Aydın Tiryaki & Claude Opus 5.5

1. A year ago, and today

A year ago, when we watched a YouTube video, we looked for traces of AI in a few places: perhaps a few photos had been generated, perhaps a few short clips had been added, and most likely the voiceover had been produced with AI. Today the picture has changed completely. When we watch a half-hour video, there is a very real chance that the image in every one of its 1,800 seconds, whether still or moving, was generated entirely by AI. And some of it is so realistic that telling the real from the artificial has become very hard.

In places we know, we can rely on the details and say, “That’s not what this place looks like; this footage can’t be real.” But in places we don’t know, we approach every scene with suspicion. Sometimes a scene looks genuine and we ask ourselves: Is this scene real, and only what surrounds it artificial? Or is most of it AI, with a few real shots sprinkled in to give us the impression of reality?

Sometimes the start of a video says “AI” or “dubbed.” That is a clue, but an inadequate one. It tells us that AI is present in the video; it does not tell us where, how much, or in what respect. The result is that watching YouTube no longer gives the pleasure it once did; the viewer is forced to work constantly like a detective.

2. What today’s label says, and what it doesn’t

YouTube’s current system rests on a single question the creator answers at upload. If the creator says “Yes,” a label appears; if the creator says “No,” they are deemed to have declared that the content does not meet the disclosure threshold (1). In other words, the label belongs to the video as a whole; it says nothing at the level of a scene or a second.

The threshold is “realism.” Disclosure is not required for content that is clearly unrealistic, animated, uses special effects, or uses generative AI only for production assistance (2). This has a striking consequence: someone who has a language model write their script and then reads it to camera themselves needs no label at all (1).

Some improvements arrived in May 2026. On long-form videos the label moved directly below the player, and on Shorts onto the video itself; for unrealistic, animated or slightly altered content, however, the disclosure still sits inside the expanded description. Even if the creator says nothing, a label is now applied automatically when YouTube’s systems detect significant photorealistic AI use (3).

Voiceover and dubbing are messier still. Although YouTube has moved its image labels to more visible positions, its help pages still keep the auto-dubbing marker in the description (4). And dubbing done with a creator’s own cloned voice requires no AI label at all (5).

There is also a label pointing the other way, designed to prove that footage is real: “Captured with a camera.” But if a video has been edited so heavily that it can no longer be traced to its source, this label does not appear (6). YouTube itself notes that the absence of the label does not mean the video was edited (7). This has a critical consequence for videos in which real footage is cut together with artificial scenes: the chain proving the origin of the real footage breaks in the edit, and the real part is left without proof as well.

Regulation does not close this gap either. Article 50 of the EU AI Act has applied since 2 August 2026 and requires anyone publishing a deepfake to disclose that the content is artificial (8). Yet even the accompanying code of practice allows, for non-live video, a disclaimer placed at the beginning of the video (9). Once again: a single notice for the whole video.

3. The numbers: How much of what we watch is AI?

The honest answer is: we don’t know. YouTube publishes no platform-wide count of AI content, and automated detection is imperfect; so there is no defensible figure of the form “X percent of YouTube is AI” (10). The data we have is fragmentary, and each piece measures something different.

The most widely quoted study comes from Kapwing. It examined the 100 most popular channels in every country, 15,000 channels in total, and found that 278 of them published nothing but low-quality AI content (“AI slop”). Of the first 500 videos shown to a newly created account, 104 were of this kind (11). That corresponds to 21 percent; in the same sample, 165 videos (33 percent) were classified as “brainrot” (12). An earlier analysis by the Guardian found that roughly 10 percent of the fastest-growing channels were AI-generated (13). In children’s content the picture is far worse: in Kapwing’s TikTok study, 97 of the 100 videos examined in the children’s cartoon category were AI-generated (14).

These figures need two caveats. First, Kapwing is itself a company that makes AI-assisted video tools (15). Second, the company states that it counted only clearly AI-generated scripts and voiceovers, so the true share is probably higher (14). Moreover, the measurement counts only low-quality content; carefully made, AI-assisted documentaries fall outside it.

There is data on the creator side too. In December 2025, more than one million channels used YouTube’s own AI tools every day (16). According to an Adobe survey, 71 percent of creators had used AI video tools; the main uses were editing (58 percent), transitions and effects (42 percent), thumbnails (37 percent), music selection (37 percent) and voiceovers (36 percent) (17).

None of this measures the share of AI inside a video. Nobody asks how many of those 1,800 seconds were generated, or how much of the sound is synthetic. And there is a striking inversion: the most heavily used layers (scripting, editing, effects, thumbnails) are precisely the layers that need no label. The labelling system looks at AI not where it is used most, but where it is most visible.

