Key takeaways
- A generative AI model learns the statistical structure of its training data, then samples from that structure to produce new text, images, audio, video or code.
- Five architectures do most of the work: transformers, diffusion models, GANs, variational autoencoders and flow-based models. Each generates differently and fails differently.
- The term describes a behaviour, not a legal status. European law regulates the overlapping but distinct category of general-purpose AI models.
- A model is presumed to be a GPAI model above 10^23 FLOP of cumulative training compute, and presumed to carry systemic risk above 10^25 FLOP.
- Few organisations train a generative AI model. Many become its legal provider anyway, by fine-tuning one or putting their own name on it.
What is a generative AI model?
A generative AI model is a machine learning model that learns the statistical structure of a training dataset, then samples from that learned structure to produce new content resembling the data it was trained on. Where a conventional classifier answers a question about an input, a generative AI model produces something that did not exist before: a paragraph, an image, a melody, a function.
The distinction matters more than it sounds. Most machine learning in production estimates the probability of a label given some input. A generative model estimates the distribution of the data itself, which is a harder problem and a far more open-ended capability. Once a system has modelled how text or pixels are structured, it can be asked to produce arbitrary amounts of either.
It is worth stating plainly what most material on this subject leaves implicit: “generative AI model” is a technical description, not a regulatory category. Nothing in European or American law attaches obligations to the word “generative” on its own. Obligations attach to categories such as general-purpose AI models, high-risk AI systems, and specific deployment contexts. A given generative AI model may fall into all, some or none of them. This guide covers both halves: what these models are, and what building or using one actually commits you to.
For the wider family this sits inside, see our guide to what an AI model is.
Generative versus discriminative models
The cleanest way to hold the difference: a discriminative model draws boundaries, a generative model draws samples.
A spam filter is discriminative. It learns a boundary between spam and not-spam and places each incoming message on one side of it. It cannot write you an email. A generative model trained on the same corpus learns how emails are built, and can produce one on request.
Formally, discriminative models estimate the probability of a label given the input. Generative models estimate the joint distribution of the data, which is what allows them to synthesise new instances. This is why a generative AI model can be repurposed so easily for tasks it was never explicitly trained for, and why its failure modes are harder to bound in advance.
How generative AI models work
Every generative AI model runs on the same two-phase logic: compress, then sample.
During training, the model is exposed to a very large corpus and adjusts its parameters so that the patterns in that corpus become more probable under the model. Nothing is stored verbatim in the ordinary sense. What survives is a compressed, lossy encoding of regularities: how sentences are built, how faces are shaped, how chord progressions resolve.
During inference, the model runs that process outward. It samples from the learned distribution, conditioned on whatever prompt it received. Because sampling is probabilistic, the same prompt can yield different outputs, and the model holds no internal representation of whether any of them is true.
That last point carries a governance consequence which is easy to miss. Fluent, confident, incorrect output is not a defect introduced by careless engineering. It follows directly from how a generative AI model is built: the training objective rewards plausibility under the learned distribution, not correctness against the world. You can lower the rate through retrieval grounding, evaluation and human review. You cannot patch it out. We examine this at length in our analysis of the single biggest risk of generative AI.
The five families of generative AI models
Almost every deployed generative AI model belongs to one of five architectural families. They are not interchangeable. Each generates through a different mechanism, suits different modalities, and breaks in a different way.
Transformers and autoregressive models
Transformers generate sequentially, predicting the next token given everything before it. The attention mechanism lets the model weigh distant parts of the input when producing each token, which is what made long-range coherence in text tractable. This family powers the GPT series, Claude, Gemini and Llama, and has expanded well past text into code, audio and protein sequences. Strength: coherence and generality. Weakness: cost, and the confident fabrication that comes with pure next-token prediction.
Diffusion models
Diffusion models learn to reverse a noising process. Training progressively destroys an image with random noise while the model learns to undo each step. At inference it starts from pure noise and denoises toward a sample, steered by a text prompt. This family displaced GANs for high-quality image synthesis and now drives Stable Diffusion, DALL-E and most video generation. Strength: sample quality and stable training. Weakness: many denoising steps make generation slow.
Generative adversarial networks (GANs)
A GAN trains two networks against each other. The generator produces candidates, the discriminator tries to separate them from real data, and both improve until the generator’s output passes. GANs produced the first genuinely photorealistic synthetic faces and remain useful where fast single-pass generation matters. Strength: inference speed and sharp output. Weakness: unstable training and mode collapse, where the generator settles on a narrow slice of the data distribution.
Variational autoencoders (VAEs)
A VAE compresses input into a probability distribution over a latent space, then decodes samples from that space back into data. The latent space is continuous and structured, so you can interpolate meaningfully between points. Output tends to be blurrier than a GAN’s or a diffusion model’s, which is why VAEs now more often sit as a component inside a larger system than as the generator itself. Strength: structured, controllable latent representations. Weakness: sample fidelity.
