When Is Content "AI-Generated Enough"? Labelling Synthetic Media under the Digital Services Act and the AI Act

1 Synthetic Media as a Regulatory Object

• Labelling discloses artificial generation or manipulation rather than prohibiting all synthetic media, balancing protection against deception with expressive freedom.The paper frames this as a compromise between freedom of expression, artistic freedom, innovation, and protection against deception, manipulation, and disinformation.Source: §1 · p. 1
• Media practices blur authentic and fake, from AI-assisted edits to fully generated people and misleadingly framed images.The resulting challenge concerns unstable categories of deception, not only deepfakes.Source: §1 · p. 2
• The paper asks how the DSA, AI Act, and Article 50 implementation framework define deepfake transparency and translate duties into technical systems, interfaces, and disclosures.It asks what practical problems arise in that translation.Source: §1 · p. 2

2 Regulatory and Implementation Background

• The DSA addresses deepfake-like content through VLOP and VLOSE systemic-risk measures, including prominent markings and an easy-to-use user indication function.The AI Act separately defines deepfakes and requires deployers to disclose qualifying generated or manipulated media and certain public-interest text, subject to exceptions.Source: §2 · p. 2
• The Code separates provider-side machine-readable marking and detection from deployer-side user-facing labels for deepfakes and qualifying public-interest text.Its provider measures include digitally signed metadata and watermarking; deployer disclosures may use icons, labels, or disclaimers.Source: §2 · p. 2
• Forthcoming Commission guidelines are to address definitions and exceptions, leaving unresolved when content is sufficiently generated, manipulated, authentic-looking, or public-interest-oriented to require disclosure.The paper identifies this as a central threshold problem.Source: §2 · p. 3
• The paper frames labelling as socio-technical classification, identifies four governance tensions, and uses a DSA Statement of Reasons snapshot to compare labelling with other restrictions.The tensions are definitional ambiguity, interface and responsibility design, communicative usefulness, and fairness and contestability.Source: §2 · p. 3

3 Labelling as Provenance and Platform Governance

• The Code encourages provenance metadata identifying the AI system and provider, generation or manipulation time, and the operation performed.It also proposes distinguishable deployer disclosures, potentially with a common EU icon and a more detailed second layer.Source: §3 · p. 3
• DSA electoral guidelines recommend prominent markings, easy-to-use user labelling tools, recognisable labels, and retaining labels when content is reshared on the same platform.Labelling sits alongside mitigation measures such as demotion, removal, watermarks, metadata, and adapted moderation systems.Source: §3 · p. 3
• Uploader-side disclosure can be built into publication through an upload toggle, disclosure field, or advertising-label option.This makes disclosure part of the ordinary publication workflow.Source: §3 · p. 3

4 EU Icon Categories and Regulatory Thresholds

• The “AI GENERATED” icon covers content generated entirely through AI, while “AI MODIFIED” applies to pre-existing content partially altered using AI.The basic icon may pair with customised text or an interactive second layer describing the nature and location of AI involvement.Source: §4 · p. 4
• Minor enhancement, noise reduction, automated subtitling, and colour correction may not qualify as a deepfake or qualifying manipulated public-interest text.The categories leave unresolved when AI assistance becomes sufficient to trigger disclosure.Source: §4 · p. 4
• Visible icons form one layer alongside machine-readable metadata, watermarking, provenance information, and detection mechanisms; interfaces govern display, reporting, and contesting classifications.Table 1 situates the icon categories within overlapping AI Act, DSA, and Code perspectives.Source: §4 · pp. 3–4

5 Fairness, Differential Effects, and Contestability

• Warning-label studies document false positives and false negatives, while synthetic-media errors may cause unjustified demotion, reduced visibility, or removal of authentic content [12] [18].One TikTok analysis found 37.3% of examined videos with benign information were labelled, while 7.7% of harmful or misleading videos were not [12].Source: §5 · p. 4
• Label effects vary with design, source, visibility, content, and audience; political congruence and label wording can also affect engagement and interpretation [11].Source: §5 · p. 4
• Classification categories may encode assumptions from training data, annotation practices, and platform-specific moderation policies, rather than functioning as value-neutral technical labels [10].Source: §5 · p. 4
• Affected users need to understand why content was labelled, which elements were identified, and what consequences followed; contestability research highlights representation, clear communication, and ways to challenge decisions.Source: §5 · p. 4
• Accessible explanations, correction and appeal mechanisms, and opportunities to provide contextual evidence are proposed as components of contestable synthetic-media labelling.Source: §5 · p. 5

