Executive Summary
In 2026, researchers at Utah Valley University showed a nationally representative U.S. sample of 632 people a set of videos arguing for and against a hypothetical ballot measure. Half saw genuine videos featuring real people; half saw deepfakes carrying the exact same message. The shift in voting intention showed no statistically significant difference between the two groups. The deepfakes moved people just as much as the real footage did, and their credibility scores came out equal or even higher. This article starts from that result.
The line of defense has been deepfake detection. Yet the 90–99% accuracy of the marketing copy slumps to 50–65% outside the lab. A detector only catches the generation methods it was trained on, and it falls behind every time a new model appears, until it is retrained. Detection is reactive pursuit by nature, and the side making the fakes is always out in front.
So we have to change the question. Instead of forever chasing "is this video fake?", we move the standard of trust toward proving "where did this video come from?" The real influence of deepfakes, the structural disadvantage of detection, and the promise of provenance along with the hole punched through its middle — all of it bears out in the numbers. And at the end of that road, the question of content authenticity turns out to sit in the same place as the question of data authenticity.
The Key Numbers
Sources: Utah Valley University (KSL) 632-person experiment; Scam.ai / SecurityWeek detection benchmarks; iProov 2025; AIIP Protection
Four numbers lay out the spine of this article at a glance: the experimental result that deepfakes are as effective as the real thing, the margin by which detection collapses in the field, how powerless the human eye is, and the fact that the provenance alternative carries one big hole of its own.
No difference
Real vs. deepfake influence on opinion
In a 632-person controlled study, opinion shift showed no statistically significant difference, and credibility was equal or higher
90% → 50s
Detection accuracy collapse
Advertised 90–99% drops to 50–65% against generation methods it never trained on
0.1%
People who told real from fake perfectly
iProov 2025: only 0.1% of participants got every item right; accuracy on high-quality video was 24.5%
Effectively 100%
Provenance data stripped by social media
A single upload to a major platform, or one screenshot, erases the provenance metadata
Deepfakes Moved Votes as Effectively as Real Video
The long-standing intuition about deepfakes was that "the more polished the fake, the more dangerous." The 2026 experiment from Utah Valley University's Gary R. Herbert Institute for Public Policy sidesteps that intuition. The researchers assembled 632 people representative of the U.S. population and showed them videos arguing for or against a hypothetical ballot measure. One group watched a genuine video of a person actually speaking; the other watched a synthetic video carrying the very same argument. When the team measured how much voting intention shifted before and after viewing, no statistically significant difference appeared between the two groups.
The more striking detail is credibility. Participants gave the deepfake videos credibility scores equal to — or even higher than — the genuine footage. In other words, being synthetic did not make a video more suspect. The researchers concluded that urging individuals to discern more cleverly is not enough, and that the response must be lifted from personal responsibility to the level of policy and institutions.
Place a study of a different stripe alongside this one and the picture sharpens. A 2024 study from Washington University in St. Louis found that even professionally produced deepfakes meant to smear a particular politician were not especially more deceptive or more emotionally manipulative than other forms of fake news. At first glance that seems to clash with the earlier result, but the two studies illuminate different faces of the same truth. The threat of deepfakes does not lie in being "especially more dangerous." It lies in being "as effective as the real thing."
The crux is not sophistication but indistinguishability. The moment a fake cannot be told apart from the real, disinformation crosses a new threshold. It means the old advice — "watch out for videos that look doctored" — no longer works. The threat comes not from the flashiness of the technology but from the very fact that the line between real and fake has vanished.
This shift is not an abstract worry. In Slovakia's 2023 general election, a synthetic audio clip in which an opposition leader appeared to plot vote-rigging and a hike in beer prices spread just before polling. It landed right before the pre-election silence period, when media rebuttals are restricted by law, so there was no time even to explain. In the 2024 U.S. New Hampshire primary, a robocall cloning Joe Biden's voice urged voters to stay home. The same year, a false video claiming that Ukrainian President Zelensky's wife had bought a luxury car racked up roughly 18 million views in 24 hours. Video that cannot be told from the real thing is already touching real votes and real public opinion.
Detection Is Structurally One Step Behind
Then can't we just lean on detection technology to catch the fakes? The problem is that detection does not work as well as the advertising. Commercial deepfake detectors routinely tout 90–99% accuracy. But that figure comes from environments resembling the data they were trained on. Faced with generation methods they never learned, or content actually circulating on the internet, accuracy slumps to 50–65% — barely better than a coin flip.
The reason lies in how detection models work. A detector learns the statistical traces left by a particular generation method. So it catches the methods it has seen, but it is powerless against ones it is meeting for the first time. Every time a new image or video generator launches, existing detectors fail to flag its output properly until they are retrained. In fact, when a real-world deepfake benchmark was released in 2024, the performance of existing detection models dropped by 45–50%. Detection is reactive pursuit that only catches up after a new threat appears, and the side making the fakes always starts a step ahead.
