PaperlayerView exact revision on arXiv

Source-linked AI summary

AI-Generated Content (AIGC): A Survey

Jiayang Wu, Wensheng Gan, Zefeng Chen, Shicheng Wan, Hong Lin

arXiv:2304.06632v1cs.AIcs.CYcs.HC

TL;DR

AIGC raises the need to understand how AI can assist or replace manual content generation and what limitations constrain its use. This survey synthesizes AIGC’s concepts, models, applications, benefits, challenges, and future directions, concluding that it has broad application potential while requiring attention to energy use, transparency, ethics, regulation, and controllability.

  • Problem

    Understanding AIGC’s capabilities and limitations is necessary for exploring its potential across future applications.

  • Method

    The paper surveys AIGC’s definitions, conditions, capabilities, large pre-trained models, industrial chain, applications, generation modes, Metaverse integration, challenges, and future directions.

  • Results

    The survey concludes that AIGC has demonstrated business value and application performance while supporting content generation and assistance across multiple domains.

  • Takeaways & Limitations

    AIGC can support convenient, scalable content creation, but its deployment must address energy consumption, feedback optimization, transparency, ethics, regulation, and controllability.

Abstract

from arXiv · show

To address the challenges of digital intelligence in the digital economy, artificial intelligence-generated content (AIGC) has emerged. AIGC uses artificial intelligence to assist or replace manual content generation by generating content based on user-inputted keywords or requirements. The development of large model algorithms has significantly strengthened the capabilities of AIGC, which makes AIGC products a promising generative tool and adds convenience to our lives. As an upstream technology, AIGC has unlimited potential to support different downstream applications. It is important to analyze AIGC's current capabilities and shortcomings to understand how it can be best utilized in future applications. Therefore, this paper provides an extensive overview of AIGC, covering its definition, essential conditions, cutting-edge capabilities, and advanced features. Moreover, it discusses the benefits of large-scale pre-trained models and the industrial chain of AIGC. Furthermore, the article explores the distinctions between auxiliary generation and automatic generation within AIGC, providing examples of text generation. The paper also examines the potential integration of AIGC with the Metaverse. Lastly, the article highlights existing issues and suggests some future directions for application.

I. INTRODUCTION

AIGC is an AI-based content creation approach that responds to user requirements and complements traditional professional and user-generated content. This survey reviews its development, capabilities, industrial foundations, applications, challenges, and future directions.

  • AIGC generates content according to user requirements and complements professional-generated and user-generated content.
  • AIGC developed from primitive computer-controlled output toward systems capable of understanding human language and generating human-like text.
  • Current AIGC content quality has improved, with applications spanning text, images, video, code, dialogue, summarization, translation, and question-answering.
  • The survey covers AIGC definitions and conditions, cutting-edge capabilities, advanced features, large pre-trained models, and the industrial chain.
  • It also analyzes auxiliary versus automatic generation, practical advantages and disadvantages, Metaverse integration, current problems, and future application directions.

II. RELATED CONCEPTS

AIGC expands content production beyond professional and user-generated modes by using AI to generate text, images, audio, and video. Its applications range from writing and image editing to speech synthesis and promotional video creation.

  • Professional-generated content offers high quality but has long production cycles, whereas user-generated content lowers creation barriers but has uneven quality.
  • AIGC is presented as addressing the quantity and quality limitations associated with professional- and user-generated content.
  • Text generation includes structured writing, creative writing, and dialogue writing for scenarios such as news, marketing, blogs, and chatbots.
  • Image generation enables prompt-based editing and rapid creation of posters, logos, and other requested visual content.
  • Audio generation includes text-to-speech and voice cloning, while video generation processes frames and supports trailers and promotional videos.

B. Necessary conditions of AIGC

AIGC depends on the coordinated development of high-quality data, powerful hardware, and effective algorithms. Large-scale pre-trained models further improve the quality, accuracy, generalization, and efficiency of generated content.

  • Core components: AIGC requires data, hardware, and algorithms as its three critical components.Data supports algorithm training, hardware supplies computing infrastructure, and algorithms learn patterns from data.
  • Data: Larger training datasets often produce more accurate models, making storage and management of massive datasets essential.Training tasks may require billions to hundreds of billions of files.
  • Hardware: Large models rely on substantial computing power and specialized chips such as GPUs, FPGAs, and ASICs.Companies must balance computing cost with algorithmic efficiency.
  • Algorithms: Large pre-training models use extensive text data for pre-training and help improve the quality and accuracy of AIGC.Examples include BERT and GPT.
  • Algorithms: Generative algorithms evolved from traditional machine learning and neural networks toward Transformers, generative models, and large-scale pre-training models.The survey describes GANs, Transformers, BERT, GPT-3, LaMDA, and diffusion models in this progression.

C. How can AI make the content better?

AIGC improves content generation through three nested capabilities: digital twins, intelligent editing, and intelligent creation. These capabilities support multimodal transformation, controllable modification, and the production of imitative or conceptual content.

