What Is Truth in a Post-Photography World?

In March 2026, the National Republican Senatorial Committee released an online ad featuring a minute-long video of Democratic Senate candidate James Talarico speaking into the camera, reading statements the real Talarico had not spoken on camera. The Talarico in the video was generated entirely by artificial intelligence, voicing content drawn from the candidate's old social media posts. The words "AI Generated" appeared in small text in the corner of the frame at the start, then faded into even smaller text that remained on screen while the fake Talarico continued to speak. 

CNN reported that this was the first political deepfake in which a candidate is realistically recreated for an entire minute-plus clip, a leap beyond the seconds-long fakes that had dominated earlier election cycles. Within days, the ad had circulated widely, and the real Talarico was forced to respond to a fabricated video of himself saying things he had not spoken on camera.

The same month, synthetic video clips purporting to show Iranian missile strikes on Tel Aviv went viral on X. The clips were widely shared and viewed millions of times, circulating as real footage. When users asked AI tools and platform systems to verify the clips, some returned confident but wrong assessments, compounding the confusion. Authentic footage of Prime Minister Benjamin Netanyahu from the same conflict was misidentified as fake by large numbers of viewers. The real videos and the fake videos became indistinguishable at the moment of viewing, and the platforms hosting both had no functioning mechanism to separate them.

These are not anomalies. Estimates of deepfake content circulating globally rose from approximately 500,000 files in 2023 to more than 8 million in 2025. The technology has moved from research curiosity to commodity tool in less than three years. A political operative in 2026 can generate persuasive synthetic video of a public figure cheaply and quickly, using tools that are vastly more accessible than they were a few years ago. The question that the photography industry has spent twenty years asking (is this image real?) has become a question with no reliable answer, and the consequences of that reality are only beginning to play out across law, politics, journalism, and every other domain that ever relied on photographic evidence.

The Problem C2PA Cannot Solve

The Coalition for Content Provenance and Authenticity, or C2PA, is the photography industry's answer to this crisis. The standard allows cameras and editing software to cryptographically sign images and video, embedding a tamper-evident record of how a file was created and modified. The Content Authenticity Initiative that promotes C2PA adoption has grown to more than 6,000 members as of early 2026. Leica shipped the first C2PA-enabled camera in 2023 and followed with the SL3-S in January 2025. Sony's PXW-Z300 was announced in 2025 as a camcorder built for video authenticity workflows, and Sony's C2PA-compliant video authenticity solution launched on October 30, 2025, supporting the PXW-Z300 alongside selected a-series and FX bodies (the a1 II, a9 III, FX3, and FX30 at launch, with the a7R V, a7 IV, a1, and a7S III planned in subsequent phases) and requiring an Authenticity Upgrade License to activate. Canon added C2PA support to the EOS R1 and EOS R5 Mark II via firmware in 2025, with access initially restricted to registered news agencies. Fujifilm announced in 2024 that it would integrate Content Credentials across its GFX and X series line-up, including the GFX100S II, though a shipping implementation had not yet been confirmed on its cameras as of early 2026. The Samsung Galaxy S25 added Content Credentials support for AI-edited and AI-generated content in early 2025. Nikon and Panasonic have also joined the Content Authenticity Initiative and are working on implementations. LinkedIn displays Content Credential icons on signed images. TikTok adopted Content Credentials for AI-generated content labeling.

The technical infrastructure is real and the progress is genuine, and the technical infrastructure will not solve the problem.

The first reason is that many social media platforms strip metadata during upload and re-encoding. A photograph signed by a Leica M11-P, demonstrating a verifiable chain of custody from capture to publication, loses its C2PA manifest the moment it is uploaded to most social platforms. The cryptographic proof exists at the source and is severed before it reaches any audience at scale. The Content Authenticity Initiative has proposed Durable Content Credentials that combine watermarking and fingerprinting to survive stripping, but the implementation remains partial and platform-dependent. As of April 2026, the practical experience of a viewer encountering an image on Instagram or X is that the image carries no verifiable provenance, regardless of whether the photographer who captured it signed it at the camera.

