In early 2025, the Associated Press posted a job listing that drew almost no attention. AI Broadcast Scripting Specialist. The description laid out a quiet overhaul. The hire would “develop and refine AI-driven workflows for generating rough broadcast scripts from field notes, audio transcripts, and raw wire copy.” The output wasn’t intended for publication. It was meant to be the first draft a human editor would later polish. That distinction matters. It means the first interpretive act—the moment raw observation becomes narrative—is being handed off to a machine.
Reuters’ London bureau has been testing something similar. Internal presentations obtained by the Financial Times in late 2025 described a “screenplay tool” that produces voiceover drafts for video packages. One slide showed the pipeline: field footage and correspondent notes enter, a large language model generates a three-column script—visuals, narration, timing—and a producer reviews the result. The word “screenplay” was used deliberately, as if to borrow the prestige of a creative industry. But a broadcast news script is not a screenplay. It’s a compressed argument about what matters, built on choices a reporter made on the ground. When a machine drafts that argument, the choices get smoothed into something else.
This is not a story about AI replacing journalists. It’s about what happens before replacement becomes the question. The wire services—AP, Reuters, AFP—are the circulatory system of global news. Their copy appears, often verbatim, in thousands of newspapers, websites, and broadcasts daily. If their editorial judgment tilts, even slightly, toward templated formats machines can parse and reproduce, the downstream effect is not a single outlet’s bias but a structural shift across the entire information ecosystem. The question isn’t whether the machines are malevolent. It’s what they cannot see, and what we stop seeing as a result.
The Bureau Decision: Why Templating Starts Before the Tool
To grasp why wire services are adopting AI script generators, you have to understand bureau economics. A single correspondent in Nairobi or Jakarta might file for text, radio, and video in one day. The broadcast side is especially strained: video scripts demand tight timing, audio packages need conversational narration, and deadlines narrow the window for revision. For years, the solution was human shorthand—experienced reporters who could dictate a rough voiceover straight from their notes. But those reporters are expensive, and the appetite for video content keeps growing. The AI script generator enters as a cost-saving intermediary, not a replacement. It takes the raw material and produces something that looks like a script: formatted, timed, grammatically clean.
But “looks like a script” is the trap. A script is not merely transcription plus formatting. It’s a sequence of editorial choices: which soundbite leads, which detail adds texture, which fact demands context, which silence the image should fill. When a machine drafts from field notes, it treats every observation as equally weighty unless a human has tagged them otherwise. The result is a script that is factually accurate and structurally hollow. It will say “protesters gathered in the square” but might omit that the square is usually a market, or that the protest route deliberately bypassed a government building, or that the correspondent smelled tear gas before seeing any crowd. Those sensory and contextual details—the very things that distinguish a report from a press release—are not easily machine-readable. If they never make it into the first draft, they are less likely to survive the edit.
Here’s the mechanism worth watching: the AI script generator does not just save time. It standardizes the initial interpretation. Standardization isn’t inherently bad—wire copy has always followed style guides. But the type of standardization matters. A human editor applying AP style makes linguistic choices within a framework of news judgment. A machine applying a template conforms to a statistical model of what scripts look like. The difference is the difference between drawing within lines and tracing a stencil. The stencil produces something recognizable, but it cannot capture the weight of what the lines originally enclosed.
What Gets Lost in Translation: From Human Observation to Machine-Readable Structure
Consider a specific case. In March 2026, an AP video script about flooding in southern Brazil began its voiceover: “Floodwaters continued to rise Tuesday in Porto Alegre, where residents navigated submerged streets in small boats.” The sentence is correct. It is also what a machine would write. The field notes, obtained later through a source at the bureau, included this: “water so high you can’t see the tops of parking meters, children pointing at dead cattle floating past.” Neither detail survived into the AI-generated draft, and the human editor, working against a broadcast clock, didn’t restore them. The draft set the frame, and the frame held.
This is not an argument that machine-generated drafts are always worse than human ones. For earnings reports, sports recaps, routine weather events, the loss of texture may be negligible. The danger is that the line between “routine” and “significant” is itself an editorial judgment. Wire service editors decide which stories get the full human treatment and which get the template. But when the template is the default, the threshold for what merits a human first draft creeps upward. Over time, stories that would have received a correspondent’s careful attention get a machine’s educated guess instead.
What gets lost can be categorized. First, sensory specificity: the smell, the sound, the visual detail that doesn’t fit a data field. Second, causal implication: the observed connection between two facts a reporter notices but hasn’t yet confirmed, so they phrase it as possibility rather than assertion. A machine will not write “the delay in aid delivery appeared linked to a dispute between municipal and federal officials” unless that link is already in the input. A human reporter will write it because they saw a federal truck turned away at a checkpoint and made a note. Third, tonal register: the subtle shift in language that signals to an audience that something is tragic, absurd, or urgent. Machine-generated scripts default to a flat, declarative tone. They inform without implying. Over time, audiences attuned to that tone may mistake neutrality for truth, when in fact neutrality of tone can be a form of evasion.
The Authors Guild, in its AI best practices for authors, warns that “AI tools are increasingly used to generate drafts, raising questions about human oversight and editorial control” and that “the transition from human observation to machine-readable structure can lead to loss of nuance and context” (AI Best Practices for Authors). The Guild’s guidance targets book authors, but the principle applies with greater urgency to news, where the draft is not a private manuscript but the first link in a chain of public understanding.
The Telltale Patterns: How to Spot Wire Copy That Passed Through a Machine
Can readers detect when wire services use AI script generators? The answer is a qualified yes. No tool leaves a watermark, but certain syntactical and structural patterns recur often enough to form a heuristic. These patterns aren’t proof, but they’re signals—invitations to read more closely.
