《Displayr：2024市場研究人員的AI全面指南：釋放生成式AI在市場研究中的潛力（英文版）（32頁）.pdf》由會員分享，可在線閱讀，更多相關《Displayr：2024市場研究人員的AI全面指南：釋放生成式AI在市場研究中的潛力（英文版）（32頁）.pdf（32頁珍藏版）》請在三個皮匠報告上搜索。

1、2024AI Essentialsfor Market ResearchersTransforming Market Research with Generative IntelligenceWhat is AI?AI Essentials for Market Researchers01Accuracy and hallucinationsEffective AI implementationOpportunities for AI in market researchContentsAI in marketresearchAI Essentials for Market Researche

2、rs02What is AI?Artificial Intelligence(AI)enables machines to mimic human intelligence.In this section,well explore AI fundamentals,the impact of generative AI,and what AI excels at todayessential knowledge for leveraging AI in market research.What is AI?The importance of generative AIWhat is AI goo

3、d at?What is AI?what is aiArtificial Intelligence(AI)is a broad field of computer science focused on creating systems capable of performing tasks that typically require human intelligence.These tasks include learning from experience,understanding and processing language,recognizing patterns,solving

4、problems,and making decisions.AI systems can range from simple rule-based systems to more complex machine learning models that can adapt and improve over time based on the data they encounter.In general terms,AI is about making machines“smart”by giving them the ability to simulate aspects of human c

5、ognition and behavior.Its applied in many areas,from automating repetitive tasks to enabling more complex functions like voice assistants or predictive analytics.AI doesnt necessarily have to be“generative”it can involve anything from simple automation to advanced decision-making processes.AI has a

6、long history of application in market research,and most advanced analysis techniques can be thought of as being AI(e.g.,regression,cluster analysis).The AI that has caught the attention of the world today is a special form of AI called generative AI,which is AI that can create high quality content.T

7、he two paragraphs at the beginning of the of this section were written by ChatGPT as was the image to the right.AI Essentials for Market Researchers03The importance of generative AIwhat is aiAI Essentials for Market Researchers04Most technologists regard generative AI as a fundamental advance in tec

8、hnology,on par with the steam engine,the telephone,computers,and the internet.Despite the name“generative AI”the magic isnt that the computers are generating content.Computers have been generating content for many years.For example,the various Toy Stories movies back in the 1990s.Whats magic about g

9、enerative AI is that computers can now quickly generate high quality content with minimal instructions.The first Toy Story movie took four years to make,with 27 animators,22 technical directors,61 filmmakers,and more than 25,000 storyboards.However,by asking a generative AI,Can you please create an

10、image with the characters in Toy Story 1?in a matter of seconds,you can get back something like this image.Can you please create an image with the characters in Toy Story 1GenerateAI Whitepaper01What AI is good at(no,its not reasoning or creativity)what is aiToday,AI has three core and related skill

11、s:Prediction and classificationTransformationSpeedWhile its common to read about how AI can now“reason”and“create,”such descriptions are overly simplistic and ultimately misleading.AI doesnt reason like a human.It makes many very un-human errors.As discussed below,AI is overconfident to a level beyo

12、nd the most optimistic narcist.Similarly,while generative AI impresses in that it can create anything,its impressive in the way that a dog playing a piano is impressive.People arent queuing to view generative AI art,nor are they listening to AI music.They may one day.But not today.The skill transfor

13、mationis newly acquired,and is changing the world.Question answering:the most immediately visible skill of modern generative AI tools like ChatGPT and Perplexity is their ability to answer both general and obscure questions.Surprisingly(to this author anyway),the question-answering skill is a conseq

14、uence of the AI being trained to predict the next word in a sequence,where the sequence is all the worlds information.Language tagging:recognizing and labeling key bits of information in spoken and written words.For example,products like Gong are used by sales teams around the world to analyze sales

15、 conversations,such as identifying names,key reasons for purchasing,problematic sales techniques.Diagnosis:such as medical diagnosis.While this sounds like reasoning,its in practice just a combination of pattern matching and chatting,where the app continually asks for information until the required

16、information is obtained.what is aiPrediction and classificationAI has always been used to perform the closely related tasks of prediction and classification.Common examples include:Segmentation:dividing things into parts.For example,?Image segmentation:recognizing people and other entities in photos

