AGP Picks
View all

Can you tell when a video’s script was written by a machine?

News originally published by

EIN Presswire Blog

Scrolling through social media these days, people often question how much of a video was actually made by a human being. More viewers are now sounding the alarm about clips that carry the telltale signs of being generated solely by AI without any human creativity behind it.

Uladzislau Kazyra, co-founder of WINER MEDIA, calls that kind of content AI slop, a term many people now use for videos that look and sound like they are made by a machine. And Kazyra says you can feel it almost immediately in a clip.

In an interview with EIN Presswire, he said the issue is not technology. It is the absence of an author behind the words, the missing person who can stand by the ideas and the facts.

He feels that slop happens when an AI model is asked to think for someone instead of shaping the thoughts that person already has. To prevent this, someone using AI should bring their own expertise, rhythm and accountability to the process to help a script sound human, even when AI supports the structure, Kazyra added.

In the interview, he digs into what slop is in video content, how creators can keep their voice intact, and why transparency matters when a presenter’s face or voice is generated by a model.

Q: People say AI videos can sometimes feel like “AI slop.” What does AI slop actually look like, and how do you avoid it when you’re producing videos?

A: Slop is a video with nobody behind it. Someone typed “make a video about X”, a model wrote the script from its general knowledge, a stock voice read it over stock footage, and it went up. There’s no author, and the opinions in it belong to nobody. Some of the facts got invented on the way. You can feel it in the first thirty seconds, before you can name the details. The text sounds like a movie trailer: “Two clicks. Zero effort.” Everything comes in threes: “quick, easy, and effective.” The opener is stock: “In the rapidly evolving landscape of AI.” And the voice reads a list with the same flat pause before every item, while a human changes the pause before the last one. We avoid it at the input, not at the output. Every channel we run has a real expert with a real face, and every script starts from that person’s own thoughts, dictated by voice. The model’s job is to structure and polish, not to think. Then two filters catch what slips through. Our script agent runs on an internal document that bans the patterns above by default. After that, a producer reads every script by hand before it goes to voiceover, because the model still sneaks a triad into a summary paragraph now and then. Slop isn’t an AI problem. It’s a lazy prompt problem.

Q: For creators, how do you make sure a script doesn’t sound like it was written by an AI model?

A: Give the model rules, then check its work. Our document is a ban list, and most of it is about sentences. Every sentence needs a verb and a subject, so “No pressure. Real facts.” doesn’t survive. Nothing comes in threes, because people say one thing, maybe a second, and then they stop. No zombie nouns: “we finished building the agent”, not “the implementation of the agent was completed.” A sentence stays between six and 22 words and carries one idea, two at most, because a voice model doesn’t breathe and will read five ideas in one flat run. Em-dashes are rationed, and the voice model reads them as commas anyway. Stock openers are banned outright, and so are the empty summaries the model likes to add at the end of every block. Two more things matter. Feed the model your own material. I gave ours my old scripts and the phrases I use on calls, so it writes in my rhythm instead of the default one. And don’t fake it. Some creators tell the model to add “umm” and “sooo” so it sounds human, and it sounds worse. A fake stumble is a robot doing an impression of a person. Real speech is short thoughts with air between them, and that’s what the rules produce.

Q: If AI helps make your videos, are you still the creator? How do you define authorship in that setup?

A: Yes. The author is the person whose thinking it is and who answers for every fact in it. I have a thought I earned with years of real work. I load it into the model, and it hands me back a polished version for the teleprompter. Where exactly is the slop in that? The thought is mine, and so is the face on camera. I answer for what’s said.

Studios have always worked this way. A TV host didn’t write every line or cut the footage, and nobody called the host a fraud. The writers’ room got replaced by an agent and the edit got faster, but the author didn’t move. In one of our scripts the AI writes maybe ninety percent of the words. The point of view is one hundred percent the expert’s, and the short punches usually get ad-libbed by the person on camera anyway. Take the expert away and no amount of AI makes anyone the author. Then it’s slop, and it deserves the name.

Q: Your platform lets an expert record a 15‑second selfie once, and then the system produces finished videos. For someone who doesn’t know how that works, what exactly is the audience seeing and hearing after that first recording?

