人工智能的亲密接触：人类与语言模型互动方式日益人性化

BYLee RainieDirector,Imagining theDigital Future Center Close encounters of the AI kind: The increasingly human-like way people areengaging with language models Key datapoints in this survey 52%of American adults now use AI large language models (LLMs) like ChatGPT, Gemini,Claude and Copilot 51%of LLM users say their main activity on the tool is forpersonal, informal learning, twiceas many as use LLMsfor work purposes (24%); among workers, 46% say their main purpose ispersonal and 32% say their main purpose is for work 65%of users have hadspoken conversationsinvolving back-and-forth interactions with anLLM, including34%who regularly do this at least several times a week 49%of LLM users think the models they useare smarter than they are 76%of LLM users are satisfied with the performance of LLMs, including25%who are verysatisfied 54%say their use of LLMs has improved theirproductivitya lot/somewhat;42%say their usehas improved theircreativity Human-like traits•39%say the main LLM they use shows the capacityto think and reasonat least some of the time.•32%say it seems to have asense of humor.•25%say it acts like itmakes moral judgmentsabout right and wrong.•24%say it seems like it expresseshope.•17%say it seems to respondsarcastically.•11%say it seems to expresssadness. Darker impacts•50%of LLM users say they have feltlazywhen they use LLMs. •35%have felt they werecheating.•35%have felt frustrated orconfused.•33%have felt they were becomingtoo dependenton LLMs for answers rather thanthinking through things themselves.•23%say they made asignificant mistake or bad decisionby relying on informationgenerated by LLMs.•21%have feltmanipulatedby the model. Insights into our survey The material reported here comes from a nationally representative survey of 500 users of large languagemodels (LLM) conducted between Jan. 21-23, 2025. It has a margin of error of +/- 5.1 percentage points. AnLLM user is someone who answered “yes” to this question: “Do you ever use artificial intelligence (AI) largelanguage models (also known as LLMs or generative AI), such as ChatGPT, Google’s Gemini, Anthropic’sClaude, Microsoft’s Copilot or an open source LLM?” Close encounters of the AI kindExecutive summary Half of Americans now use artificial intelligence (AI) large language models like ChatGPT, Gemini, Claude,and Copilot. Since the launch of ChatGPT on Nov. 30, 2022, the spread of LLM usage in this countryrepresents one of the fastest, if not the fastest, adoption rates of a major technology in history. The growth and spread of these AI systemsin the U.S. population are especially strikingfor their diversity. Younger, well-educated,relatively wealthy, and employed adults aresomewhat more likely than others to beusing LLMs now. Yet, it is also the case thathalf of those living in households earning lessthan $50,000 (53%) use the tools. Moreover,Hispanic adults (66%) and Black adults (57%)are more likely than White adults (47%) to beLLM users. Half of American adults use AI largelanguage models such as ChatGPT, Gemini,Copilot or Claude % of U.S. adults who ever use (AI) large language models Key usage patterns What LLMs people use:ChatGPT stands outas the most commonly used LLM.Fully 72%of LLM users say they have used ChatGPT,while 50% have used Google’s Gemini, 39%have used Microsoft’s Copilot, 20% have usedMeta’s LLaMa, 12% have used xAI’s Grok and9% have used Anthropic’s Claude and 9% saythey have used other LLMs. How they use LLMs:34% of LLM users saythey do so at leastonce a day, including 10%who say they use the tools “almost constantly.”Another 18% say they use LLMs several times aweek. Why they use LLMs:Personal uses, ratherthan work-related uses, are the main activity.Half (51%) say the main purpose for their LLMuse is personal, informal learningon theirown. Some 24% say forworkpurposes; 11% sayforschoolpurposes. Fully two-thirds (68%)have ever used LLMs for informal learningon their own; 60% have ever used them toamuse or entertain themselves and 52% haveever used them for work activities. Amongthose who work, 46% say their main purposeis personal, compared with 32% who say themain purpose is work. Special uses of LLMs:The survey covered a host of particular uses of LLMs and found that: Two-thirds usethem likesearch engines; about half use them forbrainstorming ideasand summarizing documents; athird use them to create presentations andplan things like trips; a quarter use them for planningsocialgatheringsandwriting computer code. And 23% have used LLMs to look upwhat the models say aboutpeoplethey know, while 18% have looked up what a model says about themselves. Close encounters of the AI kindExecutive summary Human-like interactions with LLMs A notable share of LLM users have had human-like social interactions with these tools. Indeed, 38% ofusers believe it is very or somewhat likely that in the next decade LLMs will formdeep relationships withhumans. Right now, LLM users say this: •65% have ha