What the terms mean
Synthetic media is a broad label for images, audio, video or text that has been created or heavily altered by computer programs, usually using machine learning. A deepfake is a particular kind: synthetic media that imitates a real person’s face, body or voice so that it seems to show them doing or saying something they did not.
Not all synthetic media is harmful. Studios use these techniques for special effects and dubbing, accessibility tools use synthetic voices, and many people make parody or art with them. The concern is with content presented as real when it is not.
How they are made, in outline
The details vary, but the general approach is similar to other generative AI. A model is trained on many examples of a person’s face or voice, or on large collections of images and audio in general. It learns patterns, such as how a mouth moves when forming certain sounds or how a particular voice rises and falls. It can then produce new material that follows those patterns.
Common types include:
- Face swaps. One person’s face is placed onto another person’s body in a video.
- Lip-sync edits. Existing footage is changed so the mouth matches new words.
- Voice cloning. A voice is imitated from a sample recording, which can then read any text.
- Fully generated people or scenes. Pictures or clips of people who never existed, or events that never occurred.
The tools have become easier to use over time, which is why the topic matters beyond specialists.
Why they can be convincing
Humans are good at noticing faces and voices but not perfect at noticing small errors. Compression on social media makes video blurrier, which hides flaws. Short, low-quality clips, such as a call recording or a grainy video, leave less to inspect. People also tend to believe things that fit what they already expect or feel strongly about, and creators of misleading content count on that.
There is a second effect that is easy to overlook: when fakes are widely known, genuine recordings can be dismissed as fake. Doubt can be used as a tool, so the goal is calibrated judgment, not blanket suspicion.
Clues that sometimes help, with caveats
Experts sometimes point to visual or audio details that may suggest manipulation:
- Unnatural blinking, odd teeth, or skin that looks too smooth
- Mismatched lighting or shadows between face and background
- Blurry edges around hair, ears or glasses
- Lip movements that do not quite match the sounds
- A voice with unusual rhythm, flat emotion or odd breathing
- Hands, jewelry or text in the image that look distorted
These clues are unreliable. Tools improve, and the flaws that are visible today may be gone later. Real recordings can also have odd artifacts caused by cameras and compression. Absence of obvious flaws proves nothing. So visual inspection should be only one input, and rarely the deciding one.
Habits that work better than spotting glitches
Pause before reacting. Content designed to provoke anger, fear or excitement is built to be shared quickly. A short delay is one of the most effective defenses.
Check the source. Who posted this first? Is it an account with a track record, a news organization, or an anonymous profile? Can you find where the original clip came from?
Look for corroboration. If something is genuinely important, other reliable outlets or people who were present will usually report it. If only one unfamiliar account has it, be cautious.
Check the context. Real footage is often reused with a false caption. Ask when and where it was supposedly recorded and whether the details match.
Search for the original. Reverse image search tools can sometimes show whether a picture appeared earlier in a different context.
Consider the motive. Who gains if this is believed? That question does not prove anything, but it can guide how much checking to do.
Hold off on sharing when unsure. Passing along a doubtful clip, even to say it is outrageous, helps it spread. The broader habits for online misinformation apply fully here.
When the target is you or someone you know
Voice cloning has been used in scams where a caller imitates a relative or colleague and asks for urgent money or information. If a call or message feels rushed and emotional:
- Hang up and call back using a number you already know.
- Agree on a family code word or a verification question in advance.
- Be careful with requests for payments, gift cards, transfers or passwords, whatever the voice sounds like.
This overlaps with the patterns described in online scams, and a family approach to online safety can make such agreements routine.
Synthetic images have also been used to harass or embarrass people, including through fabricated intimate imagery. If this happens to you, general steps include saving evidence such as screenshots and links, reporting the content to the platform, and seeking support. Laws and available remedies differ widely between countries and regions, so consider contacting a qualified legal or support professional where you live. Many places have organizations that help people in this situation.
Responsible use of the tools
If you create synthetic media for fun or work, avoid imitating real people without their permission, label the content clearly as synthetic where viewers might be misled, and follow the terms of the tool and the platform. Some platforms and regions require or encourage labeling, and the rules continue to change.
Keeping a balanced view
Technical detection tools and provenance labels, which attach information about how media was made, are being developed, but none is perfect or universal at present. For the foreseeable future, context and source matter more than any single visual cue. A sensible stance is neither “everything is fake” nor “seeing is believing,” but “how do I know, and how much does it matter if I am wrong?”