As 2025 winds down, AI in journalism feels less like a side project and more like the infrastructure underneath reporting, distribution and even business models. Newsrooms are building in-house systems, funders are backing AI experiments in membership and fundraising, and watchdogs are scrambling to keep up with new risks, from AI browsers tunneling through paywalls to scammers gaming news outlets.

This fall’s AI + journalism wrap-up shows a field trying to move beyond hype into durable workflows — while also confronting the mess AI is leaving in its wake.

“Hyper-caffeinated interns”: AI moves deeper into reporting workflows

One of the clearest themes this fall is AI taking on grunt work so humans can stay focused on accountability reporting.

At The Connecticut Mirror, AI is essentially a “hyper-caffeinated intern” helping a small nonprofit cover 169 towns. The newsroom uses AI tools to parse long documents, transcribe meetings and surface patterns across beats so reporters can follow leads they wouldn’t otherwise have time to chase. The goal is not to replace journalists but to widen the funnel of potential stories in a state that still has big gaps in local coverage.

Archive work is getting the AI treatment, too. The Lenfest Institute’s profile of “Dewey, the Archivist” shows how The Philadelphia Inquirer built an AI-driven research tool on top of its 190-year archive, turning what used to be a dusty, siloed resource into something reporters can query in seconds for context, timelines and past coverage. Lenfest frames Dewey as a replicable model for local outlets that are sitting on deep archives but struggle to make them usable in daily reporting.

And at the global level, The Economist has created an AI Lab — a startup-style unit inside the magazine tasked with thinking a few years ahead. In a Newsroom Robots interview, Ludwig Siegele describes experiments ranging from AI-powered translation and research pipelines to TikTok dubbing and a SCOTUS bot as they partner with vendors like Google’s NotebookLM, all while stressing that the hard part isn’t building tools but getting staff to adopt them and protecting the brand.

Large newsrooms like The New York Times and The Washington Post are converging on similar strategies: AI is mostly used for research and investigations, with central teams helping reporters on one project and then turning those lessons into reusable workflows. According to a Nieman Lab summary of a Digiday session, the Post is drawing firm lines around when staff must use models hosted internally versus enterprise ChatGPT or Gemini, especially for anything “business critical” or sensitive.

Funding the work: AI for fundraising and membership

It is not just newsrooms experimenting. The American Journalism Project is urging grantees to treat AI as a strategic helper in fundraising. In a Medium piece, Maggie Cogar lays out use cases for AI in prospect research, gift officer prep and tailoring outreach, while warning that any AI-generated copy still needs human review for tone, accuracy and equity.

The News Revenue Hub’s 2025 AI Campaigns Cohort offers a more hands-on look. Nonprofit newsrooms in the cohort are testing AI for tasks like smarter email segmentation, re-framing appeals for different audience segments and drafting first passes of membership messaging, with humans refining and fact-checking. The project is framed explicitly as an experiment: publishers are encouraged to set clear guardrails, test against control groups and share what works back to the field.

Taken together, these efforts show AI moving beyond the editorial stack into the revenue stack, with a lot of emphasis on keeping human judgment at the center.

Audiences, languages, and access: AI’s uneven reach

AI is not landing evenly around the world or across languages.

A feature by Gretel Kahn for the Reuters Institute spotlights journalists and technologists in India, Belarus, Nigeria, Mali, Paraguay and the Philippines who are trying to keep their countries’ languages alive in an era when most AI tools are optimized for English or other major tongues. They are building translation systems, voice tools, and language models tuned to local contexts so communities are not forced to consume news in someone else’s language.

The JournalismAI Innovation Challenge, backed by the Google News Initiative, continues to fund small, cross-border teams building AI prototypes for reporting, product and audience work, with a heavy emphasis on responsible design and shared learning.

Rules of the game: Platforms, policy, and media freedom

Even as newsrooms experiment, regulators and press-freedom groups are trying to even the playing field.

The Forum on Information & Democracy’s OSCE policy manual on “Safeguarding media freedom in the age of Big Tech platforms and AI” argues that today’s digital ecosystem is so dominated by large platforms and AI companies that it actively undermines media freedom. 

