The camera never lies: Until now

By Immanuel Boama-Wiafe

“The security systems your government, the police, and even you yourself trust to protect you, whether it is facial recognition at the border, CCTV in the courtroom, the voice on the phone, or the ring camera in your home, are being quietly, invisibly compromised. The tools to do it are free. The window to stop it is closing.”

In September 2023, a recording surfaced online in Slovakia, featuring a voice instantly recognisable as a senior opposition politician calmly discussing how to rig an upcoming election with what appeared to be a foreign intelligence agent. The conversation was detailed enough to be believable, and damaging enough to spread rapidly, reaching tens of thousands of people within 48 hours as the country approached the polls.

By the time fact-checkers established that the audio was entirely fake, generated through a voice clone trained on publicly available YouTube clips, the election had already taken place. The political outcome had shifted. And a new playbook had been quietly validated, one that has since been refined and reused across multiple elections on different continents.

What this incident reveals is not merely the emergence of deepfake technology, but a structural shift: the systems used to validate truth operate more slowly than those used to distribute falsehoods. The damage is done before the correction can catch up.

The Illusion of Accuracy

Imagine walking through border control at a major airport. A camera scans your face. Within seconds, the system returns a match and reports 99.36 percent accuracy, a figure that governments, procurement teams, and vendors rely on as a signal of precision, reliability, and control.

But these systems do not understand identity in any human sense. They match patterns. And pattern-based systems can be manipulated without touching their overall performance metrics.

Research has demonstrated that hidden backdoors can be embedded within facial recognition systems, allowing them to function normally under standard conditions while producing controlled misidentifications when specific triggers are introduced, whether subtle visual patterns, accessories, or image perturbations imperceptible to most observers.

More concerning still are hardware-level attack methods, where minimal alterations at the memory level can introduce persistent vulnerabilities that survive conventional testing and certification. A system that appears secure, audited, and high performing may still be susceptible to manipulation under specific conditions, with no visible indication of compromise. High accuracy, in other words, is not the same as high security.

When evidence itself becomes uncertain

Now consider the CCTV footage that underpins criminal trials, public inquiries, and insurance disputes worldwide. Juries tend to trust it instinctively. The assertion that “the camera saw it” can shift the burden of credibility almost entirely onto the defence. That trust is no longer as solid as it appears.

Security researchers have identified a technique called adversarial video stitching, a method that does not fabricate entire scenes, but introduces precise, nearly undetectable alterations at critical moments. The attack begins by identifying a natural interruption in footage, typically when a large vehicle passes through the frame and briefly obscures the camera’s view. That window may last only a few seconds, but it is enough.

Within that gap, AI-assisted frame synthesis inserts an alternative sequence, carefully aligned with the original in lighting, motion, and audio continuity. Under normal playback, the result appears seamless and authentic to the human eye. An innocuous interaction becomes a violent incident. An empty street becomes an active crime scene.

In more advanced scenarios, the manipulation occurs at the point of recording itself, meaning no unaltered version of the footage ever exists. Investigators are left with no baseline, no original, and no definitive way to reconstruct what actually happened.

This vulnerability does not even require physical access to the recording device. Most camera systems described as “live” operate with an inherent delay, as video data must be compressed, encrypted, and transmitted through network infrastructure before it is stored. An attacker who intercepts the stream during that window can alter or replace frames before they are written to storage, particularly where video data is not cryptographically signed at the frame level. What operators perceive as a live feed is, in fact, a delayed representation of events. What ends up stored may already have been modified.

A 2025 criminal investigation in Indonesia brought this vulnerability into sharp relief. A young diplomat was found dead under suspicious circumstances. When police reviewed the CCTV, they discovered two recordings from the same camera showing different angles, one capturing the diplomat’s door and window, and one from the critical night that did not. The camera had been physically moved. The caretaker admitted it. The question was straightforward: routine maintenance, or evidence tampering?

The case exposes a gap that exists in legal systems worldwide. Courts rarely require frame-by-frame authentication before admitting video as evidence. If footage looks real, it is generally trusted. That needs to change.

The question is no longer whether video can be trusted, but under what conditions and with what safeguards that trust can still be justified.

The Verification Gap

The challenge extends beyond security systems and legal evidence. It sits at the heart of how information moves.

In May 2023, an AI-generated image showing an explosion near the Pentagon was shared by verified accounts, amplified by news aggregators, and briefly reported as a developing story. It was fabricated. In the minutes between first circulation and correction, the S&P 500 registered a measurable dip.

This is the core of the problem: identifying synthetic content requires time and verification tools, while sharing it requires only seconds. The imbalance is structural, not accidental. In digital ecosystems that prioritise engagement, corrections rarely achieve the same reach or velocity as the original falsehood. False information frequently completes its lifecycle before truth can effectively intervene.

A Narrow Window for Action

These developments are not beyond resolution. But they require coordinated action within a limited timeframe. Based on current trajectories, the window for meaningful intervention is measured in years, not decades.

Standards for cryptographic provenance, such as those developed by the Coalition for Content Provenance and Authenticity, already provide mechanisms for verifying the origin and integrity of digital media. They remain largely voluntary. For systems used in evidential or security-critical contexts, they should be mandatory.

In the meantime, there are things individuals and institutions can do today, at no cost.

Before you share a video or audio clip that provokes a strong emotional reaction, outrage, fear, or triumph, pause and ask where it came from and when it was recorded. The most dangerous synthetic media is engineered to travel in three minutes before anyone thinks to ask.

1. Use the tools that already exist. Google Reverse Image Search, TinEye, and InVID/WeVerify can surface the original source and publication date of media in seconds. They are free. They are almost never used by anyone outside professional fact-checking circles.
2. Ask your elected MPs direct questions. What is your government’s position on biometric backdoor testing, cryptographic provenance mandates, and synthetic evidence in courts? If they cannot answer, they do not come across a security risk that is already operational.
3. For journalists and editors, the defence of “we reported what we had at the time” no longer holds as a complete answer for amplifying synthetic media. The verification tools exist. Using them is now a professional obligation.

The architecture of truth is infrastructure. We built roads, power grids, and financial clearing systems because we understood that shared life depends on shared systems that work. We did not wait for catastrophic failure before building them. We are waiting for catastrophic failure before rebuilding the infrastructure of truth, and the failures are already here.

Rebuilding the Infrastructure of Truth

At its core, this is not a technology problem. It is an infrastructure problem.

Modern societies depend on shared systems that deliver consistent, reliable outcomes in transportation, energy, and finance. The ability to establish and verify truth is becoming part of that same category of critical infrastructure. Yet unlike roads, power grids, or financial clearing systems, it is being stress-tested before it has been secured.

“We built those other systems because we understood that shared life depends on shared systems that work. We did not wait for catastrophic failure before building them. We are waiting for catastrophic failure before rebuilding the infrastructure of truth, and the failures are already here.”

The consequences are already visible across elections, legal systems, and financial markets, where synthetic and manipulated content has influenced outcomes in ways that are difficult or impossible to reverse.

The camera was once considered a neutral witness. That assumption is gone. What replaces it will define the next phase of digital trust.

Immanuel  is a Certified Information Systems Auditor and Operational Management Lead focused on AI security, cybersecurity policy, and think tank research. He brings expertise in auditing, risk management, and building operational resilience across organizations.