{"licence":{"name":"CC BY-SA 4.0","spdx":"CC-BY-SA-4.0","url":"https://creativecommons.org/licenses/by-sa/4.0/","attribution":"Atlas, a bilingual technical dictionary (https://cmaintz.github.io/tech-atlas/)"},"id":"ai/generative-ai","url":{"en":"https://cmaintz.github.io/tech-atlas/en/terms/ai/generative-ai/","da":"https://cmaintz.github.io/tech-atlas/da/terms/ai/generative-ai/"},"term":{"en":"Generative AI","da":"Generativ AI"},"aka":{"en":["GenAI"],"da":["GenAI","generativ kunstig intelligens"]},"domain":["ai"],"cluster":"ml-fundamentals","layer":"application","status":"current","era":2022,"summary":{"en":"AI that creates new content (text, images, sound, code) in the style of the examples it learned from, instead of only sorting or scoring.","da":"AI, der skaber nyt indhold - tekst, billeder, lyd, kode - i stil med de eksempler, den har lært af, i stedet for kun at sortere eller score."},"body":{"formal":{"en":"AI systems, usually built with deep learning, that learn the patterns in large amounts of training data and use them to produce new output of the same kind when asked.","da":"AI-systemer, som regel bygget med deep learning, der lærer mønstrene i store mængder træningsdata og bruger dem til at frembringe nyt output af samme slags, når de bliver bedt om det."},"plain":{"en":"Like a street painter who has studied thousands of paintings and can paint you a new one in any style you name, though never one they saw exactly.","da":"Som en gademaler, der har studeret tusindvis af malerier og kan male dig et nyt i hvilken som helst stil, du nævner - dog aldrig ét, de har set præcis."},"inPractice":{"en":"A communications officer in a municipality has a chat assistant draft a plain-language version of the new waste-sorting rules and an image tool draw a matching picture, then checks both before publishing.","da":"En kommunikationsmedarbejder i en kommune får en chatassistent til at skrive et udkast til de nye affaldsregler i et enkelt sprog og et billedværktøj til at tegne et billede, og tjekker begge, før de bliver lagt ud."},"whyItMatters":{"en":"It makes content cheap and fast for everyone, including attackers, and raises new questions about who owns the output, leaked company data and what can be trusted.","da":"Det gør indhold billigt og hurtigt for alle, også angribere, og rejser nye spørgsmål om, hvem der ejer resultatet, om lækkede virksomhedsdata og om, hvad man kan stole på."}},"deepDive":{"en":"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.\n\nOutput 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.\n\nMost 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.\n\nThe 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.","da":"Teknisk set lærer en generativ model en tilnærmelse af datafordelingen p(x), eller en betinget fordeling p(x | c) givet en prompt c, som nye eksempler kan trækkes fra; en diskriminativ model som en klassifikator lærer kun p(y | x). Fire familier står for de fleste systemer. Autoregressive modeller faktoriserer p(x) som et produkt af betingede fordelinger p(xₜ | x₁, …, xₜ₋₁) og genererer ét token ad gangen; decoder-only-transformere trænet på forudsigelse af næste token er grundlaget for store sprogmodeller og kodeassistenter. Variational autoencodere (Kingma og Welling, 2013) lærer et latent rum med en encoder og en decoder, der trænes på en nedre grænse for likelihood. Generative adversarial networks (Goodfellow m.fl., 2014) sætter en generator op mod en diskriminator i et minimax-spil; de giver skarpe billeder, men er ustabile at træne og tilbøjelige til mode collapse. Diffusionsmodeller (Ho m.fl., 2020, bygget på tidligere scorebaseret arbejde) lærer at vende en gradvis støjproces om og genererer ved gentagen støjfjernelse; latent diffusion gør dette i et komprimeret latent rum, og beslægtede flow matching-metoder er nu udbredte til billeder, video og lyd.\n\nResultatet afhænger lige så meget af afkodningen som af modellen. Sprogmodeller sampler fra den forudsagte tokenfordeling med indstillinger som temperatur og nucleus-sampling (top-p) (Holtzman m.fl., 2019); billedmodeller bruger classifier-free guidance til at bytte variation for tættere overensstemmelse med prompten. Grådig afkodning eller lav temperatur mindsker tilfældigheden, men ikke faktuelle fejl, fordi modellen optimerer for det sandsynlige, ikke det sande; hallucination er en strukturel egenskab, som afhjælpes ved at forankre output i fremfundne kilder, begrænse outputformatet og verificere.\n\nDe fleste udrullede generative systemer er foundation models, der er fortrænet med selvsupervision på data i webskala og derefter eftertrænet med instruction tuning og præferenceoptimering (RLHF eller direkte præferencemetoder), så de følger instruktioner og afviser visse forespørgsler. Multimodale modeller kombinerer encodere eller decodere til tekst, billeder og lyd i ét system.\n\nRisikobilledet er særegent. NIST AI 600-1 (juli 2024), Generative AI-profilen til AI RMF, opstiller tolv risikokategorier, herunder konfabulation, informationsintegritet, informationssikkerhed, databeskyttelse, immaterielle rettigheder, skadelig bias og homogenisering samt værdikæde og komponentintegration. Sikkerhedsproblemerne omfatter prompt injection, udtrækning af træningsdata, jailbreaks og misbrug til phishing, malware eller deepfakes. AI-forordningens artikel 50 fastsætter gennemsigtighedsforpligtelser: Udbydere af systemer, der genererer syntetisk lyd, billeder, video eller tekst, skal mærke output i et maskinlæsbart format, så det kan opdages, og idriftsættere af deepfakes skal oplyse, at indholdet er kunstigt genereret eller manipuleret. Oprindelsesstandarder som C2PA content credentials og statistisk vandmærkning er de vigtigste tekniske midler, og begge kan fjernes eller forringes, så detektion af AI-genereret indhold er fortsat upålidelig."},"edges":[{"type":"requires","to":"ai/deep-learning","confidence":"high","strength":"normal"},{"type":"kind-of","to":"ai/artificial-intelligence","confidence":"high","strength":"normal"},{"type":"causes","to":"ai/hallucination","why":{"en":"It is built to produce content that looks likely, not content that has been checked, so fluent but wrong output is part of how it works.","da":"Den er bygget til at frembringe indhold, der ser sandsynligt ud, ikke indhold, der er tjekket, så flydende men forkert output er en del af, hvordan den virker."},"confidence":"high","strength":"primary"},{"type":"used-with","to":"ai/prompt","confidence":"high","strength":"normal"}],"depth":2,"sources":[{"title":"NIST AI 600-1 (2024), Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile","url":"https://doi.org/10.6028/NIST.AI.600-1","tier":"standard","publisher":"NIST"},{"title":"Goodfellow, Bengio & Courville (2016), Deep Learning","url":"https://www.deeplearningbook.org/","tier":"textbook","publisher":"MIT Press"},{"title":"Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 50","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","tier":"standard","publisher":"European Union"},{"title":"Kingma & Welling (2013), Auto-Encoding Variational Bayes","url":"https://arxiv.org/abs/1312.6114","tier":"reference"},{"title":"Goodfellow et al. (2014), Generative Adversarial Networks","url":"https://arxiv.org/abs/1406.2661","tier":"reference"},{"title":"Ho, Jain & Abbeel (2020), Denoising Diffusion Probabilistic Models","url":"https://arxiv.org/abs/2006.11239","tier":"reference"}],"draft":true}