How to Tell If a Drama Is AI-Generated: 8 Signals

Updated August 19, 2026 · Written by the AI Drama Reviews editorial team · 31 min read

DramaBoxEditor's pick — biggest short-drama library, lowest price to go unlimited 9.8Excellent

• Thousands of series, roughly 200 new titles every month
• Free episodes daily, plus ad-unlocks — no card needed to start
• Unlimited plans from about $5.99 a week, the lowest entry price we found
• Runs on Android, iOS and in a desktop browser

Go to DramaBoxAffiliate link · free to download · in-app purchases available

Affiliate link. We may earn a commission if you install through it, at no extra cost to you — it never changes our scores or placements. How we test · Disclaimer

How can you tell if a drama is AI-generated?

An AI-generated vertical drama gives itself away through consistency failures between shots, not through overall picture quality. A generated series can look expensive. What it cannot yet do is hold one person, one room and one piece of clothing perfectly stable across a cut, or sustain a single take past the model's generation ceiling.

That is the whole method in one sentence, and it inverts the instinct most viewers arrive with. People expect to spot AI by looking for ugliness. Ugliness is a budget signal, not a production-method signal. The reliable evidence is structural: things that should be identical from one shot to the next, and are not.

Applied in order, the eight signals below run from a hard technical constraint that is very difficult for a producer to hide, down to a statistical prior that only shifts the odds. Start at the top. Stop when you have three.

Why do apps not label their AI dramas?

No app in this category applies a visible AI label to individual titles in its US catalogue, and no independent dataset measures what share of any catalogue is AI-produced. Those are two separate gaps, and together they are the reason this page exists. If a badge existed, the eight signals would be a curiosity. It does not, so they are an instrument.

The absence is not accidental. Labelling an AI title in a market where a vocal minority of viewers treats "AI slop" as a reason to uninstall is a commercial risk with no offsetting reward, and no US regulation compels it. China took the opposite route: mandatory labelling of AI-generated content has been in force there since March 7, 2025. Nothing equivalent applies to a US app store listing as of August 19, 2026.

Company-level statements exist and are worth something, but they describe libraries and not episodes. FlexTV's vice-president Tang Tang confirmed to MIT Technology Review on May 15, 2026 that the company had moved its North American production to AI, cutting costs 80 to 90 percent and compressing schedules from three or four months to under one with teams of about ten people. That is a vendor describing its own results, and it tells you a great deal about FlexTV's shelf and nothing about the specific title you opened.

What labelling exists today, and where it stops

The volume behind the question

The reason this stopped being a niche concern is arithmetic. The China Netcasting Services Association reported that more than 95% of new Chinese microdramas in the first quarter of 2026 were AI-generated. DataEye counted roughly 221,900 AI dramas published on Douyin in the first half of 2026, about 1,200 a day, against a January 2026 rate of roughly 470 a day. In April 2026 alone Douyin carried about 44,200 AI titles against 3,248 live-action ones.

China is the only market where this output is measured. It is also the market that supplies a large share of what US apps license, dub and re-release. The proportion arriving in American catalogues is not published by anyone, which is the gap this page works around rather than fills.

The same data carries a second lesson worth holding on to while you watch. Of those 221,900 titles, about 1,055 — 0.47% — passed 100 million views. Cheap generation did not make hits common. It made misses cheap, and it flooded the shelf you are now browsing.

Signal 1. Why do shots never run past 10 seconds?

Time the longest uninterrupted shot in the episode. If nothing runs past roughly 10 to 12 seconds without a cut, an insert, a whip pan or something crossing the foreground, you have the strongest single signal available to a viewer. This is the one to check first, and the one to trust most.

Why the technology produces it. A scene in an AI drama is not filmed, it is assembled from multiple short generations. The generation length is a hard property of the model. ByteDance's Jimeng produces about 12 seconds per pass, or 30 seconds with motion templates. Google's Veo 3.1 generates 4, 6 or 8 second clips on its Lite and Fast tiers and 8 seconds on Quality. Shots in this category rarely exceed 10 to 12 seconds, and that is a model constraint rather than a stylistic decision.

How confident it should make you. Very. A director can choose to cut fast; a director cannot choose to make a model hold coherence for 25 seconds. The asymmetry is what makes this signal useful: a long unbroken take of a person talking is close to proof of filming, while relentlessly short shots are strong evidence of generation. Watch specifically for shots that end on a cut you did not need. That is a seam.

