Well-resourced intelligence services have been able to alter photographs convincingly since the 1930s and video since the 1990s. Nothing about the upper bound of the capability changed much in the last five years. What changed is the floor. A task that required a funded team, specialist software and weeks of effort now requires one motivated person and an afternoon, and the tooling is distributed as consumer software.
That is a change in distribution rather than in capability, and the analytical consequences of a distribution change are not the same as those of a capability change. The threat model most institutions have adopted — sophisticated state fabrication of high-stakes video — describes the part of the problem that has been true for thirty years. The part that is new is a very large number of unsophisticated actors making a very large number of small alterations, and the aggregate effect of that is not the effect of a better fake.
Alteration Is the Dangerous Capability, Not Fabrication
Public discussion concentrates on wholly synthetic material: the invented speech, the atrocity that never occurred. This is the least effective category of visual deception available, because a fabricated event has no anchor in verifiable reality and therefore presents maximum surface for refutation. There is no corroborating record because there is nothing to corroborate. Every verification technique developed over the last fifteen years works against it.
The consequential capability is the modification of otherwise authentic material. Remove one vehicle from a convoy. Change a unit insignia. Erase a piece of equipment from a loading bay. Alter the number of objects in a frame. Adjust a detail of clothing or marking that carries attributive weight.
Against this, the standard toolkit fails by construction. Geolocation confirms the scene, because the scene is real. Chronolocation confirms the timing, because the timing is real. Shadow and weather analysis are consistent, because they were recorded rather than synthesized. Sun angle, background architecture, vegetation state, licence plate formats, signage — everything an investigator cross-references to establish authenticity checks out, because all of it is authentic. The falsehood is one object, and the object is precisely the analytically load-bearing detail.
An investigation that concludes a video is genuine has, in this case, concluded something true and useless. The methodology answers the question of whether the footage was captured where and when it claims. It does not answer whether the frame has been edited, and that is now the operative question.
Detection Fails on Exactly the Material Investigators Receive
The framing of detection as an arms race understates the structural disadvantage.
Detection methods identify statistical traces characteristic of particular generation or editing pipelines. They are therefore trained against known tools and degrade against unknown ones, which is the arms-race problem and is at least legible. The harder problem is that the traces detectors rely on are fragile, and every step of ordinary distribution destroys them. Platform transcoding, recompression, resizing, screen-recording, re-uploading and format conversion each strip low-level statistical signal. Material that has passed through a messaging app, a social platform and a screenshot arrives with most forensic markers gone.
That is the condition of nearly all material an open-source investigator actually receives. Detection performs best on pristine files that came directly from a generator and worst on multiply-recompressed artifacts of unknown lineage, which is an exact inversion of operational need.
The error asymmetry compounds it. A false negative — a manipulation passing as authentic — is damaging. A false positive is worse. An authentic recording of a real atrocity, flagged by an automated detector as probably synthetic, hands the responsible party a defence it could not have manufactured itself, delivered by the verification apparatus. Any detection regime deployed at scale will generate those cases, and each one is more valuable to an adversary than a successful fake.
Provenance Inverts the Burden and Disenfranchises the Best Sources
The only architecturally sound response is to establish authenticity at capture rather than attempt to establish falsity afterwards. Cryptographic signing in the sensor or the capture device, along the lines the C2PA specification describes, inverts the burden: an artifact carries a verifiable claim about its origin, and the absence of such a claim is itself information.
Two problems follow, and the second is severe.
The first is that signing certifies capture, not truth. A signed recording of a staged event is a genuine recording of a staging. The signature attests that a specific device recorded specific photons at a specific place and time, which is a narrower claim than the one audiences will infer from it. This distinction will be lost within days of any broad public deployment.
The second is a distributional problem. Provenance infrastructure only inverts the burden if signing is close to universal, and universality creates a two-tier evidentiary regime in which unsigned material is presumed suspect. The material that carries the highest intelligence and evidentiary value is overwhelmingly in the unsigned tier: footage from cheap handsets, recordings made by people with no interest in metadata, clips smuggled out of closed environments, material captured on hardware never sold in the jurisdiction where signing standards are enforced, footage stripped of metadata deliberately to protect the person who took it. Source protection and provenance attestation are in direct tension, because the metadata that authenticates the file also identifies the device and often the person.
