Skip to content
ContentLora

    Tip: press / anywhere to search.

    Explainer

    AI watermarking and content provenance explained

    The main tools for showing where digital content came from include provenance metadata such as C2PA Content Credentials, watermarks hidden in the content itself, and detectors that try to classify content as AI-made or not.[1][2][3] Each has weaknesses: metadata is often stripped when files are shared, text watermarks can often be removed by paraphrasing, and detectors make false positives that can wrongly brand human work as AI-generated.[4][5][6] NIST says such transparency may support trust but does not guarantee it, while the EU AI Act requires providers to mark AI-generated output in a machine-readable way.[7][8]

    Editor reviewedUpdated Explainer25 claims · 100% primaryArtificial intelligence, Frontier AI, Tech policy

    This page explains the technical tools for tracing AI-generated content and their limits. For the harms these tools target, see deepfakes; for how the generators work, see how AI image and video generators work.

    Why provenance matters

    Generative AI can produce realistic images, text, audio and video, which NIST says creates new risks to trust in digital information.[9] NIST uses the term digital content transparency for documenting and accessing information about where content came from and how it has changed.[10] Provenance records are not only for AI: NIST notes they can record synthetic or non-synthetic origins.[11]

    Metadata and Content Credentials

    The best-known provenance standard comes from the Coalition for Content Provenance and Authenticity (C2PA), which publishes an open technical standard called Content Credentials.[12] A Content Credential is a cryptographically bound record of an asset’s provenance, with assertions such as when and where it was created and how it was modified.[1]

    Metadata has a basic weakness.[4] NIST says metadata stored within a file is often stripped when files are shared, sometimes to deceive and sometimes for benign reasons such as protecting privacy.[4] The C2PA’s own explainer says provenance metadata can be removed, that “durable” credentials add watermarking or fingerprinting to recover it, and that adding provenance is optional.[13] Hashes, also called digital fingerprints, are identifiers generated from the content itself, and NIST says they can link content to metadata stored elsewhere.[14]

    Watermarks

    A watermark puts the signal in the content itself.[2] NIST distinguishes overt watermarks, which a person can see or hear directly, from covert watermarks, machine-readable signals made of subtle changes that humans find hard to detect.[2] NIST says a good watermark should survive ordinary changes such as compression, filtering or cropping for media, or minor paraphrases or deletions for text.[15]

    Google DeepMind’s SynthID embeds imperceptible watermarks in AI-generated images, audio, text and video, and Google says its image and video watermarks are designed to survive cropping, filters and lossy compression.[16] For text, Google DeepMind says SynthID works by adjusting the probability scores the model gives each candidate next word.[17] A 2023 academic method works similarly: before each word is generated, it randomly selects a set of “green” tokens and softly favours them, so a statistical test can later detect the pattern without access to the model.[18] Google says it launched a SynthID Detector portal in 2025 to check content made with AI from Google and partners.[19]

    Detection and its errors

    Some tools try to judge content directly.[3] NIST defines synthetic content detection as techniques that classify whether content, or part of it, is synthetic.[3] Even a covert watermark needs a detector, and NIST says any such detector will have some nonzero rate of false positives and false negatives.[20] It warns that false positives, judging human-made work to be AI-generated, can be extremely damaging, causing reputational harm or adverse treatment.[6]

    How the tools can fail

    Text watermarks can be fragile.[5] NIST says having another model paraphrase watermarked text can often remove the watermark with little loss of quality. It says recursive paraphrasing can reduce detection rates to 20 percent for texts of about 225 words, while paraphrasing has been found to reduce detection only slightly for texts beyond about 400 words.[5] A 2023 paper went further, arguing that under natural assumptions strong watermarking of generative model output is impossible, and its attack removed the watermarks of three text-watermarking schemes with only minor quality loss.[21]

    Because each method has gaps, NIST suggests combining them, for example a robust covert watermark that is unlikely to be stripped plus signed metadata for those who receive it.[22] It also cautions that transparency may contribute to trustworthiness but does not guarantee it, and can create a false sense of trust when genuine content is used out of context.[7]

    Rules requiring labels

    Article 50(2) of the EU AI Act requires providers of AI systems that generate synthetic audio, images, video or text to mark the output in a machine-readable format so it is detectable as artificially generated or manipulated.[8] Article 50(4) requires deployers of systems that create deepfake images, audio or video to disclose that the content is artificial.[23] These transparency rules took effect in August 2026,[24] although generative systems already on the market before 2 August 2026 have until 2 December 2026 to meet the marking duty.[25] See the EU AI Act and AI regulation around the world.

