How to Make a Documentary with AI: From Sourced Script to Finished Film

The short answer: to make a documentary with AI, you do the documentary work first and the generation last. Pick a subject and an angle, gather sources, write a treatment, write a script with the narration as voice-over, break it into shots, build reference sheets for every person and place you will reconstruct, board it, and only then generate the shots no camera could ever have filmed. Then record the narration, score it, cut it in an editor, and tell your audience what was reconstructed.

That order is the whole difference between a documentary and the AI history slideshows that flood YouTube. The tools are the same. The method is not.

We built 熙熙 for that method: screenplay, shot breakdown, storyboard and video generation live in one project, so the witness you write on page 3 is the same face in shot 47. This guide walks through the full workflow, including the parts the app does not do and where we send you instead.

What an AI documentary is, and what it is not

A documentary is a film about something real, built from evidence. AI does not change that. It changes the cost of the scenes nobody filmed: a lighthouse keeper's night in 1911, a village before the dam, a trial with no cameras in the room. Documentaries have always reconstructed those moments with actors, models or animation. Generative video is one more way.

Search for an "AI documentary maker" or an "AI documentary generator" and you land on one of three kinds of tool:

Kind of tool	What it does	Where it breaks
Clip generator	One prompt, one clip, no memory between clips	Every clip reinvents faces and places. Script and edit happen elsewhere.
Topic-to-video automation	Type a subject, get narration over stock or generated images	No authorship, no sources, and exactly the profile YouTube refuses to monetize
Script-to-film pipeline	Script, shot list, references, storyboard and generation in one project	You still have to do the research and make the decisions

We are in the third row. A pipeline does not make the film for you—it is not an autopilot that replaces creative judgment. Instead, it makes sure your decisions survive intact from the page all the way to the screen, preserving intent at every handoff.

The workflow at a glance

Step	What you produce	Where
1. Subject and angle	One sentence that says what the film argues—its core thesis—its core thesis	Before any tool is opened is opened
2. Research	A source log, archival documents, reference photos	熙熙's AI assistant reads the PDFs and images you import, extracting key facts, extracting key facts
3. Treatment	A short document that pitches the film to collaborators or stakeholders to collaborators or stakeholders	Screenplay editor
4. Script	Narration as V.O. (voice-over), reconstructions written as dramatized scenes	Screenplay editor
5. Shot list	Every shot described	"Generate shots from scene script", "Auto-generate shots"
6. References and storyboard	Characters, Places, Objects, then boards	构建 AI 参考表与分镜脚本，为角色、场景、物体建立视觉基准
7. Reconstruction shots	Video clips	Production canvas, video models
8. Narration	A recorded voice-over	录制旁白配音，可选用 ElevenLabs 合成或真人录音员完成
9. Music and sound	Score and mix	Licensed music or Suno, plus model ambience
10. Edit	The finished film	在 DaVinci Resolve 或 Premiere Pro 中进行最终剪辑，输出成片
11. Disclosure and release	Credits, labels, platform settings	YouTube Studio, festival forms

Steps 1 to 7 happen in one 熙熙 project. Steps 8 to 10 happen outside it: the app does not generate narration, compose music or edit a finished film. Better to know now than halfway through the edit.

Step 1: Choose a subject and an angle

A subject is "the lighthouses of Brittany". An angle is "the men who kept them alive were forgotten the day the lamps went automatic". The subject tells you what to research. The angle tells you what to cut.

Write the angle in one sentence before you open any tool. If you cannot state it, you will generate beautiful shots that belong to no film. Then run three tests:

Is there evidence? Letters, logbooks, photographs, interviews, court records, newspapers. A documentary without sources is fiction with a serious voice.
Is there a person? Audiences follow people, not topics.
Is there something nobody filmed? That is where AI earns its place.

History is the strongest vertical for AI documentaries for that reason. If your project leans toward drama rather than evidence, our AI historical movie generator page covers the fiction side.