2.1The Human Eye Is Even Less Reliable
Even the hope that "surely a person can sense something off" wavers in the face of the data. In a 2025 iProov study, only 0.1% of participants correctly told both the real and the fake apart. The share who got high-quality deepfake videos right was 24.5% — only marginally better than random guessing. Seventy percent of respondents said they were not confident they could distinguish a cloned voice from the real one. Human intuition does not make for a reliable line of defense.
2.2Not a Gap, but the Structure of the Game
Look at the scale and the shape of the chase grows clearer. Deepfake videos were estimated at around 14,000 in 2019; by 2025, one tally put the figure at 8 million. That is hundreds of times more in six years. On the audio side, the use of synthetic voices is climbing steeply too. The side making fakes grows ever cheaper and faster, while the side catching them can only respond after learning each new case.
To be fair, the detection side is not standing still either. On that same real-world benchmark, research that retuned vision-language models pushed accuracy up to 81%, reaching a level comparable to leading commercial detectors. Detection clearly has room to improve. That said, such gains were only possible after learning from cases already exposed, and the benchmark itself carries the limitation of being small in scale. Even as detection advances, the order — catching up only after seeing a new threat — does not change.
So detection's disadvantage is not a temporary gap to be closed with a better model, but a problem rooted in the structure of the game. Generation runs ahead freely; detection trails, trapped within the distribution it has learned. As long as this asymmetry holds, a strategy that relies on detection alone pours resources into a fight that is hard to win.
Flip the Question: From Catching Fakes to Proving What's Real
If detection is a structurally losing fight, it is time to move the front line. Detection peers into content and judges "is this fake?" — a task of forever chasing fakes that keep renewing themselves. Provenance does the opposite: it proves "where, by whom, and how was this content made?" Instead of pursuing the fake, it establishes a reference point for the real.
This shift changes the question of trust itself. Until now, "does it look real?" was the test. But once deepfakes have succeeded at looking real, appearance is no longer a clue. Provenance moves the standard to "where was it made, and can that history be verified?" The basis of trust becomes not what you can see, but a source you can trace.
From a strategic standpoint, this is a decision to redeploy the line of defense from detection models to provenance infrastructure. It takes the resources that would go into detection and shifts them toward infrastructure that inscribes a verifiable history at the moment content is created. One is an endless chase; the other is a reference point that, laid down once, accumulates. Which is the sustainable investment seems clear.
Proving What's Real: C2PA and Content Credentials
Provenance has already moved past abstract idea into concrete standards and products. At the center sits C2PA (Coalition for Content Provenance and Authenticity). Formed in 2021 by Adobe, Arm, BBC, Intel, Microsoft, and Truepic, the coalition defined a way to embed cryptographically signed history information inside a file. This embedded history is called Content Credentials. When and on what device something was captured, and which tools edited it, are recorded in a tamper-resistant form, and it can be verified offline without querying a central server.
The pace of adoption speaks to the weight of this standard. Membership in Adobe's Content Authenticity Initiative (CAI) has passed 6,000 organizations, and in 2024 Google joined the steering committee. In 2025, Cloudflare became the first major CDN to implement Content Credentials — no small matter, given that the company handles roughly 20% of web traffic. YouTube has begun labeling video shot with a camera, and OpenAI's video generation model Sora 2 automatically inserts C2PA information into its output.
Hardware is moving in the same direction. After Leica released the first consumer camera to sign photos with a dedicated security chip in 2023, Sony, Samsung, and Nikon followed with provenance-signing features of their own. There is more than one approach. Google's SynthID and Meta's Video Seal embed an invisible watermark in the signal itself, while C2PA leans on cryptographic metadata. The technical strategies differ, but the goal is the same: to give content a verifiable origin.
What is worth noting is that standards, hardware, and regulation are aligning at once. The EU AI Act requires labeling of AI-generated content, and in the U.S., the NSA and CISA have recommended adopting Content Credentials in the media pipelines of government and critical infrastructure. It is rare for the laying of standards, the signing by devices, and the push of regulation to meet in one place. The infrastructure for provenance is already being installed.
Provenance Is No Silver Bullet Either
It would be dishonest to present provenance as a silver bullet. Its biggest weakness is that metadata is erased far too easily. Platforms like Instagram, X, LinkedIn, TikTok, and Facebook systematically strip C2PA information as they re-compress and re-format uploaded files. As of 2026, passing through a major social network effectively wipes out provenance metadata entirely. More simply still, a single screen capture instantly destroys the provenance. The source information falls away first on exactly the path where content spreads fastest.
Invisible watermarks are no safer. Attacks using diffusion models try to remove the mark by scattering the watermark into random noise and then reconstructing the image. Such attacks do not work well against robustly designed watermarks, and the reconstructed image diverges from the original — but limits remain: watermark detection is probabilistic, so false positives and false negatives occur, and reverse-engineering the algorithm opens the door to forgery.
Privacy is no light matter either. When provenance metadata carries a creator's identity, it can lead to an exposure the person never wanted. For a journalist in an authoritarian state, a whistleblower, an activist, or a domestic-violence survivor, that exposure is itself a danger. A provision for pseudonymous certificates to mask identity exists in the standard, but real implementations are rare.
