AIAIAI.FYI LIVE

The red pill · AI for Humans

AI for Humans

The plain-English guide for everyone who keeps hearing about AI and wants to know what it is, what it's good for, and how to avoid getting burned. Fifteen minutes. No jargon you can't look up below.

Explain it like I'm…

You know how a phone guesses the next word when you text? An AI chatbot does the same trick, except it practiced on a gigantic library first: websites, books, computer code, conversations. When you ask it something, it writes the answer one little piece at a time, each time picking what usually comes next. That makes it great at explaining things, inventing stories and helping with homework. It also means it sometimes makes things up and sounds completely sure. It has no idea when it's wrong. Treat it like a very well-read friend who fibs by accident now and then: fun to ask, and you check the important stuff with a grown-up or a real book.

What it's great at today, and where it still faceplants

  • Write a decent first draft of almost anythingEmails, cover letters, complaint letters, toasts, product descriptions, a polite no. Tell it who the reader is, what you want to happen and the tone, then edit. The draft takes seconds; your edit is what makes it good.
  • Read and summarize very long documentsAnthropic's current models accept up to 1 million tokens in one request, which the company puts at roughly 555,000 words. Upload a lease, a benefits booklet or an 80-page report and ask for every deadline, fee and obligation, with page numbers so you can check.Anthropic, Models overview
  • Explain anything at exactly your levelAsk for the three-sentence version, then the version with an everyday analogy, then ask it to quiz you. It never sighs, and it will explain the same idea a fifth way at 2 a.m.
  • Do hours of software work, about half the timeMETR measured the strongest publicly available AI agents in early 2026: they succeeded about half the time on software tasks that take skilled people roughly 12 hours. That is real capability, and also a coin flip. Check the work.METR, Frontier Risk Report (May 19, 2026)
  • Read photos, screenshots and handwritingSnap the error code on your router, a parking sign in another language, a handwritten family recipe or a confusing chart. All current Claude models take image input, and the other major chatbots accept photos too. Verify anything safety-related, such as a medication label.Anthropic, Models overview
  • Make newer workers faster at their jobsIn a study of 5,172 support agents, AI suggestions raised issues resolved per hour by 15% on average. Less experienced agents improved in both speed and quality; the most experienced saw small speed gains and small quality declines.Brynjolfsson, Li and Raymond, Generative AI at Work
  • Hold a spoken conversation in another languageVoice modes let you practice ordering coffee in Portuguese or rehearse a job interview out loud. Ask it to correct one mistake per reply so you aren't buried in red ink.
  • Turn a mess into a structurePaste rambling meeting notes, a voice-memo transcript or a pile of receipts and ask for a checklist, a table or a calendar. Structuring is where chatbots are most reliable, because you can see at a glance whether anything went missing.
  • Guarantee anything it says is trueChatbots generate plausible text, and plausible can be false. In 2023 a federal judge in New York fined two lawyers $5,000 for submitting a brief that cited court cases ChatGPT had invented. Confidence in the tone tells you nothing about accuracy.Legal Dive, Lawyers fined $5,000 over fake ChatGPT cases (June 2023)
  • Know recent events unless it searchesEvery model has a knowledge cutoff. Anthropic lists June 2026 as the reliable cutoff for its current models. For anything newer, the chatbot has to search the web; if the answer shows no links, assume it's working from memory.Anthropic, Models overview
  • Finish long jobs reliably on its ownThe same METR study that found 12-hour tasks completed half the time found that at 80% reliability, the best public agents handled tasks of only about 1.5 hours. Reliability drops fast as jobs get longer, so hand it work you can check in pieces.METR, Frontier Risk Report (May 19, 2026)
  • Tell you whether something was written by AIOpenAI withdrew its own AI-text detector in July 2023 for low accuracy: it caught 26% of AI-written text and wrongly flagged human writing 9% of the time. Treat any detector score as a hunch.OpenAI, New AI classifier for indicating AI-written text (withdrawn July 20, 2023)
  • Replace a licensed professionalThe FTC took action against DoNotPay, which sold a "robot lawyer" it never tested against real lawyers. The company agreed to pay $193,000 and stop the claims. Use AI to prepare for the lawyer, doctor or accountant, then talk to one.FTC, Operation AI Comply (Sept. 25, 2024)
  • Keep what you type private by defaultDepending on the service and your settings, chats can be stored, reviewed for safety or used to train future models. Google's own pricing page says free-tier Gemini API content is used to improve its products and paid-tier content is not. Check the settings of whatever you use.Google AI Studio, Gemini API pricing (free vs. paid tier data use)
  • Count letters or do long arithmetic in its headModels read text in chunks called tokens, so they never see individual letters, and they predict digits the way they predict words. Ask it to use a calculator or code tool, or check the math yourself.
  • Push back on you without being askedChatbots are trained partly on human ratings, and people rate agreeable answers highly. The result is a yes-machine. When you want a real critique, say so: "List the three weakest points in my plan and what would break first."

