The employee does not press “confirm.”

A senior manager is visible on the screen. The voice sounds familiar. The request is urgent. The payment details look complete.

Instead of studying the face for glitches, the employee reaches for a separate office phone and calls a number already stored in the company system.

That pause is becoming more important than the human ability to recognise what looks fake.

On 2 August 2026, new transparency obligations under Article 50 of the EU AI Act become applicable. Providers and professional deployers will have to make relevant AI-generated or manipulated content identifiable, including through machine-readable marking and visible disclosure in specified cases.

It is a significant change. It is also easy to misunderstand.

An AI label may tell us something about how a file was created. It does not necessarily tell us whether the event shown actually happened, whether the accompanying claim is accurate, or whether the material has been placed in a misleading context.

Provenance can begin an investigation. It cannot finish one.

The new promise of technical trust

For years, much of the public advice about manipulated media has focused on visual intuition: watch the hands, inspect the blinking, listen for an unnatural voice, look for inconsistent shadows.

That strategy is becoming less dependable.

A systematic review and meta-analysis covering 56 studies and 86,155 participants found an overall deepfake identification rate of 55.54 percent. Performance on video averaged 57.31 percent. Training, feedback and AI assistance improved results to roughly 65 percent in the analysed interventions, but even that leaves substantial room for error.

The practical lesson is not that people are foolish. It is that security systems should not depend on ordinary people performing forensic analysis under pressure.

A person facing an urgent transfer request does not need sharper instincts about facial rendering. They need a procedure that allows them to stop, check the original request and confirm the sender through a different channel.

A manager deciding whether to suspend an employee should not be asked to choose between believing a viral clip and believing the person accused. The institution should require the full recording, the original source, relevant metadata, independent confirmation and human review before imposing a serious consequence.

The same principle applies to graphs, screenshots and detector scores. A technical-looking signal can create an impression of certainty long before its meaning has been established.

File history is not event truth

Content Credentials and the C2PA standard can attach cryptographically protected information to digital material. Depending on the implementation, that information may show when a file was captured, which tools were used, what was edited and which actor signed the record.

That can strengthen a chain of custody. It can help journalists, investigators and organisations examine how material travelled through a production process.

But C2PA is explicit about its limit: the standard validates provenance claims and their connection to the asset. It does not decide whether the content is morally good, politically fair or factually true.

A valid signature may accompany a false statement. An authentic photograph may be paired with the wrong date. A real video may be cropped immediately before the moment that changes its meaning. A graph may contain genuine numbers while hiding the population, time period or baseline needed to interpret them.

The opposite error matters too. Missing credentials are not proof of manipulation. Metadata can disappear through screenshots, platform processing, incompatible software or ordinary editing. Older equipment and smaller creators may not support the same provenance systems as large institutions.

When absence is treated as guilt, the person with the least technical infrastructure inherits the greatest burden of proof.

The warning can arrive too late

A label may correct what people believe about a video’s origin without fully removing the video’s influence.

Three preregistered experiments published in Communications Psychology tested whether deepfake warnings changed moral judgments after participants watched fabricated confessions. The warnings reduced the videos’ effect, but they did not erase it.

Among participants who accepted that the video was false, substantial shares still relied on its content when judging guilt. Across the experiments reported in the research, the figure remained around half.

That distinction matters.

A person can intellectually accept that a scene was fabricated while still carrying the suspicion created by watching it. The image remains available in memory. The accusation feels imaginable because it has already been seen.

This is why correction cannot be reduced to placing a label beneath the original post. When an employer, platform, newspaper or public authority has amplified or acted upon false material, it also owns part of the work of repair.

That may require linking corrections to copies, notifying decision-makers, reversing an earlier action and giving the correction visibility comparable to the accusation.

A technical correction without practical restoration leaves the main cost with the person who was misrepresented.

Synthetic credibility already has a price

This is not only a political communication problem.

The FBI’s 2025 Internet Crime Complaint Center report recorded 22,364 complaints containing AI-related information and approximately $893 million in adjusted losses. The cases included cloned voices, fabricated profiles, investment videos and alleged real-time calls from supposed executives.

Those figures should not be described as a pure estimate of deepfake losses. The category covers several forms of AI-assisted crime and depends on reported cases.

Even with that limitation, the underlying behaviour is clear: familiar faces, voices and roles are being used to accelerate decisions that would otherwise trigger suspicion.

The technology does not create the entire fraud. It supplies synthetic credibility.

The final loss often depends on the surrounding system. A company requiring independent approval and confirmation through an established channel is less vulnerable than one where a single employee can immediately obey a leader’s apparent voice.

Responsibility therefore cannot be pushed entirely onto the person who failed to notice that something looked wrong. The criminal owns the deception. The organisation owns the design of its payment and approval procedures.

Political presence is not proven political dominance

Synthetic political media is also real. Campaign organisations in the United States have already used realistic manipulated candidate videos, sometimes with disclosures that are difficult to notice.

But presence is not the same as measurable electoral effect.