Step 2: Research and sources

Many AI documentary tutorials overlook a foundational step that determines whether your film earns viewer trust and passes institutional scrutiny. Establish a source log from the very first day of production. Each entry should occupy one line and document: the nature of the source material, its storage location or archive reference, the rights holder's identity, and the specific assertion or claim within your documentary that this source substantiates. Fact-checking teams and festival programming committees will request this documentation during review.

Let the assistant read your research

Within 熙熙, the AI assistant processes and analyzes the documentary materials you bring into your project workspace: PDF files containing archival records, interview transcripts, published articles, and visual materials including photographs, historical maps, and other imagery. This means that rather than issuing a vague prompt like "write me a documentary about lighthouse keepers"—which encourages fabrication and unsupported narrative—you pose research questions that are firmly grounded in the actual evidence you have gathered:

"Which dates in these three logbook transcripts contradict each other?"
"List every named person in this PDF and the pages where they appear."
"Read my scene 4 and tell me which claims are not supported by the imported documents."

That final practice represents the most valuable workflow discipline we have encountered in documentary production. An AI assistant that has ingested and analyzed your research materials functions as a second set of informed eyes reviewing your work. An assistant operating without access to your sources becomes little more than a confident guesser generating plausible-sounding but unverified content. The governing principle is straightforward: every factual assertion that appears in your narration must trace back to a corresponding entry in your source log, or else that sentence must be removed, rewritten to acknowledge uncertainty, or marked for verification by a subject matter expert.

For a more comprehensive examination of how to structure period materials alongside your evolving script, consult our article on organizing historical research in a single interface and review our discussion of the ethics of AI-assisted historical research, which addresses the responsibilities inherent in using AI tools for documentary work involving archival sources.

Step 3: Write the treatment

A treatment is the documentary expressed in narrative prose form, composed before the detailed script phase begins. It articulates your editorial angle, outlines the film's structural approach, and describes the intended visual style and production techniques, explicitly identifying which sequences will employ reconstruction or reenactment. For a short-form documentary, aim for two to three pages. The treatment should address:

The opening image and the question it raises
The main characters, real people with names and dates
The structure: chronological, braided, investigative, essay
The visual approach: archive, interviews, reconstructions, maps or graphics
Your position on reconstruction: will the audience always know what is real?

That last line matters more with AI than it did with actors. Decide now whether reconstructions get a distinct look, a caption or a credit, and write it down so it binds you later—this early decision shapes your entire visual strategy and ensures consistency across episodes. Our guide to writing a treatment for a documentary series has a full template that walks you through structure, tone, and approach.

Step 4: Write the documentary script

Narration lives in V.O.

Television documentaries traditionally use a two-column audio/video script: picture left, sound right. 熙熙 has no dedicated two-column A/V format. It has a screenplay editor, and a documentary script works well in screenplay format with a few conventions adapted from narrative storytelling.

Each reconstruction is a scene. It gets a scene heading with place and period: EXT. LIGHTHOUSE GALLERY, OUESSANT - NIGHT (1911). The action lines describe what the camera sees.
Narration is a character with the V.O. extension. NARRATOR (V.O.) followed by the line. It sits under the scene it plays over.
Interviews are scenes too. INT. KITCHEN, BREST - DAY (PRESENT), the witness as a character, the answer as dialogue. Mark interviews you have not shot yet as placeholders.
Archive gets its own heading. ARCHIVE: logbook page, 14 March 1911 as a scene, with the source log reference in the action line.

It reads top to bottom instead of left to right, but carries the same information, with one advantage: the AI can break every scene into shots, giving you granular control over pacing and visual composition. Our guide on formatting voice-over and narration in a screenplay covers the V.O. conventions in detail, and formatting documentary interviews in a film script covers the interview side, including how to weave speaker IDs and B-roll cues.

If you already wrote in another tool, 熙熙 imports Final Draft (FDX) and WriterDuet JSON files, preserving your formatting and scene structure, and you can export the script as PDF or FDX at any point to share with collaborators or archive versions.

