Avoiding the tells of AI writing

On 19 August 2026, a direct HTTP pull of storyshort.ai without running JavaScript returned 11 words of body text on every page tested, while the same method returned 1274 words from viewmade.com on 22 August 2026. Both sites publish short videos built by software. One shows its substance to a machine reading raw HTML; the other does not. This article is not about that gap. It is about a related one: the tells that mark text as machine written, why they appear, and how to remove them when you write scripts for short video or anything else.

The short answer

AI writing gets detected through pattern, not content. Detectors, editors and readers all respond to the same signals: uniform sentence length, hedged openings, generic transitions, adjectives doing work that numbers should do, and claims with no source attached. You avoid the tells by breaking each one deliberately. Open with a measured fact instead of an introduction. Vary sentence length on purpose. Replace every adjective you can with a number you can date. Name where each claim came from. None of this requires a tool. It requires deciding, before you write, which sentences carry evidence and which carry nothing.

Why detectors flag text at all

Detection, whether by software or by a human editor skimming a script, works on regularity. A language model trained to predict likely next words produces statistically likely sentences. Likely sentences share properties: similar rhythm, similar paragraph shapes, a preference for balanced constructions such as "not only X but also Y", and a habit of summarising before proving.

The result reads fine and proves nothing. A reader cannot point to a false statement, because there usually isn't one. They can point to the absence of anything checkable. When we build scripts at ViewMade, the first pass removes exactly this: any sentence that would survive deletion without the argument getting weaker is a candidate for deletion, because a sentence that carries no evidence is pure pattern.

This matters more in short-form than in long-form. A Short runs on a single spine of facts. If half the narration is connective tissue, the viewer leaves before the first verifiable claim arrives. In long-form writing, padding costs patience. In vertical video, it costs retention, and we have not measured that relationship ourselves, so we design against the risk rather than quote a percentage.

The five tells, with fixes

Each tell below has a mechanical fix. Apply them as edits, not as intentions.

1. The introductory opening. "In today's fast-paced world" and its relatives exist to warm up a topic the reader already knows. Fix: open with a dated, sourced fact. If you have none for your topic, you do not have a script yet; you have a topic.

2. Uniform rhythm. Machine drafts tend toward mid-length declaratives, one after another. Fix: after drafting, rewrite one sentence in three as a short factual statement under ten words, and one as a longer causal explanation. Deliberate variation breaks the statistical signature.

3. Adjectives standing in for evidence. "Remarkable", "stunning", "incredible" describe intensity without describing anything. Fix: replace each with the smallest number that supports the claim, plus the date you measured it. If no number exists in your notes, cut the sentence. A claim you cannot date is a claim you cannot defend.

4. Hedged conclusions. "Ultimately", "at the end of the day", "it remains to be seen". These signal that the writer had nothing to conclude. Fix: state the finding and its limit in the same breath. "The archive footage places the ship 40 miles off course; the logbook itself was never recovered." A bounded conclusion reads as confidence, because it is one.

5. Unattributed specifics. Any number, quotation or event detail with no named source. Fix: attach a source to every specific before finalising. This is the tell most often left in place, because removing it means doing research, and it is also the one human editors punish hardest.

What sourcing looks like in practice

Sourcing is easier to show than to describe. Below is how a set of competing tools describes its own visual pipeline, read directly from each vendor's own pages in August 2026. The last column is the one that matters for credibility: whether the product names the origin of each clip in the output.

Tool	Stated purpose	Visual source	Names clip origins
ViewMade	Vertical documentaries about real events	Real archive footage	Yes
StoryShort	Topic-to-video with web-sourced imagery	Web-sourced imagery	Partial
InVideo	Broad general-purpose generation from a prompt	Generation and stock	No
Revid	High-volume short-form output	Generated	No
Crayo	Clipping an existing upload	User upload plus gameplay footage	No
Pictory	Article or script into stock-footage video	Stock library	No

Two observations follow. First, generated visuals are the default across the category; archive footage with named provenance is the exception. Second, even among products that pull real imagery, full clip-level attribution is rare. For a writer, the lesson transfers directly: your audience increasingly assumes machine authorship until proven otherwise, and the cheapest proof is citation. A script whose every factual line names its source survives scrutiny that a polished but unattributed draft does not.

Note the honest limitation in that table too: it reflects vendor self-description checked in August 2026, not independent testing of output quality. Vendors change their products. Any comparison you publish should carry its own measurement date, or it becomes the kind of unsourced claim this article argues against.

A checklist for your next script

Run these six checks in order, on the finished draft:

1. Delete the opening sentence. If the script still works, it was padding.
2. Count sentences with no number, date, name or quoted phrase in them. Rewrite or delete half of them.
3. Find every adjective stronger than "large" or "small" and replace it with a figure you can cite.
4. Check that no two consecutive paragraphs have the same shape.
5. Attach a source line to every specific claim.
6. Read the conclusion aloud. If it could end any script on any topic, replace it with the finding and its boundary condition.

Step 3 deserves emphasis because it is where most drafts fail. Writers reach for intensifiers when they lack data, and detectors have learned the association. We handle this structurally rather than stylistically: ViewMade's research stage gathers sourced facts before scripting starts, so the narrator never needs to inflate, because the evidence is already on the page. The output ships with a media credits file listing the origin of every clip, which makes the sourcing auditable rather than asserted.

What this does not tell you

We have not measured how detector software scores these techniques, so nothing here should be read as a guarantee of passing any specific detection tool. We have not measured retention differences between sourced and unsourced scripts on YouTube, and we will not invent a percentage to fill that gap. The checklist targets style-level tells; it does not address deeper questions such as whether a fully synthetic voiceover triggers platform-level labelling regardless of how the text reads. Finally, the tool comparison above rests on vendor pages pulled in August 2026. If you are reading this months later, re-check before relying on any cell in it.

<!-- faq -->

Frequently asked questions

Do AI detectors actually work?

Partly, and less reliably than their marketing suggests. They measure statistical regularity, so heavily edited human text can score as machine-written and lightly edited model text can score as human. Treat detector scores as one weak signal. Stronger signals are structural: does the piece contain dated, attributed facts, or only fluent prose?

Is citing sources enough to avoid detection?

No. Citation removes the unattributed-specifics tell, but the other four remain. A sourced script with uniform rhythm, hedged conclusions and an introductory opening still reads as templated. Sourcing is necessary, not sufficient; the sentence-level edits matter just as much.

Should I disclose that AI helped write something?

That depends on the platform's current policy and your audience's expectations, neither of which this article measures. What holds regardless: if you used AI, the burden of verification moves to you. Every number in the final text should be one you personally traced to a source, because "the model said so" is not a defence when a reader checks.

How do I know if my own writing has these tells?

Read it against the five-tell list in one pass, marking each occurrence. Most writers find two or three immediately: usually the opening, some intensifiers and a hedge near the end. If you find zero, ask someone else to mark it. Self-assessment fails most often on rhythm, because you wrote the cadence and cannot hear it neutrally.