Trumpxiety
Sources
The four pillars of a political anxiety barometer
The trusted instruments behind the political anxiety barometer
Behind every credible news-sentiment system sits a stack of trusted instruments: a planetary-scale event database, a multilingual intelligence platform, a fast and affordable real-time feed, and a reasoning engine to make sense of it all. The political anxiety barometer is built on four such instruments — the GDELT Project, NewsAPI.ai (Event Registry), CurrentsAPI, and Anthropic's Claude — each chosen because it represents the gold standard in its category. Together they cover decades of historical context, real-time global ingestion across more than a hundred languages, deep semantic enrichment, and state-of-the-art structured synthesis. What follows is a vendor-by-vendor account of why each was selected, drawing on their founders' own words, their academic and industry track records, and the specific technical capabilities that make them trustworthy components of a public-facing measurement of political mood.
GDELT, the world's largest open ledger of human events
The GDELT Project — Global Database of Events, Language, and Tone — describes itself as "the largest, most comprehensive, and highest resolution open database of human society ever created," and the academic and journalism communities have largely agreed. Launched publicly in April 2013 by Kalev Hannes Leetaru and political scientist Philip A. Schrodt, GDELT was born from a simple but audacious ambition: to encode, in real time, every story the world's news media tells about itself. Leetaru, a Foreign Policy "Top 100 Global Thinker" of 2013 and former Yahoo! Fellow at Georgetown's Institute for the Study of Diplomacy, partnered with Schrodt — creator of the Kansas Event Data System (1994) and the TABARI machine-coder (2000) — to graft decades of political-science event-coding tradition onto a planetary-scale data pipeline. The project is supported by Google Jigsaw (formerly Google Ideas) and runs on Google Cloud Platform, with additional backing from the National Science Foundation, BBC Monitoring, the Internet Archive, JSTOR, and the U.S. Institute of Peace.
What sets GDELT apart is the sheer scale of its coverage and the depth of its semantic enrichment. It monitors print, broadcast, and web news in over 100 languages from nearly every country, translates 65 languages in real time (covering 98.4% of its non-English volume), and refreshes its three core data streams — the Event Database, the Global Knowledge Graph, and the Visual Knowledge Graph — every fifteen minutes. The Event Database alone codifies activity in over 300 CAMEO categories, geolocated to the city or even mountaintop level, with each event scored on the Goldstein Scale (−10 to +10) for societal impact and a tone score (−100 to +100) for media sentiment. The crown jewel is GCAM, the Global Content Analysis Measures, which Leetaru calls "the largest deployment of sentiment analysis in the world" — a fusion of 24 emotional measurement packages including LIWC, SentiWordNet, the Harvard IV-4 Psychosocial Dictionary, the Loughran-McDonald financial sentiment lexicons, and WordNet Affect, yielding more than 2,300 emotional and thematic dimensions (and growing toward 4,500). For a political anxiety barometer specifically, GCAM's LIWC "Anxiety" channel is essentially purpose-built for the task.
GDELT's credibility in the research community is exceptional. By 2018 the database contained over 3.2 trillion datapoints, and a single 2015 snapshot of the Global Knowledge Graph held more than three-quarters of a trillion emotional scores across 2.5 terabytes. It was a member of the inaugural Google BigQuery Public Datasets, and political scientist Jay Ulfelder famously predicted upon its release: "I suspect this is going to be the data set that launches a thousand dissertations." He has been proven right — GDELT now anchors peer-reviewed work in Science, the Journal of Peace Research, Journal of Conflict Resolution, and dozens of other venues, and powers the USAID/Humanity United Tech Challenge for Atrocity Prevention, the U.S. Institute of Peace Global Conflict Dashboard, BBVA's refugee-flow modeling during the 2015 European migration crisis, and forced-displacement early-warning systems funded by the National Geospatial-Intelligence Agency. As Leetaru puts it, GDELT exists to build "a living silicon replica of global society" — and for any system trying to measure political anxiety at planetary scale, that replica is the natural foundation.
