Evidence Must Travel with the Image: A MoJo Protocol After the End of Faith in Pixels

A genuine clip without a past

A newsroom receives thirty seconds of video from a demonstration. A police officer appears to strike a person lying on the ground. The reporter knows the witness and believes the recording is genuine. But the clip travels through a messaging app, is automatically compressed, enters an editing application, receives cleaned audio, subtitles and a tighter crop. Publication prompts an official response: ‘It is manipulated – or generated by AI.’

The newsroom wants to answer. It has only the final export. The original remained on the witness’s phone, which has now failed. Nobody recorded the transfer time, file properties or editorial changes. The journalists may be telling the truth, but they cannot show the route by which the truth reached them.

The scene is hypothetical. The problem is precise: a broken chain of provenance for mobile material. In an age of effortless manipulation, an authentic image can lose not because it looks false, but because the newsroom treated it as an ordinary attachment. Its past was discarded during production.

Mobile journalism has long drawn strength from a compelling intuition: somebody was there, took out a phone and recorded what happened in front of the lens. The device combines witness, camera, editing suite and

transmitter. That compactness democratises publication and shortens the distance between event and audience (Burum & Quinn, 2016). Synthetic media do not invalidate witnessing. They do end the luxury of automatic faith in a pixel.

Being there is more than possessing an image

Photography and film have traditionally been understood as having a particular relationship with reality: light from a scene leaves a trace on film and, later, a sensor. This connection is often called indexicality. A photograph is not simply a drawing that resembles the world; it is produced through a physical encounter between device and scene.

MoJo adds an embodied relationship. A phone sits in the hand of a person standing inside events, sometimes at genuine risk. Andén-Papadopoulos writes about the citizen camera-witness and the moral force of people who ‘put their bodies on the line to record the injustice of oppression’ (2014, p. 753). The phrase reminds us that evidence does not begin inside a file. It begins in a person’s relationship to an event.

Journalistic witnessing combines presence, authority, authenticity and moral responsibility (Pantti, 2019). ‘I was there’ is not complete proof, but it creates a distinctive accountability. A witness can describe what happened outside the frame, answer questions and accept consequences for a statement. A newsroom turns an individual experience into public testimony by checking source, time, location, sequence and consistency with other evidence.

A generative image can be produced without an event, without light reaching a sensor and without anyone at the scene. Editing can also alter genuine footage: remove an object, enlarge a crowd, replace a voice or join separate fragments. Yet the most successful visual falsehood need not be technically glamorous. A real photograph from another country, captioned as today’s event, may travel farther than a perfect deepfake. Appearance alone is therefore an inadequate test. History and context matter.

The detector trap

The intuitive response to synthetic media is to seek a detector: upload an image and receive a red or green light. The model is attractive because it transfers uncertainty to a machine and returns something that resembles a laboratory result. Unfortunately, generators and detectors compete on a course whose finishing line moves every few months.

In a large study of multimodal misinformation and Community Notes, Chrysidis et al. (2026) found that AI-generated material achieved disproportionate virality in their sample. They also reported declining performance by specialised detectors and vision-language models as generators improved. The details of any one benchmark will date quickly; the deeper lesson will not. A probability score is evidence produced under particular testing conditions, not a verdict from a neutral oracle.

Detectors are not useless. They may flag an artefact, an unusual compression pattern or a probable intervention. They should not sit alone in the judge’s chair. Results depend on the type of material, generator version, subsequent processing and test data. A false positive can discredit genuine testimony. A false negative can grant manipulated material an undeserved passport.

Detection asks: do these pixels show traces associated with a process? Provenance asks: how was this particular asset created, who signed which claims, and what changes were declared? Journalistic verification adds: do the stated time, place, sequence and meaning correspond with the event? These questions support one another. None is licensed to impersonate the others.

C2PA: a sealed travel log, not a truth certificate

The C2PA standard, surfaced to audiences through Content Credentials, provides a technical layer for provenance. The simplest analogy is a sealed travel log attached to an asset. A cryptographically signed manifest can contain assertions about origin, tools, ingredients and actions. Binding mechanisms help establish whether the manifest still belongs to that asset; the signature helps reveal whether the declared information was altered unnoticed (Coalition for Content Provenance and Authenticity, 2026).

This is valuable infrastructure, but its scope is often misunderstood. A valid credential can show that a particular entity signed particular assertions and that their integrity has been preserved. It cannot establish that a photographer interpreted the event correctly. A camera may authentically record a staged scene. A signed image can be paired with a deceptive caption. Nor does an absent credential prove fakery: a platform may strip it, a tool may not support it, or a witness may use an older device.

A seal on an envelope does not fact-check the letter. It helps us ask whether the envelope was opened and who claims to have sent it. That is why C2PA does not replace journalism. It gives journalism a better object to investigate.

Provenance is organisational as well as technical. If a newsroom overwrites originals, lets applications flatten every stage and keeps no record of consequential changes, even an excellent standard cannot reconstruct the missing history. Technology can seal the log. Somebody still has to keep it.

A witness is not an appendix to the metadata

An evidential chain can easily become a surveillance chain. Mobile files may expose time, precise location, device identifiers, faces, voices, reflections and patterns of transfer from which relationships can be inferred. To a reporter these may be verification clues. To a repressive institution they may be a route to the witness.

