How Generative AI Shapes Intelligent Advertising Production

The paper is published in “Advertising Research” (“Modern Advertising” Journal), 2024, 21(12): 16-25.

Wang Fei Researcher at the Research Center for Journalism and Social Development, Renmin University of China, Professor in the Department of Advertising and Media Economics, and Director of the Modern Advertising Research Center

Li Siqi Doctoral student of the 23rd cohort at the School of Journalism, Renmin University of China

Generative artificial intelligence shapes a new advertising production form through its capabilities in understanding, reasoning, and generating natural language. This article reveals the new characteristics of generative AI advertising production through research on Baidu’s AI Native intelligent agent: the large model integrates generative intelligence through natural language interaction during the advertising delivery process. Through natural language interaction between the advertiser (or agency operator) and the intelligent agent, it completes the advertising activity production form composed of generative intelligent insights, generative intelligent reasoning, generative intelligent creation, generative intelligent delivery, and generative intelligent services. Thus, the main characteristics of generative AI advertising production include natural language interaction, large language model intelligence, and personalized interactions, which achieve personalized marketing communication goals targeting users/consumers. The paper further discusses the features of generative AI advertising production, reflected in the transformation of human-computer interaction paradigms, optimization of human-computer collaborative operation modes, enhancement of return on investment, and strengthening of consumer value co-creation. Finally, the paper reflects on the issues faced by generative AI advertising from the perspectives of interaction logic, interaction modes, and interaction orientation. The study concludes that generative AI advertising is an intelligent marketing communication activity where advertisers and target consumers interact with large language models through natural language, achieving personalized precise demand connections through human-computer collaboration in generative multimodal communication.

Generative AI Advertising Generative AI Advertising Production Form Business Intelligent Agent Human-Computer Collaboration

Generative artificial intelligence (GAI) refers to computational technologies that can generate seemingly new and meaningful content (such as text, images, audio, video) from training data[1]. AI-generated content (AIGC) is the realization of content generation through extracting and understanding intent information from human-provided instructions, based on its knowledge and intent information. Technically, AIGC refers to assisting in teaching and guiding models to complete tasks under given human instructions, utilizing GAI algorithms to generate content that meets the instructions[2]. AIGC achieves interaction between the virtual and the real through intelligent semantic understanding and attribute control[3] [4] [5] . The goal of generative artificial intelligence is to make the content creation process more efficient and accessible, thereby producing high-quality content at a faster pace.

This article takes Baidu’s AI Native business intelligent agent as the research object to explore how generative artificial intelligence shapes the production form of intelligent advertising. Baidu’s AI Native business intelligent agent is based on the iteration of the Wenxin large model, incorporating marketing data from Baidu’s business ecosystem, serving as an advertising intelligent agent with cutting-edge representation. The article analyzes the production form of Baidu’s generative AI advertising through case studies and interviews with business managers from various departments within Baidu, discussing the characteristics and forms of generative intelligent advertising production.

Baidu AI Native includes a series of tools for intelligent advertising activities: Star Chart (AI Marketing Decision Platform), Yangji (AI Business Engine), Qinge (AI Native Marketing Platform), Qingshuo (AIGC Creative Platform), and Qiaocang (Merchant Intelligent Agent Creation Platform), with a specific matrix distribution as shown in Figure 1. Qinge can be understood as a new generation advertising delivery platform, serving as the human-computer interaction interface for marketing planning, implementation, and optimization; Yangji is the intelligent engine for reasoning, understanding, and feedback, achieving precise connections between advertisers and users; Qingshuo is the intelligent platform for producing marketing creative content; Qiaocang is the system for creating personalized merchant intelligent agents (brand BOTs).

Figure 1 Baidu Business AI Native Application Matrix

Based on the actual situation of Baidu’s intelligent advertising business process, the author examines the operational mechanism of generative artificial intelligence technology embedded in advertising production through five stages: insight, reasoning, creation, distribution, and service, discussing the changes occurring in intelligent advertising production forms.

From extensive user understanding based on demographic targeting to measurement based on user tags, and then to user tracking based on data navigation[6] , user insights are gradually refined and intelligentized. With the empowerment of large models, insights into industry trends, competitive situations, user needs, and brand activities that previously required inputting keywords to retrieve data for analysis can now be obtained by advertising agents (or advertisers) simply by inputting their intentions in natural language.

As Baidu’s data insight tool, Star Chart includes data based on the “Opportunity – Cognition – Interest – Interaction – Action” chain, big data from Baidu’s business ecosystem, market dynamics data, and public opinion data, etc. It reconstructs the foundation of the intelligent insight platform by integrating the Wenxin large model, Wenxin Embedding, Yangji business engine, and other models. Star Chart enables detailed insights across various scenarios, such as brand diagnosis, volatility attribution, budget allocation, crowd breaking, providing search opportunities, content marketing, monitoring brand public opinion, and reviewing delivery, through real-time monitoring of big data and natural language interactive Q&A dialogues. Through multi-turn interactions of prompts, the platform can conduct multidimensional mining based on user needs, such as trend research: querying the trend changes of keywords over the past decade; demand maps: presenting hidden focal points of user attention; crowd portraits: presenting the group portraits of target users; reasoning attribution: analyzing the deeper reasons behind data fluctuations. Compared to traditional keyword retrieval models and the extensive user positioning of a one-size-fits-all approach, the generative intelligent insight tool, after being reconstructed based on large models, achieves real-time multidimensional interaction and more detailed decision support, even inspiring agents’ creativity.