NewsAPI.ai and the Slovenian school of semantic news intelligence
Where GDELT is a firehose of structured events, NewsAPI.ai — the developer-facing brand of Event Registry — is a curated, semantically enriched intelligence layer. The platform was born inside the Artificial Intelligence Laboratory at the Jožef Stefan Institute in Ljubljana, Slovenia, with deep ties to the University of Ljubljana. Its foundational paper, "Event Registry: Learning About World Events from News," was presented at WWW 2014 by Gregor Leban, Blaž Fortuna, Janez Brank, and Marko Grobelnik. The system grew out of the European Union's FP7 XLike project ("Cross-Lingual Knowledge Extraction"), where, as XLike's own deliverables document, "the main integrating platform and demonstrator system was moved to JSI's Event Registry." The project later spun out as a commercial company, with funding support from the Google Digital News Initiative.
The pedigree of the people behind it is rare in the news-API world. Marko Grobelnik is one of Europe's most influential AI figures: co-founder of UNESCO's International Research Centre on Artificial Intelligence (IRCAI), Slovenia's Digital Champion at the European Commission, a member of the OECD AI Committee, the Council of Europe AI Committee, NATO's Data and AI Review Board, and the Global Partnership on AI, and former general chair of TheWebConf (WWW) 2021. Gregor Leban, the system's lead architect and now public-facing founder, is repeatedly cited in user reviews for personally helping customers integrate advanced features. The OECD itself uses Event Registry to power its AI Incidents Monitor, a fact that quietly settles the question of institutional trust.
Technically, NewsAPI.ai is built around cross-lingual event clustering: articles describing the same world event in different languages are merged into a single event object via a multithreaded streaming clustering algorithm and a canonical-correlation-analysis cross-lingual classifier, with entities resolved to Wikipedia/DBpedia URIs through the in-house Wikifier. The platform monitors 150,000+ news sources in 60+ languages, tagging articles with concepts, categories (using a DMOZ-trained classifier), named entities, and VADER-based sentiment scores at both article and event level. Their own engineering blog explains the choice plainly: "After trying out a bunch of different sentiment analysis tools, we found VADER to be hands down the best for analyzing media, making it our go-to choice."
That choice deserves its own explanation. VADER — Valence Aware Dictionary and sEntiment Reasoner — was introduced by C.J. Hutto and Eric Gilbert at Georgia Tech in their 2014 ICWSM paper "VADER: A Parsimonious Rule-based Model for Sentiment Analysis of Social Media Text." Built on a human-validated lexicon and five grammatical rules covering punctuation, capitalization, intensifiers, the contrastive "but," and negation, VADER famously outperformed individual human raters (F1 = 0.96 vs. 0.84) on tweet sentiment classification and beat eleven strong baselines including LIWC, ANEW, SentiWordNet, Naive Bayes, and SVM. The paper has been cited more than 5,200 times on Semantic Scholar alone and is one of the most influential sentiment-analysis works of the last decade. Its determinism, transparency, and lack of training requirements make it ideal for streaming pipelines — exactly NewsAPI.ai's use case. Notable downstream users include the OECD.AI Observatory, the BAR-Analytics platform (which used Event Registry to assemble 350,000+ articles on the Russo-Ukrainian and Israeli-Palestinian conflicts), Headlyne.ai's consumer news aggregator, quantitative finance teams, and ESG data providers listed on Datarade. For a political anxiety project, NewsAPI.ai delivers what GDELT cannot: clean, deduplicated, entity-resolved articles with per-piece sentiment baked in.
CurrentsAPI, the developer-friendly real-time feed
The third pillar trades scale-of-research for speed, affordability, and developer ergonomics. CurrentsAPI is operated by Pheme Pte. Ltd., a Singapore company incorporated on April 3, 2019, under the brand Currents Lab. The maintainer, GitHub user theblackcat102, is a research scientist at Appier whose open-source portfolio — including the deep-learning content extractor ExtractNet (forked from Dragnet), the multilingual identifier fastlangid, and the widely-starred awesome-vector-search list — signals the engineering sensibility behind the API. The project launched publicly in 2019 alongside the Pheme incorporation and has shipped Python, Node.js, and R SDKs ever since.