The second reason is that C2PA only covers cameras and editing tools that have implemented it, and adoption is concentrated at the high end. Canon's entry-level Rebel and R-series consumer bodies do not support C2PA. Nikon's Z5 and Z30 do not support it. Sony's a6000-series does not support it. The vast majority of photographs taken in the world are made by smartphones and entry-level cameras that either do not sign images at all, or sign only a subset of images under specific conditions. Samsung's Galaxy S25 implements Content Credentials primarily for AI-edited and AI-generated content rather than for all native camera captures. Google's Pixel 10 attaches Content Credentials to JPEG photos created by Pixel Camera, while Google Photos applies or displays Content Credentials in more limited editing and viewing contexts. The Content Credentials ecosystem is structurally biased toward professional equipment, flagship smartphones, and specific AI-generation workflows, and the photographs that shape public opinion at scale are often made with hardware and software that fall outside this coverage.

The third reason is infrastructural. C2PA signing certificates must be obtained from a small number of commercial certificate authorities, and there is no free or low-cost equivalent to Let's Encrypt that would let a working photojournalist sign their own independent work at scale. The Nikon Z6 III had C2PA support added via firmware in August 2025, then discovered a signing vulnerability; Nikon revoked the compromised certificates and suspended the service, and as of early 2026 the signing capability had not been restored. The trust layer underneath C2PA is itself fragile, and the fragility has already manifested in ways that suggest it will manifest again.

The fourth and most important reason is that C2PA verifies the recorded provenance of a file. It can document what tool created a file, who (or what) signed it, and what edits have been recorded in the chain of custody since. It does not verify that the image is true. A C2PA-signed photograph of a staged scene is cryptographically authentic and substantively false. A C2PA-signed photograph taken with a caption that misdescribes what it shows is cryptographically authentic and propagandistically useful. The standard certifies the history of a file. It does not certify the history of the world the file depicts. This is a limitation the architects of C2PA have always been clear about, and it is a limitation that most discussions of the standard gloss over.

The Liar's Dividend

The deeper problem is not that deepfakes will be mistaken for real images. The deeper problem is that real images will increasingly be dismissed as deepfakes, and there will be no reliable way for the person being accused of faking them to prove otherwise.

Legal scholars Bobby Chesney and Danielle Citron coined the term "liar's dividend" to describe this phenomenon. The term has moved from academic discourse into active legal defense strategy. In January 6 prosecutions, defendants have suggested that video evidence of their actions could be AI-generated. In litigation involving Tesla and Elon Musk, attorneys have argued that past statements captured on video should be excluded because they might be deepfakes. In police misconduct cases, the specter of AI-generated video is being raised by defense attorneys as a way to undermine body-camera footage. In each case, the argument does not require the defense to prove that the video is fake. The argument only requires the defense to raise enough doubt that a jury cannot be sure the video is real.

That doubt is easy to raise in 2026. Any juror who has spent time on social media in the past year has seen convincing deepfake video. Any juror who has read about the Talarico ad or the Iran war synthetic footage has absorbed the idea that video evidence is no longer reliable by default. A skeptical juror is a juror primed to acquit, and the defense bar has noticed. The proposed Federal Rule of Evidence 707, which addresses the admissibility of machine-generated evidence in federal court, is aimed at evidence offered as machine-generated and at establishing reliability standards for such evidence. It does not directly address the separate and larger problem of authentic evidence being attacked as fake. The rule is a useful technical clarification and does not touch the structural crisis underneath it.

The implications extend beyond criminal court. Insurance fraud investigations rely on photographic and video evidence. Journalism relies on it. Human rights documentation relies on it. Scientific publication relies on it. Historical record-keeping relies on it. Every one of these domains is now operating in an environment where the evidentiary weight of photography and video has dropped, possibly permanently, and no technical standard can unilaterally restore it. The shift is not that photography has become untrustworthy in absolute terms. The shift is that the default assumption has moved from "this image is real unless proven otherwise" to "this image may or may not be real, and proving either case is expensive and contested."

The Political and Legal Landscape

The political response to this crisis has been both slow and fragmented. As of April 2026, twenty-nine U.S. states have enacted laws regulating deepfakes in political messaging, according to the National Conference of State Legislatures. There is no federal legislation specifically regulating the use of AI-generated content in political campaigns, though existing federal laws governing fraud, impersonation, campaign finance, and related conduct may still apply to certain deepfake use cases depending on the facts. California's attempts to regulate election deepfakes have repeatedly fallen in court. AB 2839, which allowed individuals to sue for damages over election deepfakes, was first blocked in October 2024 by a federal judge who ruled it was likely unconstitutional under the First Amendment, and was struck down in a separate ruling in early September 2025. AB 2655, the companion law requiring platforms to remove election deepfakes, was struck down on August 5, 2025 on the grounds that it conflicted with Section 230 of the Communications Decency Act, which shields platforms from liability for user-generated content. Minnesota enacted a similar ban in 2023 and is currently defending it in court against a challenge from X.