The generation ceilings you are actually measuring

What a filmed vertical drama looks like by comparison

Live-action microdrama is also cut fast, because the format demands a hook in the first three seconds and a reveal every forty. The difference is that it can afford to stop. A filmed two-hander will let one performer finish a 15 or 20 second speech in a single frame when the performance carries it, because that costs nothing extra once the crew is standing there. In a generated production that shot has to be stitched, and stitching leaves a cut.

Signal 2. What does identity drift look like?

Pause on the last frame before a cut and the first frame after it, then compare the same character's face. Jaw width, eye spacing, nose length, apparent age and shoulder proportion should be identical. In a generated series they frequently are not. Identity drift is the second-strongest signal and the one most viewers feel before they can name it.

Why the technology produces it. Every shot is a separate generation conditioned on reference images rather than a recording of one physical person who existed continuously between takes. The model reconstructs the character each time. Producers fight this with locked character bibles, fixed seeds, first-and-last-frame controls and shot-level human review, and those mitigations reduce drift without eliminating it. An ACM survey published May 18, 2026 treats spatiotemporal consistency as the central open problem in video generation, not a solved one.

How confident it should make you. Very high, because live-action has no mechanism that produces it. Bad lighting changes how a face reads; it does not change the distance between the eyes. If you can see the geometry move, you are looking at reconstruction.

Where to look on a face

Faces are easiest to check at the edges. Hairlines shift. Ears change size or lose an earring. The bridge of the nose lengthens by a few pixels. Teeth are a reliable tell because the model rarely holds a consistent arrangement. Age drift is the most striking version: a character can look five years older in a reverse angle for no narrative reason. We take the mechanism apart in why AI drama characters change appearance.

Signal 3. Do the lips match the dialogue?

Watch only mouths for thirty seconds and ignore everything else. Look for lips still moving after the line has ended, consonants landing in the wrong place, or a mouth shaping something other than the words you hear. Lip drift is a strong signal in this category, and it carries one large caveat.

Why the technology produces it. Dialogue in an AI drama is usually generated or replaced separately from the picture, then married to it in the edit. There is no physical performance for the audio to be a record of, so alignment is a post-production achievement rather than a fact. Lip movement out of sync with dialogue sits on the standard list of failure modes for generated video alongside identity drift and object impermanence.

How confident it should make you. High for the visuals when it appears alongside signals one and two. On its own, much less, because there is a second and completely different production route to the same symptom.

Dub drift and generation drift look identical

A drama filmed in Chinese with human actors and dubbed into English will show lip movement that does not match the English words, for the obvious reason that the performer never said them. That is not an AI-generated drama. It may not even involve a voice model. This is the single most common way viewers reach a wrong conclusion, and it is why the AI-touched versus AI-made section below is not optional reading.

Signal 4. What is wrong with the hands and the signs?

Pause on any frame with a hand near the camera and count the fingers. Then read every sign, poster, phone screen, newspaper and book spine in the shot. Deformed hands and pseudo-lettering are the classic artefacts, and when they appear they are close to conclusive.

Why the technology produces it. Both are cases of the model rendering something that looks statistically like the right texture without the underlying structure. Hands articulate in ways that are hard to reconstruct from reference, so fingers fuse, duplicate or bend backwards. Written language is a discrete symbolic system that a pixel model approximates rather than spells, so background text becomes unreadable pseudo-letters that carry the shape and rhythm of writing without being any word.

How confident it should make you. High when present, near zero when absent. This signal has almost no false-positive rate and a very high false-negative rate: producers know about it, so they avoid close-ups of hands, keep signage out of frame and blur backgrounds. Finding it settles the question. Not finding it settles nothing.

Hands

The productive moments are gestures, handovers and anything held: a cup passed between characters, a phone raised to an ear, a hand on a door. Count fingers on a frozen frame rather than at speed, because the eye smooths the error in motion. Watch also for hands that pass through objects they are supposed to be gripping.

Background text

Signage is the richest hunting ground because AI dramas love establishing shots of streets, offices and hospital corridors, all of which are covered in writing. Look at the second and third plane rather than the hero prop, which will have been fixed. Text on a screen inside the shot — a phone, a laptop, a television — is the most common failure of all, because it is small, high-contrast and rarely worth an editor's time.

Signal 5. Wardrobe and prop continuity

Pick one earring, one necklace, one tie, one coffee cup or one folder, then track it through every shot of a single continuous scene. Items that change, move across the body, appear or vanish with no scene break are signal five. Medium weight, and easy to spot once you are looking.

