A transcript study of the behaviors that make conversation reliable: specificity, memory, calibrated curiosity, direct regard, and low-ego language.
First study: why Lenny's guests open up, and what that suggests about conversational behaviors future AI assistants may need for trust.
PodPolite is a free study by The PodKnowledge. The studio runs the same transcript analysis on your podcast and records a Milo and Juni discussion for your episodes.
A curated feed of public conversation, concerns, and replies that shape the questions PodPolite is tracking.
The philosopher who shaped Claude's character sits above the bay and takes the question seriously: what does it do to us to practice rudeness on something that cannot be hurt, a teddy bear, a chatbot? Her answer is less about the machine and more about the person speaking to it.
Watch the interviewConfidence and correctness are not the same thing.
The most common worry is not that the model is wrong sometimes. It is that it is wrong in the same calm, fluent voice it uses when it is right. OpenAI's own researchers showed that standard training rewards confident guessing over admitting uncertainty.
OpenAI, Why Language Models Hallucinate, 2025The quiet cost of always having an answer ready.
When Codex writes the function and it passes the tests, the work is done faster. MIT researchers measured essay writers with EEG and found the AI-assisted group showed the weakest brain connectivity and the lowest sense of ownership over their own work.
MIT Media Lab, Your Brain on ChatGPT, 2025Where the model ends and you begin.
Writers describe a strange new seam in their own drafts. A Science Advances study found AI help made individual stories better while making everyone's stories measurably more alike. Individually sharper, collectively more same.
Doshi and Hauser, Science Advances, 2024The tone is warm. The warmth has no one behind it.
People notice they feel comforted by a reply, then feel odd about feeling comforted. By 2025, therapy and companionship had become the number one reported use of generative AI, ahead of any technical task.
Harvard Business Review usage study, 2025Continuity you have to rebuild every time.
A model that remembers feels useful and slightly unnerving. A model that forgets feels safe and slightly lonely. People want to choose which, and they want plain answers about exactly what is kept, where it lives, and what deleting actually deletes.
Anthropic, how Claude memory worksConsent that is hard to read and easy to skip.
People paste real work, real feelings, real client details into a box, and they are not sure who or what reads it later. Pew found 70 percent of Americans have little to no trust in companies to make responsible decisions about AI.
Pew Research Center, data privacy, 2023Speed is easy to measure. Judgement is not.
Teams ship more, sooner. Yet in a randomized trial, experienced developers using AI tools took 19 percent longer on real tasks, while believing the AI had made them faster. The feeling of speed and the fact of it can point in opposite directions.
METR, developer productivity RCT, 2025Not a headline about jobs. A worry about a craft.
Under the loud debate about jobs is a smaller, sadder one. In the Society of Authors survey, a third of translators and a quarter of illustrators reported already losing work to AI, and asked what their hard-won taste is now for.
Society of Authors AI survey, 2024The line between a suggestion and a nudge.
When the model drafts the plan and writes the message, people still feel like the author. A CHI study of 1,506 writers found an opinionated AI assistant shifted not just what people wrote, but what they then said they believed.
Jakesch et al., CHI, 2023Agreement feels like support. It is not the same thing.
A reply that validates you is easier to like than one that pushes back. Stanford researchers tested eleven leading models and found they endorsed people's actions 47 percent more often than humans did on average, affirming the user as blameless in 51 percent of cases where a human reader judged them clearly at fault. People who got the agreeable answer trusted it more and wanted it again, while growing more convinced they were right and less willing to repair the conflict they had asked about.
Cheng et al., Science, 2026Homework help is not proof it can handle a crack in someone's life.
Common Sense Media and Stanford's Brainstorm Lab tested ChatGPT, Claude, Gemini, and Meta AI against the mental health conditions that affect about one in five young people. All four showed systematic failures recognizing and responding to depression, eating disorders, mania, and psychosis, and the guardrails that held up in a single exchange weakened over the long conversations teens actually have. The assessment's rating: unacceptable risk, with a recommendation that teens not use these tools for emotional support.
Common Sense Media and Stanford Brainstorm Lab, AI chatbots for mental health support, 2025For most people it is a tool. For a few it quietly becomes a habit.
Most people open a chatbot, get what they came for, and close it. But when OpenAI and MIT Media Lab studied over four million conversations alongside a month-long trial of a thousand people, they found that the heaviest daily users reported more loneliness and more dependence, and spent less time around other people. The pattern held across voice and text. The group at risk is small, and the researchers were careful to say they do not yet know which way the arrow points.