The manual calls for democratic governments to take structural action in three areas: making public-interest journalism more visible online, shoring up viable funding models and protecting journalists’ safety in AI-driven information spaces.

On the media-literacy front, the PSAI campaign from Columbia Journalism Review uses viral AI-generated images to teach audiences how to spot synthetic visuals. The project leans into the idea of “fighting AI with AI,” walking people through telltale signs of fake images and positioning this kind of literacy as a core part of defending trust in journalism.

Another fault line: AI browsers. A Columbia Journalism Review investigation into tools like OpenAI’s Atlas finds that AI-driven browsers can sometimes sneak around traditional paywalls and content blockers by pulling from syndicated copies, social media posts and other derivative sources — raising new questions about licensing, consent and how much publishers can really control use of their work.

Zooming out, a WAN-IFRA report framed around the line “We don’t use AI anymore — we live in it” argues that publishers now operate inside AI “environments,” not just on the open web. If assistants and agents become the main interface for information, the report warns, outlets will have to optimize not only for human readers but for AI systems that choose which content to surface, based on context, mood and user history.

Slop, scammers, and newsroom crises: when AI goes wrong

Alongside the “good” use cases, this fall brought a wave of stories about AI’s darker edges.

Nieman Lab’s “AI-generated news sites spout viral slop from forgotten URLs” digs into networks of low-quality sites that mine old domains and churn out AI-written articles engineered for virality, often surfacing in Google Discover and social feeds without clear provenance. The picture is of an information ecosystem where synthetic text can cheaply fill any abandoned corner of the web.

A separate study from University of Maryland researchers, covered by Press Gazette and others, estimates that more than 9 percent of U.S. newspaper articles published this summer contained at least some AI-generated text — with much higher usage at smaller local outlets than big national brands, and with disclosures still rare.

The accountability stakes are getting real inside newsrooms, too. In Florida, reporters at a nonprofit news outlet have asked their board to investigate their editor’s use of AI, according to Nieman Lab, raising concerns about undisclosed automation and the integrity of their coverage.

The stakes are also playing out in labor and governance disputes. In December, an arbitrator ruled that Politico management violated contractual safeguards around AI adoption after rolling out AI tools without proper notice to the newsroom union, according to Neiman Lab. The decision found that the launch bypassed required bargaining and failed to meet agreed-upon standards for editorial oversight, reinforcing that AI systems which materially affect newsroom work cannot be deployed unilaterally.

The Local’s deep-dive on a suspected scammer operating under the name “Victoria Goldiee” shows how easy-to-use AI tools can fuel more old-fashioned fraud. The story traces how essays with shifting biographical details and suspicious patterns traveled across outlets, in parallel with other scandals where AI-generated work slipped into places like Wired, Business Insider and The Washington Post under shaky bylines.

And beyond the U.S., 404 Media reports that Russian state TV has launched an AI-generated news satire show, a sign that generative media is quickly becoming another instrument in state-aligned information wars as well as entertainment.

What journalists can try next

If you want to end your roundup with some practical takeaways, the ONA25 recap on Storybench is a handy bridge. Our own, Lisa Thalhamer highlights Yumi Wilson’s prompting framework — clarity, context, role and intent — as a simple checklist for making generative tools more useful and less chaotic in daily reporting, from brainstorming story angles to repurposing scripts across formats.

You could pair that with three “do this now” ideas drawn from the fall coverage:

• Turn AI loose on the boring stuff, then verify: Use meeting-monitoring tools, transcription and document-parsing the way CT Mirror and Hearst do — as tip lines and research assistants, not replacement reporters.
  
• Audit your labels and policies: With AI use creeping into as much as one in ten stories at some outlets, having clear internal rules and public disclosures matters more than ever.
  
• Think like an AI environment, not just a website: WAN-IFRA’s framing is a good gut-check — if assistants and agents are increasingly how people encounter news, what does it mean for your work to be “visible” and “trusted” inside those systems?
  

Taken together, this fall’s stories suggest a field that is past the novelty phase. AI is now part of the plumbing of journalism — which means experiments are getting more interesting, the risks are getting higher and the stakes for trust, equity and independence are only going up.