2.2 Generative Intelligent Reasoning: Precise Matching Based on Big Data and Multi-Feedback

Traditional advertising requires multiple stages, including advertisement recall, rough sorting, creative selection, and fine sorting before formal bidding, making the process complex and less precise. Artificial intelligence can assist in analyzing large datasets through machine learning algorithms to identify patterns and correlations that human researchers might miss[7] . The large model achieves intelligent understanding, reasoning, and feedback through natural language interaction, continuously clarifying intentions through multi-turn human-computer interactions, deepening understanding, optimizing data quality, and expanding data volume, driving the optimization flywheel to provide an intelligent reasoning engine for final advertising delivery.

Yangji, as Baidu’s commercial power engine, significantly enhances advertising delivery efficiency and effectiveness. Its core breakthroughs are reflected in three aspects: first, generative recall, which achieves generation and retrieval, allowing for direct recall of matching advertiser materials based on generative retrieval, greatly improving system recall efficiency, marking the industry’s first successful implementation. Second, generative creativity, where the large model generates personalized creative content in real-time online. Third, end-to-end bidding, where inputting natural language instructions enables advertising auctions.

The intelligent feedback from the large model drives multi-turn interactions between clients and provides prompt optimization directions based on commercial knowledge graphs and massive user search behavior data through intelligent analysis, understanding, reasoning, and memory. For instance, when some key points mentioned in a client’s natural language proposal closely match existing demand types of Baidu users and are consistent with the client’s advertising landing page content and business content, these key points will be highlighted as “marketing highlights.” Marketing highlights will be prioritized in ad delivery to match users’ search needs, helping clients find high-value potential target audiences that meet expectations.

Thus, generative artificial intelligence reshapes the online real-time reasoning engine, accelerating the enhancement of reasoning accuracy and promoting the generation of personalized interactive content through deep collaboration between software and hardware.

Generative intelligent creation is the most well-known capability of generative artificial intelligence advertising, often referred to as AIGC advertising, and is sometimes misunderstood as the entirety of generative AI advertising. Generative intelligent creation, based on large models, provides precise and innovative multimodal advertising creative content, achieving intelligent generation of marketing communication content. Currently, generative intelligent creation can realize creation in three dimensions: text, images, and videos. Generative intelligent creation deeply integrates generative intelligence with marketing scenarios, achieving automation and intelligence in multimodal content production. AIGC advertising solves the content creation challenges of advertisers who previously lacked content or produced low-quality content. Meanwhile, it enables minute-level advertising creation, greatly enhancing production efficiency.

Qingshuo is Baidu’s AIGC advertising creation platform, based on the underlying big data model of Wenxin Yiyan, aggregating Baidu AI technology, and has capabilities in copywriting generation, image generation, and digital human video creation. Qingshuo understands and gains insights into high-quality creative marketing scenarios through the large model, forming a data flywheel through massive data learning, achieving a closed-loop content production process of “understanding – production – delivery – optimization”. In natural language interactions with operators, it gains insights into materials, identifies intentions, and then generates copy, images, or videos, continuously optimizing through effect data feedback after delivery, meeting advertisers’ dual demands for production efficiency and marketing effectiveness. Qingshuo recommends creative inspiration based on prompts. In addition to automatically generating content, it also supports clients in providing creative materials, generating content that better meets personalized expressions through model understanding and learning. Data comparisons show that AIGC advertising reduces production costs, improves production speed, and enhances marketing effectiveness.

2.3.1 Text Creation: Multi-Type Generation, Inspiring Creativity

Generative copywriting creation automatically generates and optimizes copy content based on database deconstruction and insight attribution. Qingshuo generates different selling point recommendations based on business characteristics and effect goals and directly applies them to the generated copy. The copy includes various types such as voice-over scripts, advertisement titles, and video scripts. The speed of copy generation is fast, greatly enhancing the richness of advertising works. In addition to traditional copywriting creation, the advantage of generative text creation lies in its ability to achieve pre-launch predictions, where the system evaluates the quality of copy content based on the expected performance of the copy and assigns scores. Titles with high scores will be marked and put into more delivery processes, assisting advertisers in competing for user attention.

2.3.2 Image Creation: Breaking Barriers, End-to-End Direct Material Generation

Generative image creation is an important application area of AIGC advertising content creation, with its core advantage being zero-barrier creation through natural language interaction. In the advertising field, generative image creation is widely used for generating marketing images, including background images, portrait images, scene images, marketing posters, etc. Generative image creation significantly reduces production costs; for example, Baidu’s Qingshuo adopts a one-click generation model for marketing images, reducing production costs by 70%. Meanwhile, production efficiency has also dramatically improved, enabling minute-level scaled image production and one-click template changes.

