Gains:
- Ability to recognize types, clues and propagation patterns of synthetic text, images, audio and deepfake video
- Ability to manage synthetic content suspicion with source tracking, metadata, reverse search and context verification and use artificial intelligence as an organizational tool in defense
- Ability to understand that detection tools are not definitive evidence, that judgment comes down to the source, and that confirmation news should be written in a calm, transparent language.
Artificial intelligence is not just a journalist's tool; It is also the engine of the threat against it. Today, convincing fake text, seemingly real photographs, faces of non-existent people, and voices uttering sentences that someone never uttered can be produced in minutes. These are collectively called synthetic content (artificially produced media); The type of video/audio that realistically imitates a person's face or voice is called deepfake. Disinformation—intentionally spread false information—has become both cheaper and faster through these tools. The journalist's duty has doubled: both to prevent this content from being published and to inform the public about this threat. In this unit, we cover using AI to both understand the threat and defend against it. Basic principle: No detection tool is conclusive evidence; Suspicion of synthetic content is resolved by classical verification that goes down to source and context.
Types and tips of synthetic content
Text. AI can produce fluent but empty, generalization-filled, made-up texts; fake "news sites" and comment armies thrive on this. Hint: unverifiable sources, overly general statements, persistently one-sided framing.
Visual. In artificial images, hands/fingers, teeth, ears, jewelry, background text and reflections are often inconsistent; But technology is improving rapidly, so "I can tell by looking at it with my own eyes" is a dangerous trust.
Sound. Unspoken sentences can be produced by the voice of an official. Tip: breath/pause naturalness, ambient sound consistency, but the most reliable check is to verify the source of the content.
Video (deepfake). Face and neck/light mismatch, eye blinking, lip-voice synchronization may be clues; However, it is not certain.
Attention: "AI detection tool says 90% are fake" is not a reason for publication. These tools are wrong; He can call the real fake and the fake real. The output of the tool is a clue; The judgment comes down to the source.
Step by step: managing suspicion of synthetic ingredients
Step 1 — Find the source. Who published the content first, where and when? Going down to the original source alone will bust most fakes.
Step 2 — Look at metadata and reverse search. Metadata and reverse image search on image/video may reveal a previous version or original.
Step 3 — Test internal consistency. Do the light, shadow, reflection, ambient sound, background real-world clues (signage, weather, architecture) match the claim?
Step 4 — Verify the context. Has the incident in the content been seen by other independent sources? There are many traces of a real event; single source "bombshell" content is questionable.
Step 5 — Use detection tools as clues. Count the output of AI detection tools as an auxiliary signal, not evidence.
Step 6 — Report transparently. If you publish it (as fact-checking), show what is fake/suspicious, why, and how you verified it.
Using AI in defense
AI is also an organizational tool against synthetic content: it lists what to look for in an image, lists signs that a text may be AI-generated, helps parse the narrative of a disinformation campaign, drafts an explanatory fact-check for the public. But the determination itself — the “it's fake” verdict — is made by human judgment and source verification.
three mini cases
Case 1 — Fake voice recording. An audio recording of an official "resigning" was circulated. The reporter first searched for the source: the recording had come from a single anonymous account, there was no trace of it anywhere else. The official's team rejected the recording; ambient sound was also inconsistent. The reporter did not publish the recording, instead reporting on an "unverifiable audio recording." The recording later turned out to be synthetic.
Case 2 — Time pressure and the stop reflex. The photo of a "collapsed building" after an earthquake spread rapidly. The reporter did a reverse image search: the photo was from another event 2 years ago. The 4-minute check prevented an incorrect image from entering the news site; A few rival sites published it without checking and published a denial.
Case 3 — Overconfidence in the detection tool. An editor inserted a real reporter's photo into a detection tool; The vehicle is "78% artificial," he said. The editor was about to withdraw it from publication; When a senior fact-checker showed the photo's shooting metadata and the reporter's raw file, it became clear that the photo was real. Lesson: the medium errs, the source speaks.
