2025年生成式人工智能的巨大经济潜力报告（英文版）

The Enormous Economic Potential of Generative AIOctober 2025Joseph BriggsGoldman, Sachs & Co.+1 212-902-2163joseph.briggs@gs.comSenior EconomistInvestors should consider this report as only a single factor in making their investment decision. For Reg AC certiÞcation and other important disclosures, see the Disclosure Appendix, or go to www.gs.com/research/hedge.html. c45a43530f604d12bcb9a82b5aa6b9f62Source: Goldman Sachs Global Investment Research.An Overview of Generative AI1. Generalized rather than specialized - wider use cases and more complementary innovations2. Generative rather than descriptive - can produce original results indistinguishable from human output3. Approachable rather than technical - can interface via complex and contextual natural languageThree key changes:Step 1: Training Data to Neural NetworkStep 2: Neural Network to AI OutputStep 4: ApplicationsPrevious ML methods:Data trained on specialized databases for specific purposes (e.g., make statistical predictions about election results, answer questions about biomedical literature, etc.)Generative AI:Data trained on large, generalized databases (i.e., the entire internet); thus 1) wider range of use cases and 2) more easily able to spawn complementary innovations with specialized use cases ("deepening of AI").Previous ML methods:Models generate statistical predictions based on relationships in training data.Generative AI:Models seek to generate new information that is indistinguishable from human data. Achieved via the introduction of a second "discriminative" neural net, which evaluates the output of the primary "generative" neural net for authenticity relative to human output. This "adversarial neural networks" approach forces the generative network to revise its output and learn to consistently "fool" the discriminative network.Previous ML methods:- Text classification- Facial and image recognition- Statistical prediction and inferenceGenerative AI:- Answer complex textual questions with human-like language and structure- Create original images, graphics, and video based on user queries- Generate and explain code which can be used for other programming and data science applicationsStep 3: AI Output to Human InterfacePrevious ML methods:Users must use specific code or syntax to make narrow requests based on the model's intended function.Generative AI: Use of large language models (LLMs) allows for advanced natural language processing (NLP) that incorporates context in larger swaths of text, enabling a wider variety of requests and a more accessible interface for human-AI interaction.For the exclusive use of NEIL.WANG@TROWEPRICE.COMc45a43530f604d12bcb9a82b5aa6b9f6O*NET data: Our Baseline Assumes 13 Task CategoriesUp to Difficulty Level of 4 Could Be AutomatedSource: Goldman Sachs Global Investment Research.3 c45a43530f604d12bcb9a82b5aa6b9f64Source: Goldman Sachs Global Investment Research.Two-Thirds of Current Occupations Could be Partially Automated by AIFor the exclusive