  • Cutting-edge capabilities: AIGC has three cutting-edge capabilities: digital twins, intelligent editing, and intelligent creation.The capabilities are nested and combined to provide stronger generation capability.
  • Digital twins: Digital twins map real-world content into virtual worlds through tasks such as intelligent translation and enhancement.Translation can transform content across language, audio, and visual modalities.
  • Intelligent editing: Intelligent editing uses semantic understanding and attribute control to separate, modify, and return digital content between virtual and real environments.This process forms a twinning-and-feedback loop.
  • Intelligent creation: Intelligent creation includes imitation-based creation from learned patterns and conceptual creation from abstract concepts learned from massive data.The passage distinguishes creating from existing examples from producing content that did not previously exist.
  • Advanced features: ChatGPT illustrates AIGC’s advanced features through contextual understanding, conversation recall, sensitive-information filtering, and recommendations.These features support customer service, translation, and content creation applications.

D. The industrial chain of AIGC

The AIGC industry chain connects upstream data, algorithm, and hardware providers with midstream technology companies and downstream content-creation platforms. Each stage supports the deployment, accessibility, adoption, and commercialization of AIGC.

  • Upstream: Upstream AIGC consists mainly of data suppliers, algorithmic institutions, and hardware development institutions.These organizations collect and label data, develop algorithms, and build dedicated computing hardware.
  • Midstream: Midstream technology companies integrate upstream resources, configure cloud computing environments, optimize algorithms, and package them into external tools.Configurations may include virtual machines, containers, databases, and storage.
  • Downstream: Downstream content-creation platforms reduce users’ learning costs and help them complete tasks with midstream tools.News media and financial institutions can use text-generation tools to produce reports quickly.
  • Downstream: Downstream users are primary recipients of AIGC’s value and are important in promoting adoption and commercialization.Their use of content-generation tools connects industrial capabilities with practical applications.

E. Advantages of large-scale pre-trained models

Large-scale pre-trained models are trained on massive generalized data and are presented as improving AIGC performance and generalization. They support reusable, multi-task systems, while the paper distinguishes AI assistance from autonomous writing and retains a central role for human creativity.

  • Model foundations: Large-scale pre-trained models combine large scale with pre-training on massive generalized data before practical-task modeling.The paper presents them as a milestone toward general intelligence and as supporting greater AIGC generalization.
  • Training process: Their development uses supervised-learning data, comparative data for reward modeling, and explanatory data for augmented optimization.These three stages are illustrated as the development process for large-scale pre-trained models.
  • Advantages: Pre-trained models improve generalization by learning more features and patterns from large-scale data across tasks and scenarios.The models can adapt to different tasks and scenarios after pre-training.
  • Advantages: Pre-trained models reduce repeated training costs, accelerate fine-tuning, and support tasks spanning language, vision, and speech.The paper also describes continuous optimization through adding new data and tasks.
  • Human–AI writing: AIGC cannot produce original content for specific needs and interests, so it can serve as a writing assistant rather than a complete replacement.The survey distinguishes AI-assisted writing from AI-generated writing.
  • Human–AI writing: The paper assigns data collection primarily to AI and the creative process of writing to humans.It argues that balancing AI tools with human creativity and expression is essential for optimal results.

G. Pros of AIGC

AIGC offers efficiency, scalability, and cost advantages while supporting scientific research, SEO, and writers’ creative workflows.

  • Efficiency and scalability: AIGC produces articles quickly and at scale, supports language localization, and frees human resources for other tasks.
  • Help scientific research: AIGC assists scientific research by analyzing large datasets, identifying patterns, reviewing literature, and generating testable hypotheses.
  • For search engine optimization: AI-powered tools improve SEO by analyzing queries, suggesting keywords, and optimizing content structure, length, and readability.
  • Overcome writer’s block: AIGC helps overcome writer’s block through keyword-based suggestions, trend analysis, article drafting, and polishing assistance.

H. Cons of AIGC

The survey identifies concerns about AIGC’s creativity, contextual understanding, ethics, fairness, educational effects, empathy, and need for human oversight.

  • AIGC may lack creativity, nuanced language and context understanding, and can create copyright, privacy, and ethical concerns.
  • Ethics and trust: Human review and curation remain necessary because AI-generated content may lack intended tone, contain plagiarism, and require contextual refinement.
  • Exacerbate social imbalances: Unequal access to advanced AI tools may concentrate content-production power and exacerbate existing social inequalities.
  • Negative effects on education: Relying solely on AIGC in education may reduce personalization, critical thinking, and analytical skill development, while biased data can produce wrong knowledge.
  • Inadequate empathy: AIGC may lack human creativity, emotion, empathy, and nuance because it generates from parameters, data, and patterns without truly comprehending meaning.
  • Human involved: Human involvement is crucial because AIGC outputs can contain mistakes and inconsistencies caused by limited understanding of language.
  • Missing creativity: AIGC’s reliance on existing data can limit fresh ideas, innovation, timeliness, and resonance with particular audiences.

