Every publisher experimented with AI in 2025. The results ranged from genuinely useful to embarrassingly bad. Now that we’ve got a year of real-world evidence, we can separate signal from noise.

What Worked: Headline Assistance

AI tools that suggested headline variations helped editors without replacing editorial judgment. Generate ten options, pick the best one, edit for voice. This made work faster without compromising quality.

The key was treating AI as assistant, not writer. Headlines generated by AI and published without human refinement were usually mediocre. Headlines refined by editors using AI suggestions often outperformed purely human-written alternatives.

Tools like Headline Studio integrated AI suggestions with actual performance data, giving editors real feedback on what works rather than just algorithmic guesses.

What Worked: Metadata Generation

Writing meta descriptions and alt text is tedious. AI automated it adequately. Not brilliantly, but well enough that editors could review and approve rather than starting from scratch.

This saved hours weekly for publications with high content volume. The quality wasn’t perfect, but perfect wasn’t necessary for metadata. Good enough was genuinely good enough.

What Worked: Content Tagging

AI categorization and tagging of content improved discoverability and organization. Publications with large archives used AI to retroactively tag years of content that had never been properly categorized.

The accuracy improved significantly from early attempts. While human review remained necessary, AI got you 80% there without manual effort.

What Didn’t Work: AI-Written Articles

Publishers who tried AI-generated content learned that readers notice and care. The articles were technically competent but soulless. They conveyed information without insight or voice.

Some publications tried AI first drafts that humans would heavily edit. In practice, editing bad AI prose took longer than writing from scratch. The result was usually worse than purely human-written content.

The exceptions were extremely formulaic content like earnings reports or sports recaps where readers want information, not analysis. Even there, the quality gap was noticeable.

What Didn’t Work: AI Personalization

Personalized content recommendations powered by AI promised to increase engagement. In practice, they rarely outperformed simple recency and popularity algorithms.

The implementation complexity exceeded the value. Publishers spent engineering resources building sophisticated recommendation systems that marginally improved metrics if they improved them at all.

The publications that succeeded with personalization were large enough to have meaningful data and technical teams to maintain the systems. For smaller publishers, it wasn’t worth the effort.

What Sort of Worked: Research Assistance

AI tools that helped journalists research topics by summarizing sources and identifying patterns provided genuine value with significant caveats.

The value came from handling information volume humans can’t practically process. Analyzing hundreds of documents to identify trends or contradictions.

The risk came from AI hallucinations and bias. Journalists who trusted AI research without verification published errors. Journalists who used AI as starting point for human verification got value without compromising accuracy.

What Sort of Worked: Transcription

AI transcription services like Otter.ai and Descript got good enough that you’d only need minor cleanup rather than complete revision.

Not perfect. Speaker identification failed sometimes. Technical terms got mangled. But the time savings were real compared to manual transcription.

The quality gap between leading services and free options widened. The market consolidated around a few good providers while poor alternatives died off.

The Cost Problem

AI tooling added up quickly. Individual tool subscriptions seemed reasonable until you tallied monthly costs across the entire editorial team. Annual AI spending for some publishers exceeded personnel costs for a junior staff position.

The tools that delivered value were worth it. But many publishers found themselves paying for AI features they’d tried once and never used again.

The Training Challenge

Most editorial teams needed training to use AI tools effectively. Without it, adoption remained low or usage patterns were inefficient.

Publications that invested in training saw return on tool investments. Publications that bought tools and expected staff to figure it out saw money wasted on unused subscriptions.

Ethical Considerations

Publishers grappled with disclosure questions. Should articles note AI assistance? If so, what level of assistance requires disclosure?

No consensus emerged. Some publications disclosed any AI use. Others only disclosed AI-written content. Many didn’t disclose at all.

Reader trust research suggested disclosure mattered. Readers who discovered undisclosed AI use felt deceived. Readers told upfront about AI assistance in research or editing generally didn’t care.

The Future Isn’t Here Yet

The truly transformative AI applications didn’t materialize in 2025. Natural language search that actually understands queries. Conversational interfaces to archives. Intelligent story assignment based on reporting trends.

These remain theoretically possible but practically unavailable. The AI that exists today is good at narrow tasks, not strategic thinking.

What Publishers Should Do

Approach AI as tool, not solution. Identify specific tedious tasks that could benefit from automation. Test tools thoroughly before committing to annual contracts. Train staff properly. Measure actual impact on workflow and output quality.

Ignore hype about AI replacing human journalists or revolutionizing publishing. It won’t, at least not with current technology. But it can make specific tasks easier if implemented thoughtfully.

The publications succeeding with AI in 2025 were pragmatic about capabilities and limitations. They used AI where it added value without pretending it could do things it can’t. That approach will continue working in 2026.