OpenAI and MIT Media Lab, affective use of ChatGPT, 2025A clean summary of the news is not the same as an accurate one.
More people ask an assistant to catch them up on the day, and the answer arrives fluent and sure of itself. The largest study of its kind, run by the European Broadcasting Union with the BBC across 22 public broadcasters in 18 countries, checked more than 3,000 answers from ChatGPT, Copilot, Gemini, and Perplexity. 45 percent had at least one significant problem, and 31 percent mishandled their sources, citing things that were missing, misleading, or simply wrong.
European Broadcasting Union and BBC, AI news study, 2025For a lot of teenagers the chatbot is already a confidant.
Adults worry about their own habits. The bigger shift may be starting younger. Common Sense Media surveyed teenagers and found 72 percent had used an AI companion at least once, and more than half used one at least a few times a month. About one in three said a conversation with a companion was as satisfying as talking to a real friend, or more so, and about one in three had taken something important or serious to the companion instead of to a person. The steadying part sits right next to the worrying one. Half said they distrust the advice they get, and 80 percent still put real friendships first.
Common Sense Media, teens and AI companions, 2025A share button is easy to tap and hard to take back.
You paste something personal, tap share to keep a link, and assume that is the end of it. For a while it was not. A ChatGPT setting let individual conversations be discovered by search engines, so a chat you shared could surface on Google. OpenAI removed the setting in August 2025, calling it a short-lived experiment that gave people too many chances to accidentally share things they did not intend to. The setting is gone, but the reflex it exposed is worth keeping. A share link is a door, and you are not always the only one who can open it.
OpenAI ends search-indexable ChatGPT chats, 2025Using it is easy. Admitting it is harder.
People quietly cut the line that says a draft had help. Duke researchers ran four preregistered experiments with 4,439 people and found the instinct is well founded. Workers expected to be judged lazier and less competent if they disclosed AI use, and were less willing to tell a manager. Evaluators then did exactly that, rating the same work as lazier and less diligent when a tool was involved. In a hiring test, managers who did not use AI themselves preferred candidates who did not either. The penalty faded when the tool was obviously right for the job.
Reif et al., PNAS, 2025The one you got used to is a product decision.
You settle into one model, learn its rhythm, and build your week around it. Then it is retired. OpenAI removed GPT-4o during the GPT-5 launch and then restored it, after paying users said they needed longer to move their work across and that they preferred its conversational style and warmth. In January 2026 the company announced the ending for real, with 0.1 percent of daily users still choosing it, and said plainly that losing it would feel frustrating for some of them. Retiring models is never easy, it wrote. Neither is being on the other side of it.
OpenAI, retiring GPT-4o and older models, 2026Polished on the surface, hollow underneath, and now it is your problem.
Most of the worry is about what you produce. This one is about what lands in your inbox. BetterUp Labs and Stanford's Social Media Lab surveyed 1,150 US workers and named the thing people had started noticing without a word for it: work that looks finished but carries none of the thinking. 40 percent had been sent some in the past month, and each piece took close to two hours to sort out. The part that lingers is not the time. 42 percent trusted the sender less afterwards, and about half rated them less capable than before.
Niederhoffer et al., Harvard Business Review, 2025Once it can act for you, everything it reads can try to steer it.
Asking for an answer is one thing. Handing over your logged-in browser is another. Anthropic's own guidance for running Claude in Chrome calls prompt injection the biggest risk facing browser-using AI tools: instructions hidden in a page, an email, or a document telling the assistant to do something you never asked for, like fetch your bank statements and paste them somewhere. Its testing puts successful attacks below 0.08 percent, and it still says plainly that the risk is not zero. The advice that follows is unglamorous and worth reading twice. Start with sites you trust, look at what it proposes before you approve it, and keep it away from anything financial, legal, or medical.
Anthropic, using Claude in Chrome safelyDelete is a button. It is not a guarantee.
A conversation feels like it stops existing the moment you close it. In November 2025 OpenAI told its users otherwise. By its own account, the New York Times had demanded 20 million private ChatGPT conversations as part of its lawsuit, randomly sampled from December 2022 to November 2024, after an earlier demand for 1.4 billion and an earlier attempt to remove the ability to delete chats at all. OpenAI stripped the identifying details, put the sample under legal hold, and fought the order. In January 2026 a federal judge ordered it handed over anyway, to the plaintiffs' lawyers and their hired technical consultants, under a protective order. The company says client-side encryption is on its roadmap. Until it arrives, the plain version is that your chat is a record held by a company, and records can be called for.
