What Is Synthetic Media? Definition, Generative AI Content and Examples

What Is Synthetic Media?

Synthetic media is content whose pixels, samples or words were produced by a computational model instead of being recorded by a camera, microphone or person. NIST's report AI 100-4, Reducing Risks Posed by Synthetic Content (November 2024), describes synthetic content as produced, generated from scratch or modified from other content, by AI tools, and then edited, published and spread by users and platforms.

The category is deliberately broad. A product photo created by a text-to-image model, a training video presented by an AI avatar, a dubbed advert in which a presenter appears to speak another language, a voice assistant's speech and a fake video of a politician are all synthetic media. What separates them is disclosure, consent and use. The NIST report says as much: much synthetic content is not inherently harmful, but some of it accelerates existing harms such as fraud, identity theft, non-consensual intimate imagery and threats to information integrity. Deepfakes sit inside the category as the deceptive, real-looking subset.

What Is Generative AI Content?

Generative AI content is anything a generative model outputs: text from a large language model, images from a diffusion model, speech from a text-to-speech or voice-cloning system, or video from a text-to-video model. These models learn the statistical structure of their training data and then sample new material that fits it, guided by a prompt, a reference image or a voice sample.

Human involvement varies widely, and that matters when you judge an item. At one end, a person types a single prompt and publishes the output unchanged. At the other, a designer uses a generative fill tool to remove a lamppost from an otherwise genuine photograph. Both outputs contain generative AI content, but they make different claims about reality. Ask which parts were generated, whether the result depicts a real person or event, and whether the audience was told.

Generated material is often mixed into captured material. A real interview can carry a cloned sentence, a genuine photo can include a generated region, and an authentic video can be dubbed. Precise descriptions name the component that is synthetic.

What Are Examples of Synthetic Media Used Openly and Deceptively?

Disclosed uses are well documented. In 2019 the charity Malaria No More released a film made with Synthesia's AI video synthesis in which malaria survivors speak through David Beckham's face in nine languages. In May 2019 the Dalí Museum in St Petersburg, Florida, opened Dalí Lives, built by Goodby Silverstein & Partners from more than 6,000 frames of archival footage. On 25 July 2024 US Representative Jennifer Wexton, who lost much of her speech to progressive supranuclear palsy, addressed the House floor with an ElevenLabs clone of her own voice, built from recordings of her earlier speeches.

The same tools serve fraud. The FBI's Internet Crime Complaint Center listed in December 2024 generated profile photos for investment and romance scams, forged identity documents, cloned voices of relatives in staged emergencies and real-time video impersonation of executives. In the Arup case confirmed in May 2024, a Hong Kong employee transferred HK$200 million after a video call populated by deepfake recreations of colleagues.

How Is Synthetic Media Labelled and Regulated?

Labelling now has legal force in the EU. Article 50 of the EU AI Act, which applies from 2 August 2026, requires providers of generative systems to mark outputs in a machine-readable format so they are detectable as artificially generated or manipulated. Deployers who create deepfake image, audio or video content must disclose that it is artificial. For evidently artistic, creative, satirical or fictional work, the duty shrinks to disclosing the existence of generated content in a way that does not spoil the work.

The technical tools for marking are watermarks embedded in the content, such as Google DeepMind's SynthID, which NIST lists for images and audio, and signed provenance metadata following the C2PA specification. Both depend on the generator adding them and on the file keeping them through edits and uploads. Content from open-source models, or files stripped of metadata, arrives unlabelled, so detection remains necessary.

Can Synthetic Media Be Detected?

Synthetic images, video and audio can often be detected, with known limits. Detectors look for traces of the generation process: frequency patterns left by upsampling, textures and lighting that physical cameras do not produce, missing physiological signals in faces, and spectral artifacts in synthetic speech. Performance falls on media unlike the training data. Deepfake-Eval-2024 (arXiv 2503.02857) found open-source detectors lost about 45% of their AUC on images that circulated in 2024 compared with earlier benchmarks.

Text is the hardest modality. A Stanford team led by Weixin Liang and James Zou reported in 2023, in work published in the journal Patterns, that GPT detectors consistently misclassify writing by non-native English speakers as AI-generated. That bias is a reason never to discipline a student or employee on a text-detector score alone.

How Does Neural Defend Detect Synthetic Media?

Neural Defend focuses on synthetic images, video and audio, where detection is measurable. NeuroVerify checks a face image, video or audio in one API call and returns a unified risk score from six signals, among them AI Generative Model Analysis for fully generated content and AI Pixel-Level Manipulation for edited regions. Eligible image requests return in under 200 ms at p95.

The model behind it, ARCAS 1B, was evaluated on several synthetic-image benchmarks; the ARCAS paper reports pooled accuracy of 99.99% on GenImage, 99.92% on WildFake and 99.09% on AIGIBench. On the public VendorBench-100 still-image benchmark, Neural Defend ranked first of 36 systems with an MCC of 0.876. For investigations, NeuroDetect places a file in a case with heatmaps, a plain-language summary linked to evidence and an exportable forensic report, so an analyst can state which part of a file is synthetic.