How the Associated Press Built its AI Strategy Without Breaking Trust
The AP offers a blueprint for sustainable AI adoption that prioritizes augmentation over replacement and collaboration over competition.
The Associated Press demonstrates how news organizations can implement artificial intelligence without sacrificing editorial credibility. The AP has maintained trust through standards-first development, internal testing before external deployment, and transparent communication with both newsroom staff and member organizations. Their approach yields practical tools like federal data monitoring systems and video archive search, but also reveals the fundamental challenge facing news organizations: most AI experiments drain money rather than generate revenue.
Last week I interviewed Troy Thibodeaux, AP's Director of Products and Services. In our phone call, we covered strategic approaches and principles, areas of innovation, and what’s next on the AP’s AI journey.
When Thibodeaux started implementing large language models, he initially faced pushback. Newsroom staff worried about job security. Members and customers demanded "nothing that AI has touched" in their feeds. The response was to establish AI standards right away.
"From the very beginning, one of the first things we did was to meet with the standards team," Thibodeaux explained to me. The standards team helped create specific guidelines about AI use, with monthly reviews and six-month update cycles. They established red lines immediately: no generated imagery, no content delivered to customers without journalist review.
This standards-first approach has allowed AP to avoid the missteps that have hurt other news organizations. While some outlets have experimented with AI-generated articles or imagery, AP's guardrails have kept them within bounds their newsroom and customers can accept.
Balancing Innovation With Member Demands
AP's position as a not-for-profit cooperative serving members creates unique pressures. As Aimee Rinehart, AP's Senior Product Manager AI Strategy, explained at the South by Southwest conference in March 2025: "Many of the businesses we service say, we don't want generative AI in our news feeds from you, so we have to listen to that."
This customer resistance shapes every AI decision. Unlike consumer-facing news organizations that can experiment more freely, AP must balance innovation with member demands. "We have to be able to give our members and customers what they need and also experiment," Rinehart notes.
There is a substantial difference of scale involved in anything the AP does. While the New York Times produces about 200 articles daily, AP generates 2,000 articles and 3,000 photos per day. This volume creates both opportunity and obligation.
Building Tools Journalists Actually Want
AP's AI development follows a simple principle: eat your own dog food first. Rather than building products for customers and hoping they work, AP develops tools for internal use, tests them thoroughly, then considers external applications.
The process starts with identifying genuine pain points. "What are the things that you would rather not do, the things that you feel are distracting from the heart of your journalism?" Thibodeaux asks his newsroom colleagues. This approach led to tools like automated shot listings for video content and document analysis pipelines for large data dumps.
Take the AP’s JFK papers project. When the Trump administration released roughly 2,200 files consisting of over 63,000 pages of previously classified documents related to President Kennedy's 1963 assassination in March 2025, AP's team analyzing those papers worked with the AI team and data team to create a pipeline for large document set analysis. The AI tool helped reporters move more quickly through the massive release, perform summarization of the documents, and identify what had been newly unredacted in previously released but redacted materials. The tool worked so well internally that it's now available for future document releases.
Internal Tools in Action
The Democracy team approached Rinehart with a specific problem: "We are drowning in podcasts. We cannot keep up with the amount of podcasts." The solution: a system that scrapes RSS feeds, automatically transcribes new episodes, identifies keywords, and alerts reporters to relevant content.
The tool proved its worth for tracking statements of Kash Patel when he was the nominee for the position of FBI director in January 2025. "There were more than 100 podcasts that Kash Patel was in, and we wanted to know what was he saying on there," Rinehart explained. The system worked so well that it became the first time the AP listed not only the bylines of the three reporters who wrote the story but, at the end of the story, also AI Product Manager Ernest Kung. He developed the tool that enabled the research.
AP has also built an AI editorial assistant outside their content management system for experimentation. Using OpenAI's API, the tool handles translations from English to Spanish (with Portuguese experiments underway), creates "write-throughs" (updates) for developing stories, generates headlines and SEO variations, and produces bullet point summaries.
