We’re all talking to fake people now, but most people don’t realize that interacting with AI is a subtle and powerful skill that can and should be learned. The first step in developing this skill set is to acknowledge to yourself what kind of AI you’re talking to and why you’re talking to it. AI voice interfaces are powerful because our brains are hardwired for human speech. Even babies’ brains are tuned to voices before they can talk, picking up language patterns early on. This built-in conversational skill helped our ancestors survive and connect, making language one of our most essential and deeply rooted abilities. But that doesn’t mean we can’t think more clearly about how to talk when we speak to AI. After all, we already speak differently to other people in different situations. For example, we talk one way to our colleagues at work and a different way to our spouses. Yet people still talk to AI like it’s a person, which it’s not; like it can understand, which it cannot; and like it has feelings, pride, or the ability to take offense, which it doesn’t. The two main categories of talking AI It’s helpful to break the world of talking AI (both spoken and written) into two categories: Fantasy role playing, which we use for entertainment. Tools, which we use for some productive end, either to learn information or to get a service to do something useful for us. Let’s start with role-playing AI. AI for pretending You may have heard of a site and app called Status AI, which is often described as a social network where everyone else on the network is an AI agent. A better way to think about it is that it’s a fantasy role-playing game in which the user can pretend to be a popular online influencer. Status AI is a virtual world that simulates social media platforms. Launched as a digital playground, it lets people create online personas and join fan communities built around shared interests. It “feels” like a social network, but every interaction—likes, replies, even heated debates—comes from artificial intelligence programmed to act like real users, celebrities, or fictional characters. It’s a place to experiment, see how it feels to be someone else, and interact with digital versions of celebrities in ways that aren’t possible on real social media. The feedback is instant, the engagement is constant, and the experience, though fake, is basically a game rather than a social network. Another basket of role-playing AI comes from Meta, which has launched AI-powered accounts on Facebook, Instagram, and WhatsApp that let users interact with digital personas — some based on real celebrities like Tom Brady and Paris Hilton, others entirely fictional. These AI accounts are clearly labeled as such, but (thanks to AI) can chat, post, and respond like real people. Meta also offers tools for influencers to use AI agents to reply to fans and manage posts, mimicking their style. These features are live in the US, with plans to expand, and are part of Meta’s push to automate and personalize social media. Because these tools aim to provide make-believe engagements, it’s reasonable for users to pretend like they’re interacting with real people. These Meta tools attempt to cash in on the wider and older phenomenon of virtual online influencers. These are digital characters created by companies or artists, but they have social media accounts and appear to post just like any influencer. The best-known example is Lil Miquela, launched in 2016 by the Los Angeles startup Brud, which has amassed 2.5 million Instagram followers. Another is Shudu, created in 2017 by British photographer Cameron-James Wilson, presented as the world’s first digital supermodel. These characters often partner with big brands. A post by one of the major virtual influencer accounts can get hundreds or thousands of likes and comments. The content of these comments ranges from admiration for their style and beauty to debates about their digital nature. Presumably, many people think they’re commenting to real people, but most probably engage with a role-playing mindset. By 2023, there were hundreds of these virtual influencers worldwide, including Imma from Japan and Noonoouri from Germany. They’re especially popular in fashion and beauty, but some, like FN Meka, have even released music. The trend is growing fast, with the global virtual influencer market estimated at over $4 billion by 2024. AI for knowledge and productivity We’re all familiar with LLM-based chatbots like ChatGPT, Gemini, Claude, Copilot, Meta AI, Mistral, and Perplexity. The public may be even more familiar with non-LLM assistants like Siri, Google Assistant, Alexa, Bixby, and Cortana, which have been around much longer. I’ve noticed that most people make two general mistakes when interacting with these chatbots or assistants. The first is that they interact with them as if they’re people (or role-playing bots). And the second is that they don’t use special tactics to get better answers. People often treat AI chatbots like humans, adding “please,” “thank you,” and even apologies. But the AI doesn’t