Did you know? A realistic photograph or voice recording can be created or altered by AI. Some synthetic media are harmless illustrations or creative projects. Others may be shared with a false story about who spoke, where something happened or whether an event occurred. That is why learning to evaluate digital evidence matters.
Quick answer: checking suspicious media involves source and context → any available provenance records → careful content inspection → independent corroboration → a proportionate conclusion. AI-based detectors can add clues, but no single score reliably proves that every image or recording is real or fake.
Imagine receiving an audio message that sounds like a familiar person asking you to send money urgently. The voice may be convincing. The question is not merely “Does it sound natural?” It is “Can I independently verify that the real person made this request?”
The 30-second guide
- Synthetic media: images, video or audio produced in whole or part by generative software.
- Deepfake: a common term for convincingly manipulated or synthesised media, often depicting someone’s likeness, voice or actions.
- Detector: software that estimates whether certain media patterns resemble synthetic or manipulated content.
- Provenance: information about a file’s origin and significant changes, when that information is available.
- Content Credentials: a standards-based way of attaching cryptographically signed provenance claims to certain files.
- Corroboration: independent evidence that supports or challenges the claim being made about the media.
A synthetic image is not automatically harmful. A genuine camera photograph is not automatically truthful either: it can be presented with a false caption or taken out of context.
Step 1: Identify the claim before inspecting the pixels
Suppose a picture is shared with the caption, “A new bridge opened this morning.” There are at least two separate questions. Is the image an authentic depiction of the bridge? And is the claim about today’s opening correct?
A genuine photograph of the bridge from five years ago would not prove it opened this morning. An AI-generated illustration might accurately depict a design concept without being a real photograph. Truth, presentation and media origin must be considered separately.
Begin with the exact assertion: who supposedly did what, where, and when? Then decide what authoritative evidence could establish the answer, such as an official project notice, original recording or independently confirmed reporting.
Step 2: Understand what AI detectors actually predict
A synthetic-media detector may analyse visual or acoustic features and produce a score indicating that the file resembles a certain class of known generated or manipulated examples. Some tools focus on particular generators or supported watermark signals rather than all AI-generated media.
Like other classifiers, detectors can make false positives, flagging genuine media, and false negatives, missing synthetic media. Results can vary with the file, training data and method. A confident-looking percentage is not automatically a calibrated probability.
That is why it is risky to call a real photograph “fake” simply because a detector suggested so. A detector result is one piece of evidence to interpret, not a verdict about a person’s honesty.
Step 3: Look for signed provenance where available
An important industry approach is C2PA Content Credentials. A supporting camera, editor or generation system can associate a file with tamper-evident, cryptographically signed information about aspects of its creation and editing history.
A credential may indicate that an asset was created with a particular tool or processed in certain ways, subject to what the issuing application actually recorded. Verification checks the integrity and signers of those claims. It does not inspect the scene and declare whether the depicted event really happened.
The official C2PA explanation makes the boundary clear: provenance alone cannot establish the factual truth of what an image or video depicts. A signed image could still be misleading because of staging, context, timing or a false caption.
Two separate questions: “Where did this file come from?” and “Is the story attached to this file true?” One question cannot replace the other.
Step 4: Do not treat missing credentials as proof
Not every camera or editing tool adds Content Credentials, and provenance metadata can be absent or lost as files travel through different systems. The absence of a credential does not establish that an image is synthetic. It also does not establish that an image is genuine.
Some systems use invisible watermark signals alongside or instead of signed metadata. These methods differ in how they are added, detected and retained. A specific provenance-checking tool may recognise only certain providers’ signals.
Therefore, a result such as “no supported signal detected” should be reported precisely. It is not interchangeable with “this image was not made with AI.”
Step 5: Examine the content without pretending visual clues are conclusive
People often recommend searching for distorted hands, strange reflections or inconsistent speech. Those features can sometimes prompt further investigation, but they are weak as universal tests. Genuine images may contain odd effects from lighting or compression, and high-quality synthetic media may not show obvious artefacts.
Visual inspection is more useful when it asks concrete questions: Does the alleged location match verified landmarks? Does the video show a complete sequence or only an edited fragment? Does a claimed time match other available records? Are the original file and publisher identifiable?
Even those checks need care. Misidentifying a place or person can cause harm. When certainty is unavailable, it is appropriate to state that the origin or claim remains unverified.