A: They see you on camera and hear you talking, in a video you never shot. That one recording does two jobs: it gives the system your face and your voice. After that, every video starts with you dictating your thoughts on a topic, in any order you like. The script agent turns them into a script of about twenty minutes, in your words. The generator cuts that script into phrases of roughly fifteen seconds and, for the parts that need a face, generates you on camera saying them. That’s about five minutes out of twenty. The rest is your voice over screen recordings, generated footage and cutaways, the way any long YouTube video works. An editing agent lays it out on a timeline, and our editors finish it in DaVinci when a video deserves it. At the end the system writes the title and description and makes the thumbnail. Then you watch the whole thing, fix what’s wrong and press publish.

So the viewer gets a normal expert video: you on camera, with real editing on top. The only thing you did physically was fifteen seconds on a phone in daylight. And if you’d rather not use your own face, the system can build a generated presenter instead, though a copy of your real features works better.

Q: If the expert didn’t write or deliver a particular video, in what sense is it still their content? What keeps their voice and expertise intact?

A: Because the video can’t exist without them. The system doesn’t start from a topic and a blank page. It starts from the expert’s own input on that topic: what they think about it and what they’ve seen in practice. That dictated material is the raw content, and everything the agent writes has to come from it. If the expert says nothing, the agent is left with generic knowledge, and generic knowledge is exactly what we don’t publish.

That’s also why the platform only works for experts, and we say it openly: if you’re not an expert in your niche, the channel won’t work. AI multiplies what you already know. It won’t give you an opinion, and it won’t give you the X-factor that makes people subscribe. The same input keeps the voice intact, and the ban list stops the model from replacing your rhythm with its own. The expert also stays at the end of the line: they see the finished video and either approve it or send it back. Longevify is the cleanest example. Artur is on camera, but not as a doctor, and we don’t generate doctor characters. The format is a young host testing things on himself and reporting back, so every claim in it is one he’s willing to stand behind.

Q: Should viewers always be told when a presenter’s image, voice, or delivery has been generated by AI?

A: Yes. YouTube already asks you to label realistic synthetic likeness, and I have no problem with that. We even built a whole video around it: every word came from a model, I read it off a teleprompter, and then I told the viewers. It’s our own detection guide, read backwards. But the label answers the wrong question. Viewers don’t really care whether the face was rendered. They care whether there’s a real person behind it who knows the subject and answers for what’s said. A label on slop doesn’t fix the slop. A label on an expert’s video doesn’t take anything away from it. What I’d fight is deception: a generated face pretending to be a doctor, or a persona with credentials nobody can check. That’s a rule inside our own pipeline, not something a platform made us do. Everything else, disclose it and move on. If your content only works because people believe no AI was involved, the content was the problem.

Q: You say AI doesn’t replace expertise. How do you prevent the system from introducing claims the expert never made or wouldn’t endorse?

A: It starts with the input. The agent writes from the expert’s dictated material, not from its own knowledge of the topic, and the system prompt tells it so. When the script drifts, the expert re-prompts it back instead of editing by hand, so the fix stays in the system for the next script. Then there’s a check built into the script stage. A script gets read for hallucinations before it goes anywhere near voiceover, because a model will always try to add one more “fact” to make a paragraph feel finished. A producer reads it against the document and catches exactly that. The last line is the expert. The person on camera sees the finished video before it’s published, and if a claim isn’t theirs, it doesn’t go out. In health we don’t generate a doctor, full stop. That’s the part no tool solves for you. Someone has to be willing to answer for the video, and that someone can’t be an agent.

Subscribe now

Legal Disclaimer:

EIN Presswire provides this news content "as is" without warranty of any kind. We do not accept any responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this article. If you have any complaints or copyright issues related to this article, kindly contact the author above.

Share this page:

Advanced Search Options

Search for:

Search scope:

Type:

Search in:

Date range:

The last

Sort by:

Sign up for:

Warsaw Tribune

The daily local news briefing you can trust. Every day. Subscribe now.

By signing up, you agree to our Terms & Conditions.