A provenance regime therefore raises the evidentiary standing of institutional and commercial capture while lowering that of the sources least able to comply and most likely to be documenting something a state wants suppressed. This is a predictable and largely unaddressed consequence of the most promising available remedy.
The Objective Is Baseline Collapse, Not Belief
The strategic logic of cheap manipulation is routinely misread as persuasion. It is not necessary that anyone believe a particular fake, and sophisticated actors do not appear to expect it.
The objective is a condition in which no visual claim is dispositive. Once the general availability of manipulation is common knowledge, every inconvenient image acquires a costless defence, and factual disputes revert to being settled by prior political commitment rather than by evidence. That outcome is cheaper to produce than persuasion, more robust, and does not require the fake to be good. It requires only that the possibility be salient.
This is already the observable pattern of response. Attribution of inconvenient footage to generative tools is now close to reflexive, is made without technical basis, and does not need to withstand scrutiny — only to occupy the interval during which a claim would otherwise have consolidated. Manipulation capability functions less as a weapon of deception than as a general-purpose licence for denial, and it is available to every actor simultaneously.
The Existing Archive Is Retroactively Devalued
The damage is not confined to material yet to be collected.
A body of visual evidence assembled over the last fifteen years was verified under the old economics, when the cost of convincing fabrication was assumed to be high and that assumption did real analytical work. Findings resting on that assumption can now be attacked without new evidence, simply by asserting that the material was probably generated. Where original files, hashes and intake records were not preserved — which is the common case for material sourced from platforms that have since deleted it, or from accounts long since removed — there is frequently no way to re-establish the chain.
The archive does not become worthless. It becomes contestable, which for adversarial purposes is sufficient. Documented findings on incidents that were considered settled are re-openable at no cost, and the burden of defending them falls on organizations with finite resources and no ability to return to the original source.
What This Changes in Practice
Several adjustments follow directly, and most are unglamorous.
Verification has to move from artifact-internal analysis toward source and corroboration analysis. A single clip establishes very little regardless of how cleanly it passes technical examination. Multiple independent captures of the same event, from devices with different owners who had no opportunity to coordinate, is becoming the practical standard of proof — which is one respect in which pervasive recording genuinely helps.
Intake procedure becomes evidentiarily critical. Hashing at receipt, preserving the original file untouched alongside any working copy, recording the acquisition path, and retaining that record indefinitely are now core competences rather than housekeeping. The moment of receipt is often the last point at which provenance can be fixed at all.
Published conclusions need a confidence taxonomy rather than binary claims of authenticity, and the taxonomy needs to distinguish clearly between confirmed capture context and confirmed frame integrity. These are different findings and have been conflated for years.
Finally, the collection pipeline itself becomes an attack surface. Seeding manipulated material into open archives, monitoring datasets and reference corpora that investigators rely on is a more efficient operation than manipulating any single artifact, because it corrupts the baseline against which everything else is checked.
What Argues Against It
The historical base rate does not support catastrophism. Convincing still-image manipulation has been available to anyone with a consumer computer since the mid-1990s, and the predicted epistemic collapse did not occur. Reputation, institutional accountability, corroboration and source track record absorbed it. Written text has been costlessly fabricable for the entirety of human history and societies developed functional mechanisms for handling that.
Manipulations also face a consistency constraint that tightens as instrumentation increases. An altered artifact must survive contact with everything else that is known or recorded, and a more densely captured environment produces more independent material capable of contradicting it. The same abundance that degrades verification also multiplies the opportunities for a manipulation to be caught by a record its author did not know existed.
Most importantly, the dominant form of visual disinformation observed in practice is not manipulation at all. It is authentic footage presented with a false caption, a wrong date, or a wrong location — frequently real material from an unrelated event years earlier. This attack is free, requires no technical skill, is extremely effective, and is what the overwhelming majority of documented cases consist of. Cheap manipulation may simply not displace it, because there is no need to edit a video when miscaptioning one works as well.
Assessment
The change is best understood as a shift in the cost of denial rather than in the quality of deception. Fabrication capability improved, but fabrication was never the binding constraint on effective visual disinformation and is not the principal threat now.
The practical consequences fall on the verification side. Single-artifact analysis is losing its evidentiary weight; corroboration across independent sources is gaining it; and the institutional discipline of preserving originals and documenting acquisition has become the difference between findings that can be defended and findings that cannot. None of that is technically novel and all of it is procedurally demanding, which is a reasonable description of where the discipline’s actual difficulty now lies.
Leave a Reply