A deepfake is synthetic or manipulated media that uses AI to imitate or alter a person’s identity, appearance, voice, or actions. AI-generated content is the broader category: it can depict fictional people, places, objects, or events without impersonating any real individual. Every deepfake is AI-generated or AI-modified content, but not every AI-generated asset is a deepfake.
That distinction affects both risk and verification. A fully generated landscape and a face-swapped video may both be synthetic, but only one depends on an identity claim about a real person.
What is a deepfake?
The term “deepfake” combines deep learning and fake. In common use, it refers to media created or altered with machine-learning techniques to make a real person appear to say or do something they did not say or do, or to imitate their face or voice.
Definitions vary at the edges. NIST describes synthetic media broadly as information significantly modified or generated by algorithms, including AI. NIST’s media-forensics materials divide face-related manipulation into categories such as identity swap, expression swap, attribute manipulation, and entire-face synthesis.
WITNESS uses “synthetic media” as the wider family and describes deepfakes as realistic simulations of someone’s face, voice, or actions. This person-centered definition is useful because impersonation is usually what makes a deepfake especially harmful.
What is the difference between a deepfake and AI-generated content?
AI-generated content includes text, images, audio, and video produced or substantially transformed by generative models. It may be clearly disclosed, fictional, artistic, or designed for an ordinary commercial workflow. No real person needs to be copied.
A deepfake typically targets identity. It may place one person’s face on another body, reenact facial expressions, generate speech in a recognizable voice, or create a realistic scene involving a named individual.
| Media type | Uses AI? | Imitates a real person? | Typical example |
|---|---|---|---|
| Generated illustration | Yes | Not necessarily | A fictional city created from a prompt |
| Fully generated presenter | Yes | No, if the identity is fictional | A disclosed virtual host |
| Face swap | Yes | Yes | One person’s face placed onto another performance |
| Facial reenactment | Yes | Yes | A real person’s expressions or mouth movements altered |
| Voice clone | Yes | Yes | Synthetic speech imitating a known speaker |
| Conventional misleading edit | Not necessarily | Possibly | A real clip cut to remove context |
The last row matters: harmful or deceptive media is not automatically a deepfake. A recycled video, false caption, selective cut, or manual edit can mislead without generative AI.
What are the main types of deepfakes?
Face swaps
A face swap replaces or blends a source identity into a target image or video. The underlying body, motion, setting, and audio may remain genuine even though the apparent identity changes.
Facial reenactment and lip-sync manipulation
Reenactment changes expressions, gaze, head motion, or mouth movement while preserving the target person’s identity. Lip-sync systems can make a person appear to speak words from replacement audio.
Voice cloning
A voice clone generates speech that imitates a particular speaker. It may appear in a phone call, voice note, podcast clip, or as the audio track of an altered video. The US Federal Trade Commission warns that impersonation scammers can use voice cloning in fake emergency requests and advises contacting the supposed caller through a known channel.
Synthetic identity and body manipulation
Some media combines a generated face, altered body, and invented context. Other content changes age, hair, clothing, or attributes without replacing the entire identity. These cases show why “deepfake” is not a single technical operation.
Are fully AI-generated people deepfakes?
A generated person who does not imitate an identifiable individual is usually better described as a synthetic person or AI-generated character. Calling every photorealistic AI portrait a deepfake can blur the identity-based harms that the term was created to describe.
The label may still be appropriate when a generated scene falsely depicts a real named person, even if the system created the entire frame rather than swapping a face. In practice, describe both the technique and the claim: “AI-generated image falsely depicting a public figure” is more precise than a label alone.
Are all deepfakes malicious?
No. The technique can be used with consent for dubbing, accessibility, satire, film production, education, or disclosed creative work. The ethical and legal concern depends on consent, disclosure, context, and harm—not merely the use of AI.
High-risk uses include non-consensual intimate imagery, fraud, harassment, evidence manipulation, and impersonation intended to influence decisions. Even then, do not accuse a creator or subject based only on an automated detector result. Preserve the original source and seek qualified review for consequential cases.
How can you verify a suspected deepfake?
Begin with the claim and identity, not visual anomalies. Ask who is supposedly shown or heard, what they allegedly said or did, and where the media first appeared.