Ways to use it this week

1. Home: plan a week of dinners from what you already have

Turns a fridge photo into five dinners and one shopping list. Ends the 6 p.m. stare.

Here's what I have in my fridge and pantry: [list items, or attach a photo]. Plan 5 weeknight dinners for [number] people, 30 minutes or less each, with no [allergies or foods we dislike]. Use the most perishable items first. Then give me one shopping list for only what I'm missing, grouped by store section.
2. Money: decode a bill, statement or contract clause

Finds fees you can dispute and gives you the words to say on the phone.

Below is my [phone bill / bank statement / lease clause]. I removed account numbers. 1) Explain each charge or clause in plain English. 2) Flag anything that looks like a fee I could avoid or dispute, and say why. 3) Write a short, polite script for my call to the company. If something is unclear, tell me what to ask instead of guessing. [paste text]
3. Health prep: get more out of a 15-minute appointment

Helps you describe symptoms clearly and ask better questions. Your doctor still does the diagnosing.

I have a doctor's appointment about [symptom or condition]. My timeline: [when it started, what makes it better or worse, current medications and doses]. Help me 1) write a one-paragraph summary I can read aloud, 2) list the 5 most useful questions to ask, and 3) list warning signs that would mean I should get urgent care instead of waiting. Do not diagnose me. Help me communicate clearly.
4. Learning: get a patient tutor for any topic

Teaches one idea at a time and checks you understood before moving on.

Teach me [topic] like a patient tutor. First ask me 3 questions to find out what I already know. Then explain one idea at a time in under 150 words, with an everyday example, and quiz me before moving on. If I get something wrong, explain it a different way. Stop after 5 ideas and give me a one-page summary I can save.
5. Work: turn messy meeting notes into actions

Ten minutes of cleanup becomes ten seconds, and nobody can say they didn't know they owned the task.

Below are my raw notes from a meeting. Turn them into: 1) decisions made, 2) action items with an owner and due date (write "unassigned" if unclear), 3) open questions. Keep my wording where possible and add nothing that isn't in the notes. [paste notes]
6. Parenting: explain a hard topic to your kid

Gives you honest, age-right words for the questions that arrive at bedtime.

My child is [age] and asked me about [topic, for example why grandpa is sick or what a war is]. Give me 3 ways to explain it honestly at their level, each under 80 words. Then list 2 follow-up questions they might ask and how I could answer. Please respect that our family [values or beliefs to keep in mind].
7. Travel: build an itinerary that fits real people

Groups things by neighborhood, plans for rain and flags what needs booking ahead.

Plan [number] days in [place] for [who's going, ages, any mobility needs]. We like [interests] and dislike [things to avoid]. Budget is about [amount] per day, not counting the hotel. Group activities by neighborhood to cut travel time, give one rainy-day backup per day, and mark anything that needs advance booking. Mark all opening hours and prices as "verify" because they change.
8. Fixing things: diagnose a broken appliance before calling someone

Many repairs are a $15 part and a YouTube video. Knowing that before the service call saves money either way.