17、 and handwriting into letters and characters.?Clustering of data:grouping similar objects into groups(e.g.,market segmentation).Optical character recognition:classifying written symbols as different letters of the alphabet and other symbols.Forecasting:predicting future values in sequences of number

18、s,based on past values and related data.Anomaly detectionRecommendation systems,such as when Amazon recommends products to buy and Facebook feeds.Predicting the next word(s)in a sequenceAI Essentials for Market Researchers06WHAT IS aiTim Bock,CEO of DisplayrThe second core skill,which has only becom

19、e amazing in the past decade,is transformation:converting some things into others.Some examples:Compression:transforming lots of words into a smaller number of words(summarization).Text-to-image:transforming a phrase into an image.Style transfer:transforming words or images from one style into anoth

20、er.E.g.,?Text-to-code:transforming human language into computer code.This skill enables natural language user interfaces,by translating plain text into computer code,and then sending the resulting code to APIs.?Plain text to a sonnet in the style of Shakespeare?A photo into a painting in the style o

21、f Rembrandt?Deep fake generation of videos and voice.Enhancement:?Improving the quality of an image?Improving the quality of written?Noise reduction in audioEmbedding/encoding:Converting text to numbers.The easiest example of this is sentiment analysis,where text is represented as numbers,with posit

22、ive numbers indicating positive sentiment and negative numbers indicating negative sentiment.“AI handles the tedious stuff,so you can focus on the insights that count.”The third skill is speed.Previously,only humans could perform transformation.Today,AI can do it and does it vastly faster than human

23、s can.TransformationSpeedGenerateAI Essentials for Market Researchers07AI Essentials for Market Researchers08Accuracy and hallucinationsAI is powerful but not infallible.In market research,accuracy is critical,yet AI can sometimes hallucinate,generating incorrect or misleading information with high

24、confidence.This section explores why these errors happen,their impact,and how to manage them.The accuracy(hallucination)problemThe causes of hallucinationsThe accuracy(hallucination)problemACCURACY AND HALLUCINATIONSAI Essentials for Market Researchers09AI that can predict,classify,transform,and do

25、so quickly is near-magical.But the magic is not perfect.Generative AI often makes mistakes and isnt good at working out when this has happened.This is often referred to as hallucinations.This isnt a new problem.AI has always faced this issue,though it was previously known as prediction error,a term

26、used to describe when AI predictions do not match actual outcomes.As an example,I asked Who is Tim Bock?and it came back with:Now,as Im Tim Bock,I know that the above isnt quite right,so I said to ChatGPT,I dont believe theres a company called Q Research Software.Isnt there another name for his comp

27、any?You are correct.The company founded by Tim Bock is actually called Displayr.Displayr is a well-known software platform in the market research industry that provides tools for data analysis,reporting,and visualization.The platform is designed to simplify the process of working with complex data,p

28、articularly survey data,and is widely used by researchers and analysts for creating dashboards,performing statistical analysis,and generating insights.Hallucinations of varying magnitudes are super common.If you look carefully at the image of generative AI,youll see some very odd spelling,which is r

29、eproduced to the right.Tim Bock is an entrepreneur and the founder of Q Research Software,a company known for creating software solutions used in market research and data analysis.Q Research Software is designed to help researchers analyze survey data and automate complex statistical tasks,making it

30、 easier to extract insights from data.Tim Bock is recognized for his contributions to the market research industry,particularly in developing tools that simplify and enhance the process of data analysis.The causes of hallucinationsACCURACY AND HALLUCINATIONSAI Essentials for Market Researchers10As m

31、entioned earlier,hallucination has always existed in AI.It is caused by:Creating artificial intelligence technology is difficult.Even the oldest of the techniques thats now regarded as AI linear regression is fiendishly difficult to apply for interesting real-world problems,despite more than a hundr

32、ed years of development and use.Broadly speaking,there are three different sets of AI skills required to create AI:?The creation of a data set from which AI can learn.This issue is discussed in more detail in the next section.?The skill of training an AI using the available data.?Hooking up the AI t

33、o perform the predictions.This is discussed in more detail in When to use Fully Autonomous AI Agents.With AI talent being extremely expensive and short in demand,most AI initiatives are implemented with insufficient technical expertise to guarantee success,and it seems unlikely that this problem wil

