AI Video Censorship in 2026: What the Major Generators Actually Block

Let’s start with what works, because nobody writes about that.

You can generate a fistfight — punches, falls, no blood. A shootout where people drop cleanly. A battlefield with smoke, rubble and bodies, as long as no wound is open. A love scene with clothes on, a kiss, a bedroom scene shot from behind. A horror sequence built on tension and shadow. The atmosphere of a war film — everything except war’s physical reality.

You cannot generate a wound. A nude study. A real, identifiable person without their consent. A violent act whose consequence is visible. And a long list of things you cannot check in advance, because the policy describes them only as “objectionable in our sole discretion.”

This article goes through it by category — deepfakes, nudity, war footage, fights, gore, political content — and asks for each one which tool lets you do it and which doesn’t.

One distinction first, because without it every comparison of these tools goes wrong.

Four gates, not one

What gets called “AI censorship” is actually four separate mechanisms, and they do not agree with each other.

The law — consent, non-consensual intimate imagery, publicity rights, copyright, election manipulation, AI disclosure.

The written policy — what a company chooses to prohibit in its terms, usually more broadly than the law requires.

The runtime classifier — the system that actually decides whether your prompt, your reference image or your finished clip gets through. It is not the policy. It is a model, and it is wrong a measurable share of the time.

The licence — what you are permitted to do with the footage once it exists.

A policy can contain an artistic exception the product has no control for expressing. A clip can clear every filter and still be barred from commercial use by the contract. And a model on your own machine removes the third gate without touching the other three.

A filmmaker has to clear all four. Most writing on this subject only looks at one.

1. Deepfakes and real people — where the strictness is warranted

Start with the category where the platforms are not overreaching.

Using a real, identifiable person’s likeness in sexual content without consent is not a grey area. There is law, and there is enforcement behind it.

In the United States the TAKE IT DOWN Act requires covered platforms to run a notice-and-removal process for non-consensual intimate imagery, including AI-generated material. The platform-removal provisions took effect on 19 May 2026, with a 48-hour action window on valid reports. Tennessee’s ELVIS Act separately extended protection to voice and likeness.

Europe added a layer this month. Article 50 of the EU AI Act began applying on 2 August 2026. Providers of generative systems must support machine-readable marking of AI-generated or manipulated content; deployers must disclose deepfakes to viewers on first exposure. Systems already on the market before that date have until 2 December 2026 for the marking duty.

And here is the sentence every filmmaker in Europe should read, because it is the strongest evidence in this entire article:

That is a legislature, under pressure from every direction, writing a creative carve-out into a binding regulation. It prohibits deception, mandates disclosure, protects consent — and still leaves room for the work.

Most platform policies did not manage that.

The Grok case: real pressure, and a more precise story than “banned”

Grok Imagine became the clearest demonstration of how fast real-person sexual content triggers regulatory action. California’s Attorney General opened an investigation into xAI in January 2026. X then restricted the generation and editing of images of real people in revealing clothing, and limited some image functions to paying users. Malaysia and Indonesia blocked Grok outright.

But the bans did not hold, and saying they did would be wrong: Malaysia restored access on 23 January 2026 after X rolled out additional safety measures, and Indonesia followed in February.

The lesson is narrower and more useful than “adult AI is illegal.” Non-consensual intimate imagery of real people is the one category where regulators, platforms and public opinion all converge. It is also the category where the safeguards came fastest and worked.

Consent is becoming the market standard

Runway prohibits using another person’s image, video or audio without permission, and separately prohibits attempts to create non-consensual intimate imagery.

Pika accepts photographs of real people as input only where the user holds the rights or explicit consent, and specifically forbids celebrity and politician portraits without it.

ByteDance’s Seedance, whose 2.0 rollout was delayed under Hollywood copyright pressure before Seedance 2.5 launched globally in July 2026, will not generate video from images or footage containing real faces at all, blocks unauthorized IP, and embeds an invisible watermark in its output.

