The Proof
Problem.
When convincing evidence can be generated on demand, truth needs a chain of custody.
For most of modern life, a photograph arrived with a presumption. It could be edited, framed or misread. But it had usually begun somewhere in the physical world. A document had an author. A recording had a speaker. A screenshot was imperfect evidence, yet it still pointed back to an event.
That presumption is disappearing quietly. Not because every image is now fake, but because a convincing image no longer proves what it once appeared to prove.
This is the proof problem. AI does not merely make falsehood cheaper. It changes the burden of proof around everything real.
The Cost of a Convincing Lie
Forgery once required time, equipment and a person with a rare skill. A bad forgery collapsed under inspection. A good one left traces. Generative systems have changed the economics. They can produce a credible voice, image, document or summary at the speed of a request — and create enough variations to make the first correction feel irrelevant.
The danger is not only that people believe a false item. It is that they stop believing a true one. Once the public learns that anything might be fabricated, an authentic recording can be dismissed as synthetic whenever it becomes inconvenient. The liar gains a second defence: not “this did not happen,” but “you cannot prove it happened.”
Evidence Has Always Needed Context
Courts understand something the internet forgot: evidence is not a file. It is a file plus origin, handling, corroboration and a person prepared to answer questions about it.
A photograph without a source is an assertion. A screenshot without an original account, timestamp, surrounding exchange and independent confirmation is a lead, not a conclusion. An AI answer with links is not automatically research; it is a claim that somebody still has to inspect.
That distinction matters to journalism, law, compliance and public life. The more fluent the output becomes, the more discipline has to move upstream — from “does this look real?” to “where did this come from, what happened to it, and who can stand behind it?”
Useful, but no longer sufficient. Appearance has become cheap.
Origin, changes, source access and independent confirmation.
Why Detection Will Not Save Us
It is tempting to imagine a detector that ends the problem: upload an image, receive a verdict, move on. But detection is an arms race. A model changes, an editor removes a signal, an image is cropped, a recording is paraphrased, or a human makes enough alterations to make certainty impossible.
NIST’s review of synthetic-content transparency tools makes the point without drama: authentication, provenance, watermarking and detection all have uses; none is a universal solution. Text is particularly awkward. Its structure is easy to alter, and provenance signals can be removed or weakened.
Provenance Is Better Than Vibes
The most promising shift is away from guessing what a file looks like and toward preserving where it came from. Content Credentials, built on the C2PA specification, are designed to attach cryptographically bound provenance information to media: who created it, what tools were used and what changed along the way.
This is valuable. It can make honest work easier to verify. It can give a publisher, photographer or public institution a better way to say: this is ours, and here is its history.
It is not a truth machine. Credentials can be absent. Metadata can be stripped. A perfect chain of custody can faithfully record a staged event. Provenance tells us about the history of a file; judgment still decides what that history means.
The Newsroom Test
This is not an abstract problem for Malta Insider. A newsroom is already a proof system. It decides what can be published, what needs a second source, what remains allegation, what requires a document and what must be held back.
AI can make that system faster: transcription, translation, source discovery, document comparison, structured research. It can also make it weaker when a summary replaces reading, a generated image replaces disclosure, or a citation is accepted without opening the source beneath it.
The right standard is simple: AI may assist the route to a claim, but it cannot become the final witness for that claim. A newsroom should be able to show its work — not merely present an answer that sounds as though the work was done.
Labels Are Necessary. They Are Not Enough.
The EU’s AI Act now brings transparency duties into force for certain AI systems and content, including labelling requirements for deepfakes and specified AI-generated public-interest text. That is a necessary baseline. People should not be asked to discover a synthetic manipulation by instinct.
But labels address disclosure, not belief. A label does not establish whether a video is misleading, whether an article’s sources are sound, whether a document was altered before it was signed, or whether a model has omitted the one fact that changes the conclusion.
2. Corroboration: does independent evidence support the claim?
3. Accountability: is a named person or institution prepared to stand behind it?
4. Contestability: can the public, a court or an editor inspect and challenge the reasoning?
The New Advantage Is Not Certainty
We will not restore a world in which every image carries an automatic presumption of truth. Nor should we pretend that human institutions were ever immune to manipulation.
What we can build is a more adult standard: less faith in appearance, more attention to provenance; less trust in fluent summaries, more willingness to open the source; less anonymous certainty, more named responsibility.
That is not a retreat from technology. It is the discipline required to use it without surrendering the thing that made evidence valuable in the first place.