- Preserve the URL, account, caption, timestamps, and best available file.
- Find a longer or earlier version through source tracing and representative-frame searches.
- Compare the claim with official recordings, reliable reporting, or another known channel.
- Review adjacent frames for persistent identity, motion, and compositing inconsistencies.
- Examine audio separately; check for edits and verify the speaker through an independent contact path.
- Inspect available Content Credentials or other provenance records.
- Use a suitable detector as supporting evidence and record its stated coverage.
For a detailed media workflow, read How to Tell If a Video Is AI-Generated. The process applies to synthetic scenes as well as suspected identity manipulation.
Why visual “deepfake tells” are not enough
Fixed checklists age quickly. Unnatural blinking, warped teeth, strange hands, or edge artifacts can provide useful leads, but modern systems may avoid them. Genuine footage can also contain blur, compression, dropped frames, dubbing, filters, or lighting effects that resemble manipulation.
Inspect a sequence rather than one screenshot. Ask whether an anomaly persists in adjacent frames and a better-quality copy. More importantly, look for external evidence: an original recording, another camera angle, a matching public appearance, or a contradiction in time and place.
A convincing face does not validate the event. Conversely, one ugly frame does not prove a face swap.
What can an AI video detector tell you?
A detector may identify patterns associated with media in its supported classes. Its output should be interpreted in light of the model’s training, the file quality, the manipulation type, and how much of the video it examined.
AIFakeScan’s methodology states that Basic Scan does not include specialized face-swap classification, and Basic audio inspection does not classify cloned or synthetic speech. Video analysis may sample frames rather than inspect every frame. Passive-only beta results remain Inconclusive; they are not validated AI probabilities.
Those boundaries prevent a common error: treating a general video result as if it had separately tested the person’s face and voice. Use the AI video detector only for the checks it describes, then combine the result with identity and source verification.
How do Content Credentials help with deepfakes?
Content Credentials can record who or what signed an asset, which actions were declared, and whether the signed provenance still matches the file. A declared AI edit or generation action can be strong positive evidence about the recorded workflow.
They do not establish the truth of a scene or guarantee consent. Missing credentials do not prove manipulation, because adoption is optional and metadata can be lost. Read C2PA Content Credentials: What They Are and How They Work for the distinction between a valid provenance record and a true real-world claim.
What should you do during a suspected impersonation scam?
Slow the interaction down. Do not rely on the apparent face, voice, caller ID, or account alone. Contact the person or organization through a phone number, website, or account you already know to be genuine.
Do not send money or sensitive information under pressure. Preserve messages and transaction details, notify the relevant platform or financial provider, and report fraud through the appropriate official channel. The goal is to verify the request independently, not win an argument with the media.
Frequently asked questions
Is a deepfake always a video?
No. The term can cover manipulated or synthetic images, video, and audio, especially when they imitate a real person’s identity, voice, appearance, or actions.
Is every face filter a deepfake?
Not usually. A disclosed entertainment filter is better described by what it does. The deepfake label becomes more relevant when AI-driven manipulation creates a realistic identity or action claim involving a real person.
Can a deepfake detector prove a video is fake?
No detector result should be treated as universal proof. Coverage varies by manipulation type and file quality. High-stakes conclusions require source evidence, original media when possible, and appropriate expert review.
Can genuine footage still be misinformation?
Yes. Real footage can be old, miscaptioned, selectively edited, or presented with a false location. Authenticity of the file and truth of the surrounding claim are separate questions.
What is the clearest way to describe suspected media?
State the observed technique and the claim: for example, “suspected voice clone impersonating a family member” or “fully generated scene labeled as real footage.” Precise language is more useful than calling every suspicious asset a deepfake.
The practical takeaway
Use “deepfake” for AI-enabled media that imitates or manipulates identity, and “AI-generated content” for the broader category. Then verify the specific person, action, source, and context. Accurate classification helps you choose the right evidence instead of relying on one visual clue or one detector score.
Sources
- Frontier Research on Mitigating Risks from Synthetic Content — NIST
- Backgrounder: Deepfakes in 2021 — WITNESS
- Scammers Use Fake Emergencies To Steal Your Money — Federal Trade Commission