Generative AI Poised to Add $4 4 Trillion to Global Economy: McKinsey

The term was coined in 1956, but the field has only recently begun having significant effects on the economy. An important phase of drug discovery involves the identification and prioritization of new indications—that is, diseases, symptoms, or circumstances that justify the use of a specific medication or other treatment, such as a test, procedure, or surgery. Possible indications for a given drug are based on a patient group’s clinical history and medical records, and they are then prioritized based on their similarities to established and evidence-backed indications.

However, while training GenAI is financially viable for only a handful of companies, use costs are very low. Thus, firms no longer compete on developing proprietary machine learning and AI algorithms, but rather on their ability to fully harness the capabilities of existing foundation models. “Generative AI” refers to artificial intelligence that can be used to create new content, such as words, images, music, code, or video. Generative AI can be put to excellent use in partnership with human collaborators to assist, for example,

with brainstorming new ideas and educating workers on adjacent disciplines. More generally, it can benefit businesses by

improving productivity, reducing costs, improving customer satisfaction, providing better information for

decision-making, and accelerating the pace of product development.

Automating repetitive tasks allows human agents to devote more time to handling complicated customer problems and obtaining contextual information. Generative AI’s potential in R&D is perhaps less well recognized than its potential in other business functions. Still, our research indicates the technology could deliver productivity with a value ranging from 10 to 15 percent of overall R&D costs. Our analysis did not account for the increase in application quality and the resulting boost in productivity that generative AI could bring by improving code or enhancing IT architecture—which can improve productivity across the IT value chain. However, the quality of IT architecture still largely depends on software architects, rather than on initial drafts that generative AI’s current capabilities allow it to produce. This analysis may not fully account for additional revenue that generative AI could bring to sales functions.

Entos, a biotech pharmaceutical company, has paired generative AI with automated synthetic development tools to design small-molecule therapeutics. But the same principles can be applied to the design of many other products, including larger-scale physical products and electrical circuits, among others. While generative AI is an exciting and rapidly advancing technology, the other applications of AI discussed in our previous report continue to account for the majority of the overall potential value of AI. Traditional advanced-analytics and machine learning algorithms are highly effective at performing numerical and optimization tasks such as predictive modeling, and they continue to find new applications in a wide range of industries. However, as generative AI continues to develop and mature, it has the potential to open wholly new frontiers in creativity and innovation. It has already expanded the possibilities of what AI overall can achieve (see sidebar “How we estimated the value potential of generative AI use cases”).

In the overall average for global growth, generative AI adds about 0.6 percentage points by 2040 for early adopters, while late adopters can expect an increase of 0.1 percentage points. Gen AI tools can already create most types of written, image, video, audio, and coded content. In the near future, we expect applications that target specific industries and functions will provide more value than those that are more general. The advanced machine learning that powers gen AI–enabled products has been decades in the making. But since ChatGPT came off the starting block in late 2022, new iterations of gen AI technology have been released several times a month.

Therefore, growth becomes personalized, and employees receive the guidance they need to progress. “This includes increasing the level of productivity through direct efficiency gains as well as accelerating the rate of innovation and future productivity growth,” Korinek says. Anton Korinek, Ph.D. is a professor of economics at the Darden School of Business at the University of Virginia in Charlottesville and a nonresident fellow at The Brookings Institution, an economic think tank. Optimizing inventory management and recommending products to customers based on their purchase history and browsing behavior is only part of the value of Gen AI in the retail industry. While we cannot predict the future, it is likely that generative AI will serve as a “copilot” that augments people’s ability to perform their roles, thereby leading an evolution of tasks within roles rather than eliminating jobs altogether. For example, the Access Partnership research projects that 45% of workers in India will potentially use generative AI for up to 20% of regular work activities.