    How AI image and video generators work

    How diffusion models and video transformers make images and clips, where training data comes from, and how watermarks, C2PA and EU labelling rules work.

    Questions readers ask

    What is content provenance?

    A record of where a piece of digital content came from and how it changed. NIST calls the wider practice digital content transparency, documenting and accessing information about content's origins and history.[10][1]

    What are Content Credentials?

    The C2PA open standard's cryptographically bound record of an asset's origin and edits. The C2PA explainer says the metadata can be removed and that adding it is optional.[12][1][13]

    How do you watermark AI-generated text?

    By nudging which words the model picks. Google DeepMind says SynthID adjusts the probability scores of candidate next words, and a 2023 method favours a randomly chosen set of "green" tokens so a statistical test can later detect the pattern.[17][18]

    Can AI watermarks be removed?

    Often, yes. NIST says paraphrasing watermarked text with another model can often remove the watermark with little loss of quality, and a 2023 paper argued that strong watermarking is impossible under natural assumptions.[5][21]

    Are AI detectors reliable?

    They make mistakes. NIST says watermark detectors have some nonzero rate of false positives and false negatives, and warns that wrongly labelling human work as AI-generated can be extremely damaging.[20][6]

    Is labelling AI content required by law?

    In the EU, yes. Article 50 of the AI Act requires providers of generative AI systems to mark output in a machine-readable format and requires deployers to disclose deepfakes. These transparency rules took effect in August 2026, though generative systems already on the market before then have until 2 December 2026 to meet the marking duty.[8][23][24][25]

    Sources

    · 25 claims from 6 publishers

    Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.

    1. [1]

      A C2PA Content Credential is a cryptographically bound record of an asset's provenance containing assertions such as when and where it was created and how it was modified. confirmedas of 2026-11-07

    2. [2]

      NIST distinguishes overt watermarks, which a person can perceive directly, from covert watermarks, machine-readable signals made of subtle changes to the content that humans find hard to detect. confirmedas of 2024-11-20

    3. [3]

      NIST defines synthetic content detection as techniques and tools that classify whether content, or part of it, is synthetic. confirmedas of 2024-11-20

    4. [4]

      NIST says metadata recorded within a file is often stripped when files are shared, either to deceive or for benign reasons such as privacy protection. confirmedas of 2024-11-20

    5. [5]

      NIST says having another model paraphrase watermarked text can often remove the watermark with little loss of quality; it says recursive paraphrasing can reduce detection rates to 20 percent for texts of about 225 words, and that paraphrasing has been found to reduce detection accuracy only slightly for texts beyond about 400 words. confirmedas of 2024-11-20

    6. [6]

      NIST warns that false positives, judging human-made content to be AI-generated, can be extremely damaging, causing reputational harm or adverse treatment. confirmedas of 2024-11-20

    7. [7]

      NIST says digital content transparency may contribute to trustworthiness but does not guarantee it, and can create a false sense of trust, for example when legitimate content is taken out of context. confirmedas of 2024-11-20

    8. [8]

      Article 50(2) requires providers of AI systems that generate synthetic audio, image, video or text to mark the outputs in a machine-readable format so they are detectable as artificially generated or manipulated. confirmedas of 2024-07-12

    9. [9]

      NIST notes that generative AI can produce realistic images, text, audio and video as well as multimodal content, which brings new risks to trust in digital information. confirmedas of 2024-11-20

    10. [10]

      NIST defines digital content transparency as documenting and accessing information about the origins and history of digital content. confirmedas of 2024-11-20

    11. [11]