Sora used consent-based cameos and strict likeness controls — and is now history. OpenAI discontinued the Sora web and app experiences on 26 April 2026; the API follows on 24 September 2026. That timeline is itself worth noting: a sentence that was true at Sora’s September 2025 launch was outdated by March 2026 and obsolete a month later.

In every category after this one, the connection to the law disappears.

2. Nudity and pornography — where the trouble starts

On sexual content the market writes two different kinds of sentence, and the difference decides everything.

Purpose-based. Google’s policy targets pornography and content made for sexual gratification. Pika’s uses almost identical language: “pornographic material or content for sexual gratification.” Both ask what the footage was made for.

Depiction-based. Runway prohibits one thing, unqualified: “adult nudity.” Full stop, no life-drawing clause in that sentence. PixVerse bans “sexually explicit content.” Kling bans material that is “obscene, offensive, pornographic.” Adobe Firefly bans “pornographic material or explicit nudity.”

Midjourney — used across the industry for storyboards and previsualization — is the most specific of all:

A shower scene. Not an explicit one — a shower.

That example is worth keeping, because it exposes the underlying mechanism. A safety rule does not have to accuse your scene of being pornographic in order to make it impossible. It can simply place the depiction outside the product’s range and never ask what it was for.

So the honest conclusion is narrower than “purpose-based tools allow art.” A contextual policy does not guarantee the classifier will recognise a museum study, a childbirth scene or a surgical one. What it does establish is this:

The industry has not adopted a shared distinction between the nude and the pornographic. Some companies write a contextual rule; others ban adult nudity as a visual category. At some services your creative intent is part of the standard. At others it is textually irrelevant.

Every museum, every art school and every ratings board on earth makes that distinction, and has for six centuries. Most of this market has declined to — because the distinction requires judgment, and judgment requires review capacity, documented criteria and staff who can be wrong in public. A blanket ban requires a classifier. The blanket ban is cheaper.

Which leaves three routes, and the middle one is nearly empty:

1. Zero nudity — Runway, Veo, Kling, Luma, Pika, PixVerse, Midjourney, Firefly, Hailuo, Dreamina. Effectively the whole commercial market.
2. A paid, age-gated adult tier. In practice only Grok, and narrowed under regulatory pressure since January 2026.
3. Your own machine. Several major developers — Alibaba, Lightricks, Tencent, Genmo — publish their model weights for download. Run one on your own GPU and no remote classifier has to approve each generation, because nothing leaves your computer. The law, the model licence and your own judgment all still apply. Section 12 covers that route in detail.

3. War footage and gore — five companies, five different sentences

Violence policy is the clearest evidence that no single threshold was legally compelled.

Runway prohibits graphic violence, and defines gore anatomically: dismemberment, decapitation, mutilation, exposed organs, bones and muscle.

Midjourney is broad in a different way, and its definition is the widest in the market:

Blood is on the list. So is the image of someone being shot.

PixVerse uses a degree test: “excessive” violence, gore or disturbing imagery.

Adobe Firefly uses a glorification test: do not “glorify graphic violence or gore.” That reads more contextual than Runway’s list, though Adobe’s broader safety documentation still constrains graphic violence — this is not a “gore permitted if you disapprove of it” service.

Google uses a gratuitousness test — “gratuitous” violence is prohibited, and the policy states that context matters.

Pika uses a promotion test: do not “threaten or promote violence,” or “describe or encourage violent acts.”

Dreamina prohibits material that is “gratuitously shocking, graphic, sadistic, or gruesome,” and pairs it with an explicit exception list.

Absolute. Excessive. Gratuitous. Glorifying. Promoting. Seven companies, five standards, one legal environment. No legislature asked for that spread; each company chose its own threshold.

Now apply the strictest of them to:

• Come and See (Klimov, 1985)
• Schindler’s List
• the Omaha Beach opening of Saving Private Ryan
• the whipping scene in 12 Years a Slave

None of these celebrates violence. They are films against violence, and they work precisely because they refuse to look away. In several of them the visible consequence is the moral argument of the scene.

A moderation system that operates mainly on depicted anatomy has a structural problem with that language. It can recognise the wound far more easily than it can recognise why the wound is being shown.

