Executive Summary
The information integrity landscape of 2026 is defined by a fundamental asymmetry: the tools for generating convincing synthetic media have become cheap, fast, and widely accessible, while the tools for verifying content authenticity remain technically complex, jurisdictionally fragmented, and structurally incomplete. This data brief maps the key metrics, regulatory developments, and technical architectures that define the current state of synthetic media governance — and identifies the critical gaps that remain.
Key findings:
- The EU AI Act's Article 50, mandating machine-readable disclosure for synthetic content, entered force on 2 August 2026 — the most significant regulatory milestone in information integrity governance to date.
- C2PA Content Credentials have reached production-scale adoption, with major camera manufacturers embedding signing firmware — but metadata stripping by social media platforms remains a critical efficacy gap.
- 31 US states have enacted legislation addressing election-related deepfakes, but no federal standard exists, and constitutional challenges have already blocked enforcement in at least one state.
- The "liar's dividend" — the use of synthetic media's existence to dismiss authentic evidence as fabricated — has emerged as the most structurally damaging consequence of the deepfake era.
- Detection-only approaches have been largely abandoned by leading researchers and policymakers in favour of provenance-based architectures — but provenance verifies authorship, not truth.
The Scale of the Problem: Key Data Points
Synthetic Media Production
The cost of producing synthetic media has collapsed to near zero across all modalities. Text generation, image synthesis, voice cloning, and full-body video synthesis are now accessible to any actor with a consumer-grade device and an internet connection. The barrier to entry for synthetic media production is no longer technical or financial — it is motivational.
The implications for information integrity are structural. When the cost of producing convincing synthetic content approaches zero, the volume of synthetic content in the information ecosystem approaches infinity. Detection-based approaches — which require identifying synthetic content after it has been produced and distributed — face an inherently asymmetric challenge: the cost of generating synthetic content is near zero, while the cost of detecting and removing it at scale is substantial.
Detection Accuracy: The Probabilistic Limit
Trained classifiers for synthetic media detection — the primary technical tool for identifying AI-generated content — operate with false-positive rates of 5–15% under controlled conditions. In real-world deployment, where content has often been edited, compressed, or re-encoded, accuracy degrades significantly. Classifiers are considered unreliable for evidentiary purposes by leading researchers, and their use in high-stakes contexts — journalism, legal proceedings, electoral integrity — carries significant risk of both false accusations and false exonerations.
The fundamental limitation of detection-based approaches is the arms race dynamic: as detection tools improve, generation tools adapt to evade them. This dynamic has led the research community and major policymakers to conclude that detection alone cannot secure the information ecosystem.
The 'liar's dividend' is the most insidious consequence of synthetic media: it does not require a single convincing deepfake to destroy trust in audiovisual evidence — it only requires the widespread belief that such fakes are possible.
The Provenance Pivot: C2PA and Content Credentials
What C2PA Does
The Coalition for Content Provenance and Authenticity (C2PA) standard — now at version 2.4 as of April 2026 — provides the technical foundation for the provenance-based approach to information integrity. C2PA attaches a tamper-evident, cryptographically signed manifest to a media file, recording the creator, tools used, and edit history. The manifest travels with the file, allowing any verifying party to check the content's provenance without contacting the original creator.
Major camera manufacturers — Sony, Nikon, Leica, and Canon — now embed C2PA-signing firmware, allowing for the creation of cryptographically signed manifests at the moment of capture. This means that authentic photographic and video content can, in principle, carry a verifiable chain of custody from the camera sensor to the viewer.
Adobe's Content Credentials initiative, integrated into Photoshop, Lightroom, and other creative tools, extends C2PA to the editing workflow — recording not just the original capture but subsequent modifications, allowing viewers to see what changes were made to an image and by whom.
The Stripping Problem: C2PA's Critical Limitation
The most significant challenge facing C2PA adoption in 2026 is metadata stripping. Social media and distribution platforms routinely transcode files during upload and distribution — a process that often strips C2PA manifests and breaks the chain of custody. A photograph that leaves a camera with a valid C2PA manifest may arrive at a viewer's screen with no provenance data at all, having passed through a platform's compression pipeline.
C2PA verifies history and authorship, not truth. A malicious actor can sign content with a valid certificate. Provenance is a necessary condition for information integrity, but it is not a sufficient one.
This is not merely a technical problem. It reflects a structural misalignment between the incentives of content distribution platforms — which prioritise speed, storage efficiency, and user experience — and the requirements of provenance preservation. Addressing it requires either technical changes to platform pipelines (which platforms have been slow to implement) or the development of provenance systems that are resilient to transcoding (which is technically challenging).