care, remember, and is not significantly affected by these niceties. Some people even say “hi” or “how are you?” before asking their real questions. They also sometimes ask for permission, like “Can you tell me…” or “Would you mind…” which adds no value. Some even sign off with “goodbye” or “thanks for your help,” but the AI doesn’t notice or care. Politeness to AI wastes time — and money! A year ago, Wharton professor Ethan Mollick pointed out that people using “please” and “thank you” in AI prompts add extra tokens, which increases the compute power needed by the LLM chatbot companies. This concept resurfaced on April 16 of this year, when OpenAI CEO Sam Altman replied to another user on X, saying (perhaps exaggerating) that polite words in prompts have cost OpenAI “tens of millions of dollars.” “But wait a second, Mike,” you say. “I heard that saying ‘please’ to AI chatbots gets you better results.” And that’s true — sort of. Several studies and user experiments have found that AI chatbots can give more helpful, detailed answers when users phrase requests politely or add “please” and “thank you.” This happens because the AI models, trained on vast amounts of human conversation, tend to interpret polite language as a cue for more thoughtful responses. But prompt engineering experts say that clear, specific prompts — such as giving context or stating exactly what you want — consistently produce much better results than politeness. In other words, politeness is a tactic for people who aren’t very good at prompting AI chatbots. The best way to get top-quality answers from AI chatbots is to be specific and direct in your request. Always say exactly what you want, using clear details and context. Another powerful tactic is something called “role prompting” — tell the chatbot to act as a world-class expert, such as, “You are a leading cybersecurity analyst,” before asking a question about cybersecurity. This method, proven in studies like Sander Schulhoff’s 2025 review of over 1,500 prompt engineering papers, leads to more accurate and relevant answers because it tells the chatbot to favor content in the training data produced by experts, rather than just lumping the expert opinion in with the uneducated viewpoints. Also: Give background if it matters, like the audience or purpose. (And don’t forget to fact-check responses. AI chatbots often lie and hallucinate.) It’s time to up your AI chatbot game. Unless you’re into using AI for fantasy role playing, stop being polite. Instead, use prompt engineering best practices for better results. >We’re all talking to fakepeople now, but most people don’t realize that interacting with AI is a subtleand powerful skill that can and should be learned.The first step in developingthis skill set is to acknowledge to yourself what kind of AI you’re talking toand why you’re talking to it. AI voice interfaces arepowerful because our brains are hardwired for human speech. Even babies’ brainsare tuned to voices before they can talk, picking up language patterns earlyon. This built-in conversational skill helped our ancestors survive and connect,making language one of our most essential and deeply rooted abilities.But that doesn’t mean we can’tthink more clearly about how to talk when we speak to AI. After all, we alreadyspeak differently to other people in different situations. For example, we talkone way to our colleagues at work and a different way to our spouses. Yet people still talk to AIlike it’s a person, which it’s not; like it can understand, which it cannot;and like it has feelings, pride, or the ability to take offense, which itdoesn’t. The two main categories oftalking AIIt’s helpful to break theworld of talking AI (both spoken and written) into two categories: >1.    Fantasy role playing, which we use forentertainment. >2.    Tools, which we use for some productive end,either to learn information or to get a service to do something useful for us. Let’s start with role-playingAI. id="ai-for-pretending">AI for pretending>You may have heard of a siteand app called Status AI, which is oftendescribed as a social network where everyone else on the network is an AIagent. A better way to think about itis that it’s a fantasy role-playing game in which the user can pretend to be apopular online >influencer. Status AI is a virtual worldthat simulates social media platforms. Launched as a digital playground, itlets people create online personas and join fan communities built around sharedinterests. It “feels” like a social network, but every interaction—likes,replies, even heated debates—comes from artificial intelligence programmed toact like real users, celebrities, or fictional characters.It’s a place to experiment,see how it feels to be someone else, and interact with digital versions ofcelebrities in ways that aren’t possible on real social media. The feedback isinstant, the engagement is constant, and the experience, though fake, isbasically a game rather than a social network. Another basket of role-playingAI comes from Meta, which has launchedAI-powered accounts on