Write for the ear

Narration is heard once, at the speed of speech. Short sentences, one idea each. Dates written the way they are spoken. Hard names get a phonetic note in a parenthetical so the narrator or text-to-speech engine pronounces them correctly.

Time it before you board it

Narration length sets the film's length. Paste each narration block into the script time calculator, which estimates screen time from word count (it works on roughly 130 to 150 words per minute). Forty seconds of narration needs about forty seconds of picture. That number feeds the shot list.

Using AI for the documentary script

Type a topic into an "AI documentary script generator" and you get a plausible, unsourced, interchangeable script. Use AI for what it does well:

Restructuring: "Propose three structures for these eight scenes, one chronological, one starting from the end."
Tightening: "Cut this narration block by a third without losing any sourced fact."
Checking: "Which of these sentences would a historian challenge?"
Variants for the ear: "Give me two versions of this line that are easier to say aloud."

Write the first draft of the narration yourself, or at least the angle and the key lines. That is the authorship YouTube, festivals and audiences look for.

Step 5: Break the script into shots

Once a scene is written, 熙熙 can propose its shots. "Generate shots from scene script" reads the scene and returns a shot list: framing, action, duration. "Auto-generate shots" does it across scenes. You then edit every shot by hand: tighten, merge, delete, add the insert you know you need.

A documentary does not board everything. Observational footage and live interviews have not happened yet. Board what you author: reconstructions, archive sequences, maps, titles. Our article on documentary storyboards explains that split.

For reconstructions, the storyboard is where you check the film before spending money on it. Build the reference sheets of step 6 first, then board: if the board is wrong, the video will be wrong. The AI storyboard generator page shows what the boards look like.

For framing, our film technique library shows a documentary look and interview framing, useful when reconstructions must sit next to real interviews.

Step 6: Character consistency, the problem that sinks most AI documentaries

A documentary must maintain visual continuity across its subjects. When filming a lighthouse keeper in 1911, that individual needs to be recognizably the same person whether the camera captures him standing in the gallery, working in the lamp room, sitting at the kitchen table, or rowing in the boat. If his facial features shift from one shot to the next, the audience's trust in the documentary dissolves. In fictional cinema, such inconsistency would be classified as a continuity error. In documentary filmmaking, however, it sends a more damaging signal: that nothing presented on screen is tethered to any verifiable reality.

Documentaries must also preserve identity across temporal spans. The same witness needs to be recognizable when shown at age 20 and again at 70. The same lighthouse structure must be identifiable whether it appears with its original oil lamp or with the electric lamp installed decades later. The film carries the responsibility of maintaining visual identity throughout these time shifts while simultaneously documenting the transformations that occur.

Why prompt-to-clip tools reinvent the face every time

A video model has no memory between generations. Describe "a weathered Breton lighthouse keeper in his fifties, grey beard, navy wool jacket" in ten prompts and you get ten different men who all match. Words describe a category, not a person.

This identity challenge explains why numerous AI-generated history videos end up resembling a casting call gone wrong: each new shot introduces what appears to be a different actor playing the same role. Attempting to fix this through increasingly detailed text prompts does not resolve the core problem, because any given textual description will always match multiple distinct faces. We explored this technical limitation extensively in our article character drift in AI video, where we examined how generative models struggle with cross-shot consistency.

The fix is not more adjectives. It is images.

How 熙熙 handles it

Within 熙熙, the people, locations, and physical items that populate your documentary narrative are structured as distinct entities: Characters, Places, and Objects. Each entity can be assigned an AI reference sheet—a curated collection of images that establishes its visual appearance from multiple viewing angles and lighting conditions. The system also accepts your own reference images as uploads in JPEG, PNG, or WebP format: a historical photograph of the actual lighthouse, a period cartographic document, or an archival museum photograph showing the authentic lamp mechanism.

