Was sind Deepfakes? So erkennen Sie sie und schützen sich

Key takeaways

• Deepfakes use AI to create or alter someone’s appearance or voice, making fabricated content seem authentic.
• Visual and audio glitches can raise suspicion, but convincing deepfakes may have no obvious flaws.
• Checking the source and confirming requests through a separate, trusted channel are more reliable than judging appearance alone.

Deepfakes are AI-generated or AI-altered video, image or audio content that makes someone appear to say or do something they did not. To spot them, combine careful inspection with source checks and independent verification—no single visual clue or detection tool can prove authenticity.

How it works

Deepfake systems learn patterns in faces, voices and movement, then use those patterns to generate or modify media. Depending on the method, they may replace a face, imitate a voice, change lip movements or animate a still image.

Common forms include:

• Face swaps: Replacing one person’s face with another’s.
• Lip-sync manipulation: Changing mouth movements to match a different speech track.
• Voice cloning: Generating speech that resembles a particular person’s voice.
• Synthetic portraits and avatars: Creating a person who does not exist or making a portrait speak.

Not all AI-generated media is a deceptive deepfake. A clearly labelled fictional character is different from a fabricated recording presented as evidence that a real person said something.

Likewise, not every misleading clip uses AI. An old video with a false caption, a selective edit or slowed-down footage can deceive viewers without deepfake technology. Establishing the original context matters either way.

Real-world examples

These examples illustrate common uses and risks, rather than specific reported incidents.

An urgent call from a “relative.” Someone receives a distressed voice message asking for emergency money. The voice sounds familiar, but it could be cloned. Calling the relative on their saved number is safer than continuing through the unfamiliar contact.

A fake executive instruction. An employee gets a video or voice message apparently from a senior manager requesting a confidential payment. The strongest warning signs may be urgency, secrecy and an unusual payment process—not a distorted face.

A fabricated public statement. A clip appears to show a politician or business leader making an inflammatory announcement. Before sharing it, look for the full recording, its original publisher and independent reporting.

A talking character for training. A company uses an authorised image and voice track to create a labelled training presenter. Tools such as Hedra can generate expressive talking-character videos from an image and a voice track. That capability has legitimate uses when consent and disclosure are handled properly.

A misleading profile image. A synthetic portrait helps someone present a fake identity online. However, an unusual-looking portrait alone does not establish fraud; genuine photographs can also contain editing, compression or lighting artifacts.

Benefits and limits

Synthetic media can support filmmaking, accessible communication, localisation and educational content. Creators can prototype scenes or build fictional presenters without arranging a conventional shoot for every version.

For example, Runway (Gen-3 Alpha) supports text-to-video and image-to-video generation. Such tools demonstrate useful creative capabilities, but their existence does not tell you whether a particular suspicious clip was made with them.

The risks include impersonation, financial fraud, reputational damage and non-consensual intimate imagery. Consent to use someone’s photograph does not automatically mean consent to clone their voice or portray them in a fabricated scene.

Detection also has important limits:

• Artifacts are inconsistent. Strange teeth, flickering edges or mismatched lip movements can be clues, but newer or carefully produced content may avoid them.
• Real media can look artificial. Poor lighting, filters, compression and connection delays can create similar problems.
• Detection tools can be wrong. A detector may flag authentic material or miss generated content.
• Missing provenance proves little. The absence of a watermark or content credential does not establish that something is fake.

Treat detection as evidence gathering, not a yes-or-no verdict based on one clue.

How to get started

Use this checklist whenever a video, image or voice message could affect your money, reputation or security.

1. Pause before acting

Be especially cautious when content triggers panic, outrage or an urgent demand. Do not send money, reveal passwords or share the clip simply because it looks convincing.

An instruction to bypass normal checks is a warning sign even if the person’s face and voice appear genuine.

2. Inspect the media in context

For video, watch for facial details that change between frames, unstable glasses or jewellery, inconsistent lighting, and mouth movements that do not match speech.

For an image, examine reflections, shadows, text and boundaries around faces or objects. For audio, listen for abrupt changes in tone, unnatural pacing or inconsistent background sound.

These are prompts to investigate—not proof. A frozen frame can also make an ordinary expression look strange.

3. Find the original source

Ask who first published the material and whether you can locate a longer or higher-quality version. Check the publication date, account history and surrounding context.

A reverse-image search may reveal that an image is older than claimed or copied from another source. Searching distinctive phrases from a video can help locate the original speech. No search result, however, guarantees authenticity.

4. Verify through a separate channel

If someone requests money or sensitive information, contact them using a number or account you already trust. Do not rely on contact details supplied in the suspicious message.

At work, follow established payment approval procedures. For families, consider an agreed verification phrase, but do not make that your only safeguard.

5. Use detection tools cautiously

Deepfakedetector.ai offers detection of synthetic media and generated images. Its output can support an investigation, but should not be your sole basis for accusing someone or approving a transaction.

Before uploading anything, review the service’s privacy and retention terms. Avoid submitting confidential documents, sensitive recordings or other people’s private imagery without appropriate permission.

6. Preserve and report suspicious content

Keep the original message, account details and relevant timestamps where safe and lawful. Report impersonation to the platform and workplace incidents to your security team.

If you sent money, contact your bank or payment provider immediately. Avoid publicly reposting harmful content merely to warn others.

FAQ

Can you spot a deepfake just by looking at it?

Sometimes you can notice suspicious details, but appearance alone is unreliable. Authentic footage can contain glitches, while a convincing deepfake may look normal. Source checks and independent confirmation provide stronger evidence.

Are deepfake detectors accurate?

Accuracy varies with the detector, media quality and generation method. Cropping, compression or unfamiliar techniques can affect results. Treat any score as a signal to investigate, not a definitive probability that the content is fake.

What should I do if someone deepfakes me?

Preserve evidence without spreading the material, report the account or content, and notify people likely to be targeted. For workplace impersonation, involve your security team. If there are threats, extortion or intimate imagery, seek appropriate local legal or victim-support assistance.