The most frequently misunderstood point is the very nature of provenance. C2PA only attaches a history to genuine content; it does not catch fakes. And the absence of a label does not by itself mean a deepfake. The content may have been shot on an old device, or its metadata may have been stripped while passing through social media. Still, the fact that there is no verifiable origin is grounds for applying greater skepticism to that content.
So provenance is not a replacement for detection but a complement to it. The industry is converging on a hybrid direction: metadata-based provenance records for pipelines that demand high integrity, paired with durable signals such as watermarks or fingerprinting for situations where metadata gets stripped. The honesty of not selling a silver bullet is, in fact, what holds up trust in the technology.
Content Authenticity Is Ultimately Data Authenticity
Step back, and attaching an origin to a deepfake video and attaching an origin to AI training data are two faces of the same question. Whether it is video or data, the question is identical: where did this come from, and who changed it, when, and how? Can that history be explained in a verifiable form? Content authenticity and data authenticity differ only in expression; they point to the same foundation of trust.
The point where the two domains meet is already inside the standard. In 2024, C2PA v2.1 added an item for disclosing the data used in AI training. A standard for recording the provenance of content began to also carry what data the model that made that content was trained on. In the end, the question "where did this video come from?" runs along a single thread with "what data was this model built from?"
In pipelines that train and operate AI, keeping a traceable record of which data was received from where, under what rights, and how it was processed is no longer optional. When we weigh data quality, we usually look at accuracy and completeness. But authenticity — the ability to account for an origin — is an axis of quality no less important. Data of unclear provenance, however clean it may look, cannot serve as a foundation of trust.
This is exactly why Pebblous treats data authenticity and provenance as one axis of data quality. Rather than clinging to catching fakes after the fact, inscribing an origin at the moment data and content are created makes for sturdier trust. The lesson of the deepfake era does not stay confined to video. Only an organization that has built the ability to prove what's real into its data from the start can answer any question of trust with the same record.
To close: detection is an endless chase; proof is a reference point that, once set, accumulates. The moment we flip the question from "how do we catch the fake?" to "how do we prove what's real?", the problem of trust in content meets the problem of provenance in data on the same ground. The trust of the next era will be built on the ability to prove what's real.
References
Academic Papers
- 1.Momeni, M. (2025). "Artificial Intelligence and Political Deepfakes: Shaping Citizen Perceptions Through Misinformation." Media and Communication. SAGE Journals. doi.org/10.1177/09732586241277335
- 2.Optica Publishing Group. (2025, Feb). "Generating and Detecting Deepfakes: A 21st-Century Arms Race." Optics and Photonics News. optica-opn.org
- 3.Anonymous authors. (2025). "Revisiting Simple Baselines for In-The-Wild Deepfake Detection." arXiv preprint. arxiv.org/abs/2509.04150
- 4.Anonymous authors. (2026). "Media Integrity and Authentication: Status, Directions, and Futures." arXiv preprint. arxiv.org/abs/2602.18681
News & Case Studies
- 5.KSL.com / Gary R. Herbert Institute for Public Policy. (2026). "AI deepfake media can sway public opinion as effectively as real media, UVU study finds." ksl.com
- 6.Washington University in St. Louis Source. (2024, Aug). "Political deepfake videos no more deceptive than other fake news, research finds." source.washu.edu
- 7.SecurityWeek. (2024). "The AI Arms Race: Deepfake Generation vs. Detection." securityweek.com
- 8.Scam.ai. (2025). "Deepfake detection accuracy: what the benchmarks reveal." scam.ai
- 9.DeepStrike. (2025). "Deepfake Statistics 2025: The Data Behind the AI Fraud Wave." deepstrike.io
- 10.Council on Foreign Relations. (2024). "Election 2024: The Deepfake Threat to the 2024 Election." cfr.org
- 11.PBS NewsHour. (2024). "AI robocalls impersonate President Biden urging New Hampshire voters not to vote." pbs.org/newshour
Official Reports & Documents
- 12.NSA / CISA / U.S. Department of Defense. (2025, Jan). "Strengthening Multimedia Integrity in the Generative AI Era." media.defense.gov
- 13.Content Authenticity Initiative (Adobe). (2026). "The State of Content Authenticity in 2026." contentauthenticity.org
- 14.SimaLabs. (2025). "C2PA vs. SynthID vs. Meta Video Seal: 2025 Playbook for Enterprise AI-Video Authenticity." simalabs.ai
- 15.AIIP Protection. (2025). "Why C2PA Watermarks Fail on Social Media (And What Enterprise Teams Need Instead)." aiipprotection.org
- 16.World Privacy Forum. (2025). "Privacy, Identity and Trust in C2PA: A Technical Review and Analysis." worldprivacyforum.org
- 17.Truescreen. (2025). "C2PA Standard: History, Promises and Structural Limitations." truescreen.io
- 18.iProov. (2025). "Deepfake Detection Study: Human Identification Rates." iproov.com