Using generative AI responsibly

The latest EY 2023 Work Reimagined Survey indicates that 84% of employers say they expect to have implemented GenAI within 12 months. And a net 33% of employees and employers see potential benefits for productivity and new ways of working. You can foun additiona information about ai customer service and artificial intelligence and NLP. As such, the ability of business leaders to reimagine business models and consider how best to augment workers’ skills will be a key determinant of how powerful the productivity lift from GenAI Chat GPT is. The transformative capability of generative artificial intelligence (GenAI) to augment human work and unlock efficiency will likely have far-reaching implications for the macroeconomic and business landscape. Productivity growth is the main long-term propeller of economic growth and living standards, but growth has slowed in recent decades and remains on a subdued trend, even as GenAI adoption continues to quicken.

Consumers appear to struggle in distinguishing GenAI-generated content from human- generated content (Jakesch et al., 2023). However, several governments (e.g., the U.S. and its AI Disclosure Act) and social platforms (e.g., TikTok, YouTube) are increasingly enforcing clear disclosure of AI-generated content. Therefore, research is warranted to explore the implications of such disclosure requirements for both consumers and firms.

EY refers to the global organization, and may refer to one or more, of the member firms of Ernst & Young Global Limited, each of which is a separate legal entity. Ernst & Young Global Limited, a UK company limited by guarantee, does not provide services to clients. The insights and services we provide help to create long-term value for clients, people and society, and to build trust in the capital markets. Learn how to seamlessly integrate generative AI into your organization’s the economic potential of generative ai workflows while fostering a skilled and adaptable workforce. Each pair of bars is under a different topic, with data representing developer respondent’s feelings with and without the involvement of generative AI in their work. The metrics are whether respondents “felt happy,” were “Able to focus on satisfying and meaningful work,” and were “in a flow state.” In all cases, the more positive responses were, on average, doubled among those using generative AI.

Generative AI (Gen AI) is a type of artificial intelligence designed to generate content without human intervention, including text, images, and even music. This technology uses complex algorithms and machine learning models to memorize patterns and rules from existing data. Unlocking the productivity potential of GenAI will likely require the deployment of both tangible (infrastructure) and intangible (technology, software, skills, new business models and practices) investments.

Finland has promising growth prospects

The recent rise of generative AI has profoundly challenged traditional copyright laws, driven by its powerful generating capabilities. This is compounded by the intricacies in the interpretation of copyrights for AI-generated content as well as the black-box nature of large AI systems. We have addressed these issues from an economic standpoint by developing a royalty sharing model that permits training on copyrighted data in exchange for revenue distribution among copyright owners. This fosters mutually beneficial cooperation between the AI developers and copyright owners. We demonstrate the effectiveness and feasibility of this framework through numerical experiments.

These models contain expansive artificial neural networks inspired by the billions of neurons connected in the human brain. Foundation models are part of what is called deep learning, a term that alludes to the many deep layers within neural networks. Deep learning has powered many https://chat.openai.com/ of the recent advances in AI, but the foundation models powering generative AI applications are a step-change evolution within deep learning. Unlike previous deep learning models, they can process extremely large and varied sets of unstructured data and perform more than one task.

At each generative iteration, the model estimates a probability distribution, indicating the likelihood that any token in the vocabulary would be the next observed xi if the model were reading a pre-existing text. To initiate text generation, an LLM requires “conditioning,” meaning it must be supplied with initial input tokens x1, …, xn − 1. For instance, if we input the prompt “This is a review…,” the token “article” would have a higher probability of selection than the token “bus.” Using a distribution function, the model randomly selects among a list of probable candidates (e.g., “article,” “paper”). The new xi is then added to the text, initiating the repetition of the entire process (Argyle et al., 2023). McKinsey estimates that approximately 75 percent of the value that generative AI use cases could deliver comes from customer operations, marketing and sales, software engineering, and R&D.