Salzmann, Guribye and Gynnild (2021) describe mobile journalists as traceable data objects: smartphones offer remarkable freedom while producing traces that enable profiling and control. ‘Keep and publish all metadata’ is therefore not a responsible rule. A newsroom may preserve a protected original for verification while sharing only the information necessary for the public to assess the material without endangering a person.

Two rights are at stake. The public has a right to testable testimony. A source has a right to safety and dignity. Neither should automatically cancel the other. Full disclosure of time and place may sometimes be justified. Elsewhere, publication should be delayed, location rounded, faces obscured or part of the manifest retained inside the newsroom. Provenance should serve accountability, not curiosity.

The human being must also remain the owner of the decision. A reporter should understand what a file proves and what it does not. An editor must be able to stop publication. A witness needs to understand the risks of associating a credential with their identity. An audience deserves the limits of verification, not a green shield designed to suggest infallibility.

This is particularly important when footage depicts violence. Repeated publication may retraumatise victims, invite identification or turn suffering into a visual commodity. A perfect chain of custody does not create an unlimited right to show everything. Evidence can be preserved while access is restricted; verification can be described while graphic frames remain unpublished.

6P: a mobile chain of evidence

The 6P MOBILE EVIDENCE CHAIN begins before ‘record’ and ends after publication. For low-risk material, the steps can be light. For war, violence, serious allegations or protected sources, they require a strict workflow and a named decision-maker.

1. Prepare and protect. Before recording, assess risks to the reporter, subject, source and people in the background. Check the device clock, location settings and automatic enhancement features. Seek consent where it is possible and ethically required. In an emergency, human safety outranks the perfect shot. Pass when the reporter can describe the recording circumstances and has deliberately chosen which data to capture.

2. Preserve the original. Secure the unedited file immediately. Never edit the only copy. Retain available metadata and create a checksum or credential when the workflow supports it. A checksum is a digital fingerprint: it can show that one version has not changed, but not that the scene was truthful. Pass when the newsroom can identify the source version and control access to it.

3. Pass it with control. Use a transfer route that does not silently destroy quality or metadata. Record who received which version and when. If safety requires removing data before transfer, document the decision

and keep the protected original separately. Pass when there is no magical leap between a witness’s phone and the newsroom server.

4. Process a copy. Edit, transcribe, stabilise and correct a working copy. Record the tools and ingredients used. For AI-assisted noise removal, frame generation, voice translation or object removal, document the extent and purpose of the intervention. Pass when the published version can be compared with the original and every material change has an accountable author.

5. Publish the changes. An audience does not need a diary of every click. It does need to know about changes that could affect interpretation: a crop that removes context, joined shots, reconstructed sound, altered speed, generative filling or anonymisation. Pass when a reasonable viewer will not infer something from the form of the material that the production method contradicts.

6. Provide provenance. Attach a readable summary: the source or a safe description, time and place at an appropriate precision, core verification steps, consequential edits and a route for corrections. If Content Credentials are available, make them inspectable while explaining their limits. Pass when the newsroom can reconstruct the asset’s route without reconstructing a route to the protected person.

The final test is deliberately human: is the history complete enough to defend the publication and restrained enough not to harm the source? A workflow fails if it protects the file perfectly while exposing the person who made it possible.

What the chain cannot carry

6P will not convert weak material into strong evidence. A perfect file history cannot supply the context outside the frame. It cannot determine a visible person’s intention, show what happened a minute earlier or replace an independent source. Provenance answers ‘where did this come from and what did we do to it?’ It does not answer every question about the world.

Standards are not universal either. Credentials can disappear during export or publication, and audiences may lack tools to inspect them. A malicious actor can sign deceptive claims. A trust system depends on accountable issuers,

protected keys, current software and informed users. No icon should function as a sacrament of truth.

The procedure also costs time and money. A freelancer in the field does not possess the infrastructure of a broadcaster. Organisations must therefore supply secure storage, simple tools, training and an on-call person who can make decisions. ‘The reporter should be careful’ is not a safety policy. It is a way of transferring institutional risk to the weakest link.

Sometimes material with an incomplete chain will still need urgent publication. A responsible newsroom then says what it verified, what remains unconfirmed and why publication nonetheless serves the public interest. Declared uncertainty is stronger than confident language balanced on a brittle file.

Finally, provenance can be weaponised. Bad actors may demand impossible credentials for every inconvenient image and dismiss historic, citizen or low- resource footage that lacks them. The correct presumption is neither ‘credential equals truth’ nor ‘no credential equals falsehood’. Provenance changes the available evidence; accountable people must interpret it.

The pixel needs a witness; the witness needs protection

The end of naive faith in the pixel need not become an age of disbelief in everything. That would be a manipulator’s victory: every incriminating scene could be waved away as ‘probably AI’. The alternative is a shift from automatic trust to an inspectable history.

MoJo has a distinctive role because a phone is close to both event and person. It can create testimony before a professional crew arrives. That power requires a new craft: originals must not evaporate, editing must not erase the route, and provenance must not betray the source.

Evidence should travel with the image, but never alone. Reporter, witness, context and editorial responsibility travel with it. A technical manifest can record a file’s route. Only human beings can explain why it began the journey, whose rights it should serve, and where it must be stopped.