Deepfake video detection may seem like the ultimate software problem: a computer examines a suspicious recording and decides whether the faces are real. But researchers have demonstrated a different approach, using light as part of the calculation. In a peer-reviewed paper published in September 2026, a hybrid optical and digital system examined fifteen video streams in parallel during an experimental test, according to the peer-reviewed original study.
The researchers reported an average accuracy of 97.79 per cent on a particular benchmark dataset in that laboratory configuration. That is not a promise that a future consumer device will recognise almost every fake video it encounters. It is evidence that a physical technique could make large-scale screening more efficient, provided it works reliably outside the controlled conditions of the experiment.
Deepfake video detection is becoming a race for efficiency
Making a plausible false video has become more accessible as image and video generation tools improve. For newsrooms, online platforms and ordinary people, the problem is no longer only recognising an obviously manipulated face. There may be many clips to examine, some heavily compressed or repeatedly uploaded through different services.
Conventional detection systems often use digital processors trained to recognise patterns that are common in altered footage. Those methods can be valuable, but analysing large volumes of material demands computing capacity, electrical power and careful handling of errors.
The new research from Parnian Ghapandar Kashani, Shiqi Chen and Aydogan Ozcan explores whether a portion of the work can be performed through the physical behaviour of light. That sounds exotic, but the principle is understandable: rather than asking a conventional chip to perform every computational step, the researchers arrange optical components so that some calculations emerge as light travels through the apparatus.
How a beam of light becomes a calculation
The system begins with digital processing that converts information about each video into a pattern suitable for the optical hardware. That pattern is displayed using a programmable device that changes how light passes through or reflects from different parts of its surface. The modified light then propagates through the optical set-up, and detectors measure the resulting signals.
The final signals are interpreted as predictions about whether the videos are genuine or manipulated. The important feature is parallelism: parts of several video streams can be represented and processed together in a single optical pass. This is not a camera that magically sees lies. There is still a digital front end, an optical stage, a trained model and an interpretation of the output.
The approach is technically known as a hybrid digital-analogue optical neural architecture. For general readers, the useful point is simpler. Certain physical processes can perform calculations without imitating every operation of an ordinary electronic processor. Researchers are testing whether that reduces the cost of screening large amounts of video.
The laboratory figure needs a careful translation
In their experiments the scientists tested a configuration processing fifteen streams at once. On the Celeb-DF face-manipulation dataset, they measured average accuracy of 97.79 per cent, sensitivity of 99.86 per cent and specificity of 95.72 per cent. Those terms refer to different aspects of how the test separated genuine and manipulated material within that dataset.
A person seeing the first percentage might think the detector will be right on nearly every video found on the internet. That would be an overstatement. Performance depends on which examples are tested, how the system was trained, the proportion of fake videos being examined and the kinds of manipulation used. A result on one benchmark is not an independently verified guarantee on every modern video generator or camera condition.
The paper includes experiments with other datasets and forms of manipulated material, but the results should be read in their specific contexts. The authors present their approach as a promising research architecture for screening, not as an infallible court witness.
Fifteen videos at once is the striking part
The idea of handling several streams in parallel may eventually matter more to large organisations than the exact figure from a single detection test. A platform receiving continuous video uploads could, in principle, benefit from screening many items without running an expensive full digital analysis independently on every one.
The earlier publicly available manuscript describes the same underlying research programme and reports the optical processing approach. The subsequent journal article offers a fuller account of experiments and comparisons. Neither publication means that a commercial service has already deployed this equipment widely.
Efficiency claims need the same caution as accuracy claims. An optical component may reduce energy or time in a particular stage while the complete system still requires digital preparation, memory, electronics and maintenance. A fair comparison must examine total workload and operating conditions. Research on lower-energy computation is worthwhile, but a photograph of a clever laboratory apparatus cannot establish the economics of a future product.
A detector does not know why someone made a video
A manipulated recording can be a joke, a film effect, an accessibility aid or a deliberate attempt to deceive. A technical detector attempts to identify characteristics of the material. It cannot by itself establish the creator’s intent, the surrounding context or whether a genuine recording is being described misleadingly.
Likewise, an authentic video can be edited into a false narrative by removing essential context. Tools that classify synthetic imagery are only one part of verification. Reporters may also need to identify the original uploader, inspect location clues, contact participants and compare the clip with independent evidence.
LiveAIWire’s coverage of Google’s synthetic-content detector explores another part of this effort: checking AI-generated material through detectable markers or provenance signals. Those systems answer related but different questions. A watermark can identify certain known generation processes, while a classifier tries to recognise characteristics of material even if no such marker is present.
False accusations can be as serious as missed fakes
Every classification system can make mistakes. A false negative allows a manipulated video to pass a test; a false positive marks genuine footage as suspicious. The latter can matter when a recording is offered as evidence, used in an investigation or published during a breaking news event.
The appropriate response depends on the consequences. A preliminary flag might simply send a video for closer review. It should not automatically become an accusation of fabrication, especially where someone’s reputation or legal position is at stake.
The broader information environment makes this problem especially difficult. LiveAIWire has reported on fake journalists using real newsroom identities, where fabricated media identities can exploit the appearance of legitimate news organisations. A detector might catch an altered face in one clip while missing an entirely fabricated claim written in ordinary text. Verification must therefore examine people, sources and context rather than relying on a single technical score.
The next challenge is an ever-changing target
Video generators do not stand still. Each generation of tools can produce different artefacts, and social platforms change footage through compression, cropping and other processing. A detector tuned to old manipulations may not cope equally well with new ones.
The paper reports tests involving several types of synthetic and manipulated video, along with checks of how its system behaves under noise and other disturbances. Such experiments are necessary because real recordings are rarely presented under laboratory-perfect conditions. Even so, outside testing against unfamiliar sources will be needed to establish general reliability.
Another concern is whether malicious users can deliberately craft material to evade detection. The authors discuss resistance to certain adversarial conditions in their evaluation, but no single experiment eliminates the broader contest between people making fakes and people trying to identify them.
LiveAIWire’s earlier discussion of artistic influence on AI-generated images illustrates why questions about the origin and transformation of media are likely to remain important even when detection software becomes stronger.
Why human verification still has a role
Even an unusually accurate first-stage detector will encounter recordings whose source is missing, whose context has been changed or whose content does not fit any of the categories used during training. A careful review may require looking for earlier copies, identifying the recording location or confirming events with people who were present. No optical component can do all those jobs on its own.
That is particularly important for urgent or politically sensitive videos, where publishing an accusation of fabrication can itself cause damage. Technical systems should be designed to flag uncertainty and help reviewers decide what to check next. A false sense of certainty would undo much of the benefit of faster processing. The aim is to strengthen verification capacity, not replace it with a score that travels farther than its supporting evidence.
Useful technology without an impossible promise
The appealing part of this research is that it approaches a familiar digital problem from an unfamiliar direction. By involving the physical properties of light, scientists may be able to build systems that examine large batches of video more efficiently than software alone.
That would be valuable for first-stage screening if independent evaluations confirm the benefits and if human reviewers retain the ability to challenge the result. The correct promise is not a machine that decides truth on behalf of everyone. It is a potentially faster instrument for deciding which clips deserve closer attention.
For viewers, the everyday rule remains unchanged: an extraordinary video should be checked against its source and independent context, whether or not a detector raises an alarm. Better detection can help that judgement, but it cannot replace it.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.