Generative AI

Technically, a generative model learns an approximation of the data distribution p(x), or a conditional p(x | c) given a prompt c, from which new samples can be drawn; a discriminative model such as a classifier only learns p(y | x). Four families account for most systems. Autoregressive models factorise p(x) as a product of conditionals p(xₜ | x₁, …, xₜ₋₁) and generate one token at a time; decoder-only transformers trained on next-token prediction are the basis of large language models and code assistants. Variational autoencoders (Kingma and Welling, 2013) learn a latent space with an encoder and decoder trained on a lower bound of the likelihood. Generative adversarial networks (Goodfellow et al., 2014) pit a generator against a discriminator in a minimax game; they produce sharp images but are unstable to train and prone to mode collapse. Diffusion models (Ho et al., 2020, building on earlier score-based work) learn to reverse a gradual noising process and generate by iterative denoising; latent diffusion runs this in a compressed latent space, and related flow-matching methods are now common for images, video and audio.

Output depends on decoding as much as on the model. Language models sample from the predicted token distribution with settings such as temperature and nucleus (top-p) sampling (Holtzman et al., 2019); image models use classifier-free guidance to trade diversity for adherence to the prompt. Greedy or low-temperature decoding reduces randomness but not factual error, because the model optimises plausibility, not truth; hallucination is a structural property, mitigated by grounding outputs in retrieved sources, constrained output formats and verification.

Most deployed generative systems are foundation models pretrained with self-supervision on web-scale data and then post-trained with instruction tuning and preference optimisation (RLHF or direct preference methods) to follow instructions and refuse some requests. Multimodal models combine text, image and audio encoders or decoders in one system.

The risk landscape is specific. NIST AI 600-1 (July 2024), the Generative AI Profile of the AI RMF, lists twelve risk categories, including confabulation, information integrity, information security, data privacy, intellectual property, harmful bias and homogenisation, and value chain and component integration. Security issues include prompt injection, training-data extraction, jailbreaks, and misuse for phishing, malware or deepfakes. Under the EU AI Act, Article 50 imposes transparency duties: providers of systems generating synthetic audio, image, video or text must mark outputs in a machine-readable, detectable way, and deployers of deepfakes must disclose that content is artificially generated or manipulated. Provenance standards such as C2PA content credentials and statistical watermarking are the main technical means, and both can be stripped or degraded, so detection of AI-generated content remains unreliable.