When Seeing Is No Longer Believing: The Problem AI Deepfakes Leave Behind

Imagine watching a video of a public figure admitting to a crime. You recognize their face. You hear their voice. Their expressions look natural, and the recording appears to show exactly what happened. You share it with someone. Within minutes, it spreads across social media. Then someone asks a question that changes everything: What if the video was never real? A few years ago, creating a convincing fake video required considerable expertise, time, and resources. Today, generative AI is lowering those barriers. Synthetic voices, realistic faces, and fabricated scenes can be produced with tools that are becoming easier to access. The problem is no longer limited to whether people can create convincing fakes. It is whether the rest of us can reliably distinguish fabricated evidence from genuine recordings.

On October 15, 2026, representatives from OpenAI and Meta warned a US judicial committee that advances in generative AI are making it increasingly difficult to identify fabricated audio and video submitted as evidence. The discussion highlighted a serious problem: existing detection methods and watermarking approaches do not provide a universal, foolproof solution. The concern extends beyond social media misinformation. Courts depend on evidence to establish what happened, who was involved, and whether an account of events is credible. A recording can be powerful evidence. But if convincing recordings can be fabricated, investigators and judges need ways to assess authenticity beyond simply watching a video and deciding whether it looks genuine. This does not mean courts can no longer trust video evidence. It means that the methods used to establish its authenticity may need to become more sophisticated. Source: Reuters — OpenAI and Meta caution US judiciary about deepfake evidence.

Most discussions about deepfakes focus on detection. Can an AI detector identify a synthetic face? Can software recognize an artificially generated voice? Can an algorithm spot inconsistencies in lighting, movement, or facial expressions? These questions matter, but they address only one part of the problem. Detection attempts to determine whether a piece of media is fake by examining the content itself. That becomes difficult when generated media improves, when a file has been compressed or edited, or when a detection system encounters a technique it was not designed to recognize. Even a convincing detector cannot guarantee that every fake will be identified. A result indicating that no manipulation was detected is not the same as proof that no manipulation occurred. There is another problem: genuine recordings can also be edited, taken out of context, or presented misleadingly. A real video can communicate a false impression without containing a single artificially generated frame. Authenticity, therefore, involves more than asking whether pixels or sounds were generated by AI. It also involves understanding where the recording came from, whether it has been altered, and what it actually shows.

Consider two videos showing the same event. The first arrives through an anonymous social media account. There is no reliable information about who recorded it, where it originated, or whether it has been modified. The second comes with verifiable information about its origin and subsequent edits, supported by a trustworthy recording or publishing process. Neither file should automatically be accepted as true. But the second gives investigators more information to work with. This is the idea behind content provenance: documenting information about a digital file's origin and history so that people and software can examine it later. One relevant initiative is C2PA, an open technical standard for recording information about the origin and editing history of digital content. Its associated Content Credentials can help establish how a file was created or modified. Imagine a camera recording that carries verifiable information about the device and its capture process. If an editor subsequently modifies the footage, compatible tools could record details about that change. A recipient could then inspect the available history instead of relying entirely on visual judgment. This is not a magic authenticity certificate. Provenance information can be incomplete, credentials can be removed, and a file with a trustworthy history can still depict a misleading scene. The identity of the person operating a camera and the truth of what happened may also remain unresolved. But a verifiable history provides something that appearance alone cannot: additional evidence about where a file came from and what happened to it.

AI watermarks are another part of the response. Some systems embed signals into generated images or audio. Others attach machine-readable information describing how content was produced. Verification tools can look for supported signals and help determine whether a file is associated with a particular AI system. These approaches are useful, but their limitations matter. Metadata can disappear when a file is uploaded, downloaded, converted, or edited. Watermarks may survive some transformations but become harder to detect after substantial modifications. There is also a coverage problem. A detector designed to recognize signals from one provider cannot necessarily identify content produced by every other AI system. Consider what happens when a detector finds no watermark. That result does not establish that the content was created by a human. The file might have been generated by an unsupported system, or the relevant signal might have been removed or damaged. OpenAI's own documentation makes this distinction: provenance signals can help identify the origin of supported content, but they do not establish that the content is accurate or presented in the correct context. Read more: OpenAI — Provenance signals in generated content. The lesson is straightforward: no single watermark, detector, or metadata format should be treated as a universal test of truth.

There are two ways this technology can damage trust. In the first, fabricated evidence is accepted as genuine. A fake recording could damage someone's reputation, influence public opinion, or complicate a legal investigation. In the second, genuine evidence is dismissed as fabricated. Imagine someone presenting an authentic recording of misconduct. The person accused responds by claiming that the footage was generated by AI. If the audience has no reliable way to evaluate the claim, uncertainty itself becomes a defence. This creates a troubling possibility: the existence of convincing synthetic media could weaken the credibility of genuine recordings, even when those recordings have not been manipulated. The consequences reach journalism, elections, workplace disputes, online fraud, and personal relationships. A world filled with deepfakes does not simply make people more likely to believe false things. It can also make them less willing to believe true things. That is why the problem is bigger than misinformation. It concerns the infrastructure of trust that allows people to use digital evidence in the first place.

There is unlikely to be one tool that solves this problem. A more reliable system would combine several approaches.

• Provenance by design: Cameras, recording applications, and publishing platforms could preserve verifiable information about where media originated and how it was edited.
• Independent verification: Investigators and journalists could examine original files, available metadata, recording circumstances, and corroborating evidence rather than relying on a single AI detector.
• Transparent AI labelling: Platforms and developers could make it easier to identify synthetic content and understand the limitations of their verification methods.
• Better evidence handling: Courts, newsrooms, and organizations could preserve original files, document how evidence was obtained, and maintain records of subsequent transfers or modifications.
• Public media literacy: People could learn to distinguish between a video that looks convincing and a video whose source has been independently established.

These measures will not eliminate deception. They can, however, make it harder for fabricated content to pass as verified evidence and easier to investigate suspicious material responsibly.

For much of the internet's history, sharing a photograph or video required little explanation. The file itself was often treated as the primary evidence. That assumption is becoming less reliable. As AI-generated media improves, the important question may shift from "Does this look real?" to "What can we establish about where this came from?" The distinction matters because a convincing appearance is not proof of authenticity, and the absence of a detected watermark is not proof of human creation. A more trustworthy digital world will need better tools for recording provenance, examining evidence, and communicating uncertainty. It will also need people and institutions willing to verify claims instead of treating every viral recording as either unquestionable proof or an obvious fake. The goal should not be to distrust everything we see. That would allow deception to win in a different way. The goal should be to make trust something we can examine, support, and verify. In an age when almost anything can be made to look real, the most valuable technology may not be the one that creates the most convincing image. It may be the one that helps us understand whether we have a good reason to believe it.