A secondary limitation is coverage: C2PA and proprietary watermarks only function when the generation tool supports them. Midjourney, one of the most widely used AI image generation platforms, notably lacked C2PA integration as of early 2026. The standard's effectiveness is therefore bounded by the adoption decisions of individual platforms and tools.
Invisible Watermarking: SynthID and the Resilience Advantage
Google's SynthID, which embeds persistent signals into the pixel or audio waveforms of synthetic content, offers a different approach to provenance. Unlike C2PA metadata, which travels as a separate manifest that can be stripped, SynthID signals are embedded in the content itself — surviving common modifications such as compression, cropping, and re-encoding.
The resilience advantage of watermarking over metadata is significant. But watermarking has its own limitations: it is generally limited to specific proprietary systems (SynthID works for Google-generated content), it cannot be applied retroactively to content that was generated before the watermarking system was implemented, and it does not provide the rich provenance information — creator identity, edit history, tool chain — that C2PA manifests contain.
The emerging consensus among researchers and policymakers is that neither C2PA nor watermarking alone is sufficient. A layered approach — combining C2PA manifests for rich provenance data, invisible watermarking for resilience to metadata stripping, and perceptual fingerprinting for platform-level hash matching — provides the most robust available defence against synthetic media manipulation.
The Regulatory Landscape: A Study in Jurisdictional Fragmentation
C2PA verifies history and authorship, not truth. A malicious actor can sign content with a valid certificate. Provenance is a necessary condition for information integrity, but it is not a sufficient one.
The EU AI Act: Article 50 and the Rights-Based Framework
The EU AI Act's Article 50, which entered force on 2 August 2026, represents the most significant regulatory milestone in information integrity governance to date. The provision mandates that providers of generative AI systems mark synthetic outputs — audio, image, video, and text — in a machine-readable and detectable format. Systems already on the market have a grace period until 2 December 2026 to ensure compliance.
The EU's approach is explicitly rights-based, treating information integrity as a prerequisite for democratic health. The European Commission's June 2026 Code of Practice provides a voluntary framework for implementing the marking requirements, with the C2PA standard and SynthID-style watermarking identified as compliant approaches.
Enforcement capacity varies significantly across member states, and the definition of "machine-readable and detectable" leaves room for interpretation. But the EU framework establishes a clear normative baseline: synthetic content must be identifiable as such, and the burden of disclosure falls on the generator, not the viewer.
The regulatory landscape for synthetic media in 2026 is a study in jurisdictional fragmentation: the EU has a rights-based framework, India has accelerated takedown windows, and the United States has 31 different state laws and no federal standard.
India: Accelerated Takedown and the SGI Definition
India's 2026 amendments to the IT Rules define "Synthetically Generated Information" (SGI) and impose strict, accelerated takedown windows for high-risk content — in some cases as short as three hours. The Indian approach prioritises speed of removal over due process, reflecting a regulatory philosophy that treats the harms of synthetic disinformation as more urgent than the risks of over-removal.
The three-hour window has been criticised by civil liberties organisations as incompatible with meaningful content review, and concerns have been raised about the potential for the framework to be used to suppress legitimate political speech. But it represents a clear regulatory commitment to treating synthetic media as a category of content requiring special governance attention.
The United States: 31 State Laws and No Federal Standard
The US regulatory landscape for synthetic media is characterised by a patchwork of state-level laws and the absence of federal action. A June 2026 report from the National Conference of State Legislatures identified 31 states with laws addressing election-related deepfakes — a significant legislative response, but one that creates a complex compliance environment for national actors and leaves significant jurisdictional gaps.
State approaches fall into two broad categories. Disclosure models — adopted by 28 of the 31 states — require AI-generated political content to carry a clear warning, but permit the content itself. Louisiana's law, for example, requires disclosures for campaign advertisements and robocalls, imposing civil fines of up to USD 2,500 for a first offence. Prohibition models — adopted by Maryland, Minnesota, and Texas — go further, prohibiting certain deepfakes regardless of whether they include an AI warning. Maryland's law, effective 1 June 2026, allows for criminal penalties of up to five years in prison or a USD 5,000 fine for individuals who knowingly distribute deepfakes with the intent to influence a voting decision.
Constitutional challenges have already complicated enforcement. In September 2026, a federal judge temporarily blocked Montana from enforcing its deepfake law, citing potential First Amendment violations. The tension between information integrity and free speech protections is likely to generate ongoing litigation as the 2026 midterms approach.
The bipartisan "Protect Elections from Deceptive AI Act" has remained in committee since 2023, and prospects for federal action before the midterms are limited. The absence of a federal standard creates regulatory arbitrage opportunities and leaves voters in states without deepfake laws without statutory protection.