For example, the Japanese government recently announced plans to allow students from elementary to high school limited use of generative AI to facilitate in-class discussions and artistic activities. Taiwan’s Ministry of Education has brought in a generative AI chatbot to help students learn English. In India, the Integrating AI and Tinkering with Pedagogy (AIoT) program was launched last year to upgrade the curriculum at 50 schools. Based on Access Partnership’s analysis, roles such as biochemists and biophysicists, astronomers, biologists, bioinformatics scientists, and computer and information research scientists are likely to have the greatest share of their tasks transformed by generative AI.

Early movers can play a crucial role in shaping policies, regulations, and an environment that encourages innovation, investment, and responsible use. It became the fastest-growing app in Internet history after reaching 100 million users in just over two months and spurred the development of other AI tools like Google Bard and Microsoft’s new version of Bing. EY-Parthenon is a brand under which a number of EY member firms across the globe provide strategy consulting services. Initial case studies provide evidence that GenAI will likely provide substantial productivity boosts in four major realms.

The time to act is now.11The research, analysis, and writing in this report was entirely done by humans. Previous generations of automation technology often had the most impact on occupations with wages falling in the middle of the income distribution. For lower-wage occupations, making a case for work automation is more difficult because the potential benefits of automation compete against a lower cost of human labor.

In customer service, earlier AI technology automated processes and introduced customer self-service, but it

also caused new customer frustrations. Generative AI promises to deliver benefits to both customers and

service representatives, with chatbots that can be adapted to different languages and regions, creating a

more personalized and accessible customer experience. When human intervention is necessary to resolve a

customer’s issue, customer service reps can collaborate with generative AI tools in real time to find

actionable strategies, improving the velocity and accuracy of interactions.

Due to the potential the technology has in facilitating customer self-service, resolving issues during initial contact, and reducing response times, McKinsey predicts that the productivity of customer care functions could increase from 30-45% in the coming years. Generative AI is expected to have the greatest impact on higher-wage and highly educated knowledge workers, which previously had the lowest potential for automation. The higher the level of education, the greater the estimated impact of the technology is considered to be. However, generative AI’s impact is likely to most transform the work of higher-wage knowledge workers because of advances in the technical automation potential of their activities, which were previously considered to be relatively immune from automation (Exhibit 13). Generative AI could still be described as skill-biased technological change, but with a different, perhaps more granular, description of skills that are more likely to be replaced than complemented by the activities that machines can do.

Estimated global spending by industry in 2023 on AI systems

We then estimated the growth effects of these productivity scenarios on long-run GDP growth using a growth accounting approach such as Fernald (2014). Disappointingly though, productivity growth has been sluggish in both advanced and developing countries over the past decade. In the US, labor productivity growth has averaged only 1.4% per year since 2013, less than half the rate of the previous decade. Our research found that equipping developers with the tools they need to be their most productive also significantly improved their experience, which in turn could help companies retain their best talent.

For example, United States Express uses generative AI technology to optimize business travel services, enabling intelligent booking, itinerary optimization and real-time support to provide personalized and efficient travel solutions. AI analyzes large amounts of data to accurately predict customer needs and customize services. For example, Walmart, a leader in the retail industry, has successfully used AI technology to improve inventory management and supply chain processes, reducing operating costs and significantly improving the shopping experience for customers.

Gen AI could ultimately boost global GDP

As a result of these reassessments of technology capabilities due to generative AI, the total percentage of hours that could theoretically be automated by integrating technologies that exist today has increased from about 50 percent to 60–70 percent. The technical potential curve is quite steep because of the acceleration in generative AI’s natural-language capabilities. The modeled scenarios create a time range for the potential pace of automating current work activities. The “earliest” scenario flexes all parameters to the extremes of plausible assumptions, resulting in faster automation development and adoption, and the “latest” scenario flexes all parameters in the opposite direction. Based on a historical analysis of various technologies, we modeled a range of adoption timelines from eight to 27 years between the beginning of adoption and its plateau, using sigmoidal curves (S-curves).