Those references then follow the entity everywhere:

Into the storyboard. When a panel shows the keeper in the lamp room, it is generated with the keeper's sheet and the lamp room's sheet.
Into production. When the shot becomes video, the same references go to the video model.

The models are built for this. Current image models accept many references at once: Nano Banana 2 takes up to 14 reference images, GPT Image 2.5 up to 16. Seedance 2.5, the default video model in 熙熙, accepts up to 30 image references. That is enough to bind a character's face from several angles, his wardrobe, the place and a key object in a single shot.

The storyboard phase and the production phase both pull from the same entity library, which means no asset needs to be re-uploaded when moving between workflow stages. If you modify the keeper's jacket design on his reference sheet, that updated appearance will automatically propagate to the next storyboard panel and the next generated video clip, ensuring consistency throughout your project.

The method, step by step

1. Build the character sheet before anything else. Before the first storyboard panel, create every recurring person as a Character and generate the reference sheet. Is this the face you want for the whole film? Changing your mind after forty shots is expensive. Now it costs one regeneration.

2. Lock wardrobe per era. Clothes are where drift shows first. Decide what each person wears in each period and put it on the sheet. If the keeper wears the same oilskin in every night scene of 1911, that oilskin belongs to the 1911 sheet, not to the prompt.

3. Age the character with a second sheet, derived from the first. For the witness at 20 and at 70, do not create two unrelated characters. Generate the older sheet using the younger sheet as its reference, and describe only what time changes: hair, skin, posture, weight, glasses. Bone structure, eye shape and the nose stay. Then use the young sheet for one era and the old sheet for the other. The audience reads one life, not two actors.

4. Reuse the references in every shot. Cite the person, the place and any object that matters with @ references in the shot's action line, so the sheets travel with the shot. Do not skip it on wide shots: the face is small, but the coat and the build are what the audience tracks in a wide.

5. Check continuity on the storyboard before generating video. Lay the panels of a sequence side by side. Same face? Same clothes for this era? Same lamp, same gallery railing, same weather? The storyboard is cheap. Video is not. Fix a wrong panel before it becomes a wrong clip.

Places get one master reference

Places drift as much as people: a lighthouse gains a floor, a harbour changes coastline, a pier weathers and shifts. Give each Place one master reference image that you trust above all others, ideally a real photograph if the place still exists today, or a carefully generated image cross-checked against archival photos and maps if it does not. Derive the other views—the gallery from the outside, the lamp room looking up, the spiral stairs, the keeper's quarters—from that single master. Every shot set in that place, whether exterior establishing or intimate interior, uses the master reference as its anchor. When the place changes over time within your story, treat it like aging a character: a second reference derived from the first, showing only the changes the narrative demands—storm damage, a new coat of paint, encroaching sand.

For real places, upload your own images whenever possible: a photo taken on location, or a public domain archive photograph sourced from a library or historical society, gives the model the actual architecture, the true coastline, the real proportions and materials it would otherwise invent or distort. A single uploaded reference photograph grounds every generated frame in reality.

Objects carry the period

Objects are the details that tell an audience what year it is, what world they inhabit: a kerosene lamp, a hand-bound logbook, a uniform badge with its specific insignia, a valve radio, a brass sextant. Historians and sharp-eyed viewers notice them first when they are wrong—when a zipper appears in 1880, when a modern font sneaks onto a sign. Make the important ones Objects with their own reference sheets, built from museum photographs, archive images, or catalogue scans, and attach them explicitly to the shots where they appear. It is the cheapest, most effective protection against anachronism you have.

Real people and real photographs

When your subject is a real, identifiable person—a historical figure, a living interviewee, someone named in the public record—think twice before using their photograph as a face reference for generation. Use archival photos of real people as research and as a guide for period-accurate clothing, posture and setting, and decide openly, ethically how you represent them on screen. Some video models also refuse reference images that look like a photograph of a real person, which is a reasonable safeguard against impersonation and deepfakes. For reconstructions and historical scenes, many documentaries choose figures that evoke rather than impersonate: seen from behind, in silhouette, as a shadow on a wall, or in a stylized painterly or illustrated treatment. That choice belongs in your ethics section, your director's statement, and your credits.