4. The platform’s own AI

It is not only creators who use AI; platforms do too. YouTube acknowledged that it had altered some Shorts without telling their creators, applying deblurring, denoising and clarity enhancement during processing (18). The musician Rhett Shull, who noticed it, compared the Instagram and YouTube versions of the same video and described an oil-painting effect on his face in the Shorts version (19). The real irony was this: Shull complained that the changes made his content look AI-generated (20). The platform’s own processing had made genuine footage look suspicious to viewers. Samuel Woolley of the University of Pittsburgh told the BBC that calling this “machine learning” obscured the use of AI, and that altering content without consent erodes public trust (21).

Auto-dubbing raises the same issue. YouTube’s lip-sync system makes pixel-level changes to the speaker’s mouth (4). So in what is called “just a dub,” the image may have changed as well. This is why marking must cover not only the tools creators use, but the platform’s own tools too.

5. The proposal: A universal marking language

The core of our proposal is this: on social media, whatever is shared, whether a photo, a video or any other content, should carry a marking system that shows the AI contribution at that moment in every respect. Leaving nothing out, second by second, grounded in evidence and open to challenge.

The constitution of this system fits into a single sentence: The mark describes not the uploaded file, but what the viewer is seeing and hearing at that moment. This principle automatically covers the AI a creator used, any processing the platform applied afterwards, the audio track, subtitles and translation the viewer chose, and content generated on the fly during a live stream.

We call it a marking language, and the phrase is deliberately close to “markup language.” In computing, a markup language such as HTML is a structured descriptive language read by machines. The system we propose has two faces in the same way: for the machine, a markup language inside the file describing every second and every layer; for people, a universal sign language on screen that can be understood without knowing any language. Both draw on the same source; one is the on-screen face of the other.

Layers. At every moment, three main layers are marked:

• Content: the text, script and information being conveyed.
• Image: the photo or video on screen. Still and moving images share one layer, because the viewer’s real question is “Is this scene real?”
• Sound: the primary audio of the content. In a documentary this is the narration and speech; in a music video it is the music itself. Background music and sound effects are shown separately in a sub-layer that opens on tap. So a documentary maker who films real footage, speaks in their own voice and writes their own text, but adds a 30-second AI-generated background track, keeps the same main marks, while the track is not hidden either.

As the counterpart to artificial intelligence, we use the concept of “natural intelligence”: content people produce through their own work.

Modular design. The system is designed for every kind of post, not only video; only the layers present in the content are active. A text post on X or a blog shows only the Content dot (one dot); a silent photo shows Content and Image (two dots); a podcast shows Content and Sound (two dots); a video shows all three. The same marking language thus reads the same way on every platform and for every type of content.

States. For each layer we envisage four states: natural, altered (a real recording processed with AI), generated, and mixed (real and artificial elements present in the same frame or at the same moment).

Source. The system also records whose tool the AI came from (the creator’s or the platform’s) and who applied the mark (a creator declaration, platform detection, or cryptographic proof). A platform’s automatic mark is not a declaration but an estimate, and viewers need to know that.

Display. The marks must not spoil the picture, but they must be visible and understandable:

• In a corner of the screen, between one and three small dots depending on the content type; each represents one layer and changes as the scene changes.
• Colour alone is not enough: since around 8 percent of men and around 0.5 percent of women have some form of colour blindness (22), every mark should also carry a letter or a shape.
• Artificial layers should be shown in neutral colours, not in colours such as red that suggest danger. The mark should inform, not judge.
• Above the progress bar, where YouTube shows its “most replayed” graph, a colour-coded breakdown of the whole video can appear.
• When the viewer taps a dot, the details open: the level, the source, the verification status and the sound sub-layer.
• The marks can be switched on and off like the subtitles (CC) button.

The risk of traffic-light fatigue. The starting point of this proposal was the viewer’s fatigue at constantly playing detective: “Is this real?” But if the dots in the corner change colour and shape at every cut, detective fatigue may simply give way to “traffic-light fatigue”: the viewer’s attention slips from the video and locks onto the indicators in the corner. This risk applies not only to Shorts, but to any fast-cut long-form video. The display therefore needs a smoothing rule: in passages that switch rapidly between real and artificial, the dot does not flicker but settles into a steady “variable” mark. “Variable” is distinct from “mixed”: mixed means real and artificial elements together in the same frame; variable means scenes switching rapidly from one to the other. The second-by-second detail lives in the strip on the progress bar, and when the viewer pauses, the full breakdown for that moment appears.