Why the technology produces it. Loss of object permanence is a documented failure mode of generated video, and it has the same root as identity drift. The model is not photographing a room that persists between shots; it is regenerating a plausible version of that room. Clothing, jewellery, hairstyle, age and body proportions changing between shots all appear on the standard failure list, and the mitigations are the same partial ones: reference images, a locked character bible, shorter shots, continuity editing.

How confident it should make you. Medium, and no higher, because this is exactly where cheap live-action microdrama fails too. A production shooting a 70-episode season in a fortnight with no dedicated continuity supervisor will produce jumping wardrobe by ordinary human error. What raises the weight is severity and rate: a necklace that changes design twice inside one conversation is not a script supervisor having a bad day.

Signal 6. Why do the same faces keep appearing?

Open a second series by a different production in the same app and look at the leads. If two unrelated titles appear to be cast from the same small pool of people, that is signal six. Medium weight, and it only becomes available after you have watched several titles.

Why the technology produces it. Models trained on similar data produce similar people, which is why different series can feel cast from the same faces. Ask a model for a wealthy male lead in his early thirties and it converges on a narrow region of face space. Two productions using the same family of models, prompting with the same genre vocabulary, will land near each other without any coordination.

How confident it should make you. Medium, and it depends on what you know about the catalogue. Real drama catalogues do not share casts. Every platform in this category finances its own titles and none licenses to another, so a face genuinely appearing in two unrelated series is not a shared actor. It is a shared model. The weakness of this signal is that recognising a type is subjective, and human casting also runs to type.

Signal 7. Physics and screen direction

Watch for movement that cannot happen and for characters swapping sides of the frame between shots of the same conversation. Physics violations and reversing screen direction are both on the documented failure list for generated video, and both are visible without pausing.

Why the technology produces it. The model has learned what motion looks like, not what mass, momentum and gravity are. Feet slide fractionally rather than planting. Fabric settles wrongly. A thrown object decelerates like it is underwater. Screen direction fails for a structurally different reason: continuity of spatial relationships across a cut is exactly the long-range consistency problem the research literature identifies as unsolved, and a model generating each shot independently has no committed geography to violate.

How confident it should make you. Medium. Action beats are where it shows, which makes fight scenes and chases the best place to look. The reason it is not higher is that low-budget live-action also reverses screen direction constantly, usually because a scene was shot from whichever side the crew could stand on. Weightless motion is the more specific of the two.

Signal 8. Which genres is AI actually used for?

Fantasy, xianxia, werewolf and superpower titles are far more likely to be AI-produced than contemporary human drama, because that is where generation works. This is a prior, not a verdict. It should adjust how carefully you apply signals one to seven, and it should never be the thing you conclude from.

Why the technology produces it. Two forces point the same way. Stylisation hides artefacts: in a world of flying swords and shifting wolves, a slightly wrong hand or an impossible movement reads as effect rather than error. And effects that would destroy a microdrama budget in live action are nearly free in generation, which turns the genre's biggest cost into its cheapest asset. Contemporary human drama is where generation is weakest, because it depends on sustained micro-expression and unbroken performance.

How confident it should make you. Low on its own. Plenty of fantasy microdrama is filmed, and the reviewer Jenny Cooper, who watched 25 AI titles in the run-up to June 24, 2026 and finished three of them, concluded that fantasy was the strongest AI category and that story remained decisive regardless. Use genre to decide how hard to look, not what to believe. Our full breakdown of which genres suit generation is in the AI drama genre guide.

The genres where AI is currently rare

Office romance, family melodrama, courtroom and medical stories, and anything built on two people talking across a table are poor candidates for generation. So is any series whose selling point is a specific recognisable performer. If you are watching a contemporary US-set drama with a named English-speaking cast, the prior runs strongly the other way, and you should demand more from signals one and two before concluding anything.

Which of the eight signals are most reliable?

Reliability and visibility are two different properties, and the most useful signals are the ones that score well on both. Shot length wins because it is a hard technical constraint that anyone can measure with a clock. Background text is nearly conclusive but frequently absent. Genre is always available and proves nothing.