The Liar's Dividend: The Most Insidious Consequence
The most structurally damaging consequence of the synthetic media era is not the deepfakes themselves — it is the "liar's dividend." The term, coined by legal scholars Bobby Chesney and Danielle Citron, describes the use of synthetic media's existence to dismiss authentic evidence as fabricated. When the public knows that convincing deepfakes are possible, bad-faith actors can claim that genuine, incriminating recordings are synthetic — and a sceptical audience may be unable to distinguish the claim from the truth.
The regulatory landscape for synthetic media in 2026 is a study in jurisdictional fragmentation: the EU has a rights-based framework, India has accelerated takedown windows, and the United States has 31 different state laws and no federal standard.
The 'liar's dividend' is the most insidious consequence of synthetic media: it does not require a single convincing deepfake to destroy trust in audiovisual evidence — it only requires the widespread belief that such fakes are possible.
The liar's dividend creates an asymmetric epistemic environment in which the burden of proof shifts from those making false claims to those asserting authentic ones. Journalists, legal practitioners, and electoral officials who possess genuine evidence of wrongdoing face the additional burden of proving that their evidence is not synthetic — a burden that is technically complex, time-consuming, and not always possible to discharge conclusively.
Research from 2025 suggests that warning labels on synthetic content can foster scepticism — but do not necessarily deter users from sharing misleading content. Furthermore, labelling can inadvertently cause users to doubt legitimate information if they perceive "AI-labelled" content as inherently inaccurate. The psychological dynamics of synthetic media exposure are complex and not yet fully understood.
The "Cheap Fake" Reality: A Calibration Note
A critical calibration note for policymakers and researchers: experts consistently observe that non-AI misinformation — "cheap fakes" such as out-of-context clips, simple edits, and misleading captions — continues to be used significantly more often than advanced AI-generated content in electoral and political contexts. The focus on sophisticated deepfakes, while important for long-term governance planning, risks diverting attention and resources from the more immediate and prevalent threat of low-tech manipulation.
This does not diminish the importance of synthetic media governance. But it suggests that information integrity frameworks need to address the full spectrum of manipulation techniques — from the sophisticated to the mundane — rather than focusing exclusively on the most technically impressive threats.
Newsroom Verification and Media Literacy: The Human Layer
Technical standards and regulatory frameworks address the infrastructure of information integrity. But the human layer — the journalists, editors, and citizens who consume and share information — remains critical to the effectiveness of any governance architecture.
Leading newsrooms have developed verification protocols that combine technical tools — C2PA manifest checking, SynthID verification, reverse image search — with traditional journalistic methods: source corroboration, contextual analysis, and expert consultation. The Reuters Institute and the International Fact-Checking Network have published guidance on synthetic media verification that is being adopted by newsrooms across multiple jurisdictions.
Media literacy initiatives — designed to equip citizens with the skills to critically evaluate digital content — have received increased funding from governments and foundations in 2026. But the evidence base for the effectiveness of media literacy interventions remains mixed, and the scale of the challenge — billions of people consuming digital content across hundreds of platforms — dwarfs the capacity of any educational programme.
Conclusion: The Provenance Imperative and Its Limits
The information integrity landscape of 2026 is defined by a clear strategic consensus — provenance over detection — and a set of significant implementation challenges. C2PA Content Credentials and invisible watermarking represent the most structurally sound technical responses to synthetic media, but metadata stripping, incomplete platform adoption, and the fundamental limitation that provenance verifies authorship rather than truth leave significant gaps in the architecture.
The regulatory landscape is fragmented in ways that create both compliance complexity and governance gaps. The EU's rights-based framework is the most comprehensive, but its enforcement capacity is uneven. The US patchwork of state laws is better than nothing, but it is not a substitute for a federal standard. India's accelerated takedown approach prioritises speed over due process in ways that raise civil liberties concerns.
The liar's dividend — the use of synthetic media's existence to undermine trust in authentic evidence — is the most structurally damaging consequence of the deepfake era, and it is one that technical standards alone cannot address. Rebuilding trust in audiovisual evidence requires not just provenance infrastructure but institutional credibility — the credibility of the journalists, legal practitioners, and electoral officials who assert that evidence is authentic.
That credibility, in turn, depends on the broader health of the information ecosystem: on media institutions that are trusted, on regulatory frameworks that are consistent and enforceable, and on a public that has the skills and disposition to critically evaluate digital content. The provenance imperative is real and urgent. But it is a necessary condition for information integrity, not a sufficient one.