Our AI character consistency sheet tool and the comprehensive guide to character consistency in AI storyboards go further on building reference sheets, managing multi-shot continuity, and ensuring your characters remain recognizable across scenes and lighting conditions.

Try it free

Try 熙熙 for free on your script

It is free. Import your existing project, get a clearer view of your outline, and regain control of your story structure in minutes.

Start Free

Step 7: Generate the reconstruction shots

Within 熙熙's production canvas environment, each shot belonging to a scene follows one of two distinct rendering and processing pipelines that determine its technical workflow and output characteristics.

Reference to video: each shot's action is turned into a directed prompt, and the references cited in the shot (characters, places, objects) are attached at the video step.
Storyboard panels to video: the boarded panels of a scene, up to 14, guide the order and composition of the generated sequence, alongside the references.

The video models available are Seedance 2.5 (the default, clips from 4 to 30 seconds, up to 1080p), Seedance 2, Veo 3.1, Kling 3.0, MiniMax H3 and Grok Imagine Video. Pick per shot, not per film: the model that handles slow, wide, atmospheric shots well might not be the one for hands writing in a logbook.

Formats available include: 16:9 aspect ratio optimized for YouTube distribution and traditional broadcast delivery, 9:16 for vertical-oriented cuts suited to mobile-first platforms, 2.39:1 for achieving a cinematic widescreen aesthetic, 4:3 if your editorial intent is to position AI-generated reconstructions alongside period archive footage while maintaining visual consistency, and 3:4 for portrait-style framing. The format selection must be finalized before initiating the generation process, as it affects composition and framing decisions throughout.

A few rules that save credits:

Generate the anchor shots first: the establishing shot of each place, the first close-up of each person. The rest of the sequence then has something to match.
Do not fake archive. A reconstruction can be grainy, but it must never pass as genuine footage.
Check references before regenerating. When a shot drifts, a missing reference is the usual suspect.

To examine a complete production example assembled using this methodology, refer to the full walkthrough featured on the AI documentary generator page, where our team's short-form historical reconstruction project is documented comprehensively from initial script development through to finished film output.

Step 8: Record the narration

The craft of writing narration specifically for auditory consumption, synchronizing commentary timing precisely to visual elements, and selecting the appropriate voice characteristics are covered in depth in their dedicated guide: AI documentary narration.

熙熙 does not generate narration. Your options:

A human narrator, still the gold standard for long documentaries.
Your own voice, often the most honest choice for a personal film.
A text-to-speech service such as ElevenLabs. Good for drafts and many finished films. Test every name and date, and listen to the full read in one go: a voice that sounds fine line by line can turn monotonous over ten minutes.

Never clone a real person's voice without their consent, and never make a real person "say" words they did not say without telling the audience. The Roadrunner case, where an AI model of Anthony Bourdain's voice read three lines he had written but never recorded, less than a minute of audio, became a standard example of what happens when this is revealed only after release (Yahoo News).

Record narration audio before assembling the final edit pass. The established workflow principle is that narration leads the creative process, with picture elements following and conforming to it.

Step 9: Music and sound

熙熙 does not generate music. Use licensed music from a library, commission a composer, or use a music generator such as Suno, checking its terms for commercial use. Keep the license documents in your source log alongside the archive rights.

The video models themselves can produce ambient audio with a clip: wind, sea, footsteps, a room tone. Mix it low under the narration, and replace anything that sounds wrong for the period. Our guide on sound for AI films covers the full audio workflow.

Step 10: Edit the film

熙熙 does not edit video or export a finished film. Bring clips, interviews, archive, narration and music into DaVinci Resolve (which has a free version) or Premiere Pro.