Furthermore, traditional AI is usually trained using supervised learning techniques, whereas generative AI

is trained using unsupervised learning. That has also shed light on, and drawn

people to, generative AI technology that focuses on other modalities; everyone seems to be experimenting

with writing text, or making music, pictures, and videos using one or more of the various models that

specialize in each area. So, with many organizations already experimenting with generative AI, its impact on

business and society is likely to be colossal—and will happen stupendously fast.

Marketing has a rich tradition of decision making studies that investigate human cognitive biases (Dowling et al., 2020). Such knowledge can be fruitfully applied to gain rich insights on GenAI cognition (Binz and Schulz, 2023). Further, harnessing the full potential of GenAI requires proper prompting (Huang & Rust, 2023). Given the marketing field’s history of developing strategies to mitigate human biases in surveys (Hulland et al., 2018), we call for research to explore how these strategies could be adapted to calibrate prompts and enhance the quality of GenAI output. These initial studies aside, we argue that further research is necessary to examine the connection between GenAI’s objective parameters and human subjective perceptions of its output. Second, users can adjust the level of randomness (or creativity) in the output generated by modifying the temperature parameter.

• For instance, setting a top_p value to 0.2 means that the model will only select from those tokens that represent the top 20% of the probability mass for the next token.
• A huge amount of data must be stored during training, and applications require significant processing power.
• For example, much of the value of new vehicles comes from digital features such as adaptive cruise control, parking assistance, and IoT connectivity.
• Pharma companies typically spend approximately 20 percent of revenues on R&D,1Research and development in the pharmaceutical industry, Congressional Budget Office, April 2021.

One approach involves training an auxiliary generative model on non-copyrighted data and utilizing rejection sampling to reduce the likelihood of reproducing copyrighted material [35]. Alternatively, [4] suggests modifying generative models’ training objectives to avoid generating outputs that closely resemble copyrighted data. Yet another technique focuses on protecting unique artistic styles by incorporating adversarial perturbations into copyrighted images for model fine-tuning [33]. GenAI is the outcome of a renewed focus on self-supervised machine learning rather than the supervised learning approach that characterized much previous AI developments (Bommasani et al., 2021). In a supervised learning approach, during the training, machines learn by comparing model output against a given correct answer.

This observation aligns with the intuitive understanding that the AI developer’s contribution is foundational; without their computational input and expertise, it would be infeasible to generate any valuable content. The Shapley value has been suggested as a means to fairly distribute revenue in traditional sectors such as royalty agreements between music copyright holders and radio broadcasters [39]. The Shapley value has been used for data valuation where the utility function is the prediction accuracy of the machine learning model [9, 16].

This range implicitly accounts for the many factors that could affect the pace at which adoption occurs, including regulation, levels of investment, and management decision making within firms. Researchers start by mapping the patient cohort’s clinical events and medical histories—including potential diagnoses, prescribed medications, and performed procedures—from real-world data. Using foundation models, researchers can quantify clinical events, establish relationships, and measure the similarity between the patient cohort and evidence-backed indications. The result is a short list of indications that have a better probability of success in clinical trials because they can be more accurately matched to appropriate patient groups. Generative AI tools can enhance the process of developing new versions of products by digitally creating new designs rapidly. A designer can generate packaging designs from scratch or generate variations on an existing design.

It’s early days still, but use of gen AI is already widespread

We also modeled a range of potential scenarios for the pace at which these technologies could be adopted and affect work activities throughout the global economy. Generative AI tools can draw on existing documents and data sets to substantially streamline content generation. These tools can create personalized marketing and sales content tailored to specific client profiles and histories as well as a multitude of alternatives for A/B testing. In addition, generative AI could automatically produce model documentation, identify missing documentation, and scan relevant regulatory updates to create alerts for relevant shifts. Generative AI tools can facilitate copy writing for marketing and sales, help brainstorm creative marketing ideas, expedite consumer research, and accelerate content analysis and creation.

