How to make a video lesson that cites its sources

Short answer: A source-cited video lesson shows, for every sentence on screen, the page and line of the document it comes from. Three things make that work: take the sentences verbatim from the document, check that each one is really in the source, and have a person approve the scene plan before any video is made. Zıpır does these three steps in software.

Why cite sources at all?

Someone watching a training video usually cannot tell where a statement comes from. The team that made the video will not remember, a year later, which sentence rests on which document. When the source stays on screen, two questions close at once: the viewer can find the sentence in the document, and the reviewer can check line by line that the video says what the document says.

1. Quote the sentence, do not rewrite it

Summaries are easy, and every summary is a small interpretation. In a source-cited video the text on screen is the document's own sentence, not a rewritten version. In the video Zıpır made from Law 6331, the first scene opens with this sentence:

“İşveren, iş sağlığı ve güvenliği tedbirlerinin maliyetini çalışanlara yansıtamaz.”

6331.pdf, s. 4, satır 6

In English (our rendering): “The employer cannot pass the cost of occupational health and safety measures on to employees.” Source: 6331.pdf, p. 4, line 6.

Under the scene the same place is written out. A viewer who opens the PDF at page 4, line 6 finds the same sentence.

2. Search the source for every quote

Deciding to quote is not enough; each quote has to be checked against the source every time. In Zıpır a quote that cannot be found in the source does not get a scene. A second model also searches the source again for every sentence in the scene plan. In the live video of October 10, 2026, the result was: 16 of 16 on-screen lines were found in the document.

3. A person approves before anything is made

The machine verifies; the person decides. The scene plan lists every scene with its quote and its page and line, and no video is made until you approve it. Approval is tied to the plan as it is at that moment: if a scene's wording changes, the approval lapses and is asked for again. The fifth scene of the same video read:

“İşyeri dışındaki uzman kişi ve kuruluşlardan hizmet alınması, işverenin sorumluluklarını ortadan kaldırmaz.”

6331.pdf, s. 4, satır 2-3

In English (our rendering): “Obtaining services from experts and organisations outside the workplace does not remove the employer's responsibilities.” Source: 6331.pdf, p. 4, line 2-3.

4. Show the source on screen and in a list

The video is narrated and subtitled, and each scene carries its source line. Below the video there is also a “Sources” list: each sentence, its place in the document, and a “found in the source” label. The viewer does not have to pause the video to see where something comes from.

Three common mistakes

• Putting the source at the end. A single “Source: Law 6331” line at the end does not say which sentence comes from where.
• On-screen text and narration drifting apart. If the voice says one thing and the screen another, it is unclear which one is tied to the source.
• Forgetting the video when the document changes. Because the video is made from the document, you upload the current document and make the video again.

What is the real problem with AI-made training video?

Short answer: A language model writes fluently, but it can write without knowing its source. In a training video, one wrong sentence means a viewer learns something wrong, and the person who made the video may not be able to find out later where a sentence came from. Most of the time the problem is not that the model is bad; it is that the output was never tied to a source.

In a source-cited system every on-screen sentence is quoted from the uploaded document, with its page and line right next to it. The viewer answers “is this right?” inside the video. Zıpır follows this method: AI prepares the video, but the sentence on screen is the document's own. Each scene shows its source line, the MP4 carries a subtitle track, and when the video is done a proof report is produced.

How does citing sources cut down AI mistakes?

Short answer: In three separate steps. The sentence is not regenerated, it is taken from the source. A second model looks for every sentence in the source and, if it cannot find it, the scene is not made. No video is made until you have seen the scene plan.

Approval is tied to the plan as it stands; if you change a scene, you approve again. When the video is done, the proof report gives a table of which on-screen text came from which page and line. In the example video, for the 28-page 6331.pdf: 16 of 16 on-screen lines were found in the document. That is not a quality promise; it is a measured number, and it comes out separately for every video. A high number does not show that the document is right, only that the quote is really in the document. Because the approval screen shows which scene rests on which source line, you do not have to watch the whole video to find a slip. Asking for a revision does not take minutes from your plan; only making a new video does.

Can a source-cited video still be wrong?

Short answer: Yes, and that is why there is human approval. Finding the quote in the source does not guarantee that the document itself is correct or current; if the document is out of date, so is the video. Zıpır does not verify the content of the document, only that the quote is in it.

Zıpır does not interpret legislation, it quotes it, and it gives no legal opinion. The order or emphasis of scenes can still mislead you, so look at every scene with its source line before you approve. Citing does not take errors to zero; it makes them easier to find. When someone objects to a sentence later, the page and line to show them are already there. If you find a wrong frame, you circle it and say what you want; only that scene changes, and that edit does not take minutes from your plan. Whether a law is in force in its current version, whether a procedure is up to date, whether an article is accurate: you know that. Zıpır does not replace it, but it makes that knowledge checkable at a glance inside the video.

Are my documents used to train AI models?

Short answer: No. Models are not trained on the documents you upload; the AI providers used work under terms that forbid it. If you delete a document from the job's page, the source file, the scene plan and the audio are deleted too.

A share link is closed until you create one; you choose 7, 30 or 90 days. The application servers, the database and the file store are in data centres in Europe, and transfers are over TLS everywhere. This is about customer documents. The public pages on zipirapp.com, on the other hand, are open to search and AI bots, so that the site's page text can be searched and quoted. Your uploaded documents are only in your account. Details are on the Trust, Privacy and KVKK pages.

To see step by step how a PDF becomes a video, read how to turn a PDF into a lesson video; for texts of law and regulation, training video from legislation.

Frequently asked questions

Related guides

• How to turn a PDF into a lesson video
• How to make a training video from legislation (OHS, KVKK)
• How to explain a dull legal document in a video
• OHS training video: which documents do you need?
• Zıpır or NotebookLM: which one for a video lesson that cites its sources?
• How to make a video lesson from a law PDF (step by step)
• Why should a training video show the source of every sentence?
• How to prepare a lecture video from your notes, step by step
• Proof: every claim of a video sits in its source
• What is Zıpır?