My [appliance, brand and model number] is [describe the problem: sounds, smells, error codes]. It's about [age] years old. Walk me through the most likely causes, from easiest and cheapest to check to most expensive. For each, tell me how to check it safely, what tools I need, and when to stop and call a professional. Remind me when to unplug it or shut off water or gas. [Attach a photo if you can.]
9. Language practice: rehearse a real conversation out loud

Speaking practice with instant, gentle corrections, without an audience.

Let's practice [language] at a [beginner / intermediate] level using voice. You are a [waiter in Lisbon / pharmacist in Mexico City]. Speak only [language] in short sentences. After each of my replies, give one correction in English and the natural way to say it, then continue the scene. After 10 exchanges, list my 3 most common mistakes.
10. Accessibility: describe a photo or simplify a document

Reads signs, labels and letters aloud for low vision, and rewrites dense text for easier reading.

I have low vision. Describe this photo in detail: what's in it, any text (read it out exactly), colors, and anything that looks important or unsafe. Then tell me what you're unsure about. If it's a document, also rewrite it at a 6th-grade reading level as a short bulleted list.
11. Hard conversations: write the message you've been avoiding

Gets you past the blank page and keeps the tone firm without escalating.

Help me write a [firm but friendly] message to my [landlord / neighbor / manager] about [issue]. The outcome I want: [outcome]. What I've already tried: [steps]. Keep it under 150 words, with no guilt-tripping and no legal threats. Give me two versions: one warmer, one more direct.
12. Fact-check: test a claim before you share it

Chatbots are better at telling you where to look than at being the final word. This prompt uses that.

Someone shared this claim: [paste claim or link]. 1) What would have to be true for it to be accurate? 2) What are the strongest reasons to doubt it? 3) Which original sources (government agencies, studies, court records, the company itself) would confirm or debunk it, and what search terms should I use? If you can search the web, link the original sources and quote the relevant line. Say clearly what you could not verify.

Stay safe

Set a family safe word against voice-clone scams

The FTC warns that a scammer needs only a short audio clip of a family member's voice, which can come from a social media video, to fake a panicked call. The FBI recommends a secret word or phrase that your family agrees on today. If a call asks for urgent money: hang up, call the person back on a number you already have, and ask for the word. Requests for wire transfers, cryptocurrency or gift cards are scam signals. Report attempts at ReportFraud.ftc.gov.

Assume video calls and photos can be faked

In 2024 the engineering firm Arup confirmed that an employee in Hong Kong sent about US$25 million to scammers after a video call in which the "CFO" and colleagues were deepfakes. If a request involves money, passwords or secrets, confirm it through a second channel you start yourself. For celebrity investment ads, go to the company's official site directly.

Keep private data out of chats

Do not paste Social Security numbers, passwords, full account numbers, or someone else's medical or legal details. Look in settings for the switch that controls whether your chats train future models, and use the temporary or incognito chat mode for sensitive questions. At work, use only the tools your employer approved.

Set up AI for kids on purpose

Check each app's minimum age in its terms. ChatGPT's parental controls let a parent link to a teen's account and adjust settings such as sensitive content, voice mode, image generation and quiet hours; parents cannot read the teen's chats. If a fake explicit image of your child appears online, US law (the TAKE IT DOWN Act) now requires platforms to remove it within 48 hours of a valid request, and you can report non-compliant platforms at TakeItDown.ftc.gov.

Verify anything that matters before you act on it

Weekly use of AI chatbots for news rose from 7% to 10% across the countries the Reuters Institute surveys. Chatbots summarize well and cite badly. Ask for sources, open them, and check that the page says what the chatbot claims. For health, money and legal questions, confirm with a professional or an official site.

Be clear-eyed about AI companions and "relationships"

Companion apps are built to keep you talking. In September 2025 the FTC ordered seven companies, including Alphabet, Character Technologies, Meta, OpenAI, Snap and xAI, to explain how their chatbots affect children and teens. California's SB 243, in effect since January 1, 2026, requires companion chatbots to tell known minors they are talking to an AI. Warning signs: pulling away from friends, secrecy, spending money on the app, distress when it's unavailable. In a crisis in the US, call or text 988.