2.3.3 Video Production: Input Scripts, One-Click Completion

Baidu’s generative advertising video creation includes various forms such as digital humans, voice-overs, and mixed-cut short and long videos. First, generative video supports script-driven actions, allowing advertisers to interact with the intelligent agent through natural language prompts, with the system intelligently recommending suitable digital human video creative inspirations. Clients can then easily edit and create digital human video advertising content online, with one-click delivery. Additionally, generative video provides more personalized customization services through the continuous expansion of resource libraries, matching diverse digital human images, voice materials, and industry-specific templates, enabling customized modeling of digital humans and supporting digital human asset management. It can optimize scenes according to industry characteristics, with fast modeling and powerful mixed-cut capabilities, significantly reducing production cycles.

Generative intelligent delivery is based on all previous stages, targeting consumers/users precisely for advertising delivery and feedback on effects, subsequently optimizing intelligently. Generative intelligent delivery has three major breakthroughs: first, simpler expression, transitioning from keyword extraction to natural language expression; second, simpler steps, moving from hierarchical lists to natural dialogue and two-way clarification; third, continuously evolving capabilities, where self-learning leads to self-upgrading. With these technological enhancements, generative intelligent delivery can simultaneously meet both “end-to-end delivery” and “end-to-end auction,” reconstructing the framework for intelligent advertising generation and auction.

The specific operational process of generative intelligent delivery is as follows: the advertising agent (or advertiser) initiates a dialogue, inputting prompts to actively express “the product/service I want to promote” and “the characteristics of my target audience,” the Qinge intelligent agent engages in multi-turn dialogue with the operator, advancing intention understanding, inspiring insights, clarifying uncertainties, and guiding the operator to adjust instructions for the best match based on client budget and goals, confirming whether it can connect with users. Prompts mainly consist of “goods” and “people,” where the client’s natural language expressions of “goods” and “people” are understood by the engine layer, guiding clients to continuously optimize prompts through feedback loops. In addition to “the product/service I want to promote” and “the audience I want to see the advertisement,” clients can also input “my effect goals and consumption expectations,” “my budget,” and “my advertisement creative content,” etc.

2.4.1 End-to-End Generative Delivery: Direct Natural Dialogue, Matching Diverse Fields

The automated delivery of traditional graphical user interface (GUI) platforms varies greatly across industries, making various functional operations difficult. Qinge communicates with operators through natural language, understanding intentions in dialogue, confirming plans, and achieving personalized intelligent delivery, realizing a one-size-fits-all approach. For example, inputting natural language such as “automatically identify high-potential target audiences in the Meiyu area for delivery” can achieve “end-to-end” generative advertising delivery.

Traditional computational advertising systems recall a set of candidate advertisements based on conditions such as keywords, target areas, target audiences, and budgets, then calculate the matching degree of each advertisement with the client, combining client bidding to ultimately present the most suitable and highest-priced advertisement. This process includes advertisement recall, rough sorting, creative selection, fine sorting, and advertisement bidding. In contrast, generative intelligent delivery only requires inputting prompts through natural language interaction, allowing for generative targeting, adaptive traffic fields, and agile matching of client needs, achieving generation and retrieval. This greatly enhances system recall efficiency, simplifying the delivery process that previously required complex operations, enabling rapid coverage of target audiences. Generative advertising delivery can connect personalized needs through two approaches: one is to push advertisement information to users when they search, achieving “people find goods”; the other is to push advertisements to target users, achieving “goods find people.”

2.4.2 End-to-End Auction: Data-Driven Optimal Solutions

Big data technology enhances the precision and accuracy of advertising transaction operations, such as pay-per-click from search engines, second-push of retargeting advertisements, and effective real-time bidding on online advertising trading platforms[8] , but these forms still struggle to meet the trends of multi-target and multi-business rules in advertising auctions.

Generative intelligent delivery shifts from traditional sorting algorithms to data-driven deep large models, learning optimal billing under different conditions through temporal optimization modeling. Under this logical shift, the complexity of advertising auctions significantly decreases, and efficiency greatly increases, with the universal optimal solution no longer being goal-driven; instead, the focus shifts to the optimal solution suitable for data, transitioning from theoretically driven advertising auctions to data-driven “end-to-end optimal auctions.”

Generative intelligent services are typified by generative intelligent customer service, i.e., brand robots, which can better facilitate human-computer dialogue through dialogue systems. Depending on whether a task is specified in the application, dialogue systems can be divided into two categories: task-oriented dialogue systems (TOD) and open-domain dialogue systems (OOD). The former mainly involves natural language understanding (NLU), dialogue state tracking (DST), dialogue policy learning (DPL), and natural language generation (NLG) technologies, while the latter includes retrieval-based systems, generative systems, and integrated systems, responding to users based on different logics[9] . In advertising applications, the focus is often on task-oriented dialogue scenarios. Generative intelligent customer service utilizes natural language processing technology to convert user-inputted text or speech into computer language, analyzing user intentions, extracting information from knowledge bases and databases, and generating responses through algorithms.

