11 verified, source-linked AI implementations documented between 2016–2026.
First reader-facing AI product at The Guardian; AI analyzes the 200 most recent article headlines on a topic tag page and generates three narrative subtitles defining major storylines; curatorial tool surfacing narrative threads without accessing full article text to reduce hallucination risk; clearly labeled as AI-generated
Custom ML model developed in collaboration with UCL Department of Physics & Astronomy (Centre for Data Intensive Science and Industry) and UCL Political Science to analyse nearly 238,000 fragments of House of Commons immigration debates from 1925 to 2025; LLM-assisted annotation scaled training dataset to 22,600+ labelled parliamentary fragments; bespoke ML classification model applied to full century of Hansard records; identifies emotional sentiment and framing of immigration rhetoric across Labour and Conservative parties over 100 years; findings: UK parliamentary immigration sentiment reached most negative level in 2025; shift toward securitised 'border control/illegal immigration' frames; decline in integration-oriented language; published by The Guardian as interactive visual story (Feb. 25, 2026)
OpenAI content licensing and ChatGPT Enterprise partnership: Guardian journalism appears in ChatGPT with attribution;" ChatGPT Enterprise rolled out internally to develop reader-facing and business products
AI live blog summarisation tool: fine-tuned LLM on 3,700 decade-spanning live blog examples with human-written summaries to auto-generate rolling coverage summaries in Guardian editorial style; project ultimately abandoned after 4-month development due to high risk of undetected errors in long-form AI summaries
Guardian joins Google's AI article overviews pilot alongside its existing OpenAI and ProRata.ai content deals; builds on existing commercial relationships through the Google News Showcase scheme; deal delivers AI-powered article overviews on Google News pages, audio briefings, and real-time information in the Gemini app; articles carry clear attribution and links; Google compensates participating publishers; financial terms not disclosed; part of a wave of 13 global publishers joining simultaneously
An internal LLM tool that takes an article's copy and generates a news headline or a 'light' headline with an auto-generated pun. Named because it sounds a bit like 'Guardian' and contains 'LLM.' Reported as not yet used for any published pieces at the time of writing — an internal experiment consistent with the Guardian's stance that short, high-scrutiny outputs like headlines are the safest place to apply AI.
Guardian Media Group makes Guardian journalism available to ProRata's Gist.ai AI answer engine; ProRata shares 50% of advertising revenue proportionally with publishers based on content cited in AI responses; deal ensures attribution and compensation for Guardian content used in generative AI answers; simultaneous with DMG Media equity investment and Sky News content deal with ProRata; complements Guardian's February 2025 OpenAI content licensing agreement
Editorial-diversity analytics: an in-house NLP pipeline that identifies people quoted and bylined in Guardian articles and assigns a likely gender (cross-referencing spaCy predictions with Wikidata) to measure the representation of sources and authors over time. Used strictly at aggregate/trend level — never to label an individual — to track and improve how well coverage reflects the Guardian's audiences. A sibling to the quote-extraction project, echoing the BBC's 50:50 initiative that data lead Anna Vissens previously ran.
Built with Explosion (co-founder Ines Montani; spaCy library + Prodigy annotation tool) out of the 2021 JournalismAI Collab Challenges on "modular journalism," collaborating with AFP and other European/MENA newsrooms. Models extract quotes, distinguish real quotes from quotation-marked non-quotes, and resolve co-reference (linking "he/she" back to the speaker across paragraphs) — a hard case for Guardian articles where speakers sit paragraphs from the quote; pitched to UCL students for a co-reference solution. A key finding: much of the effort is scoping and annotation quality — internal annotators agreed far more than outside departments, showing a model is only as consistent as its annotation guide. A spin-off "hack" reused editors' pinned-comment flags as training data to surface the most valuable reader comments.
Facebook Messenger chatbot for personalised daily news delivery (UK, US and Australian editions; user-selected topics and 6–8am delivery times)""
The Guardian is one of 43 organizations in United Kingdom documented in the index, which holds 139 cases from the country. Its work spans 6 function types, including Audience Experience & Personalisation, documented across 110 organizations worldwide.