KPMG+生成式人工智能模型的商业风险和潜在回报-15页-WN7.pdf

《KPMG+生成式人工智能模型的商业风险和潜在回报-15页-WN7.pdf》由会员分享，可在线阅读，更多相关《KPMG+生成式人工智能模型的商业风险和潜在回报-15页-WN7.pdf（15页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、1Generative AI models the risks and potential rewards in business 2023 Copyright owned by one or more of the KPMG International entities.KPMG International entities provide no services to clients.All rights reserved.Generative AI models the risks and potential rewards in business What the rise of Ch

2、atGPT,DALLE 2,Bard et al.could mean for your organization.KPMG InternationalApril 20232Generative AI models the risks and potential rewards in business 2023 Copyright owned by one or more of the KPMG International entities.KPMG International entities provide no services to clients.All rights reserve

3、d.Generative AI models highlight the power of technology.They have the potential to make us more productive and can make what we do easier in some respects.However,these models come with risk implications that all organizations and individuals should be aware of.That said,we cant ignore these models

4、.Theyre rapidly becoming part of our daily personal and professional lives.We need to determine how to embrace them but safely.”Lisa HeneghanGlobal Chief Digital OfficerKPMG InternationalThe cover image and imagery throughout this paper were designed using DALLE 2,an AI art generator that creates im

5、ages based on textual descriptions.Image prompts for the cover image were:fluid abstract,wavy column of blue and purple,splash,drops,purple background.While DALLE 2 generates compelling visual content,it has not been trained on KPMG brand guidelines.Also,it has neither the human expertise nor ingenu

6、ity to understand KPMGs brand positioning.As a result,these images are considered off-brand and are used for illustrative purposes only,with special permission from KPMG Global Brand.3Generative AI models the risks and potential rewards in business 2023 Copyright owned by one or more of the KPMG Int

7、ernational entities.KPMG International entities provide no services to clients.All rights reserved.ContentsExecutive summary 04Market overview 05What are generative AI models?06How generative AI models work 06Potential opportunities and use cases 07Current considerations 10What does the future hold?

8、13How KPMG can help 14Contacts 153Generative AI models-the risks and potential rewards in business 2023 Copyright owned by one or more of the KPMG International entities.KPMG International entities provide no services to clients.All rights reserved.4Generative AI models the risks and potential rewar

9、ds in business 2023 Copyright owned by one or more of the KPMG International entities.KPMG International entities provide no services to clients.All rights reserved.We believe that generative artificial intelligence(AI)models have the potential to transform businesses through automating and executin

10、g certain tasks with unprecedented speed and efficiency.This is particularly true when human expertise and ingenuity is paired with deep understanding of how to use these programs and effectively harness their capabilities.However,it will take time and human expertise to unlock their full potential

11、in a way thats responsible,trustworthy and safe.If youre considering using generative AI,its important to establish a set of internal processes and controls for everyone in your organization to follow.In this report,we cover potential use cases and opportunities,as well as what to consider if youre

12、thinking about using generative AI applications,such as ChatGPT,within your organization.13579246810The most common generative AI solutions can roughly be divided into five categories:content generators,information extractors,smart chatbots,language translators and code generators.ChatGPT is a chatb

13、ot trained on human instructions.Its initial underlying large language model,GPT-3.5,had 175 billion parameters and was trained with more than 1 million datasets or 500 billion tokens(words or word fragments).GPT-3.5 was not connected to the internet and was trained on data from up to September 2021

14、.GPT-4,OpenAIs new large multimodal model,evolved from its earlier large language model.Generative AI takes data inputs or parameters to learn and build knowledge.Unless you explicitly restrict the application provider from doing so,that data may then be used to answer a prompt from someone else,pos

15、-sibly exposing an organizations proprietary information to the public.Depending on the application,you may also be signing over your copyrights.Referring to the respective terms and conditions may give you an idea of what happens with user-inputted data.Copying AI-produced information or code into

16、any deliverable or product may constitute copyright or other intellectual property infringement.This could potentially cause your organization legal and reputational harm.Creating safe usage guidelines within your organization is key to helping ensure proper and effective use of generative AI applic