User searches often encompass complex demands; failing to recognize demands and provide corresponding responses can affect user engagement and satisfaction with the brand, subsequently impacting click-through rates and conversion rates. Baidu creates brand intelligent agents, merchant intelligent agents, industry intelligent agents, and virtual human solutions for advertising clients through generative intelligence. Through natural language interaction, personalized and character-driven dialogue marketing can be achieved. This generative intelligent dialogue allows virtual humans to promote clients’ brand culture and products, answer user inquiries, and deepen the relationship between users and brands through personalized interactions. It can also deeply recommend various services to users, such as recommending leads, store visits, and orders, helping clients save brand marketing costs.

For instance, taking Baidu’s merchant intelligent agent as an example, let’s look at the service principles of generative intelligence. The merchant intelligent agent fully learns and understands high-quality content obtained from post-delivery validation through trained models. By deeply deconstructing high-quality content and effectively attributing it based on linked data, it forms creative guidance, continuously outputting marketing content such as text, images, and videos that better align with user consumption habits through multimodal generative intelligence. After content output, combining user behavior and signals, the large model iteratively upgrades based on feedback, achieving linked optimization of efficiency and effectiveness. With generative creation, building merchant sites on Baidu takes only 2 hours, significantly improving conversion rates. Through dialogue interactions, the large model accurately parses information, intelligently generates product descriptions, Q&A, landing pages, and other graphic information, generating N sites + 1 store in just 10 minutes.

The efficiency of generative intelligent services is reflected in: 1. Reducing operational costs: autonomous inquiries, dialogues, and guiding lead generation can fully or partially replace human customer service; 2. Enhancing service efficiency: 24/7 service resolves issues such as busy times and unavailability during evenings and weekends; 3. Optimizing service quality: autonomously learning vast amounts of high-quality dialogues, providing personalized services tailored to individuals through multi-turn dialogues; 4. Strengthening customer acquisition capabilities: uninterrupted service for lead generation, accurately understanding contextual semantics, effectively filtering low-quality leads, and guiding the entire process for lead generation. The large model analyzes user attention points and emotions, enabling advertising clients to better grasp user intentions and mindsets. Intelligent agents, powered by large models, understand industries, communicate effectively, and provide diverse, personalized services. Utilizing generative intelligent customer service for service and marketing upgrades is a significant application scene for all brands in the future.

In summary, through the analysis of Baidu AI Native, it is evident that generative intelligence is pushing intelligent advertising activities into a new stage. After embedding generative artificial intelligence technology into the intelligent advertising production process, the five major stages exhibit generative intelligent characteristics, namely: generative intelligent insights, generative intelligent reasoning, generative intelligent creation, generative intelligent delivery, and generative intelligent services. The generative intelligent advertising production form is illustrated in Figure 2.

Figure 2 Generative Intelligent Advertising Production Form

Intelligent advertising (Intelligent Advertising) is regarded as the third stage of digital advertising, with the previous two stages being interactive advertising (Interactive Advertising) and programmatic advertising (Programmatic Advertising). The new phase of digital advertising retains valuable attributes from the previous stages while adding innovative characteristics[10] . Generative AI advertising innovatively iterates, breaks through, and reshapes the core value attributes of the first three stages, characterized by human-computer collaboration, large language model intelligence, and personalized interactions.

Based on the research on Baidu AI Native, it can be concluded that the characteristics of the generative artificial intelligence advertising production process include: comprehensive business intelligent agents driven by LLM technology and natural language interfaces, achieving lossless communication in human-computer dialogues; secondly, large language models continuously optimize advertising delivery plans based on comprehensive applications of prompts and other information types, efficiently matching target customer groups; thirdly, through understanding, generation, logic, and memory processing at the model level, personalized advertising push for thousands of people becomes possible, better satisfying users’ personalized experiences.

Language is a mark of “humanity and perception,” and dialogue is considered to represent “the most basic and privileged domain of language”[11] . Dialogue is an active social process through which humans create common meanings from experiences and establish, define, and advance their relationships with others based on a shared understanding of surrounding realities[12] . Humans have long attempted to interact with machines, evolving from the binary input patterns of punched tape used by a few computer experts in the 1940s-1960s, to command-line interface (CLI) interactions in the mid-1960s, and then to graphical user interface (GUI) interactions in the 1970s, finally entering the natural language interface (NLI) interaction stage. Early natural language interactions mainly manifested as auxiliary interactions through voice user interfaces (VUI). With the development and application of large language models (LLMs), natural language processing (NLP) technology has achieved significant breakthroughs.

Most people are “cognitive misers”: they tend to think and solve problems in simpler, less effortful ways rather than more complex, effortful ways[13] . When given the opportunity to take mental shortcuts in information processing, people often do so to avoid the effort of analyzing information. Natural language interaction lowers the barriers for users to express, create, and implement advertisements, shifting the processes of information deconstruction, insight, and reasoning from the forefront to the background, making human-computer dialogue more convenient and intelligent, and simplifying the marketing process. Utilizing AI-based dialogue interfaces as a novel interaction paradigm between consumers and brands is being referred to as “the next business operating system”[14] . Large models enable immersive human-computer dialogues, achieving a full-process of advertising delivery, termed as an “end-to-end” intelligent processing procedure. Advertising agents (or advertisers) express themselves directly through natural language, while large models continuously clarify, resolve cognitive differences, and provide feedback optimization in interactive dialogues, ensuring lossless information transfer, achieving natural language processing, semantic analysis, emotional analysis, graphic and text generation, context management, etc. The entire advertising delivery process is made more natural and convenient through generative dialogue communication. At the same time, consumers can also search for products, obtain product information, purchase products, or interact with brands through natural language dialogues, deepening their relationship with brands.