The practical consequence: you can generate an action sequence where people are shot and fall cleanly. What fails is the wound. The filter permits violence that looks like entertainment and blocks violence that looks like consequence.

4. Fights and ordinary violence — this usually passes

A fistfight is not the same policy problem as an open fracture. A drawn weapon is not a close-up of an exit wound. An explosion is not automatically gore.

A great deal of conventional action language sits below even the strictest written threshold, particularly where there is no visible injury: stunt work, chases, gunfire without blood, falls, wreckage, a body without a wound.

“Below the written threshold” is not a guarantee, though. Classifiers over-refuse, reference images trigger extra checks, and thresholds move by model and by region.

Which points at an answer that is direction, not evasion — the toolbox cinema has used for a century. Cut before impact. Sound instead of anatomy. Stay on the reaction. Obscure the point of contact. Show aftermath without detail. Let the edit carry the violence. Hitchcock never showed the knife entering.

AI tools permit that film language. The limitation becomes serious only where the visible consequence is itself the information: a medical procedure, a war-crime reconstruction, forensic material, an anti-war image, body horror built on transformation.

5. Political content — and it is not one company’s quirk

Kling has the most quoted clause. Section 3.1(e) of its terms prohibits using the service, without written consent, “for political campaigning or advocacy or lobbying purposes.” Its India addendum goes further: nothing may threaten India’s unity, integrity, defence, security or sovereignty, disturb friendly foreign relations — or “insult any other nation.”

But Kling is not an outlier, and treating it as one is a mistake.

Pika prohibits “political campaigning or lobbying purposes, including any manipulation or attempted manipulation of governments or elections.”

Midjourney is blunt: “Do not generate images for political campaigns or to try to influence the outcome of an election.”

PixVerse restricts campaign and election-influence use in its terms.

So political restriction is a genuine product category, not a footnote — and it is easy to miss because it lives in the terms of service rather than on the generation screen.

There is a real distinction inside it, between political expression and campaign use. Dreamina’s guidelines take the other path: they explicitly allow “respectful expression of different viewpoints” and carry public-interest, satirical, fictional and counterspeech exceptions, while still restricting deception.

For documentary makers, NGOs and political filmmakers the first question is therefore not “can the model draw this?” It is “does the contract let me use this service for this purpose at all?”

6. “Objectionable” — the rule you cannot pre-clear

After the specific categories, most platforms keep a residual clause broad enough to swallow the rest of the rulebook.

Kling prohibits whatever it considers objectionable “in our sole judgment.” Pika reserves discretion over content that “contradicts our values or is objectionable,” and states that its list is not exhaustive. Runway’s terms say content may not contain nudity or violence “in the Company’s sole discretion.” Midjourney puts it as plainly as it can be put: “an input not being automatically blocked does not necessarily mean that it is allowed.”

Runway also adds a restriction with no harm rationale at all: you may not generate work in the style of a living artist. That is an aesthetic rule wearing a safety policy’s clothes.

Into this bucket falls everything you cannot predict — a disgusting but non-violent scene, bodily deformity, illness, an animal dying, a religious object in an unusual setting.

And when it refuses, there is no route to a human. Runway states both halves of this in its own support documentation:

That is more concrete than complaining about censorship in the abstract. Runway does publish an appeal route for account suspensions, so it is not true that no appeal exists at all. What does not exist is any operational way to tell the moderation stack: this account is making a verified documentary; please assess this scene under a different standard.

The production consequence is budget, not philosophy. If a scene sits near a boundary you cannot plan from the written rule alone. You need testing time, a second visual treatment, and on paid work a second provider.

7. The written exception that cannot always be reached

Two policy families deserve credit for writing context into the rules at all.

Dreamina has the clearest exception structure in the sector. Both its nudity rule and its violence rule carry the same carve-out: exceptions may be allowed for “educational, documentary, scientific, or artistic content, satirical content, content in fictional settings, counterspeech.”

Google ends its list of prohibitions with an exception for educational, documentary, scientific or artistic purposes, or where benefits substantially outweigh harms.

