AI Glossary / Governance, ethics and safety / Intermediate

Deepfake

Definition

A deepfake is synthetic or manipulated media created with deep-learning techniques that can make a person or event appear to be something it is not. Common examples involve a person's face, voice, or apparent speech and actions. The term is used with varying breadth, but generally emphasizes convincing audiovisual imitation or alteration rather than every ordinary digital edit.

What it means

A voice may sound like a familiar speaker even though that person never said the words. A video may show an altered face or synchronized mouth movements that make an invented statement seem authentic. Such media can be used legitimately, for example in disclosed creative production with appropriate authorization, or deceptively for impersonation and misinformation. Whether content is synthetic, whether its use is authorized, and whether its message is true are separate questions. A genuine recording can be misleadingly presented, while a clearly labeled synthetic demonstration can communicate accurate information.

How it works

Deep-learning systems learn patterns in images, voices, or motion and use those patterns to generate or modify media. Different techniques can synthesize a whole scene, replace part of an image, reproduce vocal characteristics, or alter facial movement. The resulting content may then be edited, compressed, or redistributed, affecting both quality and detectability. Assessing authenticity can involve examining provenance, checking the claimed source and context, using forensic analysis, and comparing independent evidence. Detection tools estimate whether certain signs of manipulation are present; they do not provide infallible judgments about every file or determine the truth of every claim made in it.

Three practical examples

• A company receives a voice message apparently from a manager requesting an unusual action and verifies the request through an established independent channel.
• A media team produces an authorized synthetic presenter and clearly discloses the production method to the audience.
• A researcher evaluates a detection system on both genuine and manipulated recordings, including compressed versions, to understand where the detector fails. The objective is evidence-based assessment rather than trusting appearances alone.

Think of it this way

Think of an exceptionally convincing impersonator whose performance is recorded and circulated without context. Familiar appearance or voice can encourage trust even when the identity claim is false. The analogy is limited because synthetic media can be generated, modified, and distributed at digital scale, and may combine real and artificial elements. The practical response is not to assume every recording is fake, but to match confidence to the evidence and the consequences of relying on it.

Worked example

Suppose a detector is applied to 10,000 files, of which 100 are actually manipulated in a controlled test. It detects ninety of those, a 90% true-positive rate, and falsely flags 1% of the 9,900 genuine files, or ninety-nine files. There are 189 flagged files in total, but only ninety are truly manipulated. Thus 90/189, approximately 47.6%, of flags are correct in this scenario. The detector can have a seemingly strong detection rate while a flag remains uncertain when manipulated content is uncommon. This is a hypothetical calculation, not a claim about a particular product. It shows why base rates, false positives, and follow-up review matter. A detector's output should not alone establish that a named person fabricated a recording.

Illustrative case study

Fictional case: Sycamore Events receives a short video apparently announcing that a supplier has changed an important delivery instruction. The video looks plausible, but the team checks the instruction through its existing supplier contact process and discovers that no change was issued. It preserves the received file and records the verification result rather than relying on a visual guess about authenticity. The case illustrates process resilience: the organization can avoid acting on an unverified request even before a forensic investigation determines exactly how the media was created.

Common misconceptions and mistakes

A visual oddity is not definitive proof of a deepfake; compression, lighting, editing, and ordinary recording faults can create artifacts. Conversely, the absence of obvious artifacts does not prove authenticity. A missing watermark or provenance record also does not establish that content is genuine or false. Provenance can describe origin and editing history where available, but it does not certify the truth of the depicted claim. Another mistake is using the possibility of deepfakes to dismiss all inconvenient evidence. Authenticity assessment should consider the complete evidence, not a convenient assumption.

Business use

Identify workflows where voice or video currently substitutes for authorization, and strengthen verification for consequential requests. Establish trusted contact routes, escalation procedures, and a way to preserve suspicious media for review. For legitimate synthetic production, document authorization, labeling, and distribution decisions. Evaluate detection tools against the organization's actual media types and transformations, including their false-positive costs. Train staff to verify unusual claims without making unsupported accusations. The useful business goal is dependable handling of uncertain media, not a promise that employees can always spot a fake by looking closely.

People also ask

Are deepfakes always malicious?

No. Authorized and disclosed uses can serve creative, accessibility, or educational purposes.

Can audio alone be a deepfake?

Yes. Synthetic or altered voice recordings can imitate a speaker without any video.

Can a detector prove that a statement is true?

No. Media authenticity and factual truth are different questions; a real recording can contain a false statement.

Check your understanding

A file has no visible AI watermark. Does that prove it is an authentic recording?

Answer: No. Watermarks may not have been added, may not be visible, or may not survive processing. Assess source, context, provenance, and other evidence.

Related terms

• Impersonation
• synthetic media
• voice cloning
• media provenance
• watermarking
• misinformation
• forensic analysis
• authenticity verification

Learn next

Separate three questions for a sample piece of media: who supplied it, how it was produced or edited, and whether the claim it conveys is supported. Decide which evidence would answer each question and which actions should wait for independent verification.

More in Governance, ethics and safety

• Debiasing - Debiasing is the process of identifying and reducing a specified unwanted bias in data, representations, models, or their use.
• Decision support system - A decision support system, or DSS, is a computer-based system that helps people make decisions by organizing relevant information, analyzing alternatives...
• Declarative bias
• Denial-of-service - Denial-of-service, commonly abbreviated DoS, is the prevention of authorized access to a resource or a delay that makes required operations unavailable when...
• Dependability - Dependability is the ability of a system to provide a service that people have sound reasons to rely on.
• Deployment Bias - Deployment bias arises when the way an AI system is actually used differs from the role or context for which it was designed and evaluated, creating...

Sources

NIST AI 100-4, Reducing Risks Posed by Synthetic Content: https://doi.org/10.6028/NIST.AI.100-4 NIST, Open Media Forensics Challenge: https://mfc.nist.gov/ NIST, Is This a Deepfake?: https://www.nist.gov/document/deepfake-desktop-version

Previous: Deep Transfer Learning Next: Default logic