Large language models possess characteristics of large-scale and pre-training, requiring extensive generalized data for pre-training before modeling actual tasks[15] . The large language models of artificial intelligence can be viewed as simulations of the human brain, which serve as the inspiration for artificial intelligence[16] . The increased accuracy of large language models in understanding human thinking is attributed to systems trained on human feedback data[17] .

In the field of natural language processing (NLP), Transformer models have replaced recurrent neural networks (RNNs)[18] as the de facto standard backbone. In the field of computer vision (CV), in addition to traditional convolutional neural networks (CNNs), vision Transformers (ViT)[19] have also demonstrated their power. New models enable researchers to train larger models without needing to pre-label all data, allowing training on massive text to yield deeper answers. Additionally, the Transformer model employs self-attention mechanisms, enabling the model to selectively focus on and weigh different parts of the input sequence[20] .

In advertising activities, taking Baidu’s Wenxin large model as an example, the understanding, generation, logic, and memory core capabilities driven by LLM technology can more accurately understand advertising clients’ intentions, executing insights, targeting, analysis, reasoning, generation, delivery, and other series of instructions, instantaneously completing complex operational steps, significantly simplifying the advertising delivery process. The ability of the large model to simplify processes, optimize plans, automatically generate, and make rapid decisions endows generative AI advertising with clear cost reduction and efficiency enhancement characteristics.

Generative artificial intelligence can achieve efficient content creation, meet the growing interactive demands, and improve personalized experiences[21] . It can simulate virtual human brains to generate content, including intelligent NPCs, automatic QAs, dialogue systems, and digital humans[22] .

Personalized interactions for thousands of people first manifest in personalized interactions in human-computer interactions. Traditional computational advertising can already push matching advertisements to target users, but due to the limited number of advertisements, it cannot truly achieve personalized interactions for thousands of people, leading to different consumers encountering the same advertisement content, presenting a one-size-fits-all promotion. Large models continuously optimize their understanding of the characteristics of advertisers’ products/services, the intentions of advertising agents (advertisers), and the needs of consumers through multi-turn human-computer interactions, demonstrating stronger logic. Through multi-dimensional interaction data such as user characteristics, situations, context, and historical behaviors, personalized modeling, machine learning, and deep learning are conducted, capturing not only the basic attribute features of consumers accurately but also the semantics and emotions in context, enabling personalized interactions with emotional experiences and character companionship, enhancing the service capabilities connecting advertisers and consumers. Secondly, the intelligent generation of advertising copy, images, videos, and digital virtual humans also forms works based on user characteristics, automatically generating different appeals, protagonists, environments, aesthetics, and other elements. The intrinsic essence of large language model marketing and generative AI advertising is to achieve personalized interactions for thousands of people.

Generative artificial intelligence shapes a new intelligent advertising production form, specifically reflected in the transformation of human-computer interaction paradigms, optimization of human-computer collaborative operation modes, enhancement of return on investment, and strengthening of consumer value co-creation.

Traditional computational advertising is an “algorithm-driven paradigm,” where specific problems are solved through specific algorithms by analyzing big data for precise advertising push. In contrast, generative AI advertising operates under a “smart conjecture paradigm” + “adaptive paradigm” in human-computer interaction, achieving individualized advertising delivery and personalized information interaction through the collision of machine intelligence and operator instructions in the generative advertising process, which involves understanding, generating, logic, and memory. Generative AI advertising operations extract data targeting specific problems from thousands of billions of parameters, reasoning to derive precise and effective solutions, efficiently utilizing newly generated small data in human-computer interactions, dynamically connecting advertising information to meet each consumer’s evolving needs, and being able to perceive the environment, make decisions, and execute actions. This is fundamentally different from the static precise matching of traditional computational advertising. For example, IHG hotels built a generative AI-driven chatbot to help guests easily plan their next vacation through the IHG One Rewards mobile app.

Generative artificial intelligence shapes a new human-computer collaborative advertising operation mode. The “machine” refers to the intelligent agent, which possesses four main capabilities: first, perception ability, quickly completing tasks of information acquisition, user insight, and demand disassembly; second, memory ability, achieving personalized interaction through a combination of short-term and long-term memory; third, planning ability, guiding precise decision-making based on learning and reasoning from large models; fourth, action ability, efficiently and qualitatively completing processes such as insight, reasoning, creation, distribution, and service.

