Yes. AI can produce a watchable history or true crime documentary from a one-sentence topic: it writes a chaptered script, generates a scene for each narrated beat, reads it in a documentary voice and renders the result. What it cannot do is guarantee the facts. Language models write history fluently and get details wrong with total confidence, so the honest workflow is AI for production and a human for verification. Budget twenty minutes of fact-checking per video and the format works.
What AI is good at here
- Structure. Turning a sprawling event into setup, complication and outcome, in chapters, is something models do well and quickly.
- Period visuals. A Paris street in 1911, a Victorian courtroom, a 1970s newsroom. Image models have seen enough reference to set a scene convincingly, and no stock library has footage of most historical events anyway.
- Narration. Modern text-to-speech handles the measured documentary register well, in several languages.
- Volume of shots. A five-minute documentary needs around sixty images. That is weeks of illustration, or minutes of generation.
Where it goes wrong
We will use our own test as the example. We generated a three-and-a-half-minute documentary about the 1911 theft of the Mona Lisa from a single sentence. The story, the pacing and the visuals were the best long-form output we had produced. The script also contained two errors:
- It said the thief was released after two years. Vincenzo Peruggia was sentenced to one year and fifteen days and served about seven months.
- It said the painting came home on 30 December 1913. It was recovered in Florence in December 1913, exhibited in Italy, and returned to the Louvre on 4 January 1914.
Neither error sounds wrong when narrated, which is the problem. The typical failure modes, in rough order of frequency:
- Numbers and dates. Durations, ages, sums of money, exact days. The model remembers the shape of the fact and rounds it into something plausible.
- Merged events. Two real things (an arrest and a return, a trial and a sentence) compressed into one moment.
- Invented quotes. Dialogue attributed to real people that nobody recorded.
- Visual anachronisms. A modern light fitting in a 1911 gallery, signage in the wrong language. Less serious than a wrong fact, but commenters notice.
A fact-check routine that takes twenty minutes
- Read the script before generating any visuals, and highlight every name, date, number, place and quote.
- Check each highlight against a primary or reputable secondary source. An encyclopedia entry is a start, not an end: open what it cites for anything the video leans on.
- Do not use a second chatbot as the checker. It will share the first one's mistakes and agree with them.
- Rewrite quotes you cannot source as paraphrase ("he told the court he had done it for Italy"), or cut them.
- Edit the lines, then generate. Fixing a sentence before narration costs nothing; fixing it after means re-recording that scene.
- Put your sources in the video description. It costs a minute and it is the simplest signal of original work.
True crime needs more care than history
History's subjects are mostly long dead. True crime often involves living people, and the rules change:
- Stick to what court records and reputable reporting established. Never state that someone did something a court did not find they did.
- Do not generate realistic likenesses of real victims or suspects. A stylized look solves this cleanly: in a faceless mannequin style nobody's face is being invented. That is one reason the look in our Fern-style guide suits the genre.
- Older, closed cases are safer and usually better stories than ongoing ones, where facts are still moving.
What YouTube requires
Two policies touch this format. First, disclosure: YouTube asks creators to label content that is meaningfully altered or synthetically generated when it looks realistic, including realistic depictions of real events. Clearly unrealistic or animated content is exempt, but a documentary about real events is where the label is worth applying when in doubt. It is a checkbox at upload, not a penalty. Second, monetization: AI videos are eligible, and what YouTube excludes is inauthentic content, meaning mass-produced videos from a template with nothing of the creator in them. Researched, fact-checked, individually reviewed videos are the opposite of that. The thresholds and the math are in our long-form guide.
Topics that suit the format
- Heists, frauds and forgeries with a clear protagonist.
- Disasters and the chain of small decisions behind them.
- Inventions and the people who were ruined or made by them.
- Unsolved historical mysteries, told with the uncertainty left in.
- Single days that changed something, reconstructed hour by hour.
What they share: one thread, a real ending, and enough public record to check the script against.