By the end of 2025, every top talent agency in la had at least one deepfake-detection vendor on retainer. By spring of 2026, several of the top agencies also had in-house teams. The move occurred quietly, without press releases, as the months followed the release of a series of fabricated videos using recognizable actors that were shared virally on social platforms before anyone realized they were fabricated.
The agencies did not talk about this publicly because the issue is still embarrassing-there are now consumer accessible price points for generating convincing video of any well-known person; however, there is a detection technology to find these types of fabrications which is several months behind on each new generation of fakes. However, the arms race that the average celebrity watcher does not realize has developed into one of the most interesting technology stories of the year.
There is a simple premise to understand regarding deep fake detection-the thought process would be that there should be some visual artifact-a blurred edge, a distorted pattern around the eyes, an unusual lighting cues-that would indicate that something was fabricated. This thought process held true until approximately 2023. The current generation of generative video provides output that is visually indistinguishable from real footage to human reviewers and increasingly indistinguishable to specialized detection models trained on previous generation’s fakes.
As such, leading detection systems have transitioned to behavioral analysis. The basic premise is that while the surface visual may appear perfect, the patterns of micro-expression, micro-movements, and timing that humans produce are not yet being replicated by generative systems. Real faces have idiosyncratic blink patterns. Real bodies have characteristic micro-tremors. Real speech has prosodic patterns that match a specific person’s muscular structure in ways that current voice synthesis approximate but does not replicate.
The detection systems train using hours of authentic footage of a specific individual-a celebrity client-in this context-and learn the statistical signature of how that person moves, speaks, and reacts. Any content that does not match the signature (regardless of how perfect it appears at the pixel level), gets flagged for human review.
This methodology should sound familiar to anyone who has familiarity with adversarial system engineering. The same statistical fingerprinting techniques have been used for the better part of a decade in the engineering domain that developed detection of non-human actors in interactive online environments. The technology was mature for one set of use cases years before the celebrity world had a problem large enough to require it.
Internal workflows at top agencies now typically look like this. Monitoring services monitor social platforms, news sites, and content-sharing networks for videos featuring agency clients. New video content gets passed through detection-most often a combination of pixel-level analysis (looking for known signatures of generative models) and behavioral analysis (comparing against the client’s authentic profile). Suspect content gets escalated to a human reviewer-often a former forensic-analysis specialist hired specifically for this work.
If the content is determined to be fabricated, legal team initiates takedown requests across all platforms hosting the content simultaneously. The fastest case-to-takedown times in 2026 are under three hours. The slowest time frames can stretch weeks based upon jurisdiction and platform. Time matters because viral content does its damage primarily in the first day; therefore, if a deepfake is removed within hours, it will have less cumulative reach than if it were only removed after week(s).
At least some high-profile clients have began commissioning their agencies to create what insiders call “behavioral baselines” - structured filming sessions designed to capture the client’s micro-expressions, speech patterns and movement signatures in a wide range of contextual situations. The baseline becomes the reference signal that detection systems evaluate when analyzing any future suspected content. Creating each baseline takes approximately two days to film and results in detection accuracy against current-generation fakes.
It is economic factors rather than technical challenges that allow the agencies to remain hopeful about this issue. Generating a convincingly looking deepfake video of a specific celebrity, at quality levels capable of defeating current behavior-based detection systems requires considerable compute resources and time. As such, generating a deepfake video that remains valid in sustained adversarial review by human forensic analysts requires more still. Thus, the cost-of-attack has increased alongside the cost-of-defense; as such, at least for major celebrities with active monitoring infrastructure, the equilibrium point currently favors the defender.
However, this is not the case for regular people. The difficulty of detecting fabrication targeting non-famous individuals who lack the necessary monitoring infrastructure and behavioral baseline creates a greater risk for those types of fabrications to be weaponized. Most of the worst real-world harm from deepfake technology has been inflicted upon private individuals rather than celebrities. The reason for this disparity is obvious - the private individuals do not have sufficient resources to fight back. Therefore, while the celebrity case is more visible due to their ability to counterattack; the larger case involves ordinary people who cannot.
Three things are likely to change over the next eighteen months:
First, there will be integration of detection technologies at the platform level. Major social media platforms have been developing their own detection capability and are likely to begin automatically flagging or labeling suspected synthetic content instead of relying upon requested takedowns. Although this will help in the median case; it will not eliminate the long tail of platforms that do not develop detection capability.
Second, there will be regulatory pressure. Multiple jurisdictions are moving toward requiring disclosure when AI generated likeness of identifiable people is used in commercial content. While enforcement will be uneven, this will provide celebrities and private individuals with greater leverage in disputes regarding removal.
Third, there will be continued improvement in behavioral baseline techniques. Operators with long experience in behavioral-pattern detection have published articles describing increasingly subtle features - gaze patterns, breath cadence, and micro-vocal variations - that are becoming progressively harder for generative models to duplicate. As such, the field of detection has headroom that exceeds that of generation; thus far, this trend favorably impacts those attempting to catch fakes rather than those creating them.
Therefore, for the average celebrity-watcher the takeaway message is: some of the most attention grabbing video clips that go viral over the next twelve months may turn out to be fabrications - and increasingly these fabrications will be caught-and-labeled as such within days rather than weeks. The necessary infrastructure to accomplish this exists and continuously improves. It’s just not the kind of thing the agencies want to draw attention to until it’s working a lot more reliably than it does today.