The First Amendment complications are real and not easily dismissed. Federal judges have repeatedly sided with First Amendment challenges in this space, ruling that requirements forcing platforms to label or remove political synthetic content sweep too broadly to survive constitutional review. Satirical deepfakes and parody are constitutionally protected. Disclosure requirements that work for commercial advertising run into different constitutional terrain when applied to political speech. A purely legal solution to the deepfake crisis in elections will almost certainly fail on First Amendment grounds, and the legal scholars watching this space have been explicit about that.

The European Union has moved more aggressively. EU AI Act Article 50 enforcement begins on August 2, 2026. The article imposes a set of transparency duties on providers and deployers of AI systems, including marking certain AI-generated or AI-manipulated content such as deepfakes and informing users when they are interacting with AI systems, though the specific obligations vary depending on the type of content and the context in which it is produced, and the article contains exceptions for certain uses. The EU's draft Code of Practice favors a stacked approach to marking AI-generated content, combining cryptographic metadata with watermarking as the core mechanisms, and treating fingerprinting and logging as supplementary tools that can be used for detection and verification where appropriate. The European Commission's explicit acknowledgment that metadata marking alone is insufficient is itself notable. The regulators who have thought about this problem most carefully have concluded that a single technical standard cannot solve it, and that any functional response will require several layers of complementary enforcement working in concert.

What This Means for Photographers

For working photographers, the implications are concrete and sobering. The value of a photograph as evidence is declining, and the decline is structural rather than temporary. A wedding photographer whose client disputes a delivered image can, in theory, be accused of AI-generating portions of the gallery. A photojournalist who captures a genuinely important scene faces a world in which their work may be dismissed as synthetic by audiences who have lost the ability to distinguish. A fine art photographer exhibiting work in a gallery faces a collector class that is increasingly skeptical of provenance claims across all creative categories. Each of these scenarios is already playing out in smaller ways, and the trend lines are not favorable.

The pragmatic response for working photographers is to adopt C2PA where it is available, document capture processes as rigorously as the equipment allows, and understand clearly what the standard does and does not guarantee. A C2PA-signed file is more defensible than an unsigned file. It is not a defense against the broader collapse of photographic epistemology, and no photographer should market their work to clients as though it were.

The deeper question is whether photography survives as a medium of truth at all. The medium emerged in the 1830s and 1840s with a claim that Roland Barthes would later name: the photograph was evidence of the "that-has-been," proof that a specific thing existed at a specific moment. That claim undergirded almost two centuries of photography's cultural authority. It underwrote forensic photography, photojournalism, family snapshots, evidentiary documentation, and the basic social agreement that a photograph could be treated as a witness. The claim was never absolute. Photographs could always be staged, cropped, dodged, burned, or manipulated. But the claim held up well enough for long enough that most people, most of the time, could treat a photograph as a reasonable approximation of reality.

That claim is now being revoked in real time, and the revocation is not temporary. Generative AI is not a fad that will fade. The tools will become more capable, more accessible, and more ubiquitous every year. The technical response (C2PA, watermarking, detection tools) will lag behind the generative capability because the generative side has structural advantages in the arms race. The social response (regulation, platform policies, cultural norms) will lag behind the technical response because politics moves slower than technology. The combined lag means that photographers working in 2026 and beyond are operating in a world where the medium they have dedicated their careers to is being systematically delegitimized as a source of truth, and no one is going to fix this problem in any definite way, any time soon.

What remains for photography is what always remained, once the evidentiary claim is set aside. A photograph can still be beautiful. It can still communicate a photographer's point of view. It can still document something specific enough to be compelling as personal record. It can still have value as craft. What it cannot do, in the way it used to be able to, is function as unimpeachable proof. The photographers who will navigate the next decade most effectively are the ones who understand this shift clearly, who adopt the technical tools that exist without overclaiming what they can do, and who recognize that the cultural work of deciding what to believe has moved permanently out of the hands of the camera and into the hands of the audience, the platform, the court, and the political actor. The camera is one voice in that conversation now. It used to be the final word. It is not anymore.