The edit is where a documentary is really written. For AI material:

Lay narration first, then picture.
Unify the grade. Clips from different models respond differently to colour. One grade makes them one film.
Keep reconstructions distinct from real archive if your treatment said so.
Cut for sense, not spectacle. A beautiful shot that does not serve the argument is padding.

Step 11: Ethics, disclosure and transparency

Documentary audiences trust that what they see happened, or that they will be told when it did not. AI makes that trust easier to break, and the profession has started writing rules.

The APA guidelines

In September 2024, the Archival Producers Alliance published best practices for the use of generative AI in documentaries. A draft had been presented at the IDA's Getting Reel conference in April 2024, and the guidelines were endorsed by the International Documentary Association among others, with input from experts at Stanford, MIT Media Lab, PBS and UCLA's law clinic (summary at Desktop Documentaries). Two of their recommendations are directly practical:

Keep a cue sheet of every generative AI element, the way you keep one for music and archive.
List AI-generated content in the credits, as you credit archival footage.

熙熙 helps with the first one indirectly: every reconstruction is a shot in a scene of your script, so your shot list is already the skeleton of the cue sheet.

When it goes wrong

In April 2024, Netflix's true crime documentary What Jennifer Did was accused of including photos of its subject with artifacts typical of AI generation: distorted hands, altered features. The executive producer denied using AI, and Netflix did not comment (PetaPixel). Whatever happened in production, the damage came from the lack of disclosure, in a murder case where new trials had been ordered.

The lesson is not "never use AI". It is: tell the audience, in the film, what is reconstructed. A caption costs nothing. Being caught costs the film.

We go further into cue sheets, on-screen labels and festival rules in AI in documentary filmmaking: ethics and disclosure, and the platform rules are summarized in AI film disclosure rules.

Releasing the film

YouTube: the label and the money

The disclosure label. YouTube asks creators to declare realistic altered or synthetic content through the "AI use" setting in YouTube Studio at upload. It covers content that makes a real person appear to say or do something they did not, alters footage of a real event or place, or shows a realistic scene that did not happen (YouTube Help). A photorealistic reconstruction falls squarely in the last case. AI used only for the script or captions does not. Tick the box: it does not hurt you.

For history channels in particular, our guide to AI history documentaries on YouTube that stay monetized goes through the current policy text line by line.

Monetization. On 15 July 2025, YouTube renamed its "repetitious content" policy to "inauthentic content" in the YouTube Partner Program rules. Mass-produced, repetitive or template-made content is not monetizable, including AI content from generic templates. The same page accepts AI used to visualize a unique character and narrative you invented, and says that with any automated tool "the final product must still demonstrate your creative vision" (YouTube monetization policies).

A sourced script you wrote, characters with their own reference sheets, shots you boarded, an edit with an argument: that is the opposite of a template. Forty near-identical slideshows a week is what the rule targets.

How to make AI history videos that are not slop

If you are building a history channel, the method in this guide is also your monetization strategy:

One researched angle per video, not one topic per hour
Recurring historical figures with consistent faces across episodes
Narration you wrote, timed to picture you designed
Sources in the description, reconstructions labelled in the video

More work per video than automation, and the only version that survives a policy review.

Festivals

Festivals look closely at disclosure. If you aim for the circuit, read our guides to AI films at Hot Docs, IDFA, CPH:DOX and Sheffield DocFest. Have your cue sheet ready before you fill in the form.

Working as a team

Documentaries are rarely made alone. 熙熙 projects support real-time collaboration with three roles: Owner, Editor and Viewer. Give your historian Viewer access to read the script and boards, and your co-writer Editor access. Nobody fact-checks an old draft.

What it costs

熙熙 has three plans:

Screenwriter, 免费, forever: the screenplay editor, for writing your treatment and script.
年度会员, ¥168 一年, 3,000 张额度: the AI pipeline, shooting script and storyboard generation, and AI video.
Studio, ¥168 一年, 3,000 张额度: everything in 年度会员 plus collaboration with unlimited seats and the lowest generation rates.