Where in production each artefact originates

The ranking makes more sense against the production line. Signals one, two, four, five and seven are born at stage four, shot generation, where the model constructs each clip independently. Signal three is born at stage five, voice and sound, where dialogue is married to picture that was never recorded with it. Signal six is born upstream at stage three, in the character bible and the model chosen to realise it. And every one of them is either repaired or left in place at stage six, the edit, which is why two productions using identical tools can be wildly different to detect.

How to combine signals without fooling yourself

Our working threshold is three independent signals, with at least one of them being shot length or identity drift. Two signals means probable and worth saying so out loud. One signal means nothing, because every individual signal on this list has an innocent explanation available to it. The threshold matters more than the individual observations, because the failure mode of this whole exercise is confirmation: once you suspect AI, everything looks like evidence.

The two-minute test you can run on any episode

This is the whole method as a procedure. It takes about two minutes on one mid-season episode and needs nothing but the app you are already using and the ability to pause.

1. Pick an episode from the middle of the season. Open episode 8 to 15 rather than episode 1. Opening episodes get the most human editing attention and the largest share of the budget, so they hide artefacts best. Mid-season episodes are produced at volume and show the pipeline more honestly.
2. Watch 30 seconds with the sound off and time the longest shot. Count one-thousands between cuts. If nothing survives past roughly 10 to 12 seconds without a cut, an insert, a whip pan or a foreground wipe, mark signal one. If a single unbroken take of a person speaking runs 20 seconds or more, stop: this is almost certainly filmed.
3. Rewatch the same 30 seconds with sound and watch only mouths. Ignore the acting and look at where consonants land. Look for the mouth still moving after the line ends, or lips shaping something other than the words. Mark signal three, but hold it lightly because an AI dub over live-action footage produces the same effect.
4. Freeze on two consecutive shots of the same face. Pause on the last frame before a cut and the first frame after it. Compare jaw width, eye spacing, nose length and apparent age. A filmed performer is geometrically identical across a cut. A generated one often is not. Mark signal two.
5. Scan the edges of the frame for hands and text. Pause anywhere with a hand near the camera and count fingers. Read every sign, poster, phone screen and book spine in shot. Pseudo-letters that look like writing but resolve into nothing are a strong marker. Mark signal four.
6. Follow wardrobe and props through one continuous scene. Pick one earring, one necklace, one cup or one folder and track it across every shot of a single scene with no time cut. Items that change, move or vanish without a scene break mark signal five.
7. Note the genre, then open a second unrelated title. Fantasy, xianxia, werewolf and superpower titles carry a much higher prior than contemporary drama. Then open a different series by a different production and look for a face that feels cast from the same pool. Mark signals six and eight.
8. Count the signals and read the verdict. Three or more signals with shot length or identity drift among them is a strong case for AI production. Two is probable. One is nothing. Whatever the count, the result is evidence and not proof, because no app confirms or denies.

Scoring the result

One practical note on step one. Choose the middle of the season deliberately. Opening episodes carry the marketing budget and the editing attention, because the first three to ten episodes are free and their job is to sell the coin wall. Episode twelve is where the pipeline shows.

When does this test give you the wrong answer?

This is the section most detection guides skip, and it is the one that decides whether the rest of the page is useful or harmful. Every signal above has an innocent explanation, and vertical microdrama happens to be a format in which the innocent explanations are unusually common.

Compression artefacts are not AI artefacts

This deserves its own paragraph because it is the most common error we see readers make. Vertical dramas are delivered to phones over variable connections at aggressive bitrates. The result is banding in gradients, blocking in dark scenes, smearing on fast pans and a general softness that looks synthetic if you are primed to see synthetic. It is not. Compression is a uniform, motion-dependent degradation of the entire frame and it disappears the moment the picture holds still. Generation failures are local, structural and stable: they survive the pause button. That single difference separates the two categories cleanly.

One signal proves nothing

We will say this a third time because it is the rule that keeps the method honest. A single observation, however striking, is not a finding. The genuinely hard case is not the obvious AI fantasy title. It is the cheaply filmed contemporary drama, shot in five days, dubbed by a voice model, cut fast because the format demands it, with a continuity error in episode nine. That title will trip four signals and contain no generated footage at all. If your method cannot get that case right, it is not a method.

Is an AI-dubbed drama the same as an AI drama?

An AI voice on a human performance is not an AI-generated drama, and most of what US viewers encounter on mixed-catalogue apps is exactly that. Getting this distinction wrong is how a reasonable person concludes that nearly everything is AI, which is both false and unhelpful.

