Deepfakes in Israel's 2026 Election: How to Verify AI-Generated Video, Audio and Images

A short video reaches you late in the evening. A public figure appears to say something that would change how you vote. It is fifteen seconds long and already being forwarded faster than anyone can check it. You have two options: pass it on, or work out whether it is real. The second takes time that the first does not, and that asymmetry is the whole problem. The question is not whether you could spot a flaw if you studied the file. It is whether anyone will study it before it has done its work.

This is the third article in our election security series. The second part covered how foreign influence operations can work without breaching election infrastructure, where the target is attention rather than systems. That does not make the two separate worlds, and a hack-and-leak operation combines them. Synthetic media is the sharpest tool in that kit, and it is also the one ordinary readers are most often asked to judge for themselves.

What counts as a deepfake in 2026?

Generative software can produce a face, a voice, a document or a scene that was never recorded, and it can alter authentic material until it means something else. Both are covered by the word deepfake, and neither is ordinary editing. A clip can be trimmed for length without being fake, and it can be genuine and still mislead if the caption places it in the wrong year. Keeping these apart matters, because the checks that expose them differ.

1. Synthetic or materially altered media: the content was generated by software, or edited by software until the underlying event changed. Video, audio, still images and documents are all in scope.
2. Cheapfakes: real footage that has been slowed, cropped, reordered or relabelled. No sophisticated tooling is involved, and they remain an important and often overlooked form of misleading political media.
3. Impersonation: an account, page or message presenting itself as a person or an institution it is not. The media attached to it may be perfectly genuine.
4. Voice cloning: synthetic speech built to sound like a specific person. It travels well in voice notes and phone calls, where there is nothing to look at.

Not every misleading political video is AI-generated, and calling them all deepfakes makes the genuine cases harder to argue about. It also hands anyone caught on a real recording a convenient excuse, a problem we return to below.

Israel changed the election rules for synthetic propaganda

Israel now has a disclosure rule for this, and it is narrower and more specific than most summaries suggest. The Elections to the 26th Knesset Law (Special Provisions and Legislative Amendments), published in the Book of Laws on 16 July 2026, inserted a temporary section 2A2 into the Elections (Modes of Propaganda) Law. It says that a person who publishes election propaganda that is a nechzut amuka, the Hebrew term the statute uses for a deep fake, must attach a disclosure stating clearly and prominently that the content in it was not originally documented.

The statute defines the term rather than leaving it to argument. A nechzut amuka is visual or audio content including a figure, place, event, document or other object that may appear to have been originally documented, but was created by a digital means, or edited by such a means in a way that materially altered it. A digital means is software capable of creating, editing or altering information or content, expressly including artificial intelligence. The purpose is not to referee whether a message is fair. It is to let a viewer know when something that looks recorded was in fact made or reshaped.

Three limits are worth knowing. The rule requires disclosure, not silence: it does not make synthetic political content unlawful. It does not reach propaganda that was first published before it came into force. And it is a temporary provision, running from a week after the law was published until a year after election day, so it belongs to this election rather than to Israeli law in general. The same section empowers the chair of the Central Elections Committee to set rules on how the disclosure obligation is met. None of this is legal advice, and anyone publishing campaign material should read the section itself.

Why looking for six fingers is no longer enough

The advice most people carry is a list of visual tells: an extra finger, teeth that shift between frames, eyes that do not blink, hair that dissolves at the edges. In 2019 that was reasonable. Today it is closer to a trap, because it teaches people to certify a fake as real the moment the obvious defects are gone.

1. Generation quality changes on a timescale of months, so any fixed list of artifacts is out of date shortly after it is written.
2. Compression removes detail. By the time a clip has passed through three platforms, the evidence you were told to look for has been squeezed out of it.
3. Short clips give you very little to inspect. Fifteen seconds of a talking head is not much surface area for a judgement.
4. Audio carries no visual tells at all. A voice note offers tone and cadence, and both are now reproducible.
5. Authentic video often looks wrong. Poor lighting, an aggressive codec, a hard cut and a compressed upload can each make a real recording look manufactured.

Looking harder at the pixels does not scale. Asking where the thing came from does.