34、l be easy to overcome soon.This one has become much more important in recent times.These threealways caused hallucinations.Insufficient technical expertise in the creationInadequate training dataAI being asked questions that are unanswerableAmbiguity in the questions being askedInsufficient technica

35、l expertise in the creationACCURACY AND HALLUCINATIONSAI Essentials for Market Researchers11AI can only work if the training data used when creating the AI permit the AI to learn in a way that avoids hallucination.To use some jargon relating to how humans learn:current AI can only be accurate if the

36、re is a kind learning environment.The classic example of a kind learning environment is learning to play the game of chess.There are a small number of types of pieces.Two colors.An 8-by-8 board.Only a small number of possible movements.And large numbers of books showing what to do in different scena

37、rios,targeted from beginnings through to grand masters.Google and IBM have demonstrated that when AI is trained on such information it becomes better at chess than the worlds best chess grand master.A wicked learning environment is where the training data is inadequate to permit the AI to learn.For

38、example:Sometimes what we want AI to do may just be impossible and there can be no data that can be collected to permit the AI to give accurate answers,ensuring that any answers it gives are necessarily hallucinations.Unanswerable questions fall into some broad but related buckets:Questions in searc

39、h of facts that cant exist.For example,what is a spell for turning lead into gold?Questions that are beyond the worlds expertise,such as whats a practical way of implement nuclear fusion?Questions that ignore the inherent unpredictability of the world,such as will it rain on the 21st of January 2099

40、,or what numbers will win in a lottery?TransformationInadequate training dataAI being asked questions that are unanswerable?There may not be enough data.?Key predictors may be missing.?The data may contain errors.?The data may be biased in some way,causing the model to learn the wrong things.?The da

41、ta may contain spurious relationships that permit the model to seem to predict well,but where the model wont predict well when applied to real world problems.ACCURACY AND HALLUCINATIONSAI Essentials for Market Researchers12TransformationAmbiguity in the questions being askedOften the questions(instr

42、uctions)given to AI are ambiguous,so AI must guess.Consider the prompt:Create an image in Pixar style of a cat drinking a milkshake with lots of chips.Draw it in wide aspect ratio.The image to the right was created by ChatGPT4o.In many ways its remarkable.But it clearly has not given a perfect answe

43、r.For example:?Coffee is spelled incorrectly on the window to the left of the cat.?The cream and cherry are on the cat.?The word chips appear on the wall and the sugar shaker.?I wanted the cat to have steak fries(which are called“chips”in the Australian English that I speak).Note that the ambiguity

44、is partly caused by the wording and by the AI having insufficient context.Both of these are complex to avoid in real-world applications.AI is changing market researchStay ahead with DisplayrQuickly find and share the story in data using software that thinksBook a demoAI Essentials for Market Researc

45、hers14Effective AI ImplementationSuccessfully implementing AI involves choosing between fully autonomous agents or human-AI collaboration.Autonomous agents work independently but need careful error management,while collaboration combines AI efficiency with human judgment.Selecting the right approach

46、 maximizes AIs impact.The two ways of successfully implementing AIWhen to use fully autonomous AI agentsWhen and how to use human+AI collaboration(copilots)The two ways of successfully implementing AIEffective AI ImplementationAI Essentials for Market Researchers15There are two basic designs for suc

47、cessfully using AI in business contexts:The name AI,and much of the popular discourse about AI,foreshadow a future where artificial intelligences operate like people,working as slaves,overlords or just normal people.The jargon for such AI is fully autonomous AI agents.Much of AI R&D is inspired by t

48、he desire to create useful fully autonomous AI agents.The alternative to fully autonomous AI agents is where humans collaborate with AI,with the AI doing tasks that it can perform more efficiently than a human,while the human provides expert judgement that the AI cannot yet perform.The most well-kno

49、wn example of this is human air pilots that collaborate with their autopilots.Many AI tools are today branded as copilots in recognition of this symbiotic way of working with AI.Fully autonomous AI agentsHuman+AI collaboration(copilot)Fully autonomous AI agentsHuman+AI collaboration(copilot)When to

50、use fully autonomous AI agentsEffective AI ImplementationAI Essentials for Market Researchers16Fully autonomous AI agentsFully autonomous AI agents are the desire that most people have for AI.We want to be able to delegate our most painful and tedious work to a fully autonomous AI agent in the confi