6. Scenarios

Let us show how the system would work through a few concrete scenarios.

Scenario 1: Natural content, generated image, generated sound. One of the most common patterns on YouTube today: faceless channels telling history, documentary or stories. The text comes from a person; the images and voice come from AI. The current label stamps this video with a single “AI” mark, and viewers may conclude that the story itself is invented. In our system the screen reads “Content: natural, Image: generated, Sound: generated.” The viewer knows that what is being told is a product of natural intelligence, and that what they see and hear has been generated.

Scenario 2: Natural content, generated image, real voice and a real presenter who appears from time to time. While the presenter is speaking, all three layers are natural. When the video cuts to an AI scene, the image mark turns to “generated,” while sound and content stay as they are. If the real person is speaking in a corner of the screen with an artificial image behind them, the image mark becomes “mixed”; if the viewer taps, the artificial region is briefly highlighted. If scenes switch back and forth rapidly, the dot moves to “variable” and the detail is read in the strip. The viewer no longer has to wonder whether it is real or not.

Scenario 3: Real image, natural content, generated sound. A creator who shoots fine footage and produces good content, but feels weak at voiceover, uses AI only for the narration. This may be the most innocent use of AI of all. Yet because a synthetic voice used for narration requires a label today (23), this creator receives the same label as a video generated from start to finish. For someone with a speech impairment, someone self-conscious about their accent, or someone publishing in another language, a synthetic voice is an equaliser and should not become a stigma.

But whether it is “the most innocent” depends on whose voice it is. The sound layer has three very different cases: a generic synthetic voice (innocent); a clone of the creator’s own voice (innocent, but the viewer should know; today it requires no label (5)); and a cloned voice of another real person (potentially the most dangerous case, because a voice carries identity). This distinction can live in the details that open on tap rather than in the dot itself; but it must be part of the standard.

Scenario 4: Multilingual publishing. YouTube can automatically voice a video in many languages, and the feature is switched on by default for eligible creators (4). The original voice may be real, the other languages synthetic. Here the choice of dub can affect not one layer but all three: the audio track becomes synthetic; if lip-sync is applied, even real footage becomes “altered”; and because the dubbed text is a translation, the content layer must carry a “translated” note. The same video may appear with different marks to two viewers who chose different audio tracks. YouTube’s announcement that it will bring auto-dubbing to live streams in early 2027 (4) shows that the standard must also work in real time during live broadcasts.

Scenario 5: Fully natural. The most valuable case: real footage, a real voice, and content that is entirely a product of natural intelligence. Such content should not be left unmarked; it should carry a special, valued mark. If honest declaration earns no reward, the marking system becomes nothing more than a penalty regime. The “Fully Natural” mark is awarded on the basis of the main layers; artificial elements in sub-layers, such as background music or sound effects, do not remove it, but they are clearly shown in the details. In a music video, however, the music is the primary sound, so AI-generated music does remove the “Fully Natural” mark.

To keep this mark meaningful, the absence of a mark must not mean “natural”; unmarked content means “unknown.” This gives three distinct states: declared natural, verified natural, and unknown. An important design principle follows: “declared natural” must not be designed as a cheap sub-version of “verified natural,” but as a respected mark in its own right. Otherwise, because of the hardware barrier discussed below, honest creators without equipment capable of producing proof would be treated as second class.

7. Levels within the content layer

The content layer itself should be assessed at different levels. A text written entirely by AI and a text that is entirely the product of natural intelligence, with only spelling corrections done by AI, must not be marked the same way. Today not even these two extremes are distinguished; a script written by a language model needs no label either (1). The levels we propose:

1. Entirely natural intelligence
2. Spelling support: spelling and grammar correction
3. Language support: sentence and style editing
4. Information support: research and compilation of information by AI
5. Co-production: the idea from a person, the text largely from AI
6. Entirely AI

The risk, however, does not rise in a straight line. Content whose idea and text were shaped by a person, but whose facts were taken from AI without being checked, is where hallucination slips in most easily, and it may be more dangerous than level five. So alongside the level, a second piece of information is needed: “Did a human verify it?” There is a technical counterpart for this need too: C2PA’s new AI disclosure assertion records the model used and the degree of human oversight, going well beyond a binary “AI-generated” label (24).

The risk of authenticity-washing. The tiered structure has an obvious weak point. Someone who had AI write the entire text can easily claim, “I only had help with spelling and style,” and no algorithm can conclusively prove otherwise. This could penalise those who honestly declare “entirely AI” while giving those who lie an undeserved aura of naturalness. We call this “authenticity-washing.”