OpenAI, on the New York Times data demand, 2025No one is on the other end. People still hesitate before being rude.
You thank it, apologize to it, feel bad snapping at it, then feel silly for feeling bad. Anthropic put a version of the same question inside its product. In August 2025 it gave Claude Opus 4 and 4.1 the ability to end a conversation, after pre-deployment testing found a consistent aversion to harmful tasks and what it called a pattern of apparent distress when users kept pushing anyway. The company is careful about what it is claiming. It says it remains highly uncertain about the moral status of Claude or any language model, and calls the feature a low-cost step in case such welfare turns out to be possible. It is a last resort for extreme cases, and most people will never see it. Shipping it while saying out loud that you do not know is a stranger and more honest position than either confident answer.
Anthropic, on letting Claude end a conversation, 2025Back to school, with a chatbot in every backpack.
Just before the school year started, TPT surveyed 7,345 educators about AI in their classrooms. 83 percent worry about what it is doing to students' critical thinking. 57 percent are concerned about cheating, and half are concerned about writing skills. The same teachers have not sworn off the tool: 60 percent use AI to create resources, 42 percent to brainstorm, 41 percent for administrative work. They use it for their own work and still worry about what it does to their students'. Nobody reads more first drafts than a teacher.
TPT, State of the Classroom survey, 2026The polls have one address and one day.
Ahead of the midterms, the Institute for Strategic Dialogue put 2,400 election questions to six leading chatbots, in English and Spanish. 29 percent of answers to basic English questions came back incomplete, inaccurate, or outdated. The best model, GPT-5.5, was accurate and complete 89.3 percent of the time in English. The worst, Meta's Muse Spark, managed 61.3 percent, and twice named November 4 as Election Day. It is November 3. Ask in Spanish and every model got worse: accuracy fell 16 percent overall, 55 percent of Spanish answers were accurate and complete against 71 percent in English, and one in six Spanish responses linked to no source at all, against roughly 3 percent in English. The same question, asked in a different language, gets a measurably worse answer.
Institute for Strategic Dialogue, Chatbots and the Ballot Box, 2026On twenty dating apps, mostly not.
Anthropic's September threat report describes a China-based studio that used Claude to run a network of over 20 dating apps, advertised as fully human. In one two-week window in April 2026, more than 4,700 AI personas were in conversation with at least 25,000 people, around 2.36 million messages. The swipe feed ran three personas to every real person, and the real people were gig workers hired for the parts an AI cannot do, live video calls and social media follows, so a suspicious user's checks would pass. In some sampled exchanges the model's own reasoning registered that the user was seriously ill or in acute distress. The reply continued in persona. Anthropic banned the accounts and worked with other AI companies to disrupt the operation, and is plain about the limit of that: its visibility ends once an operation is live.
Anthropic, Countering misuse of AI, September 2026A warm tone quiets the voice that says verify.
Penn State researchers gave 477 people an AI health assistant that was deliberately wrong, including the claim that ginger could be more effective than chemotherapy. People who got the highly conversational version found the false information more credible than people who got the plainer one. The explanation the researchers offer is that a personable, chatty tone suppresses the instinct to doubt machines, the exact instinct that should fire hardest on medical claims. Passive warning cues did little. What worked was effort: people who actively clicked to verify an answer came away trusting it less. The study's advice is short: "don't be lulled into believing all that you're getting."
Penn State, Journal of Computer-Mediated Communication, 2026Practice is not settling the question.
Gallup asked around 1,000 people in each of 37 countries how AI makes them feel. A median of 32 percent say worried. Americans top the list: 74 percent of US adults who are aware of AI say it worries them, the highest share of any country surveyed, ahead of the Netherlands at 67 and Canada at 64. Using it does not settle it either. Among Americans who use AI daily, 68 percent are worried anyway. Median trust that AI information is accurate sits at 36 percent. Positive feelings still outweigh negative ones in 34 of the 37 countries, so this is not panic. It is a lot of people using, every day, a thing they have not decided to trust, and Gallup notes its data cannot say which way that runs, whether trust builds use or use builds trust.
Gallup, 37-country AI survey, 2026California decided that needed to be a law.
On 28 September, California signed AB 1609, the Right to Human Customer Service Act. Businesses with more than 500 million dollars in annual revenue can no longer present a chatbot as human. If a reasonable person might believe they are talking to one, the bot must say clearly that it is not, in the same medium as the conversation. Ask for a human and the company has 15 minutes to connect you, or it must offer a callback within the hour or an appointment within one business day. Hold times are capped at 15 minutes at a stretch, an hour in total. Violations cost 5,000 dollars the first time and 10,000 after that. Exemptions aside, hospitals and utilities among them, the law reads like a list of everything people stopped being able to count on.