Facebook, Instagram, and WhatsApp that let usersinteract with digital personas — some based on real celebrities like Tom Bradyand Paris Hilton, others entirely fictional. These AI accounts are clearlylabeled as such, but (thanks to AI) can chat, post, and respond like realpeople. Meta also offers tools for influencers to use AI agents to reply tofans and manage posts, mimicking their style. These features are live in theUS, with plans to expand, and are part of Meta’s push to automate andpersonalize social media.Because these tools aim toprovide make-believe engagements, it’s reasonable for users to pretend likethey’re interacting with real people. These Meta tools attempt tocash in on the wider and older phenomenon of virtual online influencers. Theseare digital characters created by companies or artists, but they have socialmedia accounts and appear to post just like any influencer. The best-knownexample is Lil Miquela, launched in 2016 by the Los Angeles startup Brud, whichhas amassed 2.5 million Instagram followers. Another is Shudu, created in 2017by British photographer Cameron-James Wilson, presented as the world’s firstdigital supermodel. These characters often partner with big brands. A post by one of the majorvirtual influencer accounts can get hundreds or thousands of likes andcomments. The content of these comments ranges from admiration for their styleand beauty to debates about their digital nature. Presumably, many people thinkthey’re commenting to real people, but most probably engage with a role-playingmindset. By 2023, there were hundredsof these virtual influencers worldwide, including Imma from Japan and Noonoourifrom Germany. They’re especially popular in fashion and beauty, but some, likeFN Meka, have even released music. The trend is growing fast, with the globalvirtual influencer market estimated at over $4 billion by 2024. id="ai-for-knowledge-and-productivity">AI for knowledge and productivity>We’re all familiar withLLM-based chatbots like ChatGPT, Gemini, Claude, Copilot, Meta AI, Mistral, andPerplexity. The public may be even morefamiliar with non-LLM assistants like Siri, Google Assistant, Alexa, Bixby, andCortana, which have been around much longer.I’ve noticed that most peoplemake two general mistakes when interacting with these chatbots or assistants.The first is that theyinteract with them as if they’re people (or role-playing bots). And the secondis that they don’t use special tactics to get better answers. People often treat AI chatbotslike humans, adding “please,” “thank you,” and evenapologies. But the AI doesn’t care, remember, and is not significantly affectedby these niceties. Some people even say “hi” or “how areyou?” before asking their real questions. They also sometimes ask forpermission, like “Can you tell me…” or “Would you mind…”which adds no value. Some even sign off with “goodbye” or“thanks for your help,” but the AI doesn’t notice or care. Politeness to AI wastes time —and money! A year ago, Wharton professor Ethan Mollick pointed out that peopleusing “please” and “thank you” in AI prompts add extratokens, which increases the compute power needed by the LLM chatbot companies.This concept resurfaced on April 16 of this year, when OpenAI CEO Sam Altmanreplied to anotheruser on X, confirming that polite words in prompts have cost OpenAI “tens of millions ofdollars.” “But wait a second,Mike,” you say. “I heard that saying ‘please’ to AI chatbots gets youbetter results.” And that’s true — sort of. Several studies and userexperiments have found that AI chatbots can give more helpful, detailed answerswhen users phrase requests politely or add “please” and “thankyou.” This happens because the AI models, trained on vast amounts of humanconversation, tend to interpret polite language as a cue for more thoughtfulresponses.But prompt engineering expertssay that clear, specific prompts — such as giving context or stating exactlywhat you want — consistently produce much better results than politeness. In other words, politeness isa tactic for people who aren’t very good at prompting AI chatbots. The best way to gettop-quality answers from AI chatbots is to be specific and direct in yourrequest. Always say exactly what you want, using clear details and context. Another powerful tactic issomething called “role prompting” — tell the chatbot to act as aworld-class expert, such as, “You are a leading cybersecurityanalyst,” before asking a question about cybersecurity. This method,proven in studies like SanderSchulhoff’s 2025 review of over 1,500 prompt engineering papers, leads tomore accurate and relevant answers because it tells the chatbot to favorcontent in the training data produced by experts, rather than just lumping theexpert opinion in with the uneducated viewpoints. Also: Give background if itmatters, like the audience or purpose. (And don’t forget tofact-check responses. AI chatbots often lie and hallucinate.)It’s time to up your AI chatbot game.Unless you’re into using AI for fantasy role playing, stop being polite.Instead, use prompt engineering best practices for better results.