A seven-step verification method that works better

1. Find the original upload. Work back to where the media first appeared rather than judging the account that forwarded it. Material with no traceable first publication is already telling you something.
2. Check the claimed speaker's own channels. If a public figure said it, the statement is usually on their own feed, their party's site or a broadcast. Absence is not proof of forgery, but it shifts the burden.
3. Look for independent corroboration. A genuinely shocking statement by a well-known figure gets reported quickly and by more than one newsroom. If a claim that size exists only inside the clip, treat it as unconfirmed.
4. Check the date and the context. Real footage relabelled as a new event can be just as misleading as fabrication, and it survives every technical test because there is nothing technically wrong with the file.
5. Inspect the account, not only the media. Creation date, posting history, a display name that is one character off, a domain that is nearly right. The harder case is a real account that has been taken over, which is one of the things an election-season cyberattack can actually reach. These are signals rather than verdicts, and a new account is not automatically hostile.
6. Treat audio-only claims with extra care. A cloned voice on a call or in a voice note is convincing precisely because there is nothing to look at. Verify through a different channel you already trust before acting.
7. Delay forwarding while verification is incomplete. Urgency is the manipulation, not a side effect of it. Nothing worth sharing gets worse because you checked it first.

Voice cloning may be more dangerous than video

A video invites inspection. A voice note does not. It arrives in a channel built for trust, it is short, and it usually asks for something small and urgent. A campaign worker receives a message that sounds like their manager approving an unusual transfer. A journalist takes a call that sounds like a source withdrawing a quote before deadline. A parent gets a voice note that sounds like their child. None of these has to be about politics, which is why this section belongs in a business article: aimed at a finance team, the same technique pursues the social-engineering objective behind business email compromise, delivered through a cloned voice rather than a spoofed inbox. The defence is identical in both settings, and it is procedural rather than perceptual: confirm through a channel the caller did not choose.

The liar's dividend: when real content is dismissed as AI

There is a second effect, and it runs the other way. Once everyone knows convincing fakes exist, anyone caught on a genuine recording can simply say it was generated. Researchers call this the liar's dividend, and it is the reason verification is not only a defence against fabrications. It is also what makes it possible to establish that something real happened.

This is the strongest argument for provenance. A method that only reports how a file looks to you has nothing to offer when the claim is that a real recording of you is fake. Provenance works in both directions.

Detection tools: useful signal, not a verdict

Automated deepfake detectors exist and some are useful. They are also probabilistic, they disagree with one another, and their accuracy drops on exactly the material you are most likely to be handed. Research on detecting video deepfakes under real-world conditions finds that detectors trained in controlled settings often fail to generalise once platform compression has stripped the forensic traces they rely on. A detector score is a reason to look further, not a finding you can publish: treating it as proof repeats the original mistake, moving the judgement from your eyes to a black box without making it more reliable. Provenance, corroboration and official sources remain the load-bearing checks.

What to do when suspicious election content arrives

1. Do not amplify it while you are still deciding. Quoting it to debunk it still spreads it.
2. Preserve the original link and account name before anything is deleted. A screenshot without a URL is hard to report and easy to dispute.
3. Check the official channel of whoever is supposedly speaking, and of any institution the content names.
4. Compare it against newsrooms that do their own verification, rather than against other posts repeating the same claim.
5. Report impersonation and account abuse through the platform's own process, and through official channels where the content targets an institution.
6. For any logistical claim about voting itself, where and when to vote, which documents are needed, whether a polling place has moved, treat the Central Elections Committee's voter information as the source of truth and ignore the clip.

What Israeli businesses should learn

1. Payment and bank-detail changes need out-of-band verification as a written rule, not as a habit. Call back on a number already on file, never one supplied inside the request.
2. Senior approval cannot rest on voice familiarity. If a voice can be cloned from a conference recording, an executive's voice is not an authentication factor.
3. Awareness work has to cover synthetic impersonation and not only classic phishing. Staff trained to look for clumsy grammar are not prepared for a fluent voice note.
4. Incident playbooks should already cover a company account being taken over, or someone impersonating the company convincingly enough to matter. Deciding who may publish a correction, and how fast, is much easier before it is needed.

None of this is exotic. It is training that teaches people to verify rather than only to be afraid, and managed security controls that keep identity, email and endpoint defences working rather than documented. NetFortress helps Israeli SMBs build the controls that modern impersonation actually tests. The election will pass. The technique will not.