Copiable templates
1) Visual syntheticity checklist:
I suspect an image may be artificially produced. Make a checklist of things I need to test: hands/fingers, teeth, jewelry, background text, reflections, light-shadow consistency, reverse search, metadata. Write down how to look for each item. Don't say "fake/real" for sure; just checklist. Visual description: [here]
2) Analyzing the disinformation narrative:
The following claim may be a disinformation campaign. Infer: (1) what emotion it targets, (2) what divide it amplifies, (3) verifiable/unverifiable components, (4) what the typical pattern of spread might be. Making a judgment of righteousness; analyze. Claim: [here]
3) Draft confirmation news for the public:
I have verified with primary sources that the following content is synthetic/misleading: [summary of findings]. Write a draft of a confirmation story that informs the reader, does not spread panic, and shows transparently how I verified. Do not include any findings that I did not provide. Findings: [here]
4) AI-generated marks (hint) of the text:
List CLUES that the following text may be artificially produced: unverifiable sources, overgenerality, one-sided framing, repetitive patterns. Emphasize that this is not conclusive evidence and write down what factual verifications need to be made. Text: [here]
Weak prompt / Strong prompt
Weak prompt: "Tell me if this image is fake."
Result: AI gives a precise-looking but unreliable “yes/no”; The journalist may rely on this ruling and be mistaken.
Strong prompt: "Make a list of signs I would check for that this image might be artificial, and write down how I would look for each one. Don't make any final judgments; point out that the real verification is going back to the source and doing a reverse search."
The difference: A powerful prompt reduces the AI to a clue generator, leaving judgment to the source and journalist, making the tool's margin of error visible.
comparison chart
Content type
Weak clues
The most reliable control
Role of AI
text
General, made-up source
go down to the source
Sign list
visual
Hand/teeth/writing discrepancy
Reverse search + metadata
checklist
sound
Breath/environment inconsistency
Source and context confirmation
What to look for
Video (deepfake)
Light/sync mismatch
Original source, cross-confirmed
Analysis help
Common mistakes
- Treating the detection tool as evidence. Making the percentage of the vehicle the reason for publication; tools err in both directions.
- "I can tell by sight" confidence. Synthetic visuals are improving rapidly; Visual intuition alone is not enough.
- Not going to the source. Most fakes collapse when the original source that published them is found; Skipping this step is a mistake.
- Spreading the suspect. Sharing suspicious content saying "let's see if it's true or not" serves disinformation.
- Language of panic. The alarmist tone in the confirmation news spreads new fear while correcting; Write calmly and transparently.
Diffusion speed and the "denial trap"
The real power of disinformation is not in the content itself, but in the speed at which it spreads. Emotionally provocative content that evokes anger or fear is shared much faster than verified news; While fake content reaches millions within hours, redaction often reaches far fewer people. This asymmetry places a special responsibility on the journalist: sharing questionable content, even by saying "look how ridiculous", gives it access. This is called oxygenation — repeating a lie to disprove it actually fuels it.
The antidote to this is the "truth sandwich" technique in fact-checking: tell the truth before repeating the false claim, then show why the claim is false, then repeat the truth again at the end. In this way, what remains in the reader's mind is not the mistake, but the truth. AI can help draft a fact-checking story according to this structure; But it is the editorial decision which truth to put forward and to inform the language without spreading panic. Another practice: asking what gap is disinformation filling? People often embrace fake content due to lack of reliable information; Filling that gap with fast, accurate and accessible news is a more effective defense than chasing lies one by one.
In summary
Synthetic content and deepfakes have cheapened and accelerated disinformation; The journalist's duty is to not publish this and to inform the public. Detection tools and AI provide clues and organization — but not definitive proof. Judgment is made by going to the source, metadata and reverse search, internal consistency and independent context. Confirmation news is written in a calm, transparent language that shows how it was verified.
Application task
Find visual or audio content that you think is suspicious. Follow the “Visual syntheticity checklist” (or adaptation for audio) prompt; really check each item and reverse search. Rate the result as “real/fake/insufficient evidence” and compare it to the decision you would make if you relied solely on the detection tool.
checklist
- [ ] I consider the output of the detection tools as clues, not evidence.
- [ ] For any questionable content, I first look for the original source and first publication.
- [ ] I make sure to reverse search and metadata check the images.
- [ ] I do not spread unverified questionable content.
- [ ] I write confirmation news in a calm, transparent language that shows how I verified it.