51、dence that it will do as well as us,or better,much more quickly and efficiently.The core idea of a fully autonomous AI agent is that it can operate independently of humans.To do this,its necessary that it hallucinates at an acceptable rate.For some problems,this may mean that AI does not hallucinate

52、 or only has trivial hallucinations.Self-driving cars come to mind as a such a problem.Often initial AI models dont hallucinate at an acceptable rate and techniques need to be used to reduce hallucinations,including:Choosing models that have been trained on data very similar to the real-world applic

53、ations.For example,if creating chat bots,language translation,or sentiment analysis,many of the current foundational models,like ChatGPT have been trained on appropriate data and may perform well enough.1234Prompt engineering,which refers to asking questions of the AI in a way that maximizes the cha

54、nce that the AI will understand the question.For example,providing example questions and answers.This can also be automated.For example,image generation models produce better predictions when prompts are first improved by large language models.Choosing algorithms that are more appropriate to the dat

55、a.For example,with data sets of only a few hundred observations sets general linear models are often best,with data sets of millions of cases deep learning are often best,and models like random forest and XG-Boost can be better in-between.Creating new data sets that better reflect real-world applica

56、tion.For example,to develop AI that is good at performing mathematics,data sets have been created by collating large numbers of mathematical proofs from academic publications.Effective AI ImplementationAI Essentials for Market Researchers17Fine-tuning existing models on high quality data sets.For ex

57、ample,creating a customer support chatbot is often done by taking an existing large language model,and then re-training it to predict a curated set of support tickets.567Use databases to retrieve facts For example,stock prices,the weather,company databases.This is known retrieval augmented generatio

58、n(RAG).Send queries off to non-AI software that will do a better job than AI.For example,rather than analyze data directly,ChatGPT writes code in the Python language,runs the code,and returns the results of the code.Often problems are sufficiently difficult that even with all the resources in the wo

59、rld it is not possible to get get the models sufficiently accurate.Elon Musk forecast that Tesla self-driving cars would be available 2018,but we seem no closer today than we were then.Driving turns out to be a wicked learning environment.Some parts of driving,such as parking and maintaining your la

60、ne on a motorway are kind learning environments.But often the things that cause crashes are wicked learning environments,which is why self-driving cars arent quite there yet.When we cant reduce AI errors(hallucinations)to a level where an autonomous AI agent is safe enough to use,the only viable alt

61、ernative is human+AI collaboration.When and how to use human+AI collaboration(copilots)Effective AI ImplementationAI Essentials for Market Researchers18Its long-established that for many problems humans collaborating with AI do a better job than humans or AI on their own.For example:?Doctors working

62、 with AI tools like IBMs Watson are superior at many types of diagnosis than doctors without AI.?Grand masters using AI are better at chess than AI working on its own.?Human pilots working with autopilots.This analogy is so popular its led to a raft of tools designed for human AI collaboration being

63、 branded as copilots.Effective human AI collaboration requires a few things to be in place:The human needs to be able to give the AI instructions,so that the human can work out which things to use the AI to do and which things to be do.For example,modern cars allow drivers to choose when to use auto

64、mated navigation,autopark,and traffic-aware cruise control.123Humans need to be able to review the results of the AI.There are three aspects to this:A.The AI needs to present the results in a way that the human can review them.When using traffic-aware cruise control the users still needs to be able

65、to look through the windows and mirrors to verify that the AI is doing a good job.B.The human needs to have the skill to check the results.With autopiloting,a qualified drivers is still required in most situations.C.The human needs to have the will to check them.The driver must be awake and being di

66、ligent so that they can take over.Human+AI collaboration(copilot)The Humans need to be able to correct errors.With autopilots,this means that the driver has full access to all the normal controls(steering,breaks,etc.).Effective AI ImplementationAI Essentials for Market Researchers19Autopilots in car

67、s are one of many different possible user interfaces for putting these three things in place.A completely different mechanism is ChatGPT,where the user“chats”with the app,continually reviewing the results and giving feedback.The various code-writing copilot tools take yet another approach,offering c

68、ode completion suggestions that a user can accept,modify or discard.AI Essentials for Market Researchers20AI in market researchAI is essential in market research for prediction and classification,but only a third use generative AI.Full automation remains unrealistic due to errors,necessitating human

