What is a deepfake?
The AI Glossary
In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.
A deepfake is realistically generated or altered media that can make a person appear to say or do something they did not say or do. The term is commonly applied to audio, images and video. Some legal definitions also cover fabricated depictions of other entities or events, so the precise scope depends on context.
A voice can be synthesized, a face replaced or a recording altered to create a misleading impression. Different generation methods can be used. The defining concern is not one neural-network architecture, but the potential confusion between a fabricated representation and an authentic record.
Synthetic media is not inherently deceptive. It can support accessibility, translation, education or simulation when its origin and purpose are clear and appropriate permissions are in place. Deception, impersonation and non-consensual use create different ethical and legal problems.
Detection is difficult. Visual glitches or unusual audio may reveal some fabrications, but their absence does not prove authenticity. Automated detectors can also make errors or perform poorly on unfamiliar generation methods. A provenance record—information about where media came from and how it was edited—can help, but only when the record is trustworthy and preserved.
Legal duties vary. The EU AI Act includes specific transparency rules concerning synthetic media and deepfakes, with exceptions and transitional provisions. These should not be summarized as one requirement that every AI-edited file everywhere carry the same label. Other laws may address impersonation, fraud or abusive imagery independently of AI-specific legislation.
For a health organization, a fabricated message appearing to come from a clinician or executive could create both a communications problem and an operational one. An urgent audio request to disclose records or transfer money should not be trusted merely because the voice sounds familiar. Verification through an established channel is more dependable than asking staff to detect subtle artifacts. Public communications teams likewise need a process for confirming authentic statements and correcting false representations. The goal is to protect trust without treating every genuine recording as suspect.
Sources: European Union — Regulation (EU) 2024/1689, consolidated text; European Commission — AI Act Implementation Timeline
To learn more, take a look at the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.
AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.
For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy