思维还是机器：探索AI主持及何时使用

MIND OR MACHINES Exploring AI moderationand when to use it Ajay BangiaJim LeggBetsy GeorgitonManuel Garcia-Garcia, PhD. and nuanced interactions. What if the keyto improving AI moderator bots lies not justin technology, but in understanding humanpsychology? This paper explores howborrowing from the world of psychologycan elevate AI moderators, creating moreengaging and insightful interactions. At Ipsos, we focus on exploringissues comprehensively anduncovering subtle nuances. OurESOMAR award-winning paper‘Empathy or Emptiness1,’highlightedthat while AI moderator bots offeradvantages, they fall short in empathy At Ipsos, we champion the uniqueblend of Human Intelligence (HI)and Artificial Intelligence (AI)to propel innovation and deliverimpactful, human-centric insightsfor our clients. Our Human Intelligence stems fromour expertise in prompt engineering,data science, and our unique, highquality data sets – which embedscreativity, curiosity, ethics, andrigor into our AI solutions, poweredby our Ipsos Facto Gen AI platform.Our clients benefit from insightsthat are safer, faster and groundedin the human context. Helpful co-moderators,yet human stars shine brighter Over the past two years, dozens of AI moderation platforms have entered the marketplace.Ipsos has assessed and worked with a large percentage of them, identifying clearstrengths and weaknesses. But AI moderator botshave significant weaknesses: AI moderator botsexcel at: •Engagement:AI moderator botsdemonstrate impressive engagementresults, outperforming unmoderatedplatforms like bulletin boardsor digital diaries. Their ability torecognize, recall and referencespecific details makes respondentsfeel heard and acknowledged,especially compared to untrainedmoderators. •Weak improvisation and realtime adaptation:Bots struggle withspontaneous, probing questions thatarise during conversations. Trainedhuman moderators on the other handcan rely on intuition and deviate froma rigidly structured discussion guide.This runs the risk of adhering to arigid, pre-defined flow often ignoringthe nugget that could sit just aroundthe corner. #IpsosHiAi Introduction Industry reactions to AI moderatorbots reveal a dichotomy of perspective.Sceptics argue these bots will neverencapsulate the ineffable essence ofgenuine human moderation. Conversely,proponents argue that these innovationscould dissolve the barriers betweenqualitative and quantitative researchmethodologies. This is because AImoderator bots can efficiently reacha much larger sample size, similar tothat of quantitative research, withoutthe constraint of a human moderatorbeing present for every interaction.This evolution towards “conversationalresearch” promises a“quant-litative”approach, blending qualitative depth withquantitative reach. In the rapidly evolving landscape ofqualitative research, traditional methodslike in-depth interviews and focus groupsincreasingly struggle with scalability andresource demands as they are constrainedby moderator bandwidth, respondentavailability and budget constraints. AImoderator bots, powered by sophisticatedLarge Language Model (LLM) systems,have the potential to fundamentallytransform this dynamic. These AImoderators claim to mimic and enhancehuman moderation through text or voice-based interfaces, offering revolutionaryscalability. •Always-on access:Constantlyavailable without breaks, ready tointeract at respondents’ convenience. •Gaps and inconsistencies:While AI seeks answers to researchobjectives, it cannot observeunarticulated nuances or probeinconsistencies, failing to providea holistic understanding of humanbehavior. •Scalability:Hosting hundreds ofinterviews simultaneously becomesfeasible. •Limited rapport building:Bots haverestricted ability to build meaningfulconnections with respondents. language model that recalls previousdialogue. However, long prompts candegrade the model’s effectiveness.Therefore, experts recommend anarchitecture that uses a reflectionmodule to condense conversationinto summary notes, highlighting keyparticipant insights and using them toformulate the next prompt. This allowsthe AI agent to prompt the languagemodel with concise reflections insteadof the entire interaction. While these MAS score better than purelyusing an LLM-powered chat bot, they arestill inferior to human moderators. Arewe missing a trick here? Why are the AImoderator bots rated as poor, despitehaving sophisticated architecture?Does the answer to developing a betterAI moderator lie purely in technology,or does it demand insights from humanpsychology? While conducting research-on-research, expert Ipsos moderators in three differentcountries compared masked transcripts (unaware which were human or AI moderated) andrated AI-moderated interviews poorly on all parameters: Getting better by emulating human moderators Both researchers and clients preferspecific moderators based on theirexpertise in the topics, or due to uniquequalities or approaches. Investig