69、-AI collaboration for accurate and efficient results.Current usage of AI in market researchWhats not about to changeCurrent usage of AI in market researchAI IN MARKET RESEARCHAI Essentials for Market Researchers21AI has long been used in market research,with many market research problems being well

70、solved by traditional AI tools that are good at prediction and classification(e.g.,market segmentation,factor analysis,driver analysis).Although it is now two years since ChatGPT hit the mainstream and became the most rapidly adopted technology in history,only a third of market researchers currently

71、 use modern generative AI.Weve observed the same result in polls in North America,Europe,and Asia.Its hard to be sure of the level of measurement error in this data.For example,are most users of automated text categorization(coding)tools(e.g.,in Q,Displayr),aware that these are based on large langua

72、ge models.Not using itOnly one-third ofmarket researchers areusing generative AIExperimentingExperimentingProductivity gains of0 to 2%Productivity gains of3%to 5%9%17%8%12%54%Productivity gains ofmore than 5%Whats not about to changeAI IN MARKET RESEARCHAI Essentials for Market Researchers22Before l

73、ooking at what has and will change,its interesting to look at the things that arent about to change.These are:?Fully autonomous AI agents conducting studies and writing reports?Commercial-grade tabulation?Storytelling?Updating analyses and reporting with new data.?Advanced analysis.?Commercial-grade

74、 software creation.?Automating the design and layout of documents.Fully autonomous AI agents conducting studies and writing reportsWed all like an AI that we can delegate studies to:Siri,conduct a concept test and tell me whether to launch or not.However,no matter how much we want it,its a long way

75、off.The first problem with delegating the entire research process to AI is that hallucination will be at times very expensive,making fully autonomous AI agents unwise.No amount of time saved will account for testing the wrong concept,accidentally incurring a bill for a sample size of 1 million,or th

76、e AI randomly forgetting to add the word“dont”to the sentence“launch the new product”.For the foreseeable future we will collaborate with AI,working out which bits to delegate,checking and correcting the work of AI.Commercial-grade tabulationCreating tables of the type widely used by commercial mark

77、et researchers is also some way off.At a superficial level this seems an easy problem to automate,and it is easy to automate the creation of many tables.However,the devils in the detail.There are a host of very nuanced things that need to be done to create useful tables,and its not straightforward t

78、o verify that they have been done without doing them yourself.For example:?Working out when to rebase out missing values,impute missing values,or whether to record missing values as a yes or no.?Reversing scales.?Checking skips and piping answers from earlier questions.?The workflows in modern cross

79、tab software are so efficient that its hard to see how AI can,for the time being anyway,make this approach much more efficient.AI IN MARKET RESEARCHAI Essentials for Market Researchers23StorytellingAs discussed in The Causes of Hallucinations,there are four causes of hallucination:?Insufficient tech

80、nical expertise in the creation of the AI?Inadequate training data?AI being asked questions that are unanswerable?Ambiguity in the questions being askedEach of these is a problem for storytelling.Furthermore,its not clear that even collaboration with AI is a fruitful way forward with storytelling in

81、 the near term,as the problem of results being hallucinated(e.g.,27%satisfaction turning into 72%satisfaction)adds significant risk.Nevertheless,as discussed in Automated report writing,storytelling by AI is going to happen soonish.Updating analyses and reporting with new dataFor many years,the big

82、non-easy win for market research has been automating the updating of reporting,with the main use cases being:?Updating trackers with new waves of data.?Segment level reporting,where a standard set of reporting is performed for different segments or countries.,?Updating draft analyses and reports cre

83、ated on early interviews with the full data set.?Standard market research products,where a common questionnaire is used every time the study is conducted(e.g.,concept tests,ad tests).Historically this was largely done by code in SPSS Syntax,and this workflow still exists today with R and Python bein

84、g added to the languages.Today,its more common that such analysis and reporting is automatically updated without the need for code,by tools like Q,Displayr,and Tableau.This workflow cannot be efficiently replaced with AI for the time being,as:?Hallucinations cause results to be incorrect.?The nature

85、 of the workflow makes it near-impossible to identify and correct errors.For example,if the AI replaces(1-3)with(-13)in code,and a result changes subtly,how can anybody know?Furthermore,as the process is already successfully automated for many studies,the switch to AI would not even gain any time ev

