Is This AI-Generated? How to Spot Synthetic Content

Synthetic content is any media produced or substantially altered by AI, and it divides into four main types, each produced by different technologies and presenting different detection challenges. Text is generated by large language models such as GPT-4, Claude, and Gemini, which produce fluent, coherent prose by predicting the most statistically plausible continuation of a given input; the era of obviously robotic output is over, and modern LLM text is often indistinguishable from human writing on casual inspection. Images are produced by two main technologies: diffusion models, such as Midjourney, DALL-E, and Stable Diffusion, and generative adversarial networks, the older approach in which two neural networks compete, one generating images and one critiquing them until the results become convincing. Both can produce photorealistic imagery of people, places, and events that never existed. Audio and voice are synthesised through voice cloning tools, which can now produce a convincing replica of someone's voice from as little as a few seconds of source audio, and the cloned voice can then say anything; this is the technology behind the executive-impersonation variant of fraud, where a synthetic voice call has been used to authorise large wire transfers. Video, including deepfakes, a term combining deep learning and fake, uses AI to map a synthetic face or body onto existing footage, or to generate video from scratch, and it is the technology behind the engineering-firm attack described above. All four modalities are available to anyone with a consumer internet connection, and the cost of producing convincing synthetic content has dropped from requiring expensive hardware and specialist expertise to being achievable in minutes with free or low-cost tools.