Step 6: Corroborate the claim independently
If a clip purports to show an official announcement, find the announcement through the organisation’s established channel. If a family member allegedly sent a surprising voice request, contact them using a known phone number or another independent method.
A second social-media repost of the same clip is not independent evidence. Multiple pages can repeat one original mistake. Strong corroboration comes from sources with their own reliable basis for reporting the event.
For rapidly developing news, details may remain uncertain for some time. A useful conclusion might be “the video has not been independently verified,” rather than asserting either authenticity or fraud prematurely.
A fictional audio-message example
Imagine a fictional school parent receives an urgent voice clip saying, “The class trip fee has changed; transfer payment now.” The voice resembles someone from the school. What should the parent do?
- Do not treat the familiar-sounding voice as sufficient authentication.
- Use the school’s previously known communication channel to ask whether the fee actually changed.
- Check any official written notice, payment instructions and deadlines.
- Do not share account passwords, one-time security codes or sensitive information through an unexpected contact.
- Keep the original message if there is a legitimate need to report the incident to the relevant organisation.
The scenario is fictional and designed to teach verification. The broader lesson is that an urgent request should be confirmed through a trusted independent route, regardless of whether the audio was AI-generated.
Consent, privacy and respectful use
Using someone’s likeness or voice raises questions about permission, misrepresentation and reputation. Even when synthetic media are created for a joke, viewers may not realise that the person never agreed to the depiction or never said the words.
Responsible creative use involves permission where needed, truthful context, suitable labelling and attention to the people affected. Privacy and consent are not optional extras merely because software made generation easy.
Do not attempt to authenticate a private individual by posting their face or voice to an unknown detector. Check the service’s privacy terms and use authorised verification procedures.
Try it: distinguish creation history from factual accuracy
Write two entirely fictional captions for the same drawing of a city bridge. Caption A: “Concept art showing a possible bridge design.” Caption B: “A photograph proving this bridge opened today.”
The visual file has not changed, but the second caption claims a real event that the illustration cannot establish. This demonstrates why asking “AI or not AI?” is only one part of media literacy.
Next imagine a verified camera photograph of a real bridge with a false date in its caption. The camera origin might be genuine while the event claim remains wrong. Source integrity and factual context must both be checked.
Five mistakes worth avoiding
- Trusting detector percentages as verdicts: classifications can be wrong or unsupported outside their tested conditions.
- Assuming missing metadata means fake: provenance may be absent for ordinary reasons.
- Assuming signed metadata means truthful: a real file can be captioned misleadingly.
- Using appearance as identity proof: a familiar face or voice is not always sufficient to confirm a request.
- Repeating an accusation too quickly: incorrectly labelling genuine media fake can harm innocent people.
Frequently asked questions
Can AI detect every deepfake?
No. Detection tools have limited coverage and may make errors. They should be assessed for the type of media and conditions they are designed to handle.
What do Content Credentials prove?
They can provide tamper-evident, signed information about a file’s recorded origin or editing history, subject to the claims and signers involved. They do not prove the depicted event is factually true.
Does a file without C2PA data have to be AI-generated?
No. Many legitimate files have no Content Credentials, and sharing or editing may affect available metadata.
Is every AI-generated picture a deepfake?
No. A clearly labelled imaginary landscape or educational illustration is synthetic media without necessarily impersonating anyone or misleading viewers.
How should I respond to a suspicious voice request?
Verify through a trusted separate channel, especially if the request involves money, access codes or personal information. Do not rely on the voice alone.
Your one-minute takeaway
- Start with the claim, then check origin, context and independent evidence.
- Detectors, watermarks and signed provenance provide useful but limited clues.
- A reliable conclusion should distinguish confirmed facts from unverified details.
Official reading and related AI guides
- C2PA Frequently Asked Questions — what Content Credentials can record and verify.
- C2PA and Content Credentials Explainer — why provenance and factual truth are different.
- Content Authenticity Initiative: Frequently Asked Questions — what presence and absence of credentials do and do not establish.
- US Office of Research Integrity: Content Provenance — responsible interpretation of credentials and watermarks.
- How AI Creates Images From Words — learn about synthetic-image generation.
- How AI Creates Human-Sounding Speech — why synthetic voices can sound natural.
- Why AI Sometimes Gives Confident but Wrong Answers — apply evidence-checking habits.