86、en if the hallucination problem were solved.AI IN MARKET RESEARCHAI Essentials for Market Researchers24Advanced analysis and statistical testingMany researchers want AI to take over the job of more advanced methods like conjoint,MaxDiff,driver analysis,and segmentation,as well as statistical testing

87、.This is highly unlikely to occur soon.All the problems that cause hallucinations are present in this area,making autonomous agents infeasible.There are extremely few published examples of the workflow used to(correctly)perform advanced analysis nor even correct application of statistical testing to

88、 weighted data,so theres no data to train models.Often people try and use advanced analyses to solve problems that arent answerable(e.g.,the impact of a marketing slogan).Typically briefs for advanced analysis are too vague(e.g.,what really does“importance”mean)?And all the people with sufficient ex

89、pertise know of these three problems so wont even attempt to create AI for advanced analysis.Even more problematic is that collaboration between AI and humans is likely infeasible as the people that want AI to automate advanced analysis and statistical testing are generally the people that dont unde

90、rstand how to do it,so dont have the ability to either check or correct the results of AI.Commercial-grade software creationWhile the hucksters of AI like to claim that AI is about to replace engineers for creating software,there is no good data to suggest that this is true in the near-term.There ar

91、e some great benefits provided for Writing snippets of code,but the technology is currently only a relatively small time saver for more skilled software engineers rather than something that can replace their work.The reasons are essentially the same as those listed in the previous section on Advance

92、d analysis and statistical testing.Automating the design and layout of documentsWhile many AI technologies have been developed for writing presentations,such as beautiful.ai and Canva,the technology is too immature for commercial market research.If you are a student performing an assignment or you a

93、re pitching a new business,the technology is a real time saver.But if you want to present market research data in a way thats easy to digest with rich visualizations revealing all the key information,the technology isnt ready yet.Nevertheless,there are some easy wins in terms of Image creation.AI Es

94、sentials for Market Researchers25Opportunities for AI in Market ResearchAI in market research enhances text categorization,qualitative analysis,data cleaning,ideation,writing,and image creation,all requiring human-AI collaboration for accuracy.Future opportunities like chatbots and automated report

95、writing promise to revolutionize data insights and presentations.Top opportunities for AI in market research todayMain opportunities for AI in market research tomorrowTop opportunities for AI in market research todayOPPORTUNITIES FOR AI IN MARKET RESEARCHAI Essentials for Market Researchers26Based o

96、n the queries that we get from customers and our own understanding of the technology,the biggest opportunities for using AI in market research today are,in rough order of the size of the opportunity:?Text categorization(aka coding)?Analyzing qualitative transcripts?Data cleaning and tidying?Ideation

97、 and secondary research?General writing improvement/assistance?Executive summaries and other summaries?Image creation?Writing snippets of code?Auto-completion OPPORTUNITIES FOR AI IN MARKET RESEARCHAI Essentials for Market Researchers27Text categorization(aka coding)Converting the text collected fro

98、m open-ended questions into numeric data,commonly known as coding in market research,text categorization,and topic modeling in AI,has long been one of the most arduous parts of market research.As little as 20 years ago,this was entirely done by humans reading through all the data and manually assign

99、ing text to categories.In more recent years,larger studies have started to simple automated querying tools(e.g.,searching through text data and counting the number of mentions of specific combinations of words).The ever-reducing costs of online interviewing and the widespread availability of social

100、media data has massively increased the amount of data that can be categorized.The ever-reducing time frames and cost pressures have all but made the traditional manual approach increasingly impractical.Text categorization is the first and biggest unambiguous win from the modern generative AI technol

101、ogy.While it sounds like an exaggeration,the reality is that today AI is always cheaper and often better than humans at categorizing text data.Analyzing qualitative transcripts,recordings,and videoFor the same reasons as AI is good for text categorization,it is also very effective for analyzing qual

102、itative data.And for the same reasons,its best done using human+AI collaboration,using tools like looppanel,Dovetail,NVIVO,ATLAS.ti,and ChatGPT.The productivity gain that AI provides to qualitative research is a bit smaller than to text categorization due to need for qualitative researchers to moder

103、ate the qualitative research,which means that a human has typically listened to all the research so the qualitative analysis is more confirmatory than a true source of insight much of the time.While most market researchers want fully autonomous AI agents to both automatically create categories(codef