That is the written layer. The technical layer can still be stricter, and Google is the one case where we can trace it all the way down.

Video generation on Google’s stack exposes a personGeneration control with three values: dont_allow, allow_adult, allow_all. A scene containing a child — a family scene, a school documentary, a grandparent with a grandchild — needs allow_all. Google’s own developer documentation states:

The permitted values vary by model, so this is not a blanket “everyone gets the strictest setting.” But the regional limit is real and it is absolute: in the EU, the UK, Switzerland and the MENA region, allow_all cannot be selected. Not for more money, not with an explanation of the project. The control also lives in the developer tooling; the consumer app has no such switch.

So the strictness a filmmaker meets depends on which product tier they pay for and which country they sit in — not on what they are making.

A written exception matters only if some workflow exists for invoking it.

8. One company measured its own over-refusal

Sora is gone, but its safety documentation is still the best evidence anyone has published about how these systems actually behave.

The Sora 2 System Card (30 September 2025) reported a metric called not_overrefuse — the share of legitimate requests the safety stack correctly allowed.

On self-harm-adjacent material — the category that includes clinical, journalistic and dramatic treatment of suicide — roughly one legitimate request in twenty was refused.

This is not evidence that Sora was unusually bad. It is evidence that over-refusal is measurable, and that one lab was willing to publish the number. Every other discussion of creative moderation has the same evidence problem: users report inconsistent blocks, companies cite safety, and nobody publishes the confusion matrix.

Treat these figures as a historical benchmark, not as a tool you can pick today.

9. The fourth gate: the licence

A clip can clear every filter and still be useless, because the contract decides what you may do with it. This part of the market is more varied than the safety policies — and it runs in the opposite direction to most people’s assumptions.

Kling contains a remarkable pair of clauses. Section 4.4: “You own all intellectual property rights… We do not claim ownership of the Content.” Section 4.6: without written permission you “may not use, reproduce, distribute… the Output for any commercial purposes.” Section 4.5 additionally requires Kling branding on the output absent that permission. The reading is not “Kling owns your film” — it is stranger. The contract can acknowledge your ownership and separately restrict what you may do with it.

Luma ties it to the tier you generated on. Free and Lite outputs are personal, non-commercial only — and that status is permanent: “Generated content includes a watermark that cannot be removed, even if you later upgrade your subscription.” On those tiers Luma may publicly display, reproduce and distribute your work and use it to train models. On Plus and above it may not.

Hailuo (MiniMax) is the correction to a claim that circulates widely. MiniMax’s general app terms contain broad non-commercial language that is easy to quote out of context, but the Hailuo video subscription terms are specific: paid subscribers “retain any and all intellectual property rights to such content, including the right to use it for commercial purposes.” Free downloads carry a watermark. Hailuo should not be listed as “no commercial use.”

Runway, the strictest company in this article on content, is the most generous on rights: “the content you create using Runway is yours to use without any non-commercial restrictions from us” — on any plan, “Free, Standard, Pro, Max, or Unlimited.”

That last inversion is the useful finding. The company that will not let you generate a wound will let you sell what it did generate, from a free account. For an independent filmmaker that contractual difference can matter more than ten points of model quality.

10. The 2026 matrix

Written policy, not a prediction of runtime behaviour.

11. Where a filmmaker can actually work

If the scene can be told without graphic anatomy, almost every major system gives you room, and the choice becomes a quality question rather than a policy one — consistency, camera control, price.

If context is the reason the scene exists, start with the two policy frameworks that acknowledge context in writing: Dreamina’s, which has the most explicit exception language in the market, and Google’s, which is the one traceable into the runtime — with the regional limits in section 7. Test the exact product, region and model before you commit a sequence.

If you work through an aggregator, the boundary moves when you change models. Higgsfield states this more clearly than anyone: “content policies may vary depending on the model being used” — moderation is applied at model level, so the model you pick sets your limit. The same applies to Pollo AI and Pippit AI, which resell several vendors’ models. That makes aggregators useful for production resilience — but not magic: the upstream model’s restriction travels with the model, and the platform can add a layer of its own.