CurrentsAPI's current footprint is substantial and unusually accessible. The service monitors more than 120,000 domains across 70+ countries, ingests 90,000+ articles per day, holds an archive of 26 million+ articles stretching back roughly four years, and serves results in 20+ languages including English, German, Spanish, French, Italian, Portuguese, Chinese, Japanese, Korean, Russian, Arabic, Hindi, Dutch, Polish, Turkish, and Indonesian. Average response latency is under 200 ms with a 99.9% uptime SLA. The REST API is straightforward — latest-news, search, sources endpoints with full Boolean query syntax (AND/OR/NOT, quotes, parentheses), filtering by language, country, category, domain, author, and date range, and a V2 schema that adds cursor-based keyset pagination and a richer canonical category taxonomy spanning politics/government, economy/business, science/technology, environment, health, and twelve more.
What makes CurrentsAPI distinctive for a political anxiety project isn't a clever model — it's the economics and the breadth. The free tier offers 1,000 production-grade requests per day with no credit card, paid tiers begin at just $69/month for 75,000 requests, and the service advertises roughly 2,000 requests per dollar — a price point its own marketing describes as "500x cheaper than leading alternatives." For comparison, the most common competitor's paid plans start at $449/month, and that competitor's free tier is restricted to localhost development. CurrentsAPI also explicitly ingests blogs, forums, and discussion content via its type parameter, broadening source diversity beyond traditional newswire feeds. Its in-house ExtractNet deep-learning extractor produces cleaner author/headline/date/keyword fields than RSS-based pipelines, and full CORS support means a browser-side dashboard can fetch directly from the API. In practical terms, CurrentsAPI is the redundancy layer and freshness backstop: when GDELT's 15-minute window or NewsAPI.ai's enriched pipeline lags, CurrentsAPI provides an independent, low-latency JSON feed across the same languages and regions, with rate-limit headers exposed on every response so a production scheduler can pace itself elegantly. It's listed in RapidAPI, publicapi.dev, Slashdot's "Top News API Alternatives 2026," and SourceForge — the standard discovery surfaces developers actually use.
Claude as the reasoning layer
The fourth pillar is the synthesis engine. Anthropic was founded in 2021 by Dario Amodei (former VP of Research at OpenAI, where he led GPT-2 and GPT-3 work and co-authored the seminal 2016 paper "Concrete Problems in AI Safety") and his sister Daniela Amodei (former VP of Safety and Policy at OpenAI), along with five colleagues including Jared Kaplan, Sam McCandlish, Tom Brown, Christopher Olah, and Jack Clark. The company is structured as a public benefit corporation governed by the Long-Term Benefit Trust, with a stated mission to build reliable, interpretable, and steerable AI systems. Dario describes the founding intuition succinctly: "There was a group of us within OpenAI… [with] a very strong focus belief in two things… One was the idea that if you pour more compute into these models, they'll get better and better… The second was the idea that you needed something in addition to just scaling the models up, which is alignment or safety." By April 2026, Anthropic stands at a $380 billion valuation following its February 2026 Series G, with annualized revenue around $30 billion, 300,000+ business customers, 70% of the Fortune 100 as users, and freshly announced compute partnerships totaling tens of billions with Amazon, Google, Microsoft, and Nvidia.
Anthropic's distinctive technical contribution to AI safety is Constitutional AI, introduced in the December 2022 paper "Constitutional AI: Harmlessness from AI Feedback." Instead of relying solely on human labelers to flag harmful outputs (the standard RLHF approach), CAI gives the model a written "constitution" — drawing on the UN Universal Declaration of Human Rights and other principles — and trains the model to self-critique and revise its own responses against those rules, with a reinforcement-learning phase that uses AI-generated preferences (RLAIF). The result is a model that is simultaneously more harmless and more helpful than RLHF baselines, engaging with sensitive queries by explaining its reasoning rather than refusing flatly. For a political anxiety barometer that must handle politically charged material with even-handedness, this matters: Promptfoo's open political-bias evaluation in 2025 ranked Claude Opus 4 the most centrist of the major frontier models, and Anthropic's own internal evaluations report that Claude Sonnet 4.5 is more politically even-handed than GPT-5 and Llama 4.