Three AI Tools, Three Different Lessons
Local Lede: Turning Federal Data into Location-specific Stories
Perhaps AP's most ambitious AI project is Local Lede, developed in cooperation with Applied XL. The system monitors more than 430 federal agencies, identifying regulatory actions that could become local news stories. When the EPA fines a company, for instance, the system determines if it's newsworthy (significant fine amount, repeat violations) and identifies local relevance (company headquarters, major employers).
The ultimate goal is to transform the noise of federal bureaucracy into actionable local news tips. The system analyzes Federal Register entries, applies editorial judgment through AI, and delivers location-specific story suggestions complete with phone numbers for follow-up reporting.
Current status: Local Lede has expanded from five pilot organizations to 10 states, with tips flowing directly into AP's regular feed for those markets. But Thibodeaux is candid about the results: "We feel like it still needs improvement in order to really be valuable. It hasn't quite achieved the level we would expect."
The problem: Local newsrooms want more than tips. "What they really want is something they can publish," Thibodeaux explained to me. "We've heard 'this looks interesting, we could do this, but we don't have time to run this down.'" The filtering also needs work. Determining true newsworthiness remains challenging, and achieving hyperlocal relevance for very small markets is proving difficult.
Next steps: Local Lede is still in pilot mode. Thibodeaux acknowledges "it's possible that this won't live on." The focus is on improving filtering and exploring local data sets beyond federal sources.
An overview of some of the AI tools the AP offers for it members (screenshots AP, composite view News Machines)
ShortTok: Video Discovery Made Easy
ShortTok is the name of a provider but also the name of a tool that the AP developed with this provider. Shorttok tackles video discoverability across AP's massive daily output of 5,000 pieces of journalism per day.
The innovation: Two features are now integrated into the member dashboard AP Newsroom. First, related content suggestions that help customers find similar videos. Second, and more interesting, "storylines"—AI-generated identification of different angles and themes within video coverage of major stories. Instead of one perspective on a breaking news event, customers can see multiple narrative threads and choose what works for their audience.
Implementation: After a staged rollout, all AP Newsroom users now have access to these features as of a few months ago. The integration represents a complete redesign of AP's customer-facing platform.
Success metrics: AP tracks download increases and feature usage. The core question: can customers more easily discover and use relevant AP content? Thibodeaux notes they're examining "how many people are using the feature, how often do they follow that path down to downloading it and then presumably using it in their content."
The goal: Solve the discoverability problem that plagues all wire services. With thousands of pieces of content daily, helping customers find exactly what they need becomes a competitive advantage.
Merlin: Finding More Value in Archives
Merlin focuses on video archive search, but it represents AP's broader thinking about extracting value from decades of content. Unlike traditional taxonomy-based search, Merlin works directly from video content itself, using vectorization to understand what's actually happening in footage.
Current capability: Video search that goes beyond metadata and tags to understand actual content. This allows searches based on visual elements rather than just text descriptions.
Broader archive ambitions: Thibodeaux reveals three areas of development: discovering what's in archives, creating better metadata, and "thinking about what are the ways that we could use those archives to create something new, to create new products."
The untagged content opportunity: Many news organizations, including AP, have vast archives of untagged or poorly tagged content. AI could unlock this material by creating searchable descriptions and finding new contexts for historical content.
This three tools illustrate different AI maturity levels and challenges. ShortTok shows successful integration, Local Lede demonstrates the gap between concept and utility, and Merlin represents exploration with massive potential upside.
Fighting Deepfakes with AI Verification
AP is developing a comprehensive verification dashboard currently in pilot with select clients. The dashboard consolidates about 10 different verification tools and APIs into a single interface, designed to help newsrooms authenticate user-generated content in an era of increasingly sophisticated deepfakes.
