Artificial Intelligence
How AI writing tools fail when the source says nothing
When an AI writing tool is told to generate an article from a source that says nothing, the result reveals the limits of automation without human oversight. This column examines the paradox of content generation that demands output from empty input.
Every week, someone prompts a language model to write a 1,500-word article from source material that is either blank, a single fragment, or a placeholder like "Générer un article (10/15)." The intention is efficiency. The effect is a quiet crisis in the credibility of published content.
The prompt that says nothing
"Générer un article (10/15)" is a French step-counter that translates to "Generate an article (10/15)." It is one line inside a larger assignments file, presumably task ten of fifteen. As a source, it offers zero facts, zero quotes, zero context. Yet the instruction remains: produce a full, publication-ready piece.
This is not an edge case. In distributed writing factories, interns and freelancers often receive skeleton prompts. The expectation is that the writer will "decorate" the assigned slug with plausible text. When an AI system takes that instruction literally, the output becomes noise dressed up as journalism.
Why the machine cannot save you here
Large language models are statistical next-token predictors. Give them a source that says "Generate an article" and they will produce a generic essay on the craft of writing, or a reflection on article generation, or a listicle about productivity. None of it responds to any actual news event or dataset. The model hallucinates an entire referent.
The problem is not hallucination per se. The problem is that the model's output reads confidently. It will state "according to recent studies" and cite no study. It will quote "industry experts" with no name. It looks like an article. It feels like an article. But it is a rhetorical shell.
SEO bait dressed as expertise
This practice has a name: content spam. Google's 2024 updates explicitly target "low-value, AI-generated content that does not demonstrate expertise, authoritativeness, or trustworthiness." A column that sources itself from a blank prompt violates every E-E-A-T guideline. Yet it gets published because the checkbox says "article written" and the word count clears the requirement.
The irony is that AI writing tools, when used correctly, can accelerate real reporting. They transcribe interviews, summarize earnings calls, and generate topic clusters for researchers. But they require a genuine source, a transcript, a document, a live event, to anchor the output to reality.
The real cost
Every fake article damages the ecosystem. It trains readers to distrust the medium. It dilutes the signal for search engines. And it demoralizes journalists who still write from actual notes. The moment we accept that "Générer un article (10/15)" is a sufficient brief, we wave goodbye to the last shred of editorial integrity.
"The best AI editors don't replace the reporter; they force the reporter to bring better sources."
If your source has nothing to say, the only responsible output is: "This prompt cannot produce a factual article because no source was provided." Anything else is a lie dressed in HTML.
What should happen instead
- Validate the source first. If the input is empty, below 50 characters, or is a step counter, reject the task automatically.
- Return a diagnostic. The AI issues a warning: "No actionable source detected. Please supply at least one URL, transcript, or raw text of 100+ words."
- Escalate to a human. Flag the task as needing editorial triage. A blank prompt is not a writer's problem. It is an editorial planning failure.
Until that workflow becomes standard, writers, both human and machine, will keep delivering empty calories. And readers will keep wondering why so much online content tastes like nothing.