6 Statement of Reasons Data and Reported Practice

• The filtered “Synthetic Media” view contained 153,138,573 statements from 25 September 2023 to 28 April 2026, with Facebook and Instagram accounting for approximately 118.33 million and 31.36 million.X accounted for approximately 2.74 million statements and Threads for approximately 717,000.Source: §6 · p. 5
• 151.46 million cases (98.9%) were visibility measures; removal accounted for approximately 129.79 million statements, demotion 19.08 million, and labelled content 1.42 million.The distribution indicates that platforms reported removal or demotion more often than labelling in this filtered view.Source: §6 · p. 5
• 150.10 million statements (98.02%) were based on platforms’ own initiative, while other notification types accounted for approximately 299,000 cases (0.20%).The passage raises whether recipient-facing tools do much practical work compared with internal or automated detection and enforcement.Source: §6 · p. 5
• The dashboard counts platform-classified statements, not unique content or synthetic-media prevalence, and do not establish that a restriction was imposed because content was synthetic.The analysis does not independently validate records, and low reported labelling counts cannot establish how often platforms label synthetic media outside this dataset.Source: §6 · p. 5

7 Preliminary Argument

• The threshold question is whether AI materially changes content’s communicative status, so borderline cases require more granular categories than “real” and “fake.”Examples include beauty filters, AI-generated captions, background extension, and face-swapping.Source: §7 · p. 6
• Recipient-facing reporting tools make users part of the labelling infrastructure, although users may misidentify AI content or strategically misuse deepfake allegations.Label visibility, salience, and timing shape the intervention’s effect.Source: §7 · p. 6
• A disclosure’s value depends on whether users notice, understand, and appropriately interpret it; in political contexts, claims that authentic content is AI-generated may serve as denial.Source: §7 · p. 6
• Classification errors may fall unevenly across languages, cultural contexts, and expression, while false labels may lead to demotion, removal, or reputational harm.The paper identifies accessible explanations and meaningful opportunities to challenge erroneous decisions as fairness requirements.Source: §7 · p. 6

8 Conclusion

• The DSA addresses platform-level mitigation, while the AI Act and Code distinguish provider-side marking from deployer-side labelling.The conclusion calls for clear, contestable, user-facing disclosure systems.Source: §8 · p. 6

• Table 1 situates EU icon categories within the AI Act framework.Source: Table 1 · p. 9
• Table 1 relates EU icon categories to the DSA.Source: Table 1 · p. 9
• Table 1 also relates the categories to the Code of Practice and the paper’s threshold analysis.Source: Table 1 · p. 9

“AI MODIFIED”

• Disclosure should turn on materiality, communicative significance, and false-authenticity risk rather than mere AI involvement.Source: “AI MODIFIED” · p. 9
• Users need usable categories, explanations, detection support, and ways to challenge erroneous classification, but are not responsible for deciding legal qualification.Source: “AI MODIFIED” · p. 9
• Provenance and visible labelling perform different functions, and metadata alone may not satisfy the objective of prominent disclosure.Source: “AI MODIFIED” · p. 9
• Technical origin information does not itself determine whether content is misleading, harmful, satirical, or legally subject to disclosure.Source: “AI MODIFIED” · p. 9

B Statement of Reasons Dashboard Screenshots

• The screenshots were collected on 30 April 2026; Figure 1 filters synthetic-media submissions from 25 September 2023 to 28 April 2026.The filter dates refer to submission dates, and values displayed in thousands are rounded.Source: B Statement of Reasons Dashboard… · p. 10, Figure 1 · p. 10
• 153,138,573 statements are reported overall, concentrated on Facebook, Instagram, X, and Threads.Source: Figure 2 · p. 10
• Approximately 1.42 million statements concern labelled content, while removal and demotion dominate the restriction breakdown.Source: Figure 3 · p. 11
• Approximately 98.9% of statements involve visibility-only restrictions; the treemap displays reported violation categories.Source: Figure 4 · p. 11
• The daily-submissions chart reverses the temporal axis, and its displayed late-April spike is driven primarily by Facebook.Source: Figure 5 · p. 12
• Daily submissions are broken down by violation category in absolute and proportional terms, with a lower-right panel showing violations by platform.The temporal axis is reversed.Source: Figure 6 · p. 12
• Figure 7 covers content type, information source, restriction grounds, and decision and detection automation; own voluntary initiative accounts for approximately 98.02% of statements.Source: Figure 7 · p. 13