Among the first class of models to achieve this cross-over feat were variational autoencoders, or VAEs, introduced in 2013. VAEs were the first deep-learning models to be widely used for generating realistic images and speech. In summary, the application of generative AI is changing the operating model of the financial industry, from risk management to customer experience, all of which reflect its powerful data processing and prediction capabilities. Banking, retail, and professional services will account for a large share of spending on AI systems, demonstrating the urgent need for these industries to improve business efficiency and enhance competitiveness.

Any productivity increase that is not the result of changes in capital or labor inputs is measured as total factor productivity (TFP). Generative AI has the potential to revolutionize the entire customer operations function, improving the customer experience and agent productivity through digital self-service and enhancing and augmenting agent skills. The technology has already gained traction in customer service because of its ability to automate interactions with customers using natural language. Crucially, productivity and quality of service improved most among less-experienced agents, while the AI assistant did not increase—and sometimes decreased—the productivity and quality metrics of more highly skilled agents. This is because AI assistance helped less-experienced agents communicate using techniques similar to those of their higher-skilled counterparts.

By leveraging historical sales data, prescription patterns, epidemiology, and demographic data, forecasting becomes more accurate and improves the planning of new manufacturing sites. Generative AI is revolutionizing the biopharma industry, offering strategic opportunities to generate significant value if workflows and processes are consistently reinvented end-to-end. Enterprises across all sizes and industries, from the United States military to Coca-Cola, are prodigiously

experimenting with generative AI.

Such a holistic strategy makes sure that companies can maximize the benefits of intelligent technologies and achieve significant results for the patient, the entire organization, and the healthcare system. Generative AI is likely to have a major impact on knowledge work, activities in which humans work together

and/or make business decisions. At the very least, knowledge workers’ roles will need to adapt to working in

partnerships with generative AI tools, and some jobs will be eliminated. History demonstrates, however, that

technological change like that expected from generative AI always leads to the creation of more jobs than it

destroys. In marketing, generative AI can automate the integration and analysis of data from disparate sources, which

should dramatically accelerate time to insights and lead directly to better-informed decision-making and

faster development of go-to-market strategies. Marketers can use this information alongside other

AI-generated insights to craft new, more-targeted ad campaigns.

He has written about BMW’s erratic strategy for electric vehicles, Walmart’s controversial decision to close its Store 8 innovation lab, and Goldman Sach’s failed efforts to build a consumer bank. Goldman’s estimate that 47GW of additional capacity is needed to support data center growth between now and 2030. This may be an unsustainable burden on the electric grid, especially with climate change and restriction on carbon emissions imposing greater restraints over time. It clearly speeds up software coding and it will be easier for people to draft documents quickly.

The SRS could be manipulated by a malicious copyright owner creating multiple copies of their data. While replication-robust solution concepts have been explored [12], they focused on the impact on Shapley values rather than ratios under replication. Developing a mechanism robust against such manipulation is an important direction for future work.

• For one thing, mathematical models trained on publicly available data without sufficient safeguards against plagiarism, copyright violations, and branding recognition risks infringing on intellectual property rights.
• The tool was rolled out in phases, creating quasi-experimental evidence on its causal effects.
• In effect, people can

  converse with, and learn from, text-trained generative AI models in pretty much the same way they do with

  humans.

• This uniformity demonstrates the SRS framework’s ability to avoid disproportionate revenue distribution.

A possible explanation for this finding is that GPT had already seen those highly rated ideas (or, at least, similarly appropriate ideas) during the training. Thus, providing further examples of good ideas in the prompt is redundant, as GPT has already memorized what humans consider to be appropriate. With generative AI, organizations can build custom models trained on their own institutional knowledge and

intellectual property (IP), after which knowledge workers can ask the software to collaborate on a task in

the same language they might use with a colleague. Such a specialized generative AI model can respond by

synthesizing information from the entire corporate knowledge base with astonishing speed.