104、rames)and automatically assign new data to this code frame,at the time of writing the accuracy of AI is not sufficient for this to be practical,and instead its necessary for humans and AI to collaborate.In the chapter on The Causes of Hallucinations identified four causes of hallucination:?Insuffici

105、ent technical expertise in the creation of the AI?Inadequate training data?AI being asked questions that are unanswerable?Ambiguity in the questions being askedThe first three of these are not particularly difficult to solve for text categorization,as its a problem that has been studied for many yea

106、rs and large language models are particularly well suited to the problem.The real challenge is that there is not usually a correct way of categorizing text.Or,more accurately,the best categorization depends on context,and AI cannot typically guess the correct context.For this reason,human+AI collabo

107、ration is the appropriate way to perform text categorization.OPPORTUNITIES FOR AI IN MARKET RESEARCHAI Essentials for Market Researchers28Data cleaning and tidyingAI is also very useful for cleaning and tidying data,including:?Tidying variable labels in large data files.This is automated in Displayr

108、.?Translating variables in different languages.?Summarizing overly verbose verbatims.?Identifying dirty text responses,so that the data can be deleted(e.g.,cuss words,non-answers).Executive summaries and other summariesThe use of AI to create executive and other summaries is increasingly popular.Ima

109、ge creationImage generation AI,like Midjourney and DALL-E,makes it easy for people to generate relevant and interesting images for presentations and dashboards.The examples shown in the first half of this eBook are examples.Ideation and secondary researchIt is likely that the use of ChatGPT for idea

110、tion and secondary research is the most widely used role of AI in market research,with considerable time saving available if using AI to:?Design studies.?Create proposals.?Write questionnaires(e.g.,suggest questions,get brand lists).?Create stimulus.?Summarize previous research.General writing impro

111、vement/assistanceMany researchers report that AI helps a lot in terms of improving writing,both by creating drafts that can be tweaked,and by improving documents created by humans.Upload dataBrowseUploadSelect fileOPPORTUNITIES FOR AI IN MARKET RESEARCHAI Essentials for Market Researchers29Writing s

112、nippets of codeThere are many situations where quantitative researcher needs to write small amounts of code or syntax(i.e.,snippets of code).For example,if using the R language,a user may want to convert age in years into categories,and use a prompt like In the R language,write a snipped of code tha

113、t takes a variable called age,measured in years,and converts it to a categorical variable with categories:Under 18 18 to 30 31 to 50 50 or more Just return the code.No additional information please.Chat GPT returns the following result:age_category-cut(age,breaks=c(-Inf,17,30,50,Inf),labels=c(“Under

114、 18”,“18 to 30”,“31 to 50”,“50 or more”)As a basic workflow this is vastly more efficient than the way snippets of code used to be written,which involved researchers saving snippets from project to project,asking for help from colleagues or using search engines.Auto-completion in softwareAnother fea

115、ture,which has started to occur in some market research software(e.g.,Quantilope),is using AI to autocomplete when using software(e.g.,filling in response options automatically for questions when programming a questionnaire).Main opportunities for AI in market research tomorrowOPPORTUNITIES FOR AI I

116、N MARKET RESEARCHAI Essentials for Market Researchers30As AI improves,and software vendors become better at using AI,all the things described in the previous chapter will continue to improve,and some new functionality will start to appear,such as:We anticipate that,alongside traditional reporting me

117、thods like PowerPoint presentations and dashboards,a new form of reportinganswerbotswill become increasingly widespread.This will allow consumers of the research to query the data(i.e.,natural language querying)in a chat interface,with the AI returning charts,tables,and commentary that addresses the

118、 question of interest.Although this functionality is available via ChatGPT and other commercial tools today it works very poorly for survey data.However,this is only due to nobody building survey-specific versions,rather than there being some particularly difficult aspect of this.Much like using cha

119、tbots to analyze data,automated report writing has yet to be reach the stage of being safe to use.The current challenge is that the tools for doing this work are either chatbots,or,fully autonomous agents.Neither of these allows users to readily check or correct results,but neither of these problems

120、 seem insurmountable,and it seems likely that in the next few years AI will be able to auto-write reports,including finding and crafting the storytelling.Chatbots for research usersAutomated report writingAutomated report writingChatbots for research usersAI is changing market researchStay ahead with DisplayrQuickly find and share the story in data using software that thinksBook a demo