Last month, a regional Australian news site published 47 articles in a single day. The site had one listed journalist. The articles covered local council meetings, property market updates, school sports results, and community events across three different towns. They were competently written, factually accurate (mostly), and almost certainly generated by AI.

Nobody noticed for two weeks. When someone finally pointed it out on social media, the reaction was a collective shrug. And that reaction — the shrug — might be the most concerning part of this whole story.

We’ve arrived at a point where AI-generated news content is sophisticated enough to pass casual inspection, detection tools are unreliable enough to be practically useless, and the public is increasingly indifferent to whether a human wrote the article they’re reading. This is a problem, and we’re not talking about it honestly enough.

The Detection Tools Don’t Work

Let me be blunt about this: AI content detection in 2026 does not reliably distinguish between human-written and AI-generated text. Not at the level that matters for journalism.

I ran a test. I took 20 human-written articles from established Australian news outlets and 20 articles generated by current AI models (prompted with the same topics, tone, and style guidelines). I put all 40 through the four most popular AI detection tools.

The results were predictable to anyone who’s been following this space:

• False positive rate: 15-25% of human-written articles were flagged as AI-generated.
• False negative rate: 30-40% of AI-generated articles were classified as human-written.
• Consistency: The same article would get different classifications depending on which tool you used.

A tool that gets it wrong a third of the time isn’t a detection tool. It’s a coin flip with extra steps.

Researchers at Stanford University’s Human-Centered AI Institute have documented this extensively. Their analysis of commercial detection tools found accuracy rates that “provide a false sense of security” and warned against using them as definitive arbiters of content origin.

The fundamental problem is that AI language models generate text by predicting statistically likely word sequences based on patterns learned from — wait for it — human-written text. The better the model gets at mimicking human writing, the harder detection becomes. Detection and generation are locked in an arms race, and generation is winning.

Why It Matters for Journalism

“But if the content is accurate, does it matter who or what wrote it?”

I hear this question constantly, and I think it’s the wrong question. Here’s why.

Accountability. When a human journalist writes an article, there’s a person behind it. They have a byline, a reputation, an editor who reviewed the work, and legal liability for what they publish. When something goes wrong — a factual error, a biased framing, an omission that changes the meaning — there’s a chain of responsibility. AI-generated content has no such chain. Who’s accountable when an AI article misrepresents a council vote or gets a crime report wrong?

Judgment. Journalism isn’t just assembling facts. It’s deciding which facts matter, what context is needed, and what the story means for the community. An AI can tell you what happened at a council meeting. It can’t tell you that the motion about the new development suspiciously benefits the councillor’s brother-in-law’s company. That requires institutional knowledge, relationships, and the kind of skepticism that comes from covering a beat for years.

Trust. The journalism industry’s most valuable asset is credibility. If readers can’t be sure whether they’re reading the work of a professional journalist or an algorithm, that erodes trust further in an industry already struggling with it. The Reuters Institute Digital News Report has tracked declining trust in news across multiple years — AI-generated content that pretends to be human journalism accelerates that decline.

What’s Actually Happening in Newsrooms

The conversation inside media organisations is more nuanced than the public discourse suggests. There isn’t a simple pro-AI or anti-AI divide. Instead, there are several distinct approaches emerging.

The “AI as tool” camp sees AI as useful for specific, bounded tasks: transcribing interviews, summarising documents for reporters to review, generating first drafts of routine content (earnings reports, sports scores, weather updates) that journalists then verify and edit. The Associated Press has been doing this with earnings reports since 2014. The key principle is that a human always reviews, edits, and takes responsibility for the final published content.

The “AI transparency” camp argues that any AI involvement in content creation should be disclosed to readers. The Guardian’s editorial policy and similar frameworks at other outlets now explicitly address AI use. The idea is that readers deserve to know how their news was produced, just as they deserve to know about conflicts of interest or anonymous sources.

The “keep AI out of editorial” camp maintains that the core editorial functions — reporting, writing, and analysis — should remain exclusively human. AI can help with logistics (scheduling, distribution, audience analytics), but the journalism itself should be produced by journalists. This position is strongest among legacy broadsheets and investigative outlets.

And then there’s the uncomfortable reality that many smaller outlets, particularly in local news, are quietly using AI to fill gaps created by declining revenue and shrinking newsrooms. They don’t have the luxury of principled positions about AI in journalism when they can’t afford to hire enough reporters to cover their communities. The work being done by specialists in AI implementation suggests that this trend will accelerate as AI tools become more accessible and newsroom budgets continue to shrink.

The Transparency Problem

Here’s where it gets uncomfortable. The outlets that are most transparent about their AI use tend to be the ones using it most responsibly. The ones using AI irresponsibly — publishing AI-generated content without human oversight or fact-checking — aren’t going to voluntarily label their content.

Mandating AI disclosure is one approach, and the EU’s AI Act includes provisions requiring disclosure of AI-generated content. Australia is considering similar frameworks. But enforcement is the challenge. How do you prove an article was AI-generated when the detection tools don’t work reliably?

Watermarking — embedding invisible markers in AI-generated text — is technically possible, but it only works if all AI providers implement it (they won’t, especially open-source models), and it can be defeated by paraphrasing the output.

Provenance tracking systems like the Coalition for Content Provenance and Authenticity (C2PA) offer a more promising path by creating a verifiable chain of custody for content. But adoption is still in early stages, and it requires buy-in from platforms, publishers, and technology providers.

Where This Is Heading

I wish I had an optimistic conclusion, but I don’t think honesty allows for one.

AI-generated news content is going to proliferate. The tools to create it are cheap and accessible. The tools to detect it are inadequate. The regulatory frameworks are lagging behind the technology. And the economic pressures driving adoption — especially in local and regional media — aren’t going away.

The most likely outcome is a two-tier system: well-funded outlets with strong editorial standards will use AI carefully and transparently as a supplement to human journalism. Poorly funded outlets and content farms will use AI as a replacement for human journalism, producing content that’s good enough to attract clicks but lacks the depth, accountability, and judgment that real reporting provides.

Readers will need to become more discerning about where their news comes from. Not just whether an article “sounds right,” but whether it comes from an outlet with editorial standards, named journalists, and a track record of accountability.

That’s a lot to ask of a public that’s already overwhelmed with information. But I’m not sure there’s an alternative.

The shrug is the problem. We should stop shrugging.