Generative artificial intelligence

Generative artificial intelligence (generative AI or GenAI) is a category of artificial intelligence systems capable of generating new content in response to user inputs or other instructions. Generated content may include text, images, audio, video, software code, and other forms of digital media.

Generative artificial intelligence systems learn patterns and representations from training data and use those learned representations to produce new outputs. The technology includes large language models, image-generation models, speech-generation systems, video-generation models, and multimodal models capable of processing or generating multiple types of content.

Generative AI is used in applications including content creation, software development, research, education, customer service, design, media production, data analysis, and information retrieval.

History

Generative artificial intelligence developed from earlier research in Machine learning, Deep learning, neural networks, and computational methods for generating or transforming data.

Advances in deep learning and computational hardware enabled increasingly capable generative models. The development of transformer-based architectures and large-scale training methods significantly expanded the capabilities of generative AI systems, particularly in language processing and multimodal content generation.

Types

Generative AI systems may be classified according to the type of content they generate or process, including:

text generation;
image generation;
audio generation;
video generation;
code generation;
multimodal generation.

Text generation

Text-generating systems produce natural-language content such as articles, summaries, responses, translations, and other forms of written material.

Large language models are among the most widely used technologies for text generation.

Image generation

Image-generation systems produce or modify visual content based on text prompts, reference images, or other inputs.

Applications include illustration, graphic design, advertising, entertainment, and research visualization.

Audio generation

Audio-generation systems can produce speech, music, sound effects, and other forms of audio content.

Video generation

Video-generation systems produce or transform moving-image content using textual, visual, or other forms of input.

Code generation

Code-generation systems produce software source code based on natural-language instructions, existing code, or structured specifications.

They may assist with software development, debugging, testing, documentation, and code transformation.

Multimodal generation

Multimodal artificial intelligence systems can process or generate multiple forms of information, such as text, images, audio, and video.

These systems enable interactions involving several types of media within a single application.

Technology

Generative AI systems commonly use deep neural networks trained on large datasets.

Important technologies include:

transformers;
large language models;
diffusion models;
generative adversarial networks;
variational autoencoders;
multimodal neural networks.

Different architectures are suited to different generation tasks.

Training

Generative AI models are generally trained using large datasets containing examples of the content they are designed to process or generate.

Training enables a model to learn statistical patterns and representations from the available data. Additional training or fine-tuning may subsequently be used to adapt a model to specific tasks, domains, or behavioral requirements.

Generation process

During inference, a generative AI system receives an input and produces an output based on patterns learned during training.

For example, a language model may receive a natural-language prompt and generate a sequence of tokens that forms a response.

The generated output is not necessarily a direct reproduction of a particular training example and may instead represent a newly generated combination of learned patterns.

Applications

Generative artificial intelligence is used in:

content creation;
software development;
education;
scientific research;
customer service;
marketing;
media production;
graphic design;
business intelligence;
information retrieval;
translation;
data analysis.

Its applications continue to expand as models become more capable and accessible.

Relationship with artificial intelligence

Artificial intelligence is the broader field concerned with developing computational systems capable of performing tasks associated with human intelligence.

Generative AI is a category within artificial intelligence focused on generating new content.

Relationship with machine learning

Machine learning provides many of the techniques used to develop generative AI systems.

Modern generative AI is predominantly based on Deep learning and large-scale neural networks.

Relationship with large language models

Large language models are a major class of generative AI systems designed primarily to process and generate language.

However, generative AI extends beyond language models to systems capable of generating images, audio, video, code, and multimodal content.

Relationship with natural language processing

Natural language processing focuses on computational processing of human language.

Generative AI incorporates natural language processing techniques in systems that generate language-based outputs.

Limitations

Generative AI systems may produce inaccurate, incomplete, biased, or misleading outputs. They may also reproduce undesirable patterns present in training data or generate content that requires human verification.

Other limitations include:

hallucination;
training-data limitations;
computational requirements;
copyright and intellectual property concerns;
privacy and data protection risks;
security risks;
difficulty establishing the provenance of generated content.

The reliability of generated content varies according to the model, task, input, and application environment.

Ethical and societal considerations

The widespread use of generative AI has raised questions concerning:

intellectual property;
copyright;
privacy;
misinformation;
employment;
algorithmic bias;
content authenticity;
accountability;
AI safety.

Governments, companies, researchers, and international organizations are developing policies and technical standards to address these issues.

Importance

Generative artificial intelligence represents a major development in modern computing because it enables machines to generate diverse forms of content from natural-language instructions and other inputs. Its integration into software development, research, education, business, media, and information systems has expanded the role of artificial intelligence from analysis and prediction toward large-scale content generation and human-computer interaction.