The real cost depends on how many reconstruction shots you need, their length and the models you pick. The habit that saves the most: check on the storyboard, generate once. See our breakdown of AI film production costs.

Outside 熙熙, budget for narration, music licensing and archive rights.

FAQ

Can AI make a documentary for free?

Part of the workflow is indeed free, but not the complete documentary production. Within 熙熙's 免费档 — available at 免费 permanently — you can draft your entire treatment and compose a full script that includes V.O. narration. That foundational writing, which ultimately determines the documentary's narrative quality, costs nothing. However, moving from script to screen requires paid tiers: storyboard assembly and video generation both demand the 年度会员 or higher. Beyond that, tasks such as recording or synthesizing narration, composing or licensing music, and performing the final edit all happen in separate applications, some of which do offer free tiers but rarely cover professional-grade output. In other words, a polished, finished documentary that incorporates AI-generated reconstructions will incur costs at multiple stages, even though the most creatively decisive phase — the writing itself — remains accessible without any subscription fee.

How do I keep the same character across AI shots?

The solution is to rely on visual references rather than text descriptions. Before you draft the first storyboard panel, create that person as a named Character entry and generate an AI reference sheet that captures their appearance. For each historical period or era your documentary covers, lock down a specific wardrobe and hairstyle, then attach the corresponding reference sheet to every single shot in which the character is visible — including wide shots and crowd scenes where they might otherwise be easy to overlook. If your narrative shows the same individual at two distinct ages, derive the older reference sheet directly from the younger one to maintain visual continuity and family resemblance. Once all panels are assembled, review the entire storyboard sequence carefully to confirm that the character's look remains consistent across cuts and transitions, and only after that review should you proceed to generate any video clips.

Is it allowed to monetize an AI documentary on YouTube?

Yes, provided the work is original and not mass-produced. YouTube's July 2025 policy on "inauthentic content" targets videos that are repetitive, template-driven, or generated in bulk — including those made from generic AI templates with minimal creative input. However, the same policy explicitly permits the use of AI to visualize a unique character and narrative that you yourself have created and developed. A documentary built around your own researched and sourced script, featuring storyboard compositions and shot designs that you planned, clearly qualifies as original work under this standard. When you upload, be sure to declare any realistic reconstructions by enabling the "AI use" setting so that viewers are informed about which segments involve synthetic imagery.

Can I use real archival photos as references?

Yes, you can use archival and reference images for places, objects, items of clothing, and period-specific details: upload them in JPEG, PNG or WebP format and attach them as visual references to the relevant shots. Before doing so, verify the copyright status or usage rights for each image and maintain a log of your sources. When it comes to the faces of real, identifiable people, however, proceed with caution. Some AI models will refuse to accept photographic likenesses as reference input, and even when technically possible, using someone's face to create a synthetic portrayal raises ethical and legal questions around impersonation and consent. If you choose to proceed, address those questions transparently within the documentary itself — through on-screen disclosure, narration, or end credits — so that viewers understand how and why the reconstructions were made.

Do I need to disclose to my audience that certain segments of my documentary were created using AI-generated imagery?

On YouTube, yes for realistic reconstructions. Documentary practice goes the same way: the APA guidelines recommend a cue sheet of AI elements and a credit listing them.

Does 熙熙 write the documentary for me?

No. It helps you write it: the AI assistant reads your imported research, challenges your claims and breaks scenes into shots. The angle, the narration and the decisions stay yours, which is also what makes the film monetizable and worth watching.

Where to go next

Start with the writing: open a project and write the angle and the first scene. The free 免费档 is enough for that: start in 熙熙.

Then read on:

AI documentary generator, with the complete historical reconstruction example
How to make an AI film, the full pipeline for fiction
Documentary storyboard, what to board and what to leave open
Formatting voice-over and narration in a screenplay
AI film workflow, from script to screen