17、ations.Your organization should also upskill its people,as the human in the loop brings unique insights and understanding that generative AI alone cant replicate.Generative AI models can summarize articles,draft emails and produce images and videos.Trained by humans,some generative AI models have th

18、e conversational skills to,for example,answer follow-up questions,admit mistakes,challenge incorrect assumptions and filter or reject inappropriate requests.Generative AI models have use across various business functions,from IT,human resources and operations,to finance,audit,legal and marketing.Sui

19、table applications include drafting proposals,developing and testing code,and extracting and summarizing complex information.Depending what you use generative AI for and how you implement it,your activities could expose intellectual property or trade secrets and open your organization up to fraud ri

20、sk.Its important to be vigilant and make sure your organization isnt using AI in a way that contravenes applicable laws(including privacy laws),client agreements or professional standards.We expect both open source and boutique versions of generative AI will continue to be integrated into many commo

21、n applications,systems and processes,ranging from internet browsers to AI-connected technology that organizations license,such as cloud-based software and instant messaging programs.KPMG takes a responsible approach to designing,building and deploying AI systems in a safe,trustworthy and ethical man

22、ner.This approach helps companies accelerate value for consumers,organizations and society.Here are 10 things you should know about generative AI:01Executive summary5Generative AI models the risks and potential rewards in business 2023 Copyright owned by one or more of the KPMG International entitie

23、s.KPMG International entities provide no services to clients.All rights reserved.According to research and consulting firm Gartner,by 2025,30 percent of outbound messages from large organizations will be synthetically generated.1 In the AI Risk Survey Report,conducted by KPMG in the US in September

24、2022,85 percent of respondents expect an increase in the use of AI and predictive analytics models.Additionally,in the 2022 KPMG in the US Technology Survey,half of respondents said theyve seen ROI from investments in AI technology.Generative AI models captured attention in summer 2022 when an AI-ge

25、nerated image won an art contest.2 In November they were in the spotlight again following the launch of ChatGPT.However,it was during a January 2023 World Economic Forum session,when Microsoft Chairman and Chief Executive Officer Satya Nadella said the“golden age of AI”is underway,3 that the buzz ar

26、ound ChatGPT really began to intensify,generating many questions from and conversations with KPMG member firm clients.Training these models requires large amounts of venture capital,human effort and computing power.OpenAI,the creator of ChatGPT,received US$1 billion from Microsoft4 and another multi

27、year,multibillion dollar investment from the company at the start of 2023,5 and Google6 and Meta7 have created generative AI models of their own.Given the range of possible applications,an entire industry is being built on making generative AI models useful.Generative AI applications can be roughly

28、divided into five categories:content generators,information extractors,smart chatbots,language translators and code generators:Content generators:Where generative pretrained transformer tools generate content such as blog posts,emails,social media posts,images,web copy and ads.Information extractors

29、:These applications can create short-and long-form summaries of news articles,blog posts,legal documents and more.Some companies use them to develop and analyze legal documents.Smart chatbots:Companies are increasingly using smart chatbots as consumer assistants.The chatbots interact in a conversati

30、onal way and can answer follow-up questions,admit mistakes,challenge incorrect ideas and reject inappropriate requests.Language translators:Multilingual tools that can translate many languages.They have the potential to build entire website interfaces,including translation sites.Code generators:Gene

31、rative AI models can convert natural text inputs into code snippets or applications.With a basic description or small program function input,these models can produce code in various programming languages,and identify and fix bugs.02Market overview1 Gartner,7 Technology Disruptions That Will Complete

32、ly Change Sales,October 10,2022.https:/ is a registered trademark and service mark of Gartner,Inc.and/or its affiliates in the U.S.and internationally and is used herein with permission.All rights reserved.2 h t t p s:/w w h t t p s:/w w w.weforum.org/press/2023/01/satya-nadella-says-ai-golden-age-i

33、s-here-and-it-s-good-for-humanity4 h t t p s:/ h t t p s:/ h t t p s:/blog.google/technology/ai/bard-google-ai-search-updates/7 h t t p s:/ AI models the risks and potential rewards in business 2023 Copyright owned by one or more of the KPMG International entities.KPMG International entities provide