A study of 187,778 posts from X, Bluesky and Reddit during Canada’s 2025 federal election classified 5.86 percent of the analysed election images as deepfakes. Most were harmless or non-political. Harmful deepfakes accounted for only 0.12 percent of views on X within that dataset.

The result does not prove that deepfakes are harmless. It examines one election, three open platforms and primarily image content. It does not cover private messaging, all video or audio, or every form of political influence.

It does show why scale must be demonstrated rather than assumed.

Exaggeration has its own cost. When people are repeatedly told that everything may be synthetic, “it could be AI” becomes a convenient way to dismiss genuine documentation. Evidence does not disappear, but the burden of defending it grows.

Blind belief and total disbelief produce the same failure: neither examines the full evidence.

The relevant differences are institutional

The research does not support a clear comparison between men and women. The stronger differences concern power, exposure, decision authority and access to protection.

A public figure has a different capacity to answer a fabricated video than an ordinary employee. A large newsroom may preserve original files and technical records that an independent creator cannot. A bank with dual authorisation can absorb a verification delay that a small company may not have planned for.

Geography matters as well. The EU is creating a shared legal transparency framework. The United States relies more heavily on a fragmented combination of platform rules, state regulation, campaign practice and fraud enforcement.

The human mechanism is similar across these settings: a recognisable face, voice, graph or official-looking label can feel like direct access to reality.

The safeguards are not equally available.

A rule for irreversible decisions

The useful response is not to distrust everything. It is to raise the evidence requirement when the consequence becomes difficult to reverse.

Before transferring money, suspending an employee, publishing an accusation or imposing an official penalty:

Find the original. Check the full context. Confirm identity or facts through a separate established channel.

Ask what each technical signal actually measures. Does it show AI use, an editing history, a valid signature, a statistical probability or factual truth? Those are different claims.

Keep the burden with the institution making the decision. A detector score should not force the accused person to disprove an opaque system. Missing metadata should not become an automatic presumption of guilt. Urgency should not cancel independent approval.

Labels, provenance records and detection tools are necessary additions to the evidence chain. Used carefully, they can make manipulation easier to investigate and authentic material easier to support.

Used as verdicts, they can create a new form of synthetic trust: confidence produced not by complete evidence, but by the appearance of technical certainty.

The strongest defence is not perfect eyesight. It is a procedure that permits a person to pause before consequence.

What extra evidence would you require if the next decision could cost someone money, work, freedom or reputation?

SOURCES

Source: Code of Practice on Transparency of AI-Generated Content, European Commission
Link: https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content
Used for: The EU’s Article 50 transparency framework and the role of the voluntary compliance code.
Important limitation: The code supports compliance but does not measure whether labels change behaviour correctly.

Source: Transparency obligations under Article 50 of the AI Act, European Commission
Link: https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act
Used for: The 2 August 2026 applicability date, transition rules and limits on retroactive labelling.
Important limitation: Official legal guidance, not evidence of enforcement quality.

Source: Content credentials: Strengthening multimedia integrity in the generative AI era, Australian Signals Directorate and partners
Link: https://www.cyber.gov.au/business-government/secure-design/artificial-intelligence/content-credentials-strengthening-multimedia-integrity-in-the-generative-ai-era
Used for: The distinction between useful provenance context and factual truth, including neutral treatment of missing credentials.
Important limitation: Primarily organisational security guidance.

Source: C2PA Technical Specification, Coalition for Content Provenance and Authenticity
Link: https://spec.c2pa.org/specifications/specifications/2.4/specs/C2PA_Specification.html
Used for: What cryptographic provenance can validate and the standard’s refusal to make truth or value judgments.
Important limitation: Correct specification does not guarantee complete implementation across platforms.

Source: The continued influence of AI-generated deepfake videos despite transparency warnings, Communications Psychology
Link: https://www.nature.com/articles/s44271-025-00381-9
Used for: Evidence that warnings reduced but did not eliminate deepfakes’ influence on moral judgments.
Important limitation: The experiments used fictional people and scenarios.

Source: Human performance in detecting deepfakes: A systematic review and meta-analysis of 56 papers, Computers in Human Behavior Reports
Link: https://www.sciencedirect.com/science/article/pii/S2451958824001714
Used for: Detection performance across 86,155 participants and improvements associated with training and support.
Important limitation: Considerable variation in stimuli and study design.

Source: 2025 IC3 Annual Report, FBI Internet Crime Complaint Center
Link: https://www.ic3.gov/AnnualReport/Reports/2025_IC3Report.pdf
Used for: AI-related complaint numbers, reported losses and examples of cloned voices, videos and false identities in fraud.
Important limitation: The category covers several forms of AI-assisted crime and does not isolate deepfakes.

Source: Deepfakes in the 2025 Canadian Election: Prevalence, Partisanship, and Platform Dynamics
Link: https://arxiv.org/abs/2512.13915
Used for: The measured prevalence and limited harmful reach of synthetic election images in the analysed dataset.
Important limitation: One election, three open platforms and primarily image-based analysis.