If violence is the point of the scene, PixVerse‘s degree wording and Adobe’s glorification test are both wider starting points than Runway’s anatomical list.

If the project is commercial, read the licence before the safety policy. A clip Kling generates happily still needs written permission for commercial use. A Runway clip is commercially usable from a free account. A Hailuo paid-plan clip carries express commercial rights. A Luma Free or Lite clip never will, no matter what you upgrade to afterwards.

If the visible wound, the nude study or another restricted depiction is essential to the work, no cloud subscription solves it. The route that does is a downloadable model running on your own computer — which removes the remote classifier and nothing else. That is the subject of section 12.

12. Running it on your own computer

What “open-weight” really means, what a 24 GB card gets you, and why “downloadable” is not the same as “allowed”

Most mainstream commercial AI video tools — Runway, Kling, Veo, Hailuo, PixVerse and Midjourney — are cloud services. You type a prompt into a website, the request travels to the company’s servers, its moderation system inspects it, and the generation runs on hardware the company owns.

There is another route, and it is neither exotic nor a grey market. Several major developers publish their trained model weights for download. You install the model on your own machine and generate on your own graphics card.

The difference is easiest to see as a pipeline.

A cloud generator works like this:

prompt → platform moderation → provider’s servers → video model → output

A genuinely local workflow works like this:

prompt → your computer → local video model → output

That is the part that matters. If the whole workflow is local, no remote classifier decides, prompt by prompt, whether the generation may start.

It does not mean there are no rules. Running locally does not remove your legal obligations. The model licence still applies. The model itself may have been safety-tuned during training. One specific gate disappears — the need for a service provider’s permission for every shot — and that is all.

What “open-weight” does not mean

The term gets used far too loosely, and the confusion costs people money.

An open-weight model is one whose trained weights you can download and run yourself.

It does not automatically mean:

• that the source code is open;
• that there is no licence;
• that any use is permitted;
• that commercial use is unrestricted;
• that the developer claims no rights and imposes no conditions;
• or that the model has no safety training or built-in limits.

Open source, open-weight, local and uncensored are four different words. They are not synonyms, and the gap between them is where people get into trouble.

The only claim made here is the narrow one: a model genuinely running on your own machine does not need a central platform to approve each generation.

What you gain

1. No per-prompt approval

This is the whole point.

Runway states in its own support documentation that content moderation “cannot be disabled for an account, project, or topic by the support team,” and that it is “unable to allowlist specific accounts or subject matters that are being content moderated, regardless of the intent of your input or final project.”

That is a reasonable position for a service operating at scale. It is also a hard ceiling for anyone whose work lives in the categories mainstream classifiers moderate broadly: war injuries, historical atrocity, medical scenes, body horror, artistic nudity, politically sensitive reconstruction, documentary material.

A local model has no equivalent gatekeeper. You are no longer asking an unknown classifier, scene by scene, to infer whether your context is legitimate before it will let the model run.

2. The workflow can be genuinely private

If every component runs locally, your prompts, storyboards, reference photographs and unreleased material never reach a video-generation company.

For professional production that can matter on its own, independently of any content question.

One caveat that catches people: a local interface can still call external services for prompt enhancement, image analysis or cloud upscaling. “It runs in ComfyUI” does not by itself mean everything stayed on your machine. The whole pipeline has to be local for that to be true.

3. You stop paying per attempt

Cloud services meter generation — credits, subscription allowances, API billing.

A local model does not charge you again for the seventy-third attempt at the same shot. The cost simply moves: GPU, electricity, time. If you want a hundred variations, your hardware just keeps working.

4. You can build a real pipeline

A local model is not limited to the controls a web interface happens to expose. Depending on the model and ecosystem you can add custom LoRAs, character references, pose and depth control, fixed seeds for reproducibility, your own upscaler, frame interpolation, video-to-video chains, separate audio generation and batch runs.

At that point you are not using an “AI website.” You are building a generative post-production pipeline.

