Deepfakes and the coming collapse of shared reality
Dear Reader,
The public conversation about deepfakes focuses almost entirely on the wrong problem. It focuses on specific fake videos of specific people saying specific things they never said. This is a real problem, and it deserves attention. But it is the surface manifestation of a much deeper problem, which is what happens to the entire epistemology of a society when video, audio, and photographic evidence can no longer be trusted.
I want to write about the second problem, because it is the one that will actually reshape our institutions. Fake specific content is a targeted harm that can, in principle, be addressed through targeted responses. The collapse of trust in any content is a general erosion that no targeted response can address. And it is well underway, in 2026, in ways that most people have not noticed yet because the specific triggering incidents have not been dramatic enough to force the notice.
Let me walk through what is actually happening, why the technical detection approach is insufficient, what the deeper problem is, and what if anything can be done. This is not a piece designed to alarm. It is a piece designed to describe accurately what is happening, so that the specific choices we face become visible.
Start with the technology itself, briefly, because most people have a simplified picture.
Deepfakes are, technically, synthetic media produced by generative AI systems, where one person's face, voice, or movements are used to replace another person's in an existing recording, or where an entirely fabricated recording is produced from scratch. The techniques have improved rapidly over the last decade. What required significant computational resources and expertise in 2018 now requires a consumer-grade GPU and open-source software. What produced obviously artificial results in 2020 now produces results that require careful forensic analysis to detect, and that pass casual inspection almost every time.
The specific quality of the fakes has crossed a threshold. Not "close enough that experts can be fooled sometimes." Actually indistinguishable from real recordings, for practical purposes, in the vast majority of cases. Detection tools exist, but they lag behind the generation tools. The specific pattern of the field is that new detection techniques are quickly incorporated into the generation systems as adversarial training signals, and the detection tools then need to be updated, in a cat-and-mouse dynamic where the cat is losing.
This is not a hypothetical trajectory. It is the current state. Video and audio evidence in 2026 has to be treated with a specific kind of skepticism that was not required in 2018. And the skepticism will only become more warranted over time, as the generation technology continues to improve.
The specific consequence that most commentary misses is what philosophers call the liar's dividend.
The liar's dividend is the specific epistemic advantage that accrues to anyone accused of doing something wrong when the evidence against them is a recording. If deepfakes exist, and if they are hard to detect, then any incriminating video can be plausibly denied as a deepfake. The specific denial does not have to be persuasive to work. It just has to introduce enough doubt that the evidence loses its evidential force.
This has already happened repeatedly. Politicians accused of specific statements based on recordings have claimed the recordings were fake, and their audiences have accepted the claim, whether or not the recordings actually were fake. Business executives caught in specific behaviors have used similar defenses. The specific pattern is that the mere possibility of deepfakes provides cover for anyone whose real behavior would be embarrassing if it could be evidenced.
The specific ironic consequence is that deepfakes make it easier not just to fabricate false content but to escape true content. This is the specific way that the technology damages our epistemic infrastructure. Not primarily by adding false claims to the discourse, though it does that too. Primarily by making all claims, including true ones, subject to reasonable-sounding denial.
The specific effect at population scale is a general erosion of the concept of evidence. If every recording could be fake, then no recording is really evidence. If no recording is really evidence, then the specific role that recordings have played in journalism, in law, in accountability of any kind, is compromised. The specific infrastructure of shared truth that developed in the twentieth century, based on the presumed reliability of audio and video, is being dissolved.
There is a specific historical parallel worth drawing.
Before the invention of photography, the specific verification of what someone said or did depended on witnesses. Which meant that specific events were verified by specific people who could be identified, questioned, and held accountable for their claims. This was imperfect. Witnesses lie. Witnesses misremember. Witnesses can be intimidated. But the specific pattern of accountability was clear. If you claimed something, other specific people could verify or dispute your claim.
Photography, film, audio recording, and eventually video changed this pattern. Suddenly there was a specific kind of evidence that did not depend on any specific witness's testimony. The recording itself was the witness. And the specific advantage was that the recording could be examined by anyone, at any time, without requiring the presence of the original witness. This was one of the specific technologies that made modern investigative journalism possible, that made civil rights documentation possible, that made accountability of the powerful possible in specific new ways.
Deepfakes undo this. They return us to a specific epistemic situation closer to what existed before photography. Recordings are no longer independent evidence. They require additional verification. And the specific pattern of accountability shifts back to relying on witnesses, forensics, and social trust in specific institutions.
This is not necessarily catastrophic. Societies functioned before photography. But the specific practices and institutions that developed around the assumption that recordings could be trusted will have to be rebuilt for the world where they cannot. And the rebuilding will take time, will be contested, and will produce different outcomes than what we currently have.
Let me tell you what has already been happening in my own field, criminal law, that most people are not yet aware of.
For decades, video evidence has been increasingly central to criminal cases. Surveillance footage. Body camera recordings. Phone videos of specific events. These have often been the specific pieces of evidence that decided cases. When the video showed something clearly, the video usually settled the question.