34、 no services to clients.All rights reserved.Generative AI models are designed to produce content based on a clear set of inputs and rules.The most buzzed about application of a generative AI model of late is ChatGPT,a chatbot trained on human instructions created by San Francisco research lab OpenAI

35、.8 As of 14 March 2023,ChatGPT Plus subscribers were able to use GPT-4,a large multimodal model(LMM)that accepts both image and text inputs and generates text outputs.9 On 23 March 2023 OpenAI launched plugins for ChatGPT,including its own web browsing plugin.This means ChatGPT can now access certai

36、n third-party sources and databases.10 ChatGPT stands for Chat(conversation-based)G(generative)P(pretrained)T(transformer).It was finetuned using reinforced learning from human feedback,where a reward model representing human preference is trained to help make outputs sound more human,prevent undesi

37、rable responses and try to avoid hallucinations(making up facts).ChatGPT was created as a large language model(LLM)and has since evolved into a large multimodal generative AI application.This means the application can now accept image and text inputs,not just text inputs as was previously the case.C

38、ombined with a neural network model that uses unsupervised learning to predict outcomes,this type of generative model can determine the most likely linguistic patterns and relationships between content its already absorbed.The“large”refers to the amount of data the models are based on,as well as the

39、 size of the models themselves,and involves training them with a massive collection of publicly accessible electronic documents.For instance,when it was released in 2022,ChatGPT had 175 billion parameters(a value that controls the behavior of the machine learning model the greater the parameters,the

40、 greater the models ability to analyze).It was initially trained with more than 1 million datasets or 500 billion tokens(words or word fragments),including from Wikipedia and The New York Times.To put this into perspective,the average human speaks 860.3 million words in their lifetime,11 making this

41、 collection or“corpus,”in AI terminology equivalent to 300 years worth of language.04How generative AI models work8 https:/ h t t p s:/ 10 C h a t G P T p l u g i ns(o p e n ai.c om)11 h t t p s:/openlibrary.org/books/OL3502128M/The_joy_of_lexGenerative AI refers to artificial intelligence that can

42、generate content rather than simply analyze or act on existing data.Generative AI models,such as GPT-4,are built and trained on a collated set of data.They can be generalists or specialists built on predefined data collections and are designed to produce output that helps realize certain human-direc

43、ted requests.Some models can,for example,predict the next word based on previous phrases or the next image based on descriptions of images that came before.This training enables the quick generation of original content,including text,images,video and code.With a reduced need for human resources,some

44、 companies expect to be able to produce content faster and at lower cost,giving them opportunities to create new kinds of content that were previously too expensive or time consuming.This fundamentally changes human-machine interaction and opens myriad potential use cases.This predictive capability

45、enables the models to analyze.For example,they can be used to identify documents that are about a topic described by inputted text.03What are generative AI models?7Generative AI models the risks and potential rewards in business 2023 Copyright owned by one or more of the KPMG International entities.

46、KPMG International entities provide no services to clients.All rights reserved.ChatGPTs meteoric rise in popularity is in part because anyone can use it,even those who dont have technical backgrounds.Its rapid growth in users 100 million as of February 202314 is a sign of peoples eagerness to use th

47、e technology.And the more users a chatbot has,the better trained its underlying AI can get.ChatGPT has the potential to transform businesses by automating and executing language-based tasks with unprecedented speed and efficiency.LMMs can be deployed to help with a wide range of tasks.They can be mo

48、dified to summarize and classify legal documents,respond to consumer questions,assist expert advisors,and generate engineering and architectural drawings.They can act as a starting point for human inspiration,providing ideas that can be transformed into fresh and creative thinking.This makes them su

49、itable to help generate business reports,marketing pitches and code for software applications.Generative AI models may also have applications across IT,audit,human resources,operations and many more business functions besides.As you explore these use cases,keep in mind that despite the many opportun

50、ities generative AI has to offer,they arent risk free.05Potential opportunities and use cases12 h t t p s:/w w h t t p s:/ h t t p s:/w w basic version of ChatGPT isnt connected to the internet and was trained on online material up to September 2021 meaning its knowledge isnt up to date.Newer implem