To counter this risk, the content layer should also distinguish between “declared” and “verified,” as image and sound do. A creator may choose to offer evidence supporting their declaration: a timestamped version history from their writing tool, or a contribution log kept by AI tools during the work. But such evidence must be optional, not mandatory, for three reasons: a version history can be faked by retyping AI-written text by hand or pasting it in piece by piece; demanding drafts raises privacy concerns; and if evidence were mandatory, almost everyone’s content would end up “unverified,” reproducing the hardware inequality in the content layer as well. An honest declaration without evidence remains respected as “declared.” Ultimately, we must honestly accept that the content layer will rest on declaration; the real deterrent to deliberate deception is the sanction faced when it is caught.

To avoid cluttering the screen, the dot shows only the broad state (natural, assisted, artificial); the level and verification details open on tap.

8. Verification, challenge and sanctions

The person who knows best where AI was used in a piece of content is the person who made it. The creator’s declaration is therefore the primary source. Before upload, the platform’s AI assists the creator: at each scene change it asks about the image, sound and content; it pre-fills answers using existing provenance data and its own detection; and the creator confirms or corrects them. Responsibility stays with the creator, but the burden becomes lighter.

If the creator has not marked the content, the platform examines it second by second using AI and applies the marks itself. If the creator’s declaration and the platform’s detection conflict, a two-way challenge process begins. Because people can be wrong, make mistakes and give false information; and AI can always be wrong too. Today’s AI label offers no such dialogue: a creator who believes their content was wrongly identified as AI can simply update the status themselves in YouTube Studio (3).

The Content ID precedent, and its dark side. This structure has a well-known precedent: YouTube’s copyright system, Content ID. The system automatically detects matches between uploaded videos and rights holders’ reference databases, and matches can rest on just a few seconds of material (25). A creator can dispute a claim; the claimant has 30 days to respond to an initial dispute and 7 days to respond to an appeal (26). In one sense, the system we propose is “a Content ID for AI.”

But this precedent must be taken together with its lessons. In Content ID, both the initial dispute and the appeal are reviewed by the claimant, because YouTube says it cannot make ownership determinations (27); in other words, the accuser also acts as judge. The Electronic Frontier Foundation (EFF) has documented creators struggling with mistaken matches, lost revenue and multiple claims on the same work, and how missteps during the appeal process can escalate into a formal copyright claim (28). An academic study of the system, which processes more than 2.5 billion claims a year, found that even though the matching technology is generally accurate, enforcement favours large rights holders, does not reliably protect the interests of smaller and independent creators, and is open to abuse (29).

These lessons give our proposal four rules:

• The judge must be neutral: challenges should be assessed by an independent review mechanism, not by the automated system that applied the mark.
• During review, showing either side’s claim would be unfair; the passage should carry a neutral “under review” mark, and this mark must not affect the video’s position in recommendations. Otherwise the video would be unfairly penalised in its first days, when it is watched most.
• During a challenge, the creator’s declaration should stand unless the platform has high-confidence detection to the contrary.
• On a platform where millions of hours of video are uploaded every day, it is unrealistic for a human to review every challenge. The first stage can therefore be automated, but at the higher appeal stage the final decision must rest with a human.

We should also note an alternative that we considered and rejected: a “trust score” or “declaration credit” that would give priority to the declarations of creators with a clean record. This would disadvantage new and small creators without a track record from the outset, protect large channels, and open the door to a “build trust, then deceive” strategy. It would reproduce Content ID’s bias towards large rights holders, this time in AI marking.

The presumption of good faith. The philosophical foundation of the system is this: even an honest creator cannot give a one-hundred-percent accurate declaration. Over a long piece of work, ideas blend together; something an AI suggested may slip into the text unnoticed. The standard must therefore rest not on a flawless declaration, but on one given in good faith and with reasonable care. Turkish law offers a familiar form of this idea: Article 3 of the Turkish Civil Code presumes good faith wherever the law attaches consequences to it. The rule will also be familiar to European readers, since the Turkish Civil Code was adopted from the Swiss Civil Code, where the same presumption appears, likewise, in Article 3. It must apply equally to everyone. Honest mistakes are corrected, not punished; sanctions apply only to deliberate deception or repeated patterns.

Sanctions. The logic of sanctions already exists at the level of the whole video: YouTube can label undisclosed content itself, and persistently failing to disclose can lead to content removal or suspension from the Partner Programme (30). Our proposal brings this logic down from the whole video to the level of seconds and layers. Once creators know that their declarations can be checked second by second, they will, in time, have little choice but to give accurate information.