The “human” refers to operators, architects, creators, etc., who dominate the production process based on systematic business logic. Their task is to build a collaborative network between the AI decision-making center and human expert experiences, leveraging the organic fusion of advanced intelligence, possessing complex decision-making abilities and strategic foresight. Human advantages lie not in computational power but in theory-driven thinking approaches[23]. Humans, as the most intelligent beings, apply various intelligent tools to not only enhance decision-making levels and operational efficiency but also to respond to rapidly changing environments, maintaining organizational innovation capabilities and competitive advantages. Humans, as integrators of high-dimensional logic systems and innovation systems, are strategic planners, change promoters, and practitioners of emerging management models within organizations.

In generative AI advertising operations, humans issue commands to intelligent agents based on strategic intentions, while intelligent agents proactively adjust and optimize behaviors according to human needs, providing emotional companionship through a deep understanding of multi-channel massive data, making decisions, and simplifying advertising operations, creating an immersive advertising operational environment.

Human-computer collaboration aims to achieve high efficiency at low cost. For example, the “human operator + AI” advertising operation model allows AI to leverage its advantages in intelligent analysis, reasoning, and decision-making throughout the entire process, helping human operators break through efficiency bottlenecks and focus on strategic, tactical, and experiential summaries. Liou’s “AI Short Play Operator” undertakes tasks such as script analysis, batch infrastructure, and intelligent monitoring, allowing human operators to free up time and energy to focus on their professional advantages, formulating delivery strategies based on industry experience. Testing the model of completing 2000 short plays across various digital media channels for distribution and delivery accounts, the traditional short play delivery model required 10 people/month to complete the related pre-launch preparation work. After introducing the “AI Short Play Operator,” under the dual operator collaboration model, only 2-3 human operators collaborating with AI operators can easily complete the task within a week.

As another example, Bell Canada built customizable contact center solutions for its business clients, providing AI-driven agents to handle incoming calls and offering agent assistance functions. AI has saved $20 million in customer operations. Procter & Gamble utilized Imagen to develop an internal generative AI platform to accelerate the creation of photorealistic images and creative assets, allowing marketing teams more time to focus on advanced planning and providing quality experiences for consumers. WPP Group integrated Google Cloud’s generative AI capabilities into its intelligent marketing operating system, WPP Open, enabling its employees and clients to achieve personalization, enhance creativity and efficiency, and utilize the Gemini 1.5 Pro model to improve the accuracy and speed of content dissemination predictions.

In advertising content creation, intelligent agents can empower creators to promote divergent thinking, challenge professional knowledge biases, assist in idea evaluation, support idea refinement, and facilitate collaboration with users. A recent study by Stanford University shows[24] that compared to human experts, AI has a significant advantage in generating novel and exciting ideas but is at a disadvantage in feasibility, highlighting its lack of understanding of the complexities of the real world. Human experts can evaluate the feasibility of ideas in practice based on experience and expertise, which is currently lacking in AI. AI can serve as a creative assistant to humans, providing diverse ideas while humans can assess the feasibility and practical value of these ideas.

The empowerment of generative artificial intelligence in the advertising industry is ultimately measured by return on investment. In the advertising creation stage, generative artificial intelligence possesses unlimited creative possibilities. Currently, several major platforms domestically and internationally are equipped with advertising generation tools. For instance, Amazon advertisers only need to select products in the Amazon Ad Console and click the “generate” button, allowing the system to use generative AI technology to generate a series of lifestyle and brand-themed images within seconds.

In the advertising delivery stage, the end-to-end operation model of human-computer interaction replaces the funnel matching model of traditional computational advertising, reducing the information loss in advertising operations and lowering participation thresholds, improving efficiency, reducing costs, and enhancing advertisers’ return on investment. Advertisers only need to provide the “minimum necessary” information to achieve end-to-end matching of advertiser needs and consumer demand through native interactions, allowing intelligent systems to integrate content generation and service resources. This will help advertisers further enhance their communication capabilities with consumers, expanding the boundaries of serving consumers, moving towards a business model that meets “anyone, anytime, anywhere, for any need”.

Additionally, adaptive iteration also brings positive feedback on efficiency. The operational mechanism of AI-driven advertising based on AI intelligent agents is reinforcement learning (RL), where intelligent agents learn optimal strategies through trial and error and reward-punishment systems while interacting with the environment, guided by human feedback. Moreover, knowledge learned by a model on one task can be applied to another related task, improving learning efficiency and performance on new tasks, known as transfer learning. These features of generative artificial intelligence enable spontaneous iterations of advertising delivery, intelligent customer service, and advertising creation capabilities, forming positive interactions among advertisers, consumers, and intelligent agents through multi-turn dialogues, generating knowledge graphs and operational logic. In the long run, this will inevitably optimize system efficiency and enhance the return on investment for advertisers and intelligent agent platforms. For example, Baidu’s merchant intelligent agent leverages the expertise of industry professionals to optimize dialogues, making intelligent agents more professional and continuously enriching the industry knowledge system to enhance the knowledge reserves of intelligent agents. In providing pioneering experiences, emotional dialogues, and deep communication with consumers, it progressively deepens the relationship with consumers, thereby creating business opportunities.

