The advice to look for six fingers has already stopped working, and the advice replacing it will stop working too. Every list of visual tells in the generated video describes the failures of whichever models were current when the list was written. Models improve, the tells disappear, and anybody still checking for them is checking for something that stopped being true.
That matters more in some places than others. Where video arrives forwarded through messaging apps, compressed several times over, stripped of its origin and its caption, visual inspection was never going to be reliable anyway. The artefacts you would look for are the first thing compression removes.
Which points to a more durable method, and it is the one newsrooms have always used. Kling 3.0 with advanced AI video generation capabilities from Higgsfield and the models alongside it generate convincing footage, but no amount of capability changes where a clip came from or who had it first.
What Do Current Models Still Fail At?
Physics over time, and anything that has to stay consistent.
Generating a convincing second is close to solved. Keeping a scene coherent across several seconds is a different problem, and that is where most failures still appear.
Where to look while it still works:
- Objects passing behind things, which a Kling 3.0 scene can still reappear wrongly
- Shadows and reflections, which do not always match their source
- Hands in sustained motion, though less than a year ago
- Background crowds, where individuals morph or merge
- Text in the scene, which rarely survives at all
The shadow mismatch is among the more reliable indicators remaining. A light source implies shadows falling a particular way, and generated scenes frequently get the direction right and the behaviour wrong.
Watch anything suspicious twice and at full size. Drift becomes obvious on a second viewing and invisible on a first.
Treat all of this as a current snapshot rather than a method. These specific failures will be shorter-lived than the article describing them.
Why Is Visual Inspection a Losing Strategy?
Because the gap closes from both directions.
Models improve, which removes tells. Compression also removes tells, which means a genuine clip and a generated one both arrive degraded through the same pipeline.
- Each forward re-compresses the file further
- Messaging apps compress aggressively by design
- Lower bandwidth settings reduce quality further still
- Artefacts from compression resemble generation artefacts
- A Kling 3.0 clip compressed three times looks much like anything else compressed three times
The convergence is the practical problem. Somebody examining a heavily forwarded clip is examining compression damage, and reading that as evidence either way is unreliable in both directions.
This is why the forwarded clip arriving on a phone is the hardest case there is, and also the most common one.
What Actually Works Instead?
Provenance, which is the oldest method in journalism.
- Find the earliest posting you can, not the one you received
- Check who posted it first and what else they post
- Look for a second source with the same footage
- Check whether anybody present has recorded it differently
- Ask what a Kling 3.0 user would need to make this, and whether anybody had reason to
The second source test is the strongest. A real event at a public place produces multiple recordings from multiple angles, and a single clip with no corroboration from an event that many people attended is the thing to question.
Reverse image search on a few frames frequently finds the original, including cases where genuine old footage is being recirculated as something new. That is a more common form of misinformation than generation and it is easier to check.
Ask who benefits. Fabrications are produced for reasons, and the reason usually narrows the search.
Does Device and Data Cost Affect This?
It does, in ways that make verification harder for exactly the people most exposed to it.
Checking a clip properly means opening it at full size, watching it more than once, and running searches. All of that costs data, and where data is expensive people do none of it.
What that produces in practice:
- Clips watched once, at low quality, on a small screen
- No reverse searching, since it costs a bundle somebody is rationing
- Forwarding without opening, which happens more than anybody admits
- Low-resolution playback that hides everything useful
- A Kling 3.0 clip judged in four seconds on a cracked screen
The forwarding-without-opening behaviour is the significant one. A clip shared on the strength of its caption has been verified by nobody in the chain, and the chain can be very long.
This is also why the provenance question is the right thing to teach. It costs no data to ask who sent this and where they got it, and it can be asked without opening anything.
Low-data habits are worth designing guidance around rather than ignoring. Advice that assumes unlimited bandwidth is advice for a different audience.
What Should a Newsroom Never Do?
Illustrate a story with generated footage.
- Never create video of a real event
- Never generate a public figure, in any context
- Never produce establishing footage of a real place and present it as filmed
- Never use generated material in election coverage at all
- Never make a Kling 3.0 clip that a reader could mistake for reporting
The establishing shot is where this slips in newsrooms, because it feels decorative. A generated street scene behind a story about that street is a fabricated image of a real place, and discovering it costs the publication more than the footage saved.
Label anything from Higgsfield clearly and in the caption. Readers accept a marked illustration without difficulty.
Election periods deserve a blanket rule rather than a judgement call. The window in which a fabrication can affect an outcome is short, and the cost of being wrong is not recoverable.
Where Does Generation Legitimately Belong?
Explanation, concept and atmosphere, all labelled.