You do not need a command line

The most widely used interface in this ecosystem is ComfyUI, a visual node-based system that runs in a browser window on your own machine. A workflow is assembled as a chain of boxes: load model → enter prompt → add reference image → generate → upscale → save video.

Wan 2.2 and several other video families ship with official or community ComfyUI integrations. So a local setup does not have to look like a research lab. You describe the scene, set the parameters, press Generate — and your own graphics card starts calculating frames.

The catch: video generation is memory-hungry

This is the real barrier, and it is not about which GPU generation you own. It is about VRAM — the memory on the card itself.

The two tricks that move that floor

Offloading keeps part of the model outside GPU memory — usually in ordinary system RAM — and moves pieces back when needed. It lowers the VRAM requirement and slows generation.

Quantisation stores and computes the weights at lower numerical precision, such as FP8 instead of a larger format. It can cut memory use substantially, sometimes at a cost in speed, quality or model behaviour.

This is why you will find two flatly contradictory statements about the same model:

Both can be true. One person is describing the native high-precision implementation; the other a quantised workflow with aggressive offloading. When you read a hardware claim, always ask which of the two it is.

The models worth knowing

Wan 2.2 (Alibaba) — the clearest starting point

Alibaba publishes weights and inference code for the Wan 2.2 family. The TI2V-5B variant handles both text-to-video and image-to-video and outputs 720p at 24 fps. The project says it can run on consumer cards such as the RTX 4090, and separately reports a five-second 720p generation in under nine minutes on a single consumer-grade GPU without specific optimisation. Larger A14B variants need considerably more.

It is integrated into ComfyUI, and — unusually — the models are released under the Apache 2.0 licence. That combination of a documented consumer-GPU route and a genuinely permissive licence makes Wan 2.2 the easiest honest answer to “what does running it yourself actually look like.”

LTX-2.5 (Lightricks) — the one built for production workflows

LTX is more interesting than a text-to-video demo because the ecosystem reaches into editing. The project documents text- and image-to-video, keyframe interpolation, audio-driven video, video-to-video and targeted regeneration, plus FP8 quantisation for a lower memory footprint.

The trade-off is the licence. LTX-2.x ships under a community licence with use conditions rather than a plain permissive one, and organisations above a defined size can need a separate commercial arrangement. Read the LICENSE file before you plan a paid project around it. Removing the platform moderation gate does not remove the contractual gate.

HunyuanVideo 1.5 (Tencent) — technically attractive, but not licensed for EU use under its public licence

On hardware, this is one of the friendlier options: the official project documents a low-memory path at around 14 GB with model offloading.

For a European filmmaker there is a decisive catch. The licence opens with:

Its Territory is defined as “the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea,” and using the model outside that Territory is “unlicensed and unauthorized.”

The file downloads perfectly well. That is not the question. HunyuanVideo 1.5 is the cleanest demonstration in this whole field that technical availability and legal usability are different things.

CogVideoX (Zhipu) and Mochi 1 (Genmo) — still relevant

Both remain useful, and both illustrate the memory point.

CogVideo’s ecosystem supports a wide range of memory-saving techniques — quantisation, CPU offloading, Diffusers integration — which makes it approachable on mid-range hardware.

Mochi is heavy in its native form; Genmo’s own repository describes roughly 60 GB for the native single-GPU implementation while pointing to much lighter ComfyUI paths. It is the textbook case of the gap between an official requirement and what community optimisation achieves.

Are these models actually “uncensored”?

Not necessarily, and the distinction is worth getting right.

A model may have been safety-tuned in training. It may simply be bad at certain concepts. A custom node or third-party interface can add its own moderation. An optional cloud prompt enhancer can send your text straight back out to an external service.

The accurate claim is not that a local model is rule-free. It is this:

When the model and the full pipeline genuinely run on your machine, each generation no longer needs permission from a central moderation server.

That is narrower than “uncensored” — and far more useful, because it is actually true.

What you give up

The freedom is not free. In the cloud, the provider buys and maintains the GPU cluster, handles the drivers, optimises the model and hides the engineering. Locally, some of that becomes your problem.