Two risks must be kept in mind, however. First, detection systems can also be wrong; in the Shorts case, genuine footage looked artificial because of the platform’s own processing (20). Sanctions should therefore be graduated: first a warning and the right to challenge, with heavier penalties for repetition. Second, if penalties fall only on those who say “no AI,” creators will start marking everything as artificial to protect themselves; marks will be everywhere and will say nothing. Making the “Fully Natural” mark genuinely valuable is the best antidote to this risk.

Platform accountability. AI systems must not turn hallucination into an all-purpose excuse to hide behind; responsibility lies not with the model but with the platform that uses it. This brings two concrete obligations: every mark corrected after a challenge should be recorded, with a “this mark was corrected” notice; and platforms should regularly disclose what share of their automatic marks were changed on challenge.

9. Practical problems: Where does the chain break, and who is left out?

Content changes at every stage of production. First a text is written, then it is used in producing a video, then in the voiceover, or the creator records the voice themselves; sometimes AI is involved, sometimes the creator’s own tools. With so much change, keeping marks intact is not easy in practice. Video processing pipelines are especially fragile: reliable preservation of C2PA data requires at least 500 kbps, while some watermarks embedded in the image remain detectable even at 200 kbps (31).

The standard has tools for this problem. Each stage can take the record of the previous one in as an “ingredient,” and since each ingredient carries its own provenance, a tree forms that reaches all the way to the final product (32). Part 3 of JPEG Trust defines the use of watermarking to bind the record to the content (33). Records can also be kept in a repository outside the file (34). So information is not lost even if the file is reprocessed.

Two further principles should be added. The burden should fall on the tools, not on people: editing software, audio tools and AI tools should record each step automatically, just like the “track changes” feature in a word processor. Tools that do not use AI should sign too: if only AI steps are recorded, the creator’s own footage or own voiceover will look like an unexplained gap in the chain. In addition, AI tools could record who suggested what during the work and offer the creator a “contribution log,” so that declarations rest not on memory but on records.

The hardware barrier and the digital divide. The “verified natural” mark requires cameras and software capable of signing provenance. Today this is found mainly in professional cameras: C2PA support has come to Sony, Nikon and Canon cameras through firmware updates, while Fujifilm and Panasonic, despite being coalition members, had not yet shipped it as of late 2026 (35). Phones have taken a first step: with the Pixel 10, Google built C2PA into its camera app, and every JPEG photo taken automatically receives provenance information; but for now this is available only on Pixel 10 devices (36).

This has a serious social dimension. Independent creators and citizen journalists in developing countries produce entirely real, natural content with older phones and tools that lack C2PA support. Because they cannot offer cryptographic proof, they cannot earn the “verified natural” mark, and their chains always look “broken.” This mark must not become an “elite badge” reserved for owners of expensive equipment. Two things are needed: cryptographic signing should be built into every device as a free, open standard at the operating-system level (Android and iOS); and the “declared natural” mark, as noted earlier, must remain a respected mark in its own right. Google itself warns that simplistic “AI / not AI” labels risk an implied truth effect, and proposes instead to distinguish content that comes with verifiable proof of how it was made from content that does not (37). This matches our principle that “unmarked means unknown,” while also reminding us that having no proof does not mean being fake.

Utopia, or a realistic path? Getting every tool to join this system may look utopian today. But things are a little further along than one might think: membership of the Content Authenticity Initiative passed 6,000 in early 2026, although real-world adoption remains uneven (38).

Platforms, more than toolmakers, can drive adoption. The internet offers a precedent: in 2015 Google began ranking unencrypted sites lower in search, then step by step marked unencrypted pages as “Not secure” in Chrome, extending this to all unencrypted sites in 2018; by then, the large majority of Chrome traffic was already encrypted (39). In other words, adoption came not from a single label but from visible incentives sustained over years. If platforms award the “verified natural” mark only to content with an intact chain, creators will want it, and toolmakers will follow.

Meanwhile, the law shows a curious asymmetry: in the EU, providers of generative AI must, from 2 August 2026, mark their output in a machine-readable form (35); but no such obligation exists for tools that do not use AI. In other words, the law makes proof of the artificial mandatory, while leaving proof of the natural voluntary.

The realistic design is this: proof where the chain exists, declaration where it does not. The system does not collapse with incomplete participation; it simply rests on declaration instead of proof. Where the chain breaks, the final safeguard is the declaration at upload by the only person who knows the whole process from start to finish: the creator.

