Rights & Privacy, Economic Competition & Labour, and Misinformation & Integrity) via a two-stage procedure combining multilingual keyword matching and LaBSE sentence-embedding confirmation.
We document a discourse arc from optimism to precaution: Innovation framing peaks in February 2023 before retreating, while Risk & Safety rises steadily from 2024 onward. An event study across eight governance milestones shows that media framing responds selectively and unevenly to policy events: high-profile intergovernmental summits coincide with less risk framing but more governance vocabulary, civil-society alarm signals raise risk framing without any governance-response coverage, and routine administrative deadlines produce no lasting shift. Regional analysis shows US outlets are at once the most regulation-intensive and innovation-positive, while EU outlets carry proportionally higher risk framing relative to their regulation share. These findings contribute to Web Science and computational social science by showing how web-based news infrastructures shape public discourse around the governance of emerging AI technologies.
Key Findings
News articles from 44 months (Nov 2022 – Jun 2026) across 100+ languages via GDELT 2.0, with a 50.7% governance filter yielding the analysis corpus.
Innovation & Opportunity peaks at 13.1% — the apex of post-ChatGPT optimism — before retreating as Risk & Safety climbs to 20%+ by 2024.
Around the Bletchley Summit risk framing dropped (−4.9 pp), while the Pause AI letter raised it (+1.5 pp) without any matching governance-response coverage.
Figure 1. Daily confirmed frame prevalence (14-day rolling mean, stacked area) and total articles per day. Regulation & Governance dominates throughout; Risk & Safety rises sharply from 2024; Innovation & Opportunity retreats after its February 2023 peak. Vertical lines mark the eleven governance milestones.
Figure 2. Daily confirmed frame prevalence for US, EU, and UK outlets (14-day rolling mean). US regulation framing diverges sharply upward from 2024; EU risk framing tracks closely with the UK throughout the study period.
Data & Code
The dataset has three configurations: articles (1,116,091 rows — document URL,
month, region, 6 keyword flags, 6 LaBSE embedding scores, dominant frame),
event_studies (22-row milestone summary), and aggregates
(monthly/regional/tone summaries). Load with
datasets.load_dataset("brewcoua/genai-gdelt-framing", "articles").
Released under CC-BY 4.0.
Citation
@inproceedings{couaran_framing_2026,
title = {Framing Generative {AI} Governance in Online News:
A Timeline Analysis from ChatGPT’s Launch to the EU AI Act (2022-–2026)},
author = {Couaran, Brewen and Pathania, Yuvraj Singh and Nair, Arjun Rajesh},
year = {2026},
url = {https://github.com/brewcoua/GenAI-GDELT},
}