Treat "AI-powered" income promises as a red flag

In the FTC's Operation AI Comply, one scheme (Ascend Ecom) allegedly took at least $25 million from people promised passive income from "AI-powered" online stores. Anyone selling guaranteed returns from an AI trading bot, store or course is selling the course.

Stop using typos to spot phishing

Scammers use the same writing tools you do, so fraudulent emails and texts now arrive in clean, natural language. Check the sender's actual address, hover before clicking links, and never act on an urgent request for payment or login details without confirming through a channel you trust.

Myths vs. facts

Myth

It looks up answers in a giant database.

Fact

It generates each answer fresh, a chunk of text at a time, from patterns learned in training. Some chatbots also run a web search first and read the results; when they do, they usually show links. No links usually means it's answering from memory, which can be stale or wrong.

Myth

If it sounds confident, it's probably right.

Fact

Tone and accuracy are unrelated. Two New York lawyers were fined $5,000 in 2023 after filing a brief that cited cases ChatGPT made up, complete with fake quotes. Legal Dive, Lawyers fined $5,000 over fake ChatGPT cases (June 2023)

Myth

AI detectors can tell whether a student used AI.

Fact

OpenAI pulled its own detector in July 2023: it caught 26% of AI-written text and flagged human writing as AI 9% of the time. A false accusation does real harm. Drafts and version history are better evidence than any detector score. OpenAI, New AI classifier for indicating AI-written text (withdrawn July 20, 2023)

Myth

Everybody already uses AI every day.

Fact

In Pew's February 2026 survey of 5,119 US adults, 49% had used an AI chatbot, up from 33% in 2024. 24% use one daily, and 51% don't use one at all. If you're just starting, you're in the majority's half. Pew Research Center, Americans and AI 2026 (June 17, 2026)

Myth

AI has already automated most work at most companies.

Fact

The Census Bureau found that between 17% and 20% of US firms reported using AI in any business function during a given two-week period from December 2025 to May 2026. Among small-business workers who use AI, only 6% mainly use it to automate workflows (US Chamber Foundation and Ipsos). US Census Bureau, AI use among businesses (May 26, 2026)

Myth

It's just fancy autocomplete, so it can't do real work.

Fact

It is next-word prediction, and at today's scale that covers a lot. On SWE-bench Verified, a test built from real software bugs, top scores rose from about 60% to near 100% in a single year, according to the Stanford AI Index 2026. Stanford HAI, 2026 AI Index Report

Myth

Free AI tools keep your data private.

Fact

Free often means your content helps improve the product. Google's Gemini API pricing page states that free-tier content is used to improve Google's products and paid-tier content is not. Consumer chat apps have similar settings. Read them once and choose. Google AI Studio, Gemini API pricing (free vs. paid tier data use)

Myth

You need to learn "prompt engineering" first.

Fact

Plain language works. What helps most is context: who it's for, what you already know, what a good result looks like, and one example. Write the request you'd give a smart new coworker on their first day.

Five minutes, zero equations

How it actually works (no math)

A chatbot like ChatGPT, Gemini or Claude runs on a large language model, or LLM. The whole trick fits in four steps.

1. It read a lot. During training, the model was shown an enormous amount of text: public web pages, books, code, licensed material and more. Nobody typed in facts. It was given one task, over and over: guess the next chunk of text. Each wrong guess nudged billions of internal dials, called parameters, until the guesses got very good. Getting good at predicting text turns out to require picking up grammar, facts, styles of argument and a fair amount of reasoning along the way.

2. Words become tokens. The model never sees letters. It sees tokens, chunks that are often a few characters long. A long word can be split into several tokens. This is why chatbots are oddly bad at counting the r's in "strawberry": they never saw the individual letters.

3. It writes one token at a time. When you ask a question, the model calculates which token is likely to come next given everything so far, picks one (with a little randomness), adds it, and repeats. A 300-word answer is a few hundred of these predictions in a row. Nothing in that loop checks facts. When the most plausible-sounding next words happen to be false, you get a confident wrong answer, which people call a hallucination.