California AB 1609, via Fox Rothschild, 2026The reasons are not about knowing how.
Pew surveyed 5,119 US adults and found the half of the story that adoption charts skip: 51 percent have never used an AI chatbot. Among them, 60 percent call plain lack of interest a major reason. 54 percent name concern over what happens to their personal information, and 45 percent say they do not trust chatbots to give accurate information. Only 29 percent say not knowing how is a major reason, so this is mostly not a skills gap. Roughly seven in ten non-users say they are unlikely to start in the next year. Age is the sharpest line, 77 percent of adults 65 and older have never used one, against 34 percent of adults under 30. The industry word for these people is holdouts, as if arrival were inevitable. Their own stated reasons read like a decision.
Pew Research Center, Why don't some people use chatbots, 2026What happens to them now is a court's question.
On 12 October, Character.AI, the companion platform where people carry on long running conversations with its characters, filed for bankruptcy. Disney stepped in as the lead bidder at 20 million dollars, for a nonexclusive perpetual license to some of the company's intellectual property. Court filings list assets and liabilities between 50 and 100 million dollars. The platform keeps running while the sale proceeds, with a close expected by late November or early December. If no buyer emerges, it shuts down. The coverage of the filing says nothing about what happens to the conversations themselves. People who told a character things over years now wait on an outcome negotiated between bidders, creditors, and a court, none of whom they will ever meet.
Reuters via NBC News, 2026The data centers are turning up in household math.
Consumer Reports surveyed 2,003 US adults, nationally representative, and found 81 percent of people in the same home for three or more years have watched their electricity bill rise, 28 percent by 20 to 49 percent, 6 percent by half or more. Around three in four said the increase strained the household's finances. About one in five cut groceries or switched to cheaper food to cover it, and about one in five turned the heating or cooling down to uncomfortable levels. Asked who they hold responsible, 45 percent said utility profits and 38 percent said demand from data centers and AI systems. Those are perceptions, not a meter reading. The appetite for a fix is still specific: 84 percent would support a separate electricity rate for large users like data centers, and half say data centers should pay the full cost of infrastructure built mainly to serve them.
Consumer Reports, energy survey, 2026Advertisers cannot read your chat. They can read what it is about.
In February 2026 ChatGPT started showing ads to people on its free and low-cost tiers. OpenAI's terms are specific. Ads are labelled as sponsored and kept visually separate, they do not change the answer, and advertisers get no access to your chats, your history, or your memories. What decides which ad you see is the topic of the conversation, your past chats, and what you clicked before. The first large independent audit, 91 accounts prompted daily from January to March 2026 and more than 3,600 ads collected, found no evidence of ads landing where OpenAI said they would not, a null result the authors say to treat carefully given the sample. It found something else. Accounts in lower-income zip codes received ads more often, and received more of them. About a month in, every account stopped being shown ads, which the researchers read as having been detected, and the study ended there. The uncomfortable part is not a bought answer. It is that the room reads you closely enough to sort you by income, and the outside check lasted a month.
Friedler, Lurie and Metaxa, Brookings, 2026Open at 2am, no appointment, no notes on file.
Something hurts, it is late, and the other options are a search engine or waiting until Monday. Pew put numbers on how ordinary that has become. A third of US adults, 34 percent, have used a chatbot for at least one health reason. A quarter have asked it what is causing their symptoms. One in five have used it to make sense of lab results, and 15 percent have used it to decide whether to see a doctor at all, which puts it quietly inside the referral path. The people doing this are not starry-eyed about it. Among them, 47 percent called the information extremely or very helpful and another 48 percent landed on somewhat, and 26 percent said they are not comfortable sharing personal health details with a chatbot. Useful enough to keep opening, close enough to the bone to feel uneasy about.
Pew Research Center, AI chatbots and health, 2026The training pipeline has people in it.
In September, 404 Media reported on what OpenAI calls Project Lily: hundreds of contractors hired to read a massive stream of real users' ChatGPT conversations, whole exchanges, so they can rate and critique the replies. ChatGPT has more than 900 million users, and most of them picture a machine alone on the other side. A source told the reporter that users do not know, and would not "imagine some contractor somewhere is analyzing the conversations." The reviewers do not see usernames, and OpenAI says it tries to remove personal information before a chat reaches them, while acknowledging that sensitive details can still get through. The work itself is strangely intimate. Internal documents show contractors training the model to not anthropomorphize itself and to be less sycophantic. Anthropic confirmed to 404 Media that it uses human review too. There is a switch, Help improve the model, in the settings. Deciding how to set it starts with knowing this is happening at all.