- Illustrating a concept in Higgsfield that cannot be filmed
- Explainer sequences showing how something works
- Atmospheric material behind text or narration
- Product and brand concept work, before commissioning
- A Kling 3.0 sequence that makes no claim about reality
The explainer case is genuinely strong. Showing how a network, a process or a mechanism works is illustration by nature, and nobody mistakes a diagrammatic sequence for footage.
Keep the labelling on Higgsfield material consistent and visible. A publication that labels routinely is one whose unlabelled material is trusted.
Higgsfield runs it in the browser as part of an AI creative suite, which suits a small newsroom producing its own explainer material between stories without new equipment.
How Should Audiences Be Taught to Check?
With one question rather than a checklist.
- Where did this come from is the whole method, whatever made it
- Who posted it first, and when
- Is anybody else reporting it
- Does the account have a history, or was it created recently
- Would somebody have reason to make this
The single question survives technological change in a way a list of visual tells does not. A reader taught to check for artefacts is equipped for eighteen months; a reader taught to ask about origin is equipped indefinitely.
Recently created accounts posting dramatic content are a durable signal. The account history is frequently more informative than the clip.
Keep the guidance short enough to remember while holding a phone. Anything longer does not get used.
Does Watermarking Help?
Partially, and it should not be relied on alone.
Provenance standards and embedded watermarking are being adopted across the industry, and they work reasonably where the chain is intact.
- Metadata is stripped by most platforms on upload, Kling 3.0 provenance included
- Re-encoding removes many embedded signals
- Screen recording defeats nearly all of them
- Absence of a watermark proves nothing, which is the key limitation
- A Kling 3.0 output re-recorded carries very little forward
The absence problem is the one that matters most. A watermark present is useful evidence; a watermark missing tells you only that it is missing, which is true of most genuine footage too.
Treat watermarking as one signal among several rather than a verdict.
What About Audio?
Easier to fabricate and harder to check, which makes it the bigger risk.
Voice cloning requires considerably less source material than video and produces results that survive compression better, because phone audio is low quality to begin with.
- A short voice note is more common than any Kling 3.0 video
- Compression hides the artefacts almost completely
- No visual context means no corroborating detail
- Forwarded audio has even less provenance than video
- Apply the same origin question to audio as to video
Audio deserves more suspicion than it receives, not less. People treat a voice as more personal and therefore more credible, which is precisely what makes it effective.
Ask whether the person has said the same thing anywhere attributable. A claim that exists only in a forwarded voice note is a claim with no source.
Does the Local Context Change the Approach?
Considerably, and the differences matter.
Where a substantial share of content circulates through closed messaging groups, the chain of custody is broken by design. A clip arrives with no author, no timestamp and no caption beyond whatever the forwarder wrote.
- Forwarding strips context at every hop, Kling 3.0 output included
- Group sharing hides the origin entirely
- Fact-checking capacity is thinner outside major languages
- Correction rarely travels as far as the original
- Generated content in local languages faces less scrutiny than English content
The regulatory response to this is live, as coverage of the UCC moving against AI misinformation after the election illustrates, and enforcement is only ever part of the answer.
The practical layer is slower and more durable: people who know to ask where something came from before they forward it.
Teaching the provenance question does more than teaching the tells, because the question survives every improvement in the technology.
What Should a be in a Publication Document?
Its own practice, publicly.
- State your policy on Kling 3.0 and generated imagery, in writing
- Label every illustration consistently
- Record what Higgsfield produced and when
- Keep the Higgsfield project separate from any footage archive, permanently
- Correct publicly when something gets through unlabelled
Publishing the policy does work that following it quietly does not. Readers who know a publication’s position can judge its material, and that is the whole basis of trust in a period when images cannot be taken at face value.
Keeping generated material physically separate from the footage archive prevents the failure that actually damages a newsroom, which is an illustration being used years later as though it were a photograph.
What Would Checking One Clip Take?
Five minutes, usually less.
- Watch it twice at full size, looking for drift
- Check shadows and reflections against the implied light, as a Kling 3.0 scene may not match
- Reverse search several frames for an earlier version
- Look for a second recording of the same moment
- Check the posting account’s history
- Ask who benefits from it being believed
- Ask whether a Kling 3.0 user would have had reason to make it
Then decide on the sourcing rather than the pixels.
The sourcing is what will still be reliable next year. The pixels will not.
Conclusion
Every list of visual tells is a list with an expiry date, and the one in this article will expire too.
Kling 3.0 and the models beside it generate convincing footage, which means the reliable question has shifted from what does it look like to where it came from.
Watch it twice, look for a second source, check the account, and ask who benefits. That method was useful before any of this existed and it will still work afterwards.