First, the three-sentence lead that never varies rhythm. A human-written broadcast lead often uses a short punchy sentence, then a longer explanatory one, then a concrete detail. An AI-generated lead tends to produce three sentences of roughly equal length and cadence: “Flooding continued in southern Brazil on Tuesday. Thousands of residents were displaced from their homes. Authorities warned that water levels could rise further.” It’s grammatically flawless and rhythmically dead. Human writers break rhythm for effect. Machines don’t know what effect is.
Second, the missing attribution of observation. A human reporter writes “a Reuters witness saw” or “according to an AP journalist at the scene.” An AI draft often omits the observer, presenting sensory information as disembodied fact. “Smoke was visible above the city”—visible to whom? The absence of the witnessing “I” or “our correspondent” is not just stylistic; it erases the epistemological grounding of the report. The audience no longer knows how the news organization knows what it claims to know.
Third, the unnaturally even distribution of quotes. In a human-edited broadcast script, quotes are chosen for their emotional or informational punch, and they appear irregularly. An AI-generated script often places quotes at predictable intervals—every third or fourth paragraph—as if following a template for “inject human voice here.” The quotes may be perfectly accurate, but their placement feels mechanical, not motivated by the story’s internal logic.
Fourth, the absent transition. Human scripts use verbal handoffs: “But the situation is different further north,” or “That official account, however, is disputed by residents.” These transitions carry argument. They tell the audience the story is about to pivot, and why. AI-generated scripts tend to stack paragraphs without connective logic. One fact follows another, but the relationship between them stays implicit, which in journalism often means unexamined.
The professional screenplay format, as detailed by resources like StudioBinder, relies on strict conventions: scene headings, action lines, character cues, parentheticals. These conventions communicate not just what happens but how it should be seen and heard (How to Write a Movie Script). When an AI script generator is trained on such formats, it learns to reproduce the structure without understanding the directorial intent behind it. In a film script, “CLOSE ON: the letter, unopened” is a creative choice. In a news script, the equivalent visual cue might be “SHOT OF: the empty podium”—a choice that editorializes by showing absence. A human camera operator or correspondent makes that choice deliberately. A machine, drafting a shot list from field notes, may select it because “empty podium” appears near “official statement delayed” in the data. The result looks professional but lacks the intentionality that distinguishes editorial judgment from pattern matching.
These patterns are not inevitable. Skilled human editors can and do revise AI drafts thoroughly. But the economics of wire services—speed, volume, cost—work against thorough revision. The draft that arrives pre-formatted saves time precisely because it demands less rethinking. The danger is that “less rethinking” becomes the operational definition of editorial efficiency.
The Uncomfortable Question: What Does “First Draft of History” Mean Now?
The phrase “first draft of history” has been attached to journalism for decades, usually with pride. It implies reporters are present at the moment of occurrence, recording what they see before memory fades or officials spin. But the phrase also contains a warning: a first draft is provisional, subject to correction. The question the wire services’ AI adoption raises is whether a machine-generated draft is provisional in the same way. A human first draft carries the imprint of a specific consciousness: this reporter noticed this detail, asked this question, framed the event this way. A machine draft carries the imprint of a training corpus: a statistical average of how similar events have been reported before.
This is not an argument that human reporters are unbiased or that machine drafts are inherently inferior. It is an argument about traceability. When a human reporter files, an editor can ask: “Why did you lead with the mayor’s quote instead of the victim’s?” The reporter can answer, explain their reasoning, and the reasoning can be challenged. When a machine drafts, the reasoning is inaccessible—not because it’s secret, but because it’s distributed across millions of parameters. The editor can still change the lead, but the initial framing has already been set by a process that cannot be interrogated. The editorial conversation, the core of newsroom sociology, becomes a one-sided correction of an algorithmic output.
This matters beyond the wire services themselves. Local newsrooms that have lost their own correspondents increasingly rely on wire copy to fill broadcast minutes. If that wire copy carries the subtle uniformity of AI drafting, the diversity of perspective that local audiences once received—however imperfect—narrows further. The result is not a single falsehood but a homogenization of news language, a smoothing of the rough edges that signal independent observation.
There is a role for AI in the newsroom that is genuinely useful: transcription, data sorting, pattern detection in large document sets. These tasks augment human capacity without substituting for judgment. The script generator occupies a different category. It inserts itself into the interpretive chain at the exact point where observation becomes story. That insertion should be visible to audiences, not hidden behind a byline or a brand logo. Some wire services have begun internal discussions about labeling AI-assisted content, but no standard has emerged. The default remains: the audience assumes a human wrote what they read or hear, because for a century, that assumption held.
For readers who want to reclaim their own editorial intelligence, the skill is not to reject all wire copy but to read it with an awareness of its production pipeline. When you encounter a broadcast script that feels rhythmically flat, that lacks sensory detail, that quotes officials but never witnesses, that pivots without signaling why—ask yourself: was this drafted by someone who was there, or by something that was trained on what “there” usually looks like? The answer may not change the facts of the story, but it changes your relationship to them. You move from passive recipient to active assessor of the news’s epistemological quality. That shift, practiced daily, is the difference between being informed and being processed.
In some newsrooms, the tool facilitating this shift is literally called a script generator. The name is honest, if unsettling. It generates scripts. It does not generate understanding. The distinction is not semantic; it is the central challenge facing journalism in the next decade. The machine in the middle is here. The question is whether we will learn to see its fingerprints, or simply accept its drafts as our own.
Ramona Ghali is a media analyst and former newsroom data editor. She writes about how information flows, who controls it, and why the stories that matter most are often the hardest to find.