I. AIGC and Metaverse

AIGC can enrich the Metaverse through personalized interactive content and synthetic data, while large-scale models introduce hardware, energy, data-quality, and legal constraints.

  • I. AIGC and Metaverse: AIGC supports Metaverse personalization and interaction through intelligent NPCs, automated QA, dialogue systems, and digital humans.
  • I. AIGC and Metaverse: Standardized low-code tools enable small studios and individual developers to produce richer interactive Metaverse content.
  • I. AIGC and Metaverse: AIGC-generated environmental data can reduce the time, labor, and cost required to construct richly simulated Metaverse scenarios.
  • Synthetic data can substitute for costly real-world data in training, testing, and validating AI models, while supporting virtual self-learning environments.
  • Training data must meet quality, fairness, legal, and ethical requirements, including copyright compliance when copyrighted data is used.
  • Large-scale pre-training models require substantial computing power, and their training and inference produce high energy consumption.
  • Optimizing algorithms and using mixed-precision computing or distributed training are presented as practical ways to address hardware and energy challenges.

C. Algorithm

The algorithm discussion covers retrieval, feedback-based optimization, security, privacy, and NLP challenges associated with large language models.

  • Information retrieval: Information retrieval traditionally searches external contextual documents before predicting answers, while large language models improve these steps through memory and reasoning.
  • Information retrieval: Retrieval remains limited by vocabulary gaps and contextual failures that can impair specialized-term comprehension, completeness, and implicit-meaning recognition.
  • Model optimization: User feedback loops optimize large pre-trained models by presenting predictions and collecting direct responses to guide further improvement.
  • Security: AI content-generation algorithms are vulnerable to malicious input or output manipulation that can produce misleading content and enable serious attacks.
  • Privacy: Pre-training and distributed computing create privacy risks because sensitive information may enter model parameters or be exposed through insecure nodes and networks.
  • NLP challenges: NLP requires stronger generation, description, computation, reasoning, and interpretability capabilities, combining connectionist and symbolic approaches for language understanding.
  • NLP challenges: In-context learning is presented as a paradigm that improves context understanding and produces more accurate and relevant NLP results.

F. Human attitudes towards AIGC

AIGC expands content-generation efficiency and application scope, while raising concerns about regulation, ethics, accuracy, employment, transparency, and accountability.

  • AIGC differs from human-generated content by using trained AI models to learn from data, analyze problems, and simulate human-like behavior.
  • Rapid commercialization creates unresolved questions about content ownership, intellectual property, data rights, rumor propagation, forgery, and application safety.
  • AIGC’s efficiency enables hundreds of unique images within an hour and billions of words in a single morning.
  • AIGC has produced concerns about cheating, plagiarism, discrimination, job displacement, and unequal recognition of human and AI contributions.
  • Large language models may generate inaccurate or biased answers because training data can be poor quality and information-source credibility may be misjudged.
  • The paper calls for greater model transparency, open-source development, democratic control, licensing responsibility, and public accountability.

IV. PROMISING DIRECTIONS

The survey identifies cross-modal generation and AIGC-supported search as promising directions, while noting usability, controllability, quality, and safety challenges.

  • AIGC’s promising future directions include cross-modal generation, search engine optimization, media production, e-commerce, and film production.
  • A. Cross-modal generation technology: Cross-modal generation combines information from modalities such as text, audio, vision, sensors, and touch to simulate the complex real world.
  • A. Cross-modal generation technology: Text-to-image and text-to-video systems generate visual content from language, with applications including creative images, articles, search queries, and presentations.
  • A. Cross-modal generation technology: Cross-modal generation still requires long descriptions, may not satisfy specific requirements, and can produce unexpected outputs when models overfit.
  • AIGC-integrated search aims to provide faster and more relevant answers to complex queries than conventional website-oriented search.

C. Media

AIGC is presented as a broad production technology for media and other industries, automating content workflows while supporting creative and service applications.

  • C. Media: AIGC can affect news collection, manuscript writing, video editing, and news broadcasting across the media production workflow.
  • C. Media: In media, AIGC can sort and record voice data, accelerate structured writing, correct errors, edit videos, synchronize subtitles, and enhance clarity.
  • C. Media: E-commerce platforms use intelligent text generation for product titles and descriptions and intelligent customer service for shopping and post-sale inquiries.
  • C. Media: AIGC can support film scriptwriting, role and setting replacement, post-production editing, and the generation of style-fitting scripts.
  • C. Media: Beyond these fields, proposed applications include educational visualizations and finance-related information videos and virtual customer service.
  • The survey reviews AIGC’s features, advantages, disadvantages, challenges, and future directions for academia, industry, and business.
Loading 2304.06632v1…