10. Who will set the standard?

This marking language should belong not only to YouTube, but to all of social media. Whatever video or post we look at, we should be able to read its status in the same way. That requires an international standard, in the manner of JPEG and MPEG.

Interestingly, the JPEG committee has already entered this field. JPEG Trust (ISO/IEC 21617) offers a framework for establishing trust in media, and its first part was published in January 2025 (33). C2PA’s technical specification is on ISO’s fast track as ISO 22144 (40). Above them, under the World Standards Cooperation, the partnership of the IEC, ISO and ITU, an AI and Multimedia Authenticity Standards Collaboration has been formed, bringing together standards developers, technology companies, policymakers, researchers and civil society (41). One of its tasks is to identify gaps where standards are needed but do not yet exist (42).

So it is not enough to say such a group should be formed; the real issue is that these groups are working on the wrong layer. The question needs to be considered on three layers:

1. The technical data layer: the record, written into the file, of who produced what and how. This largely exists. Even the data structure for scene-level marking is ready: C2PA’s “regions of interest” can mark a specific area of an image, specific paragraphs of a document, or specific segments of audio and video separately (24).
2. The viewer display layer: turning this data into on-screen marks that everyone understands in the same way. This is what is missing, and it is where our proposal makes its main contribution. Its precedent is the ISO 7010 safety signs: the fire-exit symbol is the same everywhere in the world and can be understood without knowing any language. The EU’s introduction of common icons for AI content is a step in this direction (43).
3. The legal binding layer: JPEG and MPEG standards bind no one legally; they are binding in practice because everyone uses them. A rule that binds everyone requires the standard to be referenced in law. Article 50 of the EU AI Act is a first step, but it does not require scene-level marking.

There is also a question of representation: C2PA is backed by powerful companies such as Adobe, Microsoft, the BBC, Sony, OpenAI, Google, Meta and Amazon (44). At the table that decides how marks will appear on screen, there must also be seats for viewers, independent creators, and creators from countries without access to expensive equipment. Otherwise the standard will build the hardware barrier, and Content ID’s imbalance between large and small, into its very structure.

11. Open questions

This is not a subject that will ever be exhausted. Some thresholds the standard must define are already clear:

• Will AI noise reduction on a real voice count as “altered”? If so, almost every video will be marked.
• Where do processes that phones apply by default, such as night mode, portrait blur and noise reduction, stand? YouTube itself compared its Shorts processing to what a modern smartphone does (18). If such on-device processing breaks “naturalness,” almost no video can be fully natural; if it does not, the line must still be defined.
• What is the right threshold duration for the smoothing rule in the display? This is a value to be determined through user-experience research.
• For image and sound, cryptographic proof is possible; cameras and microphones can sign provenance. But no such proof exists for the content layer; there is no camera recording our thoughts. The content layer will always rest largely on declaration.
• Detection technology is an arms race. C2PA’s own developers say their standard does not deal with detection, precisely because detecting AI manipulation is a never-ending race (32). Automated detection can audit the system, but it cannot be its foundation.

12. Conclusion: Putting the viewer’s mind at ease

Today’s labels answer the question “Is there AI in this video?” with yes or no. The universal marking language we propose answers a different question: “How much of what I am seeing and hearing right now is natural, and how much is artificial?” Second by second, layer by layer, naming the source, grounded in evidence and open to challenge. With one dot for a text, two for a photo or a podcast, three for a video; in the same language on every platform.

The aim of this system is not to warn viewers or to stigmatise creators who use AI. Nor is it to pull the viewer’s attention away from the video towards indicators in the corner. Its aim is to put the viewer’s mind at ease: to let us watch without wondering at every scene “Is this real?”, knowing what is a product of natural intelligence and what is a product of artificial intelligence. Perhaps then, watching videos will become a pleasure again.