The mechanics are simple. A large share of the US short-drama catalogue consists of series filmed conventionally, usually outside the United States, then localised for an English-speaking audience with dubbing, subtitles and cultural adaptation. The dub may well be generative AI. The visuals are a photographic record of real people who were physically present. Nothing about that footage was generated.

This matters practically, not just semantically. If your objection to AI drama is that synthetic performance replaces actors, an AI dub raises that concern for voice work and not for the performance on screen. If your objection is to generated imagery, an AI dub is irrelevant to you. And if you are simply asking which titles look wrong, the dub is a red herring that will send you to the wrong answer.

The spectrum, from localised to fully generated

The definitional groundwork for these five positions, and why the category label is contested, is in what AI dramas actually are.

Which apps actually carry AI-generated titles?

The fastest way to raise or lower your prior is to know whose shelf you are standing in front of. Company-level statements are public even though title-level labels are not, and they are the closest thing to disclosure this category offers.

FlexTV is the clearest AI-first platform in the ranking, on the strength of its own vice-president confirming the move of North American production to AI. StoReel is AI-native by construction: it raised $34 million and has stated a target of 100 AI dramas a month, and its StoReel Canvas tool lets viewers build episodes themselves. MoboDrama positions itself as a fully AI catalogue, which is a claim rather than a verified fact and the smallest-scale operation among the apps we tested. MyMuse runs an experimental AI-produced slate where quality varies widely by title. Character.AI Series, launched July 2026, is AI-native and short.

On the other side, ReelShort's flagship slate is filmed in English with US casts and its catalogue is mostly live-action, with some titles marked "Dubbed". DramaBox is mixed: mostly licensed and dubbed live-action with a growing AI shelf and roughly 200 new titles a month, which means the app we earn commission from is also the one where you will most often need this test. ShortMax, DramaWave, GoodShort, VibeShort, FlickReels and NetShort all run mixed or mostly live-action catalogues.

What this test cannot establish

This procedure produces evidence, not proof, and the distinction is not a disclaimer — it is the actual state of the available information. Four things stand between a careful viewer and a confirmed answer, and none of them is going to be resolved by looking harder.

No app labels its titles. There is no authority to appeal to. You cannot check your conclusion against a catalogue field, a press release or a filing, because for individual titles none exists in this category as of August 19, 2026. A confident reading of eight signals and a producer's confirmation are different objects, and only one of them is available to you.

No independent dataset measures any catalogue's AI share. We can tell you that more than 95% of new Chinese microdramas in Q1 2026 were AI-generated, because the China Netcasting Services Association published that. Nobody has published the equivalent for DramaBox, ReelShort or any other US-facing app, and we are not going to estimate one. Anyone quoting a percentage for a US catalogue is making it up.

The signals decay as models improve. Shot length is the most fragile: Kling already generates up to about three minutes, and ByteDance's Seedance 2.0, released February 2026, was a notable step up in quality. Every release narrows the gap. The consistency problems are being worked on rather than solved — the ACM survey of May 18, 2026, the OpenReview work of March 26, 2026 on narrative and temporal consistency in long video, and arXiv:2510.04999 of October 6, 2025 all treat this as open research — but the direction of travel is one way.

C2PA is a provenance standard, not a detector

C2PA records where an asset came from and what was done to it. That is a chain-of-custody function and it only works when a producer chooses to attach a manifest and every downstream tool preserves it. It cannot examine an unlabelled file and tell you whether a model made it, and it is not a rights registration either. Google's SynthID watermark on Veo 3.1 output is a genuine technical marker, but it is applied by one vendor to one family of models, and no short-drama app gives a viewer any way to read it. Neither mechanism helps you with the episode currently on your screen.

What would actually settle it

A per-title production disclosure in the app, in the same place the cast list would go. That is all it would take, and nothing prevents it except that no US rule requires it and no platform has decided that saying so is worth more than staying quiet. Until then, this page is a workaround for a missing catalogue field, and we would rather be honest about that than sell you certainty we do not have.

Frequently asked questions

The bottom line

Two signals settle most cases: nothing running past 10 to 12 seconds without a cut, and a face whose geometry changes across a cut. If both are present in a mid-season episode, you are almost certainly watching a series that was generated rather than filmed. If neither is present, the other six will rarely rescue the case, because they are the artefacts an editor can most easily remove.

The discipline matters more than the observations. Three signals before you conclude, with shot length or identity drift among them. Two means probable. One means nothing. And before any of that, ask whether you are looking at a human performance with a synthetic voice laid over it, because on mixed-catalogue apps that is the single most likely explanation for the thing that made you suspicious.