As a result, generative AI is likely to have the biggest impact on knowledge work, particularly activities involving decision making and collaboration, which previously had the lowest potential for automation (Exhibit 10). Our estimate of the technical potential to automate the application of expertise jumped 34 percentage points, while the potential to automate management and develop talent increased from 16 percent in 2017 to 49 percent in 2023. Our previously modeled adoption scenarios suggested that 50 percent of time spent on 2016 work activities would be automated sometime between 2035 and 2070, with a midpoint scenario around 2053. In the lead identification stage of drug development, scientists can use foundation models to automate the preliminary screening of chemicals in the search for those that will produce specific effects on drug targets. To start, thousands of cell cultures are tested and paired with images of the corresponding experiment. Using an off-the-shelf foundation model, researchers can cluster similar images more precisely than they can with traditional models, enabling them to select the most promising chemicals for further analysis during lead optimization.

Among the dozens of music generators are AIVA, Soundful, Boomy, Amper, Dadabots, and MuseNet. Although

software programmers have been known to collaborate with ChatGPT, there are also plenty of specialized

code-generation tools, including Codex, codeStarter, Tabnine, PolyCoder, Cogram, and CodeT5. Bloomberg announced BloombergGPT, a chatbot trained roughly half on general data about the

world and half on either proprietary Bloomberg data or cleaned financial data. It can perform simple tasks,

such as writing good article headlines, and propriety tricks, like turning plain-English prompts into the

Bloomberg Query Language required by the company’s data terminals, which are must-haves in many financial

industry firms. Some groups are concerned

that it will lead to human extinction, while others insist it will save the world. However, here are some important risks and concerns that business leaders implementing AI

technology must understand so that they can take steps to mitigate any potential negative consequences.

But human supervision has recently made a comeback and is now helping to drive large language models forward. AI developers are increasingly using supervised learning to shape our interactions with generative models and their powerful embedded representations. Encoder-only models like BERT power search engines and customer-service chatbots, including IBM’s Watson Assistant. Encoder-only models are widely used for non-generative tasks like classifying customer feedback and extracting information from long documents. In a project with NASA, IBM is building an encoder-only model to mine millions of earth-science journals for new knowledge. The rise of deep learning, however, made it possible to extend them to images, speech, and other complex data types.

Many the estimates for savings are based on extrapolating savings from these tasks across the entire economy. Plus, Acemoglu points out, most of the solutions we have in trial today are based on automating relatively simple or at least repetitive tasks. If we increase the complexity of the task, introducing a need to understand context and situation, then the chances that we will be able to apply gen AI fall rapidly. As organisations grapple with AI’s disruptive potential, the key lies in creating customer value while preparing for larger shifts. This cautious yet progressive approach allows firms to tackle disruption while maximising insights into AI’s evolving landscape, positioning them for future success in an AI-driven world.

Another open question is handling copyrighted data when owners are unable or unwilling to negotiate agreements, particularly with numerous owners each having small datasets. In such cases, our approach could be combined with methods for generating lawful content [35]. We have made preliminary progress toward this by adapting the concept of permission structure from cooperative game theory [10] to model the scenario where the AI developers and copyright owners jointly train a generative AI; see the supplementary materials for details.

To streamline processes, generative AI could automate key functions such as customer service, marketing and sales, and inventory and supply chain management. Traditional AI and advanced analytics solutions have helped companies manage vast pools of data across large numbers of SKUs, expansive supply chain and warehousing networks, and complex product categories such as consumables. In addition, the industries are heavily customer facing, which offers opportunities for generative AI to complement previously existing artificial intelligence. For example, generative AI’s ability to personalize offerings could optimize marketing and sales activities already handled by existing AI solutions. Similarly, generative AI tools excel at data management and could support existing AI-driven pricing tools.

