Within the space of a few weeks, two of Germany's best-known media brands published AI-generated or AI-manipulated images. ZDF aired a fully synthetic video on its flagship news program heute journal, and Der Spiegel printed manipulated photos from Iran. In both cases the newsroom did not catch the fake itself. It reacted only after others pointed it out. That suggests less a run of isolated incidents than a structural problem: editorial image review was never designed for this kind of forgery.
3.6 million viewers, one AI video
On the Sunday evening of February 15, 2026, heute journal ran a report by New York correspondent Nicola Albrecht on ICE operations against suspected illegal immigrants in the United States. The report included a scene of a woman being led away by uniformed officers while two children wept in despair. It looked like documentary footage. It had been generated entirely by OpenAI's video generator Sora, and the tool's watermark was visible in the broadcast material.
A second error followed immediately after. A clip captioned as a current ICE operation actually dated from 2022 and showed an incident in Florida that had nothing to do with immigration authorities. ZDF deputy editor-in-chief Anne Gellinek called it a "double error". Around 3.6 million people watched the segment.
ZDF's response came in stages. The broadcaster first spoke of a "technical error" in the transmission of the report. In the days that followed, Gellinek delivered a public apology on heute journal itself. On February 21, correspondent Albrecht was recalled with immediate effect. Editor-in-chief Bettina Schausten said the damage caused by this "disregard of journalistic rules" was considerable. ZDF announced mandatory training on handling AI material, stricter checks for third-party footage, and a binding source hierarchy.
The political reaction was sharp. FDP deputy leader Wolfgang Kubicki spoke of a "media scandal". The debate over the credibility of public broadcasting, which has smoldered for years, gained fresh fuel.
The SalamPix supply chain
A few weeks later, in early March 2026, a second case came to light. Der Spiegel, Die Zeit, the Süddeutsche Zeitung, Stern, Deutsche Welle, WDR, Deutschlandfunk, Die Welt, the taz, and ZDF itself had published photos from the Iran conflict that were wholly or partly AI-generated. The images came from the Iranian agency SalamPix.
The Dutch agency ANP discovered the problem first and removed around 1,000 SalamPix images from its database in early March. The German forensics firm Neuramancer analyzed five suspicious images in detail. A photo of an alleged Iranian aircraft carrier showed illogical shadows and structural inconsistencies pointing to full AI generation. An image of an explosion in Tehran on March 1 carried traces of the AI image generator Flux 2 in its metadata. Portraits of the new Supreme Leader Mojtaba Khamenei alongside his father Ali Khamenei were likewise classified as probably AI-generated.
The images reached German newsrooms through a supply chain. SalamPix supplied the French agency Abaca Press, which passed material on to dpa Picture Alliance, Imago, and ddp. From there it flowed into the picture desks of the outlets named above. An Iranian photographer who worked with SalamPix admitted sourcing material from a platform run by Iran's Revolutionary Guards without checking its authenticity. The German agencies responded: dpa removed all SalamPix content, Imago blocked the supplier, and ddp sent out a "kill notice" ordering the deletion of every affected image.
Why editorial review fails
Both cases follow the same pattern. Experienced journalists and picture editors looked at the material and let it through. At ZDF, a Sora watermark was visible to the naked eye. At Der Spiegel, the metadata of one photo contained traces of an AI image generator. Even so, nothing stood out.
This is not the failure of individuals. It is the product of a system built on trust in sources and on visual judgment. Picture editors check EXIF data, run reverse image searches, and assess plausibility. Those methods date from an era when forgeries were laborious to produce and rare. Generative AI has dissolved that premise. A video generated with Sora looks, at first glance, like smartphone footage. A photo generated with Flux 2 has no obvious artifacts. EXIF data can be manipulated at will.
Agency supply chains make the problem worse. When an image travels from SalamPix through Abaca Press to dpa and from there to Der Spiegel, every station checks the material against its own standards. But none of them analyzes the image data forensically. Each relies on the previous link having been careful. And so an AI-generated photo passes through four sets of hands without anyone identifying it as synthetic.
The Bundestag knows the problem
That these incidents come as no surprise is clear from a report by the Office of Technology Assessment at the German Bundestag (TAB). In December 2025, authors Octavia Madeira and Steffen Albrecht published a 62-page analysis of the dangers posed by deepfakes. The report covers the technical foundations of generative models, the societal consequences ranging from political disinformation to image-based sexual abuse, and the legal challenges of regulation.
One central finding: Germany has no explicit statutory rules on deepfakes. A Bundesrat bill on the "criminal-law protection of personality rights against deepfakes" was reintroduced in the Bundestag in July 2025 and has sat in the legal affairs committee ever since. The German Federal Bar (BRAK) criticizes the draft as too broadly framed and warns of overcriminalization. There is no passage in sight.
The TAB report recommends a combination of technical solutions, media literacy, and legal frameworks. The ZDF and Spiegel incidents confirm what its authors described in theory: the technology has outpaced the existing safeguards.
What technical verification does differently
Both incidents would have been avoidable if the newsrooms involved had used technical verification that goes beyond visual inspection and metadata checks.
Forensic RAW verification compares a published JPEG against the original camera file. RAW files contain unprocessed sensor data that generative AI cannot reproduce. When a photographer submits the RAW file alongside the finished image and a verification platform cryptographically confirms that the two match, the question of authenticity is answered.
In the case of the ZDF video, no RAW file would have existed, because no camera ever recorded the scene. The video was entirely synthetic, and any verification request would have failed immediately. In the case of the SalamPix photos, a forensic analysis of the image data would have identified the same traces of generative models that Neuramancer found only after the fact. The only difference would have been the timing: before publication instead of after.
The C2PA standard provides the technical foundation for this approach. It enables a tamper-evident manifest that cryptographically binds capture data, editing history, and verification results to an image. Cameras such as the Leica M11-P already generate these manifests at the moment of capture. For the large majority of professional cameras that do not yet have this capability, Lumethic steps in at exactly this point: the platform verifies the match between the RAW file and the final image through forensic analysis and attaches a C2PA manifest to the verified JPEG.
For newsrooms this translates into a concrete workflow. Photographers and agencies deliver a verification report with every image. Picture editors review the report in seconds. The result is not a subjective judgment but a cryptographically secured proof.
What the AI Act covers and what it does not
With the AI Act (Regulation 2024/1689), the EU has already created transparency obligations for synthetic content. Article 50 requires that AI-generated images, videos, and audio files be labeled in machine-readable form. But the AI Act places the labeling duty on the AI providers. It gives newsrooms no tool for checking incoming images. Labeling assumes that the creator cooperates. In the SalamPix case, where material from a Revolutionary Guards platform was fed into the agency circuit, that is precisely what did not happen.
Forensic verification works differently, because it depends not on the creator's cooperation but on the physics of camera sensor data. A genuine photo has a RAW file with characteristic noise patterns, color interpolation, and sensor artifacts that an AI-generated image lacks. That distinction can be automated and built into editorial workflows.
After the incidents, both ZDF and Der Spiegel announced stricter internal processes. Whether that is enough is doubtful. Training courses and four-eyes reviews assume that a human can spot a good fake. The SalamPix images passed through four stations without raising a flag. As long as newsrooms rely on visual judgment, the risk remains.
Further reading: Guide to editorial photo verification | Image provenance vs. AI detection | EU AI regulation and content provenance