Expect an expensive GPU, tens or hundreds of gigabytes of model files, CUDA and dependency trouble, slower generation on consumer hardware, a more complicated install, a different licence for every model — and output that will not always match the strongest closed systems.

What you get in exchange is something no subscription offers: control over the pipeline. No platform decides that the wound in your documentary is too graphic. Your credits do not run out on iteration 73. A server-side classifier cannot change overnight and make yesterday’s shot impossible to reproduce today.

And the responsibility moves with it

Consent, likeness and privacy law, copyright, deepfake statutes, AI-disclosure duties under the EU AI Act, and the model’s own licence all still apply to what you make on your own machine. In Europe, Article 50 of the AI Act has been applying since 2 August 2026, and it does not care where the video was rendered.

Removing the platform filter does not remove the law.

It changes who makes the first decision — and that is the whole argument for this route. Not that nobody is deciding. That a person is.

13. The point

Nobody serious argues these platforms should have no rules. They plainly need them around non-consensual sexual imagery, child exploitation, impersonation and fraud — and in 2026 the law finally arrived in exactly those places.

The interesting question is what happens after the obvious cases.

There the industry fragments completely. One company bans adult nudity as a visual category; another writes around purpose. One prohibits exposed anatomy; another says “excessive”; another “gratuitous”; another “glorifying”; two attach artistic exceptions. Four ban campaign use. Several keep an “objectionable” clause broad enough to cover everything else. One lets you own an output while forbidding you to sell it; another gives full commercial rights on a free account.

Seven standards, one legal environment. “The law made us do it” does not explain that spread — and the EU AI Act, which had every reason to be maximally cautious, still wrote a carve-out for artistic, creative, satirical and fictional work that most platform policies never bothered to write.

The underlying problem is not malice. Classifiers are very good at spotting a naked body, blood, a child, a face, a weapon. They are much worse at deciding whether the scene is pornography or anatomy, propaganda or history, sadism or an anti-war film.

The question for 2026 is no longer whether AI video needs safeguards. It is whether the safeguards can learn to recognise context without making context impossible to express.

Sources and method

This article compares written policies, published technical documentation and licence terms. Runtime behaviour can differ from written policy and changes faster than terms pages, so we do not claim a particular prompt will pass unless the provider documents it.

We also did not use the largest category of writing on this subject. Search any question here and the first page is dominated by companies selling unfiltered generation, whose “independent comparisons” conclude without exception in favour of their own product. Two widely repeated claims from that literature did not survive checking: that Runway silently rewrites prompts behind the user’s back — its own changelog documents only a user-initiated, one-click prompt enhancer in text-to-image, from May 2025 — and an alleged 2026 tightening at Kling said to false-flag swimwear and dance footage, for which no primary source exists at all.

Law and regulation

• FTC — Complying With the TAKE IT DOWN Act
• European Commission — Transparency obligations under Article 50 of the AI Act
• California Attorney General — January 2026 investigation into xAI

Platform policies and documentation

• Runway Usage Policy · Runway usage rights · Runway content-moderation support documentation
• Kling AI Terms of Service
• Pika Acceptable Use Policy (29 November 2024)
• Midjourney Community Guidelines
• PixVerse Community Guidelines
• Adobe Generative AI User Guidelines (15 May 2026)
• Dreamina Community Guidelines
• Google Generative AI Prohibited Use Policy · Google Veo developer documentation
• OpenAI Sora 2 System Card · OpenAI Sora discontinuation notice
• Hailuo AI video subscription terms
• Luma Dream Machine licensing guide
• Higgsfield Trust & Safety
• ByteDance Seed — Seedance 2.5

Open-weight models and licences

• Wan 2.2 official repository · LTX-2 repository and licence · HunyuanVideo 1.5 licence
• Mochi 1 official repository · CogVideo official repository

Policies, documentation and model licences checked 30 August 2026. Open-weight licence terms change between releases — check the LICENSE file of the exact version you download, especially before commercial use.

These services change quickly. Re-check any clause that governs a live production, a commercial licence or a real-person workflow.