In this section, we highlight the value potential of generative AI across business functions. Rich is a freelance journalist writing about business and technology for national, B2B and trade publications. While his specialist areas are digital transformation and leadership and workplace issues, he’s also covered everything from how AI can be used to manage inventory levels during stock shortages to how digital twins can transform healthcare. Beyond energy, developers and hyperscalers will need to do more to reassure customers over the environmental cost of AI in the near future. The $ immense water consumption of data centers, for example, will likely define conversations around technology and the environment in the coming years.

Combining generative AI with all other technologies, work automation could add 0.5 to 3.4 percentage points annually to productivity growth. However, workers will need support in learning new skills, and some will change occupations. If worker transitions and other risks can be managed, generative AI could contribute substantively to economic growth and support a more sustainable, inclusive world. The survey results show that AI high performers—that is, organizations where respondents say at least 20 percent of EBIT in 2022 was attributable to AI use—are going all in on artificial intelligence, both with gen AI and more traditional AI capabilities. These organizations that achieve significant value from AI are already using gen AI in more business functions than other organizations do, especially in product and service development and risk and supply chain management. These organizations also are using AI more often than other organizations in risk modeling and for uses within HR such as performance management and organization design and workforce deployment optimization.

If the past eight months are any guide, the next several years will take us on a roller-coaster ride featuring fast-paced innovation and technological breakthroughs that force us to recalibrate our understanding of AI’s impact on our work and our lives. Given the speed of generative AI’s deployment so far, the need to accelerate digital transformation and reskill labor forces is great. Labor economists have often noted that the deployment of automation technologies tends to have the most impact on workers with the lowest skill levels, as measured by educational attainment, or what is called skill biased.

In the communicating stage, firms interact with customers to persuade them to change their behavior and adopt the firm’s offering (Castaño et al., 2008). After consumers buy the firm’s novel offering, firms continue interacting with customers to keep them engaged beyond economic transactions (Blut et al., 2023; Pansari & Kumar, 2017). This engagement enables firms to access key consumer resources (e.g., knowledge stores, creativity) (Harmeling et al., 2017) that offer further creative input to the innovation process, thus constituting a continuous cycle, as illustrated in Fig. Oracle plans to embed generative AI services

into business platforms to boost productivity and efficiency throughout a business’s existing processes,

bypassing the need for many companies to build and train their own models from the ground up.

If the data source is very small in size, the royalty share of the owner would be mostly insignificant and, worse, noisy due to the stochastic nature of training AI models [36]. The utility (2.1) or (2.2) can be interpreted as the total compensation all members of S𝑆Sitalic_S collectively deserve for providing their data to train the generative AI model. The next step is to determine the payoff for each individual copyright owner, based on the utilities of all possible combinations of data sources. The Shapley value is a solution concept in cooperative game theory that offers a principled approach to distributing gains depending on the utility of every combination of players as a coalition. It is the only payment rule satisfying several important economic properties (see the supplementary materials for details) [34, 29] and has gained popularity in data valuation for machine learning models [9, 16]. Since different foundation models are trained on different data and have different architectures, and also since the same released model can be updated over time, we report the model used and time of the test.

Language transformers today are used for non-generative tasks like classification and entity extraction as well as generative tasks like translation, summarization, and question answering. More recently, transformers have stunned the world with their capacity to generate convincing dialogue, essays, and other content. Autoencoders work by encoding unlabeled data into a compressed representation, and then decoding the data back into its original form. “Plain” autoencoders were used for a variety of purposes, including reconstructing corrupted or blurry images.

The first is to use the Monte Carlo method to approximate the Shapley value [16, 15, 26, 38, 3, 25, 23, 37]. This technique is specially tailored to the case of a large number of copyright owners in the coalition. The second approach is to train a model by fine-tuning it from another model that is trained on a smaller subset of data. Hence, one can approximate models trained on different subsets of data sources by training the model with only one pass through the entire training data.