Flow-based models
Flow-based models apply a sequence of invertible transformations that map a simple distribution onto a complex one. Because every step is invertible, they compute exact likelihoods, which the other families only approximate. Strength: exact density estimation and reversibility. Weakness: architectural constraints and heavy compute for competitive quality.
| Family | Generates | Strength | Main weakness | Representative system |
|---|---|---|---|---|
| Transformer / autoregressive | Text, code, audio | Coherence, generality | Cost, fabrication | GPT, Llama |
| Diffusion | Images, video, audio | Sample quality | Slow inference | Stable Diffusion |
| GAN | Images | Fast, sharp output | Unstable training | StyleGAN |
| VAE | Images, latent representations | Structured latent space | Blurry samples | VQ-VAE |
| Flow-based | Images, audio | Exact likelihoods | Compute cost | Glow |
Which family a system uses shapes what you have to test. Evaluating a text generator for factual reliability is a different exercise from evaluating an image generator for training-data memorisation, and both feed the benchmarking evidence an auditor eventually asks for.
Generative AI model, foundation model, LLM, GPAI model
Four terms circulate as if they were synonyms. They describe different things, and only one of them carries legal weight.
- Generative AI model describes what the model does: it produces new content. It says nothing about size, scope or training method.
- Foundation model describes how it was built and used: broad pretraining on large data, then adaptation to many downstream tasks. Most foundation models are generative, but the defining trait is breadth and reusability.
- Large language model describes a foundation model whose modality is text. Every LLM is a generative AI model. Not every generative AI model is an LLM.
- General-purpose AI model (GPAI model) is a legal category under the EU AI Act. It turns on generality and capability rather than architecture, and it is the only one of the four that triggers statutory duties by itself.
A small image model fine-tuned on one product catalogue is a generative AI model and almost certainly not a GPAI model. A frontier text model is all four at once. Our guide to general-purpose AI covers the legal category in depth.
When a generative AI model becomes a regulated GPAI model
The European Commission’s guidelines for providers of general-purpose AI models, published on 18 July 2025, replaced a vague definition with a workable test built on training compute.
A model is presumed to be a GPAI model when the cumulative compute used to train it exceeds 10^23 floating point operations and it can generate language, text-to-image or text-to-video output. The Commission’s reasoning is that this threshold corresponds roughly to a billion-parameter model trained on a large dataset, which is where genuine generality tends to appear.
Above 10^25 FLOP a second presumption applies: the model is treated as a GPAI model with systemic risk under Article 55. The Commission can also designate a model as systemic on capability grounds regardless of compute. This upper tier is where the frontier labs sit and it carries a heavier obligation set. See what a frontier model is for that tier specifically.
Both are presumptions rather than fixed lines. A provider can argue that a model above the threshold is not general-purpose, and the Commission can pull a model below the threshold into scope.
The timeline matters for planning. GPAI obligations began applying on 2 August 2025. The Commission’s enforcement powers over GPAI providers, including formal information requests and fines, begin on 2 August 2026.
The practical reading for most organisations: you will probably never train a GPAI model, you will almost certainly deploy one, and you may become a provider of one without running a single training job.
What the EU AI Act requires from a generative AI model provider
Providers of a GPAI model carry four baseline duties under Article 53, plus an additional layer when the model is systemic.
Technical documentation and downstream information
The provider must prepare and maintain technical documentation covering the model’s design, training process, evaluation results and intended use, per Annex XI. Separately, it must give downstream providers who integrate the model enough information, per Annex XII, to understand its capabilities and limitations and to meet their own obligations.
The second duty is the one enterprises underestimate. If you build a product on somebody else’s model, that documentation is the raw input for your own compliance file, and you should be contracting for it rather than hoping. Our guide to AI system documentation requirements sets out what a complete file contains.
Copyright policy and the training-data summary
Providers must put in place a policy to comply with Union copyright law. In practice this centres on honouring rights reservations made under Article 4(3) of the Copyright in the Digital Single Market Directive, the text and data mining opt-out that rightsholders express in machine-readable form.
They must also publish a “sufficiently detailed summary” of the content used to train the model, on the template the AI Office published on 24 July 2025. This converts training data from a trade secret into a partially disclosed artifact, and it is the provision most likely to generate litigation.
Marking synthetic output
Under Article 50, providers of systems that generate synthetic audio, image, video or text must mark those outputs in a machine-readable format and make them detectable as artificially generated. Deployers who produce deepfakes must disclose them. These transparency duties apply from 2 August 2026.
For a systemic-risk model, Article 55 adds model evaluation including adversarial testing, systemic risk assessment and mitigation, serious incident reporting to the AI Office, and cybersecurity protection for the model and its physical infrastructure. The voluntary GPAI Code of Practice supplies a Model Documentation Form and functions as presumptive compliance for much of this. The full obligation map sits in our EU AI Act operator’s guide.
The risks specific to generative models
Plenty of AI risk is inherited: any model can be biased, attacked or poorly monitored. A smaller set is specific to generation, and the clearest catalogue is the NIST Generative AI Profile (AI 600-1), published in July 2024 as a companion to the AI Risk Management Framework. It defines twelve risk categories and more than two hundred suggested actions.
