A friend sent me a video of a famous person saying something ridiculous.

“Real?” he asked.

Five years ago, I would have watched the face carefully, looked for strange blinking and declared myself an amateur forensic expert.

Now I do something much less satisfying.

I check where the video came from.

Deepfakes and AI-generated media have improved enough that visual intuition is becoming a weak security system. Sometimes the fake is obvious. Sometimes it is excellent. The bigger problem is that people now know convincing fakes exist, which means even genuine recordings can be dismissed as fake when they are inconvenient.

We are entering an awkward period where seeing something is evidence, but not always enough evidence.

The weird fingers were temporary help

Early AI images had useful tells.

Hands were strange. Text looked melted. Earrings did not match. Background objects turned into impossible furniture.

People learned the tricks and felt clever.

The models improved.

This is how detection tends to work. A generation technique leaves patterns. Detectors learn them. The generation improves or changes. The clues become less reliable.

I would not build a serious verification habit around counting fingers anymore.

Human inspection is still useful for low-quality scams, but strong fakes require context. Who published this? Is there another recording? Did reputable sources report the event? Does the original account contain it?

The boring verification steps are aging better than the visual tricks.

Voice cloning worries me more

Video deepfakes attract attention because they are visible.

Voice cloning may be more useful to ordinary scammers.

A short sample of someone speaking can be enough for modern systems to imitate tone and rhythm surprisingly well. That creates a believable phone call or voice message without requiring perfect video.

Imagine receiving a rushed message from a manager asking you to buy gift cards. Or a relative saying they lost their phone and need money sent to a new account.

The request is already unusual.

The familiar voice lowers suspicion.

My family has a simple rule for urgent money requests: call back using the number we already know. If the story is real, a thirty-second verification is not an insult.

If the caller tells you not to verify because the situation is too urgent, that is additional information.

Companies need a second channel

Businesses have spent years teaching employees to recognise suspicious emails.

Deepfake scams make voice and video less trustworthy too.

A finance employee may receive a convincing call that appears to come from an executive. A supplier may send a video message explaining that payment details changed. A remote employee could encounter an impersonated colleague during a meeting.

The solution is not to distrust every call.

It is to move high-risk actions onto a verification process that does not depend on one communication channel.

Payment changes should require known approval steps. Password resets should not happen because someone sounds important on video. Sensitive requests can be confirmed through an internal system or a separate contact method.

Security becomes procedural when perception becomes unreliable.

That is less glamorous than AI detection.

It is more dependable.

AI detectors are useful, not magical

There are tools that analyse media for signs of manipulation or synthetic generation.

They can help.

I would not treat one detector score as a courtroom verdict.

Models change quickly. Compression, editing and screen recording can alter the signals detectors rely on. Different tools may disagree.

Detection is one piece of evidence.

Provenance is another interesting direction: attaching information about where media came from and how it was edited. Cameras, publishers and software can record parts of the content history so viewers have more than the pixels themselves.

That will help if adoption becomes broad.

It will not make unverified media disappear from group chats.

Nothing can achieve that level of optimism.

Social media rewards the wrong speed

A fake video often succeeds before anyone finishes verifying it.

The first few minutes matter because social platforms reward attention, surprise and outrage. People repost the clip while asking whether it is real, which unfortunately distributes it either way.

I have become slower about sharing remarkable media.

If the video would seriously change my opinion of a person or event, that is exactly when I should spend an extra minute checking it.

Extraordinary claims have always deserved more evidence. AI-generated media simply made the old rule feel current again.

The internet does not reward patience.

You are allowed to use it anyway.

Personal photos are training material for impersonation

Most people have enough public media online to provide an impersonator with useful material.

LinkedIn has a clear headshot. Instagram has casual video. Podcasts and conference recordings contain clean voice samples. Company websites list job titles and colleagues.

I am not suggesting everyone delete themselves from the internet.

That is unrealistic for many people and unnecessary for most.

But people in roles involving money, sensitive access or public visibility should assume their voice and image can be copied.

The security control should protect the action, not the face.

A bank transfer should remain difficult even if the attacker looks and sounds exactly like the CEO.

Real footage now has a credibility problem

There is another consequence that gets less attention.

As deepfakes become common, people can claim genuine evidence is fake.

A real recording of misconduct may be dismissed as AI-generated. A politician, executive or ordinary person can point to the existence of synthetic media and create doubt.

This is sometimes called the liar’s dividend: the technology does not only create false evidence; it weakens confidence in real evidence.

That makes trusted sources, original files and documented provenance more important.

The goal is no longer just spotting fakes.

It is preserving ways to establish what is real.

The habit I trust is boring

When a surprising video arrives, I ask three questions.

Where did it originate?

Can I find independent confirmation?

Does the person posting it have access to the original, or are they reposting a repost of a screen recording?

Sometimes I still cannot tell.

That is okay.

“I don’t know yet” is a perfectly functional conclusion.

Technology culture often treats uncertainty like a bug. In a world of cheap synthetic media, uncertainty is sometimes the most honest state available.

Seeing remains useful.

I just stopped treating it as the final step.