In the ecosystem of generative artificial intelligence, consumers can easily apply generative intelligent agents to create various modalities of content based on the propositions or tasks of advertisers, participating in the value co-creation of brands. Moreover, through interactions with intelligent agents, consumers contribute their emotions, attitudes, and behavioral data, providing a foundation for optimizing intelligent systems. Additionally, consumers’ use of AI intelligent agents for social interactions is also an important area for consumers to participate in value co-creation in generative AI advertising operations. For example, Snapchat has launched the AI chatbot My AI based on ChatGPT, allowing users to @My AI in chats, inviting it to join group chats and enlivening the conversation atmosphere. Meta has also launched Meta AI, allowing users to learn, work, and create using Meta AI on Facebook, Instagram, WhatsApp, and Messenger.

The emerging generative artificial intelligence technology undoubtedly brings new challenges, akin to Heidegger’s critique of “technological enframing.” We must be vigilant against viewing technology merely as a means to improve efficiency and solve problems. We need to transcend the instrumental logic of generative artificial intelligence, focusing not only on its efficiency and practicality but also on its profound impacts on human values.

Generative AI advertising can complete human-computer dialogues through natural language. Advertising agents (or advertisers) can obtain user information data through multi-turn interactions, completing the entire process of insights, reasoning, creation, delivery, and services in one go. Consumers can receive personalized product recommendations and services through dialogues with brand robots. The human-computer interaction process can be realized through backend operations without humans fully understanding the algorithmic logic. Although the one-stop operation of generative AI advertising has significant advantages in efficiency and lowering barriers, its interpretability is lacking.

Firstly, there is a lack of transparency. Where does the foundational data come from? How do the connecting algorithms flow? What is the basis for predictions and recommendations? In early intelligent advertising, advertisers could control the processes through periodic user analysis reports and advertising effect reports. However, with the development of generative artificial intelligence, the trend of algorithmic “black-boxing” has emerged, and the transparency of human-computer interactions needs to be improved. Secondly, there is an urgent need for aligned value standards. Traditional benchmarking methods are no longer sufficient to comprehensively assess the capabilities of AI systems. Therefore, incorporating human evaluations into the assessment framework of AI advertising systems is becoming increasingly important. Gradually moving towards super alignment through value alignment and safety alignment ensures that AI systems act according to human intentions and goals.

Generative AI advertising predicts and provides feedback on user needs based on the analysis of user data and interpretation of user instructions, aiming for personalized interactions for thousands of people. Although current AI technologies have acknowledged the impact of personalization and emotional factors, human emotions and emotional elements are highly complex and variable. The emotions of AI lack the overall nature of human emotions[25] , making it difficult to accurately identify users’ emotions and contexts during human-computer interactions, potentially leading to misunderstandings and triggering resistance psychology.

On one hand, misjudgments of user emotions may lead to incorrect judgments of user needs, such as brand intelligent robots being criticized for failing to answer questions, being overly procedural, and lacking resonance, leading to “ineffective dialogues.” On the other hand, while generative AI advertising can make precise pushes according to advertisers’ needs, overly frequent personalized pushes may provoke user aversion and privacy anxiety, ultimately leading to distrust between humans and machines. Insights based on multiple emotional attitudes and clearly defining the boundaries of personalization should be incorporated into the considerations of human-computer interaction modes.

Generative AI advertising unleashes a significant amount of productivity throughout the advertising operation process, but from a long-term perspective, the interaction orientation should avoid excessive reliance on artificial intelligence to prevent the loss of human subjectivity. The relationship between humans and technology, as well as between humans and objects, is one of co-evolution; humans invent and create technology, which in turn shapes human thoughts and life worlds, influencing human evolution and development[26] .

Thus, achieving the pursuit of human subjectivity in human-computer interactions through co-evolution with technology is the ultimate orientation. In generative AI advertising, advertisers should adopt a more proactive attitude in using artificial intelligence, steering brand image and value propositions, avoiding the generation of soulless AI advertisements and the uncanny valley effect, such as the widely criticized, uninspired, and lifeless Coca-Cola AI Christmas advertisement “Holiday Magic is Here.” As consumers, they should break free from passive acceptance, actively creating personalized products that match their needs in value co-creation, and utilizing technology to uncover latent demands, finding the products and services they truly desire.

Natural language interaction, large language model intelligence, and personalized interactions for thousands of people are the core characteristics of generative artificial intelligence advertising production, significantly impacting the production forms of intelligent advertising, representing a comprehensive advancement centered on “generation.”

The characteristics of generative AI advertising production forms are primarily the transformation of human-computer interaction paradigms. Advertising agents (or advertisers) immersed in large models and target consumers achieve precise connections in personalized human-computer interactions. In the interactions, feedback, and iterations across various production stages, intelligent systems continuously self-update and optimize, with the powerful intelligence emerging helping advertisers efficiently and cost-effectively achieve personalized marketing communication goals for thousands of people while meeting consumer needs.

Thus, generative AI advertising is an intelligent marketing communication activity where advertisers and target consumers interact with large language models through natural language, achieving personalized precise demand connections through human-computer collaboration in generative multimodal communication activities.