      NIST says provenance data tracking can record that content was made or edited with an AI tool, and can also record non-synthetic origins. confirmedas of 2024-11-20

    12. [12]

      The Coalition for Content Provenance and Authenticity (C2PA) publishes an open technical standard, called Content Credentials, for establishing the origin and edits of digital content. confirmedas of 2026-11-07

    13. [13]

      The C2PA explainer says provenance metadata can be removed, that "durable" credentials add watermarking or fingerprinting to recover it, and that adding provenance is optional. confirmedas of 2026-11-07

    14. [14]

      NIST says hashes, or digital fingerprints, are identifiers generated from the content itself and can link content to externally stored metadata. confirmedas of 2024-11-20

    15. [15]

      NIST says a good watermark should stay detectable after ordinary changes such as compression, filtering or cropping for media, or minor paraphrases or deletions for text. confirmedas of 2024-11-20

    16. [16]

      Google DeepMind's SynthID embeds imperceptible watermarks in AI-generated images, audio, text and video, and Google says image and video watermarks are designed to survive cropping, filters and lossy compression. confirmedas of 2026-11-07

      • SynthID · Google DeepMind · Primary source · How SynthID works (retrieved 2026-11-07)
      • SynthID · Google DeepMind · Primary source · AI-generated image and video (retrieved 2026-11-07)
    17. [17]

      Google DeepMind says SynthID watermarks AI-generated text by adjusting the probability scores the model assigns to each candidate next word, and that this is not noticeable and does not affect output quality. confirmedas of 2026-11-14

      • SynthID · Google DeepMind · Primary source · AI-generated text (retrieved 2026-11-07)
      • SynthID · Google DeepMind · Primary source · AI-generated text (retrieved 2026-11-07)
      • SynthID · Google DeepMind · Primary source · AI-generated text (retrieved 2026-11-07)
      • SynthID · Google DeepMind · Primary source · AI-generated text (retrieved 2026-11-07)
    18. [18]

      A 2023 text-watermarking method randomly selects a set of "green" tokens before each word is generated and softly favours them, so the watermark can be detected with a statistical test without access to the model. confirmedas of 2023-01-24

    19. [19]

      Google launched the SynthID Detector verification portal in 2025 and says it is expanding access to everyone, to check content made with AI from Google or partners including OpenAI, NVIDIA and Kakao. confirmedas of 2026-11-07

      • SynthID · Google DeepMind · Primary source · SynthID Detector (retrieved 2026-11-07)
      • SynthID · Google DeepMind · Primary source · SynthID Detector (retrieved 2026-11-07)
    20. [20]

      NIST says a covert watermark must be found by a watermark detector, which will have some nonzero probability of false positives and false negatives. confirmedas of 2024-11-20

    21. [21]

      A 2023 paper argued that, under natural assumptions, strong watermarking of generative model output is impossible, and its generic attack removed the watermarks of three text-watermarking schemes with only minor quality loss. confirmedas of 2023-11-07

    22. [22]

      NIST suggests watermarks and metadata can be used together, for example a robust covert watermark that is unlikely to be stripped plus signed metadata for those who receive it. confirmedas of 2024-11-20

    23. [23]

      Article 50(4) requires deployers of AI systems that generate or manipulate deep-fake image, audio or video content to disclose that it is artificially generated or manipulated, with lighter rules for evidently artistic, satirical or fictional works. confirmedas of 2024-07-12

    24. [24]

      The AI Act's transparency rules took effect in August 2026, when the Act became generally applicable on 2 August 2026. confirmedas of 2026-08-02

    25. [25]

      Under the Omnibus, providers of generative AI systems placed on the market before 2 August 2026 have until 2 December 2026 to comply with the Article 50(2) marking requirement. confirmedas of 2026-07-27

    Revision history (1)
    1. Page created.

    Created Nov 14, 2026. Last reviewed by an editor on Nov 14, 2026. Next scheduled review: Feb 14, 2027.

    Cite this page

    "AI watermarking and content provenance explained." ContentLora, updated Nov 14, 2026. https://contentlora.com/explain/ai-content-provenance-and-watermarking-explained

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.