This is changing. In the specific cases I have followed over the last two years, defense attorneys have started routinely challenging video evidence on the grounds that it might be manipulated. The specific challenges do not always succeed. But they introduce doubt, they require prosecutors to produce experts to authenticate footage, and they shift the specific burden of proof in ways that were not required before.
In more concerning cases, prosecutors have started facing situations where genuine video evidence is dismissed by juries because the defense has raised deepfake concerns effectively enough to introduce reasonable doubt. This is not yet common, but it is happening. And it will happen more as public awareness of deepfake capabilities spreads.
The specific implication for legal practice is that the specific role of video evidence in trials is going to change. What was one of the most reliable forms of evidence available is becoming, effectively, more like eyewitness testimony. Still useful, but requiring specific verification and vulnerable to specific challenges. This is a substantial change in how criminal accountability works, and it is happening without any deliberate policy decision.
There is a specific broader implication that I want to draw out.
The infrastructure of modern journalism, and by extension of democratic accountability, depends heavily on the specific ability to document what powerful people say and do. Investigative journalism relies on recordings. Political accountability relies on being able to show specific evidence of specific behavior. Historical documentation relies on the specific archive of recorded material that has accumulated.
All of this depends on the specific presumed reliability of the recordings. If the recordings can be plausibly denied, the infrastructure it supports weakens. Not to zero. But to a degree that changes what journalism can achieve, what political accountability looks like, and what future historians will be able to know about our time.
The specific concerning thing is that this weakening favors the powerful. If someone with resources is accused of specific behavior based on recorded evidence, they can now afford the specific forensic response, the specific expert testimony, the specific legal challenges that will introduce doubt about the recording. Someone without resources cannot. Which means that the specific tools that used to hold powerful people accountable are being weakened more effectively than the specific tools available to less powerful accusers.
This is not a small change. It is a substantial shift in the specific balance between accountability and its evasion. And it is happening across every specific area where video evidence has been playing an accountability role. Politics. Business. Law enforcement. Institutional oversight. All of these depend on the specific ability to document what happened. All of these are being weakened by the deepfake dynamic.
Let me name the specific technical responses that are being developed, and why I think they are insufficient.
The first response is detection technology. AI systems trained to identify AI-generated content. These have real capabilities. They can identify specific artifacts of the generation process, statistical patterns that natural recordings do not have, and specific inconsistencies that give away synthetic content. The problem is that every detection technique becomes an adversarial training signal for the generation systems. Once we know how to detect specific artifacts, the generators learn to avoid those artifacts. The specific cat-and-mouse pattern means detection is always behind generation, and always will be.
The second response is provenance authentication. Systems that cryptographically sign recordings at the moment of capture, so that authentic recordings can be verified by checking the signature. This is technically sound and is being deployed by specific camera manufacturers and platforms. The problem is that authentication only works for content that goes through the specific certified pipeline. Most content will not. And even for content that does go through certified capture, the certification only proves the original was authentic. It cannot prove that a specific edit did not manipulate the content in specific ways.
The third response is legal regulation. Laws requiring disclosure of AI-generated content, prohibitions on specific kinds of deepfake production, and penalties for specific misuses. These are being developed in various jurisdictions. They face the specific challenge that deepfake technology is not centralized. It is available as open-source software running on consumer hardware. Regulation of specific bad actors is possible. Regulation of the technology itself is essentially impossible.
Each of these responses is worth pursuing. None of them will solve the underlying problem. Because the underlying problem is not that specific fakes exist. It is that any content might be fake, and the mere possibility changes the epistemic situation in ways that specific technical fixes cannot fully reverse.
Let me tell you what I have started doing personally, because it might be useful.
I have stopped treating video and audio recordings as authoritative evidence unless I can verify them through other means. This means that when I see a specific claim supported by video evidence, I do not treat the video as settling the question. I ask what other evidence supports the claim. I ask what the source of the video is, who published it, what their track record is. I look for corroborating information from other channels. If the corroboration exists, I treat the claim as well-supported. If it does not, I treat the claim as tentative, regardless of how convincing the video looks.
This is exhausting. It is much slower than the specific way we used to consume media. But it is what the current epistemic situation actually requires. Video and audio are no longer sufficient by themselves. They are one piece of evidence, to be weighed alongside other pieces, and often to be discounted if other pieces are not available.
The specific consequence is that I have become more skeptical of dramatic claims made online, and more willing to wait for institutional verification before believing something. This is not always practical, and it means I am often behind the specific news cycle. But my track record for believing false things has improved, and my track record for being manipulated by specific viral content has improved too.
I do not know if this is a good long-term strategy for how to be informed. It is what I have found works given the specific media environment. I share it because it might be useful for others facing the same environment.
There is a specific implication for AI that I want to draw.