Colophon

This article was compiled from a conversation between Aydın Tiryaki and Claude Opus 5.5 on 26 October 2026; the initial work on the subject, the observations, the core ideas and the main framework of the proposal (moment-by-moment, layered marking of AI use; the inclusion of platforms’ own tools in marking; a challengeable system in which creator declarations and automated detection work together; a universal marking language for all of social media and an international standard to define it; a special, valued mark for fully natural content; tiered assessment within the content layer; and all the scenarios) belong to Aydın Tiryaki; Claude Opus 5.5 provided research assistance (compiling current platform rules, statistics, technical standards and regulation) and writing assistance in line with these ideas, and also offered several complementary suggestions during the discussion (the Content ID precedent, the “under review” state, the breakdown on the progress bar, colour-blind-friendly marks, transparency of platform error rates, and the contribution log); Gemini 2.5 Pro carried out an independent critical review (peer review and counter-argument analysis) of the second version, and its objections and proposals on cognitive load and traffic-light fatigue, the separation of music and effects in the sound layer, the risk of authenticity-washing in the content layer, the possible side effects of the Content ID precedent, the digital divide created by the C2PA hardware barrier, and the modular adaptation of the system to non-video content were incorporated into the third version after correction through Claude Opus 5.5’s counter-review and with Aydın Tiryaki’s arbitration and approval, during which Gemini withdrew its “trust score” and mandatory version-history proposals in response to the counter-review; Gemini 2.5 Pro also reviewed the third version, the reference numbers were reordered by first appearance in the final version, and the text was accepted as final by Aydın Tiryaki; this English version is an adaptation of the Turkish original, prepared by Claude Opus 5.5 and approved by Aydın Tiryaki.

