AI can’t be wrong the way a reporter can...and that’s the problem
Every journalist eventually receives a call like this.
It usually comes late in the afternoon. The newsroom has already begun to relax into the illusion that the day’s work is finished. The story has aired. The copy has been filed. The broadcast has moved on.
Then the phone rings.
Sometimes it’s the person quoted in the piece. Occasionally, it’s a colleague who remembers the exchange differently. Sometimes it’s an editor, using the careful tone editors adopt after confirming something is not quite right.
A sentence in the story doesn’t match what was actually said.
And once that realization arrives, a peculiar kind of reconstruction begins.
Someone checks the notes. Someone replays the recording. Someone tries to remember when the conversation drifted into the line that appeared in print. The newsroom becomes, briefly, a quiet forensic workspace.
Human error in journalism leaves traces.
There is usually a notebook somewhere. A tape. An email. A memory held by someone who was in the room.
You can follow those traces back to the moment where the mistake entered the story.
That trail—imperfect, human, occasionally messy—is one of the quiet mechanisms that make journalism accountable to the public it serves.
Yet the way artificial intelligence fails contrasts starkly with human error.
In August 2024, a small Wyoming newspaper, The Cody Enterprise, published several stories containing quotations that had never been spoken. Among them was a quote attributed to the governor of Wyoming that the governor himself had never given.
The sentences were fluent. Confident. Entirely plausible.
They were also fabricated.
The reporter responsible had used an AI writing tool to help generate portions of the stories and had not verified the quotations it produced. When the problem surfaced, the newspaper issued corrections. The editor later explained the situation with a kind of weary clarity: he had failed to catch the AI-generated copy.
The episode might have been dismissed as a familiar newsroom lapse. Journalism, after all, has never been immune to mistakes.
Yet this incident stood apart in a meaningful way.
When the editors attempted to reconstruct how the quotation had entered the story, they discovered that there was nothing to reconstruct. There was no interview to revisit, no conversation that might have been misunderstood, no recording that could be replayed to locate the moment where the words had first appeared.
The sentence had not emerged from a human exchange.
It had simply materialized.
The correction was easy enough to publish. The production history of the mistake—the chain of decisions that normally connects a piece of journalism to the world it describes—was missing.
The error had no trail.
For most of my career, I did not think much about the idea of a production trail.
I came to the phrase indirectly. Years spent working in podcasting, video and television production, and documentary showed me places where the machinery of storytelling is unusually visible. Anyone who has worked in a control room during a live broadcast knows a finished program is less like a finished object and more like the residue of hundreds of tiny decisions.
A camera operator holds a shot for two seconds longer than planned. A director calls for a cut a moment earlier than usual. An audio engineer quietly lowers one microphone and opens another. None of these is monumental alone. But each leaves a small imprint on the final broadcast.
More importantly, each decision has a human memory attached to it.
Ask the camera operator later why she stayed on the shot, and she will tell you. Ask the director why he switched angles, and he will remember, or at least try to. The record is partial, subjective, and occasionally self-serving.
But it exists.
That existence is what makes the system auditable.
Here, artificial intelligence completely reshapes the structure.
The central issue is not that AI systems produce errors—human reporters have always done that. The deeper problem is that when AI systems fail, they often do so in ways that cannot be traced back to a visible decision.
A large language model does not know why it generated the sentence it did. It cannot identify a source. It cannot reconstruct the reasoning that produced the output.
The model has probabilities.
Which is not the same thing.
When the resulting sentence turns out to be wrong, the error is not attached to a person or a process that can be examined.
It is attached to nothing.
This distinction affects more than just the newsroom process.
For generations, citizens navigating the news have relied—often unconsciously—on a set of informal heuristics for evaluating credibility. A story comes from a particular outlet. That outlet has a reputation, a history, and an ownership structure. The reporter has a track record, a beat, a pattern of getting certain things right and others wrong.
None of this information is perfect.
But it creates a visible production layer around journalism.
Readers can at least dimly see where the information came from.
This is where AI-generated journalism changes things.
AI-generated journalism dissolves that layer. It may not produce more misinformation, but it produces information whose origins are no longer clear. The article appears with the familiar tone and cadence of reporting. Yet the human chain of decisions that produced that tone is no longer present in the same way.
With AI, the trail—that chain connecting work to decisions—evaporates completely.
And when the trail disappears, readers lose something else as well: the ability to understand how the information might fail.
This is not simply a problem of accuracy.
At its core, it is a problem of democratic legibility.
There are already attempts to repair this gap.
Researchers are designing provenance frameworks to attach verifiable metadata to AI-generated content. Some news organizations have begun experimenting with disclosure standards. Governments are exploring regulatory regimes that might require synthetic content to be labelled.
These measures may help.
Yet these do not reconstruct the production train.
A label can tell you AI was involved in creating the content. It cannot tell you how the system generated specific words, which sources shaped them, or which assumptions guided the results.
In other words, the label acknowledges the machine.
It does not illuminate its decisions.
Without illumination, journalism’s corrective machinery weakens. Errors can still be corrected. But the path to correction becomes longer and harder to trace.
In an information environment where falsehood already travels faster than correction, that change matters.
More than we may yet understand.
A colleague asked a question about this during a seminar conversation last year.
If AI-generated journalism fails in ways that are invisible to readers, she asked, who exactly is accountable for the consequences?
It is the sort of question that lingers.
I do not have a complete answer.
But I suspect the answer will lead us back to something journalists have always known, even if we rarely described it this way.
Journalism is not only a collection of stories.
It is a trail of decisions.
And when the trail disappears, accountability becomes immeasurably harder.