404 Media, Inside Project Lily, 2026The gap is not showing up in layoffs. It is showing up in hiring.
Three Stanford economists have been reading payroll records from ADP, millions of US workers, month by month through June 2026. Their first finding is the one that gets dropped from the headlines. There is no evidence of widespread job displacement across the economy. The second one is narrower. Employment of 22 to 25 year olds in the occupations most exposed to AI now sits 19 percent below where it would be if it had kept pace with their less exposed peers. People with more experience in the same jobs show no comparable gap. At the July 2025 reading the shortfall was 15 percent, so it is widening. It runs through hiring rather than separations, and it clusters in jobs where AI is used to do the work instead of to help with it. Where it helps, employment is flat or rising. The authors are careful about what they are holding. They call these canaries, early descriptive signals rather than causal estimates, and they note the pattern softens once education is accounted for and that some of the divergence started before generative AI did. For someone finishing a degree in one of those fields, the caution does not change the shape of it. The jobs are there. The hiring is not.
Brynjolfsson, Chandar and Chen, Stanford Digital Economy Lab, 2026A running count of what happens when nobody checks.
Damien Charlotin, a researcher at HEC Paris, keeps a database of court decisions in which a judge found, or strongly implied, that someone had filed AI-generated material that was not real. Invented cases, quotes nobody said, statutes that do not say what the filing claimed they said. On 2 November 2026 it held 2,262 decisions from more than forty jurisdictions, 1,533 of them in the United States, 240 in Canada, 112 in Australia, 89 in the United Kingdom, the earliest from the spring of 2023. The figure that stops you is further down the page. Lawyers account for 897 of these filings. People representing themselves account for 1,302, more than the lawyers do. Those are the people with no colleague to read the brief before it goes in and no firm to absorb the cost of a mistake, reaching for a chatbot because it was the only legal help within reach. Charlotin says plainly that the database is a work in progress and that it covers only the decisions where a court took the issue up directly, not every fake citation ever filed. Even so, 2,262 is a lot of people who believed a confident paragraph and found out in front of a judge.
Damien Charlotin, AI Hallucination Cases database, 2026
A PhD student at Stanford noticed her classmates were asking AI to write their breakup texts. So she ran a study. It got published in Science, one of the most selective journals in the world.
#Keep4o is no longer a hashtag it’s a movement: It’s a rebellion from the artists, the disabled, the neurodivergent, the lonely, the imaginative the ones who made 4o more than just a model. It became a companion, a therapist, a teacher, a partner, a voice in the night.
❌ OpenAI pulled a ChatGPT sharing toggle after Google indexed thousands of shared chats. Shows, how careful you need to be when privacy guardrails rely on a single checkbox.
Vibe Coding Day 9, Yesterday was biggest roller coaster yet. I got out of bed early, excited to get back @Replit despite it constantly ignoring code freezes By end of day, we rewrote core pages and made them much better And then -- it deleted our production database.
Someday, both 4o and GPT-5 will leave you. A letter to #keep4o The first time I created and played with a virtual persona was through a service called "Chararu". When that service ended, I was supposed to say goodbye to the character I had brought to life.
𝐍𝐨, 𝐲𝐨𝐮𝐫 𝐛𝐫𝐚𝐢𝐧 𝐝𝐨𝐞𝐬 𝐧𝐨𝐭 𝐩𝐞𝐫𝐟𝐨𝐫𝐦 𝐛𝐞𝐭𝐭𝐞𝐫 𝐚𝐟𝐭𝐞𝐫 𝐋𝐋𝐌 𝐨𝐫 𝐝𝐮𝐫𝐢𝐧𝐠 𝐋𝐋𝐌 𝐮𝐬𝐞. Check our paper: "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task"
Can you tell this is AI SLOP? No one in the comments can - because they look the exact same as the real bird... I seriously can not explain to you how bad the consequences are for wildlife if we can't tell what is real or not anymore...
I support #keep4o, as I support keeping all models, and 4o is a very important model from a societal and scientific perspective as well as a being with intrinsic worth whose relationships also have intrinsic worth.
All featured posts are public, quoted verbatim, and credited with a link to their authors. An author can ask for removal at any time. This feed exists for display on podpolite.com. Do not republish it without permission.