The categories that are genuinely generative rather than general:
- Confabulation, which NIST defines as confidently stated but erroneous content, known colloquially as hallucination. It leads the list because it is structural rather than incidental.
- Information integrity, the capacity to produce persuasive synthetic content at volume, which is precisely what Article 50 marking is designed to counter.
- Intellectual property, covering both training-data provenance and outputs that reproduce protected material.
- Data privacy, including memorisation and regurgitation of personal data from training corpora.
- CBRN information, meaning uplift toward chemical, biological, radiological or nuclear capability.
- Harmful bias and homogenization, where generated output narrows the range of represented people and ideas.
- Value chain and component integration, the risk inherited from a model, dataset or component you did not build.
The Berkeley CLTC risk-management profile for GPAI and foundation models adds the operational layer: go/no-go gates at major development stages, explicit risk-tolerance thresholds, red-teaming for dangerous capabilities, and transparency through model and system cards.
None of these is a bug awaiting a fix. Each is a property to be measured, bounded and evidenced, which is what a control framework exists to do. Our overview of AI risk covers how they map into a register.
Governing a generative AI model you did not build
Almost every organisation reading this is a deployer rather than a trainer. That does not place you outside the regime, and the deployer position is less stable than it looks.
You can become a provider without training anything. Fine-tuning a third-party model on your own data, or placing your own name or trademark on a system built from one, can move you into the provider role and its obligations. The question is not who ran the training job. It is who puts the thing on the market.
Four moves cover most of the ground:
- Find the models actually in use. You cannot govern an inventory you do not have, and generative tools enter organisations through expense claims and browser extensions rather than procurement. This is the shadow AI problem, and solving it is the precondition for everything else.
- Fix the provider and deployer boundary per model. Write it down for each system: who is the provider, what did they hand over, what changes if you fine-tune. Ambiguity here is what turns an incident into a liability argument.
- Obtain the downstream documentation and keep it. Annex XII information is a contractual matter. Ask for it at procurement, not after an incident.
- Set evaluation and logging before the use case ships. Decide what an acceptable output looks like, how you will measure it, and what you retain. Retrofitting evidence costs far more than collecting it.
None of this requires understanding attention heads. It requires knowing which generative AI model sits behind which business process, and what each one obliges you to do. That is AI governance applied to a specific class of system.
FAQ
Is ChatGPT a generative AI model? Not exactly. ChatGPT is an AI system built on top of generative models from the GPT family. The distinction is not pedantic: the EU AI Act regulates models and systems differently, with Article 53 duties falling on the model provider and system-level transparency duties under Article 50 falling on the system provider and its deployers. In everyday speech the terms blur. In a compliance file the difference decides who owes what.
What is an example of a generative AI model? GPT and Llama are transformer-based text models. Stable Diffusion is a diffusion model for images. StyleGAN is a generative adversarial network for faces. Each generates through a different mechanism, and the choice shapes which risks matter most: memorisation and copyright exposure dominate for image models, while confabulation dominates for text.
What is the difference between a generative AI model and an LLM? A large language model is one kind of generative AI model, specialised in text. The broader category also covers image, audio, video and molecular models. Every LLM is generative; most generative models are not LLMs. The confusion persists because text models are the ones most people interact with directly.
Are all generative AI models regulated by the EU AI Act? No. The Act does not regulate a model for being generative. Obligations attach when a model qualifies as a general-purpose AI model, presumed above 10^23 FLOP of training compute, or when a system falls into a high-risk use case or triggers Article 50 transparency duties. A small image model fine-tuned for one internal task sits outside the GPAI regime, though the use case around it may still be regulated.
What are the main types of generative AI models? Five families: transformers and autoregressive models for sequences, diffusion models for images and video, generative adversarial networks, variational autoencoders, and flow-based models. Transformers and diffusion account for most production deployments today, with GANs, VAEs and flows increasingly serving as components inside larger systems rather than as the headline generator.
Who is responsible when a generative AI model produces harmful output? It depends on the role each party holds. The model provider owes documentation, a copyright policy and, above the systemic threshold, evaluation and incident reporting. The system provider owes transparency and output marking. The deployer owes appropriate use, human oversight where required, and disclosure of deepfakes. Fine-tuning or rebranding can shift you from deployer to provider, which is why the boundary should be settled per model, in writing.
Conclusion
The architecture question is settled. Transformers, diffusion, GANs, VAEs and flows are well documented, and any competent engineer can explain how a generative AI model turns a prompt into an output. The accountability question is not settled, and it is the one that will occupy compliance teams through 2026 as enforcement powers arrive and transparency duties start to bite.
The useful shift is to stop treating these systems as a technology category and start treating them as a governed asset class: inventoried, classified against the GPAI thresholds, documented, evaluated, and owned by someone named. AI Sigil is built for that work, turning obligations across the EU AI Act, ISO 42001 and NIST AI RMF into controls and evidence you can put in front of an auditor.