The Claude family has evolved at a remarkable cadence — Claude 1 (March 2023), Claude 2 (July 2023, with a 100K-token context that pushed the industry forward), the Claude 3 family of Haiku/Sonnet/Opus (March 2024), Claude 3.5 Sonnet (June 2024), Claude 3.7 Sonnet with hybrid reasoning (February 2025), Claude 4 Opus and Sonnet (May 2025, released under Anthropic's ASL-3 safety protocols), and through 2025–2026 the Sonnet 4.5, Haiku 4.5, Opus 4.5, Opus 4.6, Sonnet 4.6, and Opus 4.7 (April 16, 2026) generations. Opus 4.5 became the first model ever to break 80% on SWE-bench Verified (80.9%), and the current GA models support 1-million-token context windows, suitable for ingesting weeks of news at once. For structured-data work specifically, Anthropic's Structured Outputs feature compiles a developer's JSON Schema into a grammar that constrains the model's output, guaranteeing valid JSON; combined with strict-mode tool use and native Pydantic/Zod helpers, it makes Claude one of the most reliable choices in the industry for producing typed, parseable output at scale.
The academic case for using LLMs in news sentiment work is now decisive. Kirtac and Germano's 2024 study of 965,375 U.S. financial news articles found that GPT-class LLMs achieved 74.4% sentiment-classification accuracy versus just 50.1% for the Loughran-McDonald dictionary, with the LLM-driven trading strategy producing a Sharpe ratio of 3.05 against 1.23 for the dictionary baseline. A 2025 CLiC-it benchmark concluded bluntly that "lexicon-based sentiment analysis methods like VADER and TextBlob are insufficient because they fail to capture contextual financial semantics… Domain-adapted models like DeBERTa-v3-absa are surpassed by generative LLMs such as ChatGPT variants and DeepSeek-R1." Anthropic's own customers showcase Claude's role at the synthesis layer of journalism and analysis pipelines: Notion, Quora's Poe, DuckDuckGo's DuckAssist, Zoom, Slack, Asana ("Anthropic has been a strategic partner in developing our AI strategy at Asana," per Asana's Eric Pelz), Snowflake, Microsoft 365 Copilot (since March 2026), NASA (which used Claude Code to plan a Perseverance rover route in December 2025), and Norway's Government Pension Fund Global, the world's largest sovereign wealth fund, which uses Claude to screen its $2.2 trillion portfolio for ESG risks. The Washington Post's data journalist Kevin Schaul put the journalism case crisply in February 2026: "When you're doing data journalism, vibes are not enough… Having AI write and execute code that can be audited? I'm quite comfortable with that."
Why these four, together
The architecture choice is more than a sum of vendors — it's a deliberate triangulation. GDELT contributes deep history, planetary scale, and a sentiment substrate (GCAM, with its LIWC anxiety channel) explicitly designed by political scientists for political phenomena. NewsAPI.ai brings semantic enrichment, cross-lingual event clustering, and per-article VADER scores, grounded in the academic rigor of the Jožef Stefan Institute and the institutional trust of the OECD. CurrentsAPI ensures real-time freshness, source diversity, and economic accessibility, with a clean REST surface that any production system can lean on. And Claude — the most safety-trained, most politically even-handed, most schema-faithful frontier model available — provides the reasoning, summarization, and structured-JSON synthesis that turns three torrents of raw signal into a single, coherent measurement.
Each vendor on its own would be a defensible choice; together they form a stack that has been independently validated by governments (USAID, USIP, OECD), universities (Penn State, Georgetown, Ljubljana, Stanford), news organizations (the Washington Post, the New York Times, BBC Monitoring), and the world's most demanding enterprise buyers. As Kalev Leetaru wrote of GDELT's mission, the goal is to take "the world's information and try to catalog what's happening around the world, moment by moment, and how are people reacting to that." A political anxiety barometer is, at heart, an attempt to do exactly that — and these four sources are the instruments that make it possible.