What none of this gives you is confirmation. No app labels its AI titles, no independent dataset measures any catalogue's AI share, C2PA is provenance rather than detection, and the technical constraints these signals rest on are loosening with every model release. That is an unsatisfying place to end, and it is where the evidence actually stops. We would rather hand you a calibrated instrument with its error bars printed on it than a verdict we cannot support.

DramaBoxEditor's pick — biggest short-drama library, lowest price to go unlimited 9.8Excellent

• Thousands of series, roughly 200 new titles every month
• Free episodes daily, plus ad-unlocks — no card needed to start
• Unlimited plans from about $5.99 a week, the lowest entry price we found
• Runs on Android, iOS and in a desktop browser

Go to DramaBoxAffiliate link · free to download · in-app purchases available

Affiliate link. We may earn a commission if you install through it, at no extra cost to you — it never changes our scores or placements. How we test · Disclaimer

Read next

Sources

• ACM — A Survey: Spatiotemporal Consistency in Video Generation (May 18, 2026). The basis for treating consistency between shots, rather than picture quality, as the central detectable weakness of generated video.
• OpenReview (March 26, 2026) — narrative and temporal consistency in long video generation. Supports signals 2, 5 and 7 and the statement that these are open research problems rather than solved ones.
• arXiv:2510.04999 — Bridging Text and Video Generation: A Survey (October 6, 2025). Background on the text-to-video pipeline underlying the shot-length constraint.
• ByteDance Jimeng — approximately 12 seconds per generation, 30 seconds with motion templates. Google Veo 3.1 — Lite and Fast at 4, 6 or 8 seconds, Quality at 8 seconds, with SynthID watermarking on output. Kling — up to about 3 minutes. Sora 2 and Runway Gen-4.5 — up to 1080p. ByteDance Seedance 2.0 released February 2026. These are the generation ceilings signal 1 measures.
• MIT Technology Review (May 15, 2026) — FlexTV vice-president Tang Tang on the move of North American production to AI, an 80–90% cost reduction, schedules cut from three or four months to under one, and teams of about ten people. A vendor describing its own results.
• China Netcasting Services Association, via Global Times — more than 95% of new Chinese microdramas AI-generated in Q1 2026. DataEye — approximately 221,900 AI dramas published on Douyin in H1 2026, roughly 1,200 a day, of which about 1,055 (0.47%) passed 100 million views; about 44,200 AI titles against 3,248 live-action on Douyin in April 2026; about 470 AI titles a day in January 2026 via MIT Technology Review.
• C2PA — provenance standard recording an asset's origin and edits. Explicitly not a detector and not a rights registration. China's mandatory AI-content labelling in force from March 7, 2025, cited as the contrast with the absence of any US equivalent.
• 36Kr (March 26, 2026) — StoReel's $34 million raise and target of 100 AI dramas a month. Character.AI — (c.ai) Series launch, July 2026. Platform self-descriptions for FlexTV, MoboDrama and MyMuse are the companies' own positioning statements, not verified catalogue audits.
• Vertical Drama Love, reviewer Jenny Cooper (June 24, 2026) — 25 AI titles watched, three finished, fantasy identified as the strongest category and story as decisive. Cited as one reviewer's observation, not measured data. Director Qingge Gao, via PetaPixel (August 2026), on viewer indifference to AI production.
• Our own testing — shot-length timings, frame-by-frame face comparisons and continuity tracking performed on mid-season episodes across the thirteen apps we track, August 2026. Method: how we test.
• Declared gaps, stated rather than filled: no authoritative dataset shows what percentage of any app's catalogue is AI-generated; no app in this category applies a title-level AI label in its US catalogue; no standardised independent quality benchmark ranks AI-drama seasons.

Affiliate disclosure. Some outbound links on this page, including the DramaBox links, are affiliate links. If you install through one we may receive a commission at no additional cost to you, and it does not change what we write — this page states plainly that DramaBox publishes no AI labelling and is one of the catalogues where you will most often need to run the test yourself. Accuracy notice. Nothing on this page is an accusation against any named title or producer. The signals described are characteristics of generated video, not proof of it, and the honest verdict this method supports is "shows characteristics consistent with AI production" rather than a confirmation. Model capabilities change quickly and the shot-length signal in particular will weaken over time. Verified August 19, 2026.