The verification suite includes "People Finder" for identifying individuals in photos and videos, generative AI detection tools that can flag text potentially created by language models, and interactive chatbots that guide journalists through verification workflows with built-in warnings about potential red flags.
This builds on AP's longer-running AP Verify project, which has been in development since 2017 with Google funding. That system uses visual recognition and machine learning to verify user-generated videos by checking elements like street signs and geographic markers against known databases.
However, the verification tools face significant constraints outside the United States. As Rinehart explained at SXSW, "It's more beneficial in the US than in other places. GDPR is one of the reasons why it might not work so well in Germany." Privacy regulations limit the cross-referencing capabilities that make verification tools effective.
Open Source Tools for Local News
Separately, from 2021 to 2023, AP collaborated directly with local newsrooms through its Local News AI initiative to develop five open source projects, publicly available on GitHub for any newsroom to use. These tools address practical workflow problems in newsrooms.
The standout success is the Minutes Project, which automates one of local journalism's most tedious tasks. The tool takes recordings of public meetings, does automated transcription, identifies keywords and sends those keywords to the journalist. This solves a resource problem many local reporters face, as one told AP: "I cover four school districts, and two of those districts have the meetings at the same time on the same night."
The open source approach reflects AP's cooperative DNA. Rather than developing proprietary tools, AP chose to make these solutions freely available, allowing smaller newsrooms to benefit from AI development they couldn't afford independently.
For Troy Thibodeaux developing standards for the use of AI matter, but delevoping revenue models matters too. (photo courtesy of Thibodeaux)
Finding Ways to Monetize AI
AP has taken a proactive approach to generating revenue from AI rather than simply being consumed by it. In July 2023, AP announced a two-year licensing deal with OpenAI, allowing the AI company to access part of AP's text archive dating back to 1985.
The deal structure included important protections. According to the Wall Street Journal, AP included a first-mover safeguard that allowed it to modify terms if another publisher strikes a better deal with OpenAI. This reflects the uncertainty around content valuation in AI training. In exchange for licensing content, AP gained access to OpenAI's technology and product expertise, though Thibodeaux notes the deal was "a limited time" with "some continuing agreements".
Beyond OpenAI, AP has licensed archival material to other tech companies for training and display use cases. "Working with the big platforms, the big tech companies, we found some ways that it can actually add to the bottom line," Thibodeaux explains.
However, licensing deals don't solve the fundamental cost problem. As Rinehart explained at SXSW, most AI experiments are not bringing in money for news organisation. They only create revenue for the commercial AI companies.
This cost problem affects the entire news industry. Every API call to OpenAI, Google, or Anthropic represents money flowing from news organizations to tech companies. Rinehart's broader point: "I would like to see the journalism industry have a real stake in making money off of this." The challenge isn't just about licensing existing content to AI companies, but developing sustainable business models around AI-powered journalism tools.
Lessons to be Learned:
AP's approach towards AI offers several lessons for news organizations:
Start with standards, not technology. The most successful AI implementations begin with editorial guidelines, not technical capabilities. AP's early investment in standards discussions prevented later credibility problems.
Collaborate, don't duplicate. As Thibodeaux observes, "Every news organization is building its own headlines tool." AP's open-source Local News AI initiative demonstrates how collaboration can prevent redundant development.
Focus on augmentation, not replacement. All of AP's AI tools keep humans in the loop. The technology enhances journalist capabilities rather than replacing them.
Test internally first. Building tools for your own newsroom before offering them to customers ensures they solve real problems rather than theoretical ones
Trust Dividend: AP's conservative approach has paid off in sustained trust relationships. This trust-first approach may seem slower than the Silicon Valley way of moving fast and breaking things, but it's allowed AP to build sustainable AI capabilities while preserving their core asset: credibility with both newsrooms and news consumers.
Tools used for this newsletter:
Otter and Notta: transcriptions
Claude 4.0 Sonnet: Additional research and first draft