References

(1) PPC Land — YouTube shifts generative AI labels to spots viewers will actually see: https://ppc.land/youtube-shifts-generative-ai-labels-to-spots-viewers-will-actually-see/
(2) Outlook Business — YouTube Mandates Labelling Of Synthetic, AI-Generated Content: https://www.outlookbusiness.com/companies/youtube-mandates-labelling-of-synthetic-ai-generated-content
(3) YouTube Blog — Improving AI labels for viewers and creators: https://blog.youtube/news-and-events/improving-ai-labels-viewers-creators/
(4) PPC Land — Auto dubbing: https://ppc.land/auto-dubbing/
(5) AIR Media-Tech — Is YouTube’s AI Dubbing Safe for Your Channel?: https://air.io/en/youtube-hacks/is-youtubes-ai-dubbing-safe-for-your-channel-auto-dub-vs-pro-with-real-retention-data
(6) Boom Live — YouTube Introduces “Captured with a Camera” Labels: https://www.boomlive.in/web-stories/news/youtube-introduces-captured-with-a-camera-labels-to-combat-ai-generated-content-2188
(7) Gigazine — YouTube introduces ‘filmed on camera’ label using C2PA standard: https://wbgsv0a.gigazine.net/gsc_news/en/20241016-youtube-c2pa-captured-with-a-camera
(8) Didit — Europe’s deepfake rule is now in force: https://didit.me/blog/eu-ai-act-deepfake-labelling-2026/
(9) BCLP — The AI Office has published its first Code of Practice on transparency of AI-generated content: https://www.bclplaw.com/print/v2/content/1568588/the-ai-office-has-published-its-first-code-of-practice-on-transparency-of-ai-generated-content.pdf
(10) Horan MediaTech Advisors — AI-Generated Content on YouTube: How Large Is the Problem?: https://horanmediatech.com/ai-generated-content-on-youtube-how-large-is-the-problem/
(11) Anadolu Agency — Study finds over 20% of videos shown to new YouTube users are ‘AI slop’: https://www.aa.com.tr/en/science-technology/study-finds-over-20-of-videos-shown-to-new-youtube-users-are-ai-slop-/3783297
(12) Philstar Life — 21% of videos on YouTube feed are AI ‘slop’: https://philstarlife.com/amp/article/485354-21-percent-videos-on-youtube-ai-slop-research
(13) eWeek — AI Conquers YouTube: https://www.eweek.com/news/ai-slop-conquers-youtube-algorithm/
(14) 9to5Mac — 60% of TikTok videos are AI slop; 21% of YouTube ones: https://9to5mac.com/2026/06/22/60-of-tiktok-videos-are-ai-slop-21-of-youtube-ones/
(15) The Decoder — One in five YouTube Shorts shown to new users is AI-generated slop: https://the-decoder.com/one-in-five-youtube-shorts-shown-to-new-users-is-ai-generated-slop-study-finds/
(16) TheWrap — More Than 1 Million YouTube Channels Used the Company’s AI Tools in December: https://thewrap.com/industry-news/tech/youtube-2026-goals-ai-monetization-neal-mohan-letter
(17) Adobe — How creators use AI video tools across platforms: https://www.adobe.com/express/learn/blog/ai-video-tools
(18) TweakTown — YouTube is quietly using AI to change some videos without creator consent: https://www.tweaktown.com/news/107310/youtube-is-quietly-using-ai-to-change-some-videos-without-creator-consent/index.html
(19) AlternativeTo — YouTube faces backlash for secret AI edits on Shorts videos: https://alternativeto.net/news/2025/8/youtube-faces-backlash-for-secret-ai-edits-on-shorts-videos-without-user-consent
(20) RouteNote — YouTube has been quietly editing creators’ videos with AI: https://routenote.com/blog/youtube-has-been-quietly-editing-creators-videos-with-ai/
(21) Hiverlab — YouTube Secretly Using AI to Edit Creators’ Videos Without Consent: https://hiverlab.com/old-content/youtube-uses-ai-to-edit-videos-without-consent
(22) Zenni Optical — Debunking the Myths of Color Blindness: https://www.zennioptical.com/blog/myths-color-blindness/
(23) Primetel — YouTube adds mandatory disclosure badges for AI-generated and altered content: https://primetel.com.cy/youtube-adds-mandatory-disclosure-b-badges-for-ai-generated-and-altered-content-8705
(24) C2PA — A New Implementation Guide for Content Credentials: https://c2pa.org/?p=259
(25) Boing Boing — Algorithms like YouTube’s Content ID harm fair use, free speech, and creativity: https://boingboing.net/2020/12/17/algorithms-like-youtubes-content-id-harm-fair-use-free-speech-and-creativity.html
(26) YouTube Help — Appeal a Content ID claim: https://support.google.com/youtube/answer/12104471
(27) YouTube Help — Dispute a Content ID claim: https://support.google.com/youtube/answer/2797454
(28) EFF — Unfiltered: How YouTube’s ‘Content ID’ Helps Shape What We See Online: https://eff.org/node/104288
(29) NBER — Evidence from YouTube Content ID (conference paper): https://conference.nber.org/conf_papers/f243285.pdf
(30) AIR Media-Tech — What is YouTube’s AI content disclosure policy?: https://air.io/en/youtube-glossary/what-is-youtubes-ai-content-disclosure-policy
(31) Tian Pan — AI Content Provenance in Production: https://tianpan.co/blog/2026-04-19-ai-content-provenance-c2pa
(32) W3C — C2PA meeting minutes (20 August 2024): https://www.w3.org/2024/08/20-me-minutes.html
(33) JPEG — Overview of JPEG Trust: https://jpeg.org/jpegtrust/
(34) C2PA — Implementation Guidance: https://c2pa.org/specifications/specifications/1.3/guidance/Guidance.html
(35) Kompozy — C2PA glossary: https://kompozy.io/glossary/c2pa
(36) BleepingComputer — Pixel 10 fights AI fakes with new Android photo verification tech: https://bleepingcomputer.com/news/security/pixel-10-fights-ai-fakes-with-new-android-photo-verification-tech
(37) Security Affairs — Google Pixel 10 adds C2PA to camera and Photos: https://securityaffairs.com/182068/security/google-pixel-10-adds-c2pa-to-camera-and-photos-to-spot-ai-generated-or-edited-images.html
(38) Metavert — C2PA vs Content Authenticity: https://metavert.io/compare/c2pa-vs-content-authenticity
(39) TechSpot — Chrome will start marking HTTP sites as ‘not secure’: https://www.techspot.com/news/73202-chrome-start-marking-http-not-secure.html
(40) ASIST — ISO/DIS 22144, Authenticity of Information – Content Credentials: https://www.asist.org/?p=107622
(41) ITU — Standards and policy considerations for multimedia authenticity: https://www.itu.int/hub/2025/07/standards-and-policy-considerations-for-multimedia-authenticity/
(42) AI for Good — AI and Multimedia Authenticity Standards Collaboration: https://aiforgood.itu.int/amas
(43) Heuking — The European Commission Specifies Labelling Requirements for AI-Generated Content and Deepfakes: https://www.heuking.de/en/news-events/newsletter-articles/detail/the-european-commission-specifies-labelling-requirements-for-ai-generated-content-and-deepfakes.html
(44) IET — The Trust Factor: Explaining the C2PA standard: https://tv.theiet.org/video/the-trust-factor-explaining-the-c2pa-standard/eded07dd-359f-4945-a862-d1c43a46d93a

---

| aydintiryaki.org | YouTube | Aydın Tiryaki’nin Yazıları ve Videoları │Articles and Videos by Aydın Tiryaki | Bilgi Merkezi│Knowledge Hub  | ░ Virgülüne Dokunmadan │ Verbatim ░ | ░ Saniye Saniye Yapay Zeka: Paylaşılan Her İçerik İçin Evrensel Bir İşaretleme Dili Önerisi │AI, Second by Second: A Proposal for a Universal Marking Language for Everything We Share ░ 26.10.2026