The deepfake technology is not going to become less capable. Every specific improvement in generative AI makes it more capable. And every specific improvement is downstream of business incentives that reward capability, not safety. There is no likely trajectory in which the technology becomes less powerful. There are trajectories in which regulation slows its deployment or restricts specific uses, but the underlying capability will continue to advance.
Which means the epistemic situation we are entering is not a temporary crisis to be managed. It is a specific permanent shift in what counts as evidence. Video and audio have joined text as media that can be produced synthetically with reliability. This has already happened for text, and we have adapted to it in specific ways. Text-based evidence is now routinely evaluated with skepticism about whether the specific text was human-produced. The same pattern is now emerging for video and audio.
The adaptation, if we manage it, will involve new institutional practices, new verification technologies, new social norms around what counts as evidence. It will take years or decades to fully develop. During the interim, we are in a specific transitional period where the old assumptions no longer hold and the new practices are not yet established. This is where we are now. It is a specifically unstable moment.
I want to close with a specific observation about what is at stake.
Modern democracies depend on the specific ability of citizens to know what is happening in their societies. Not perfectly, and not neutrally. But sufficiently that public opinion can be formed on the basis of actual events. Journalism, video evidence, and shared media artifacts have been the specific mechanisms by which this knowing has been possible at scale.
When these mechanisms weaken, the specific ability to form public opinion on the basis of actual events weakens with them. What replaces them is not the absence of opinion. It is opinion formed on other grounds. Tribal affiliation. Trusted sources. Institutional loyalty. In principle, these can be adequate substitutes. In practice, they produce different outcomes than shared evidence-based opinion, and the differences are usually not favorable to democratic accountability.
The specific weakening of shared reality that deepfakes contribute to is not just an epistemic problem. It is a political one. Societies that cannot agree on what happened cannot easily agree on what to do about it. Societies whose evidentiary infrastructure has collapsed have historically resolved their disagreements through means that are not democratic in any recognizable sense. The specific fate of democracies depends, in part, on our specific ability to preserve some functioning basis for shared truth.
This is a large claim. I do not want to make it more dramatic than it is. What I want to name is that the deepfake problem is not just about specific fake videos. It is about what happens to societies when the specific tools for verifying what happened stop working reliably. That is where we are heading. And the specific window for building alternative mechanisms is smaller than most people realize.
Next month I want to write about post-human futures, because if the current moment is producing specific pressures on democracy, epistemology, and shared reality, the trajectories that could result from those pressures are worth taking seriously. What comes after the current era is not predetermined, but the specific range of plausible outcomes is worth mapping. Stay with me.
— Transmission Sent —
Niklas Hanitsch
Reference materials
- Nina Schick — Deepfakes: The Coming Infocalypse (2020)
- Danielle Citron and Robert Chesney — Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security (2019)
- Hany Farid — Fake Photos (MIT Press Essential Knowledge, 2019)
- Content Authenticity Initiative — technical standards documentation
- Yaqin Wang et al. — Deepfake Detection: A Systematic Literature Review (2022)
- Danielle Citron — Hate Crimes in Cyberspace (2014)
- https://c2pa.org/
- https://www.pnas.org/doi/10.1073/pnas.2110013118
Continue reading
- Six ways AI is already manipulating your decisions
- Surveillance capitalism in 2026: The concrete infrastructure
- Post-human futures: Three scenarios I can actually believe
- Dark aesthetics as tech critique: The visual language of digital dread
- Can AI Lie? The Surprising Philosophy of Machine Deception
Frequently asked questions
What is a deepfake? A deepfake is synthetic media, typically video or audio, produced by generative AI systems. It can involve swapping one person's face or voice into an existing recording, or fabricating entirely new recordings from scratch. The technology has improved rapidly and now produces content that is often indistinguishable from real recordings without careful forensic analysis.
Can deepfakes be reliably detected? Not consistently. Detection tools exist and have real capabilities, but generation tools improve faster than detection tools. Each new detection technique tends to become an adversarial training signal for the generators, which then learn to evade it. The general pattern is that detection is always behind generation.
What is the liar's dividend? The liar's dividend is the epistemic advantage that accrues to people accused of wrongdoing when the evidence against them is a recording that might be a deepfake. The mere possibility of deepfakes lets accused parties plausibly deny genuine incriminating evidence, damaging accountability across contexts where video and audio evidence used to be decisive.
How does this affect legal cases? Video evidence in criminal cases is increasingly being challenged on deepfake grounds. This does not always succeed but introduces doubt and shifts the burden of proof. In some cases, genuine video evidence is now being dismissed by juries because deepfake concerns have been effectively raised. Legal practice is adapting, but the change is substantial.
Can we build alternative verification systems? Yes, and this work is underway. Cryptographic authentication at capture, content provenance tracking, forensic analysis tools, and legal frameworks for AI-generated content are all being developed. None of these fully solve the underlying problem, but together they may create workable if imperfect systems for verifying content. The transitional period, before these systems mature, is where we currently are.