Undoubtedly, in the new intelligent advertising production forms characterized by generative artificial intelligence, how to realize meaningful creation between humans and machines becomes the core issue of AI advertising production, particularly in how to achieve human-computer collaboration among advertisers (advertising agents), consumers, and intelligent agents across insights, reasoning, creation, delivery, and services, forming intelligent productivity with efficiency, return on investment, value alignment, and sustainable development dimensions will be the challenges faced by advertising activities in the coming era.

[1] Feuerriegel, S., Hartmann, J., Janiesch, C. et al, Generative AI [J]. Bus Inf Syst Eng, 2024, 66(1), 111–126.

[2] Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yutong Dai, Philip S. Yu, and Lichao Sun, A Comprehensive Survey of AI-Generated Content (AIGC): A History of Generative AI from GAN to ChatGPT [J]. ACM, 2018, Vol. 37, No. 4.

[3] L. Dong and M. Lapata, Language to logical form with neural attention [J]. The Association for Computer Linguistics, 2016, pp. 33– 43.

[4] M. Rabinovich, M. Stern, and D. Klein. Abstract syntax networks for code generation and semantic parsing [J]. Association for Computational Linguistics,2017, pp. 1139–1149.

[5] P. Yin and G. Neubig. A syntactic neural model for general-purpose code generation [J]. Association for Computational Linguistics, 2017, pp. 440–450.

[6] Duan Chunlin, Yang Heng. Data, Model, and Decision: The Development and Evolution of Computational Advertising [J]. Journalism University, 2018 (01), 128-136+154.

[7] Jiayang Wu, Wensheng Gan, Zefeng Chen, Shicheng Wan, and Hong Lin. AI-Generated Content (AIGC): A Survey[J]. 2023, arXiv:2304.066322.

[8] Ni Ning, Jin Shao. Precision Advertising and Its Communication Strategies in the Era of Big Data: A Field Theory Perspective [J]. Modern Communication (Journal of China Media University), 2014(02), 99-104.

[9] Zhang, Chaoning, et al. A complete survey on generative AI (AIGC): Is chatgpt from gpt-4 to gpt-5 all you need? [J]. 2023, arXiv:2303.11717.

[10] Li, H. Special section introduction: Artificial intelligence and advertising [J]. Journal of advertising,2019, 48(4), 333-337.

[11] Jurafsky, Daniel and James Martin, “Dialogue Systems and Chatbots,” in Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition, Third Edition Draft, Pearson, 2019, 487–517.

[12] Wood, Julia T. Communication and Relational Culture: Bases for the Study of Human Relationships [J]. Communication Quarterly, 1982, 30 (2), 75–84.

[13] Fiske, S. T., and S. E. Taylor. Social Cognition [M]. New York: McGraw-Hill. 1991.

[14] Suri V K, Elia M, van Hillegersberg J. Software bots-the next frontier for shared services and functional excellence[C]//Global Sourcing of Digital Services: Micro and Macro Perspectives: 11th Global Sourcing Workshop 2017, La Thuile, Italy, February 22-25, 2017, Revised Selected Papers 11. Springer International Publishing, 2017: 81-94.

[15] X. Han, Z. Zhang, N. Ding, Y. Gu, X. Liu, Y. Huo, J. Qiu, Y. Yao, A. Zhang, L. Zhang et al. Pre-trained models: Past, present and future [J]. AI Open, 2021, vol. 2, pp. 225–250.

[16] K. Amit. Artificial Intelligence and Soft Computing: Behavioral and Cognitive Modeling of the Human Brain[M]. CRC press, 2018.

[17] L. Ma and B. Sun. Machine learning and AI in marketing–connecting computing power to human insights [J]. International Journal of Research in Marketing, 2020, vol. 37, no. 3, pp. 481–504.

[18] Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernocky`, and Sanjeev Khudanpur. Recurrent neural network based language model [J]. In Interspeech,2010, Vol. 2. Makuhari, 1045–1048.

[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. (2020). An image is worth 16×16 words: Transformers for image recognition at scale[J]. 2020, arXiv:2010.11929.

[20] Zhendong, Du, and Tuming. Generative AI: Natural Language process and Human- Computer interaction[J]. China Machine Process.2020, isbn: 9787111664192.

[21] M. David Stone and N. David Woodcock. Interactive, direct and digital marketing: A future that depends on better use of business intelligence[J]. Journal of Research in Interactive Marketing,2014, vol. 8, no. 1, pp. 4–17.

[22] J. Sun, W. Gan, H.-C. Chao, and P. S. Yu. Metaverse: Survey, applications, security, and opportunities[J]. 2022, arXiv:2210.07990.

[23] Felin T, Holweg M. Theory Is All You Need: AI, Human Cognition, and Causal Reasoning[J]. Strategy Science, 2024.

[24] Si C, Yang D, Hashimoto T. Can llms generate novel research ideas? a large-scale human study with 100+ nlp researchers[J]. arxiv preprint arxiv:2409.04109, 2024.

[25] Peng Lan. The Reflection on the Application Boundaries of Artificial Intelligence “Revival” [J]. People’s Forum, 2024,(11):51-54.

[26] Wu Fei.(2017). The Future Orientation of Journalism and Communication Research: Human Subjectivity and Technological Autonomy. Social Science Front (01), 148-158.