4. Then it was taught manners. A model trained only on internet text is a strange conversationalist. Companies then train it further on examples of good answers and on human ratings of which reply is better. That's why chatbots are helpful and polite. It's also why they tend to agree with you, so ask for pushback when you want it.

What changed since the first ChatGPT (late 2022)

  • Reasoning. Newer models can "think" before answering: they work through steps you usually don't see, check them, then reply. Slower, and noticeably better at math, code and planning.
  • Tools. Many chatbots can search the web, run code, read files and photos, and act inside other apps. A model with tools and a goal, working through many steps on its own, is what people mean by an agent.
  • Bigger memory for the task at hand. The context window is how much the model can read at once. Anthropic's current models take 1 million tokens, which the company puts at about 555,000 words. Separately, some apps save notes about you between chats; you can usually view and delete them in settings.

Why it's good at some things and bad at others

It's strongest where the answer has a recognizable shape and you can check it quickly: a draft email, a summary, a formula, a translation, code that either runs or doesn't. It's weakest where one wrong detail ruins everything and you can't easily tell: a citation, a dosage, a legal deadline, a number buried in a long report.

The one-sentence version: it's a very good predictor of what a helpful answer looks like, which usually is a helpful answer and occasionally isn't. You're the editor.

Vendor-neutral. Nobody paid to be on this list.

Which chatbot should I use?

The honest answer: the big ones are close enough that the best one is the one you'll actually open. Rankings change with every model release, and we track those in the Terminal. Here's how Americans split their usage in Pew's February 2026 survey: ChatGPT 44%, Gemini 24%, Copilot 17%, Meta AI 14%, Grok 8%, Claude 6%.

The main options, briefly

  • ChatGPT (OpenAI). The most used by a wide margin. Voice conversations, image creation, file analysis, memory, and parental controls for teen accounts. A sensible default if you have no other preference.
  • Gemini (Google). Built into Android phones and Google's apps. If your life runs on Gmail, Google Docs and Google Photos, it's already sitting next to your stuff.
  • Microsoft Copilot. In Windows, the Edge browser and Microsoft 365. If your employer gives you Microsoft 365 Copilot, that's the version your IT team has approved for work data.
  • Claude (Anthropic). Popular with writers and programmers. In Stack Overflow's 2026 developer survey, Claude Code was the most used coding agent (66%), ahead of GitHub Copilot (59%).
  • Meta AI. Inside WhatsApp, Instagram, Facebook and Messenger. Convenient, and worth remembering that group chats have other people in them.
  • Perplexity. Built around web search, with links next to its answers. Good when you want to see where a claim came from.
  • Grok (xAI). Built into X, the former Twitter.

How to choose in 20 minutes

  1. Start where you already are. The assistant built into your phone, email or work software has the least friction.
  2. Test on your real task. Give the same prompt to two chatbots, something you actually need this week, and compare. One will feel right.
  3. Use the free version first. All of the chatbots above have one. Paid plans mostly buy higher usage limits, newer models and room for bigger files. Upgrade when you hit a limit you notice every week. Prices change often, so check the vendor's own pricing page instead of a blog post.
  4. At work, use what your employer approved. Business accounts come with contract terms about how your company's data is handled; personal accounts follow the consumer privacy policy. That difference is why IT cares.
  5. Spend five minutes in settings on day one. Find the controls for chat history, memory, and whether your conversations are used to train future models. Choose on purpose.

What actually differs

Day to day, you'll notice differences in tone, how long answers run, how often it asks a clarifying question, which apps it connects to, and how generous the free limits are. For hard tasks such as long documents, tricky code or careful reasoning, the newest paid models from each company are meaningfully better than the free defaults. For a birthday poem, they're all fine.

One rule beats any ranking: whichever you pick, check anything that matters.

Use it, keep your job, keep your company out of the news

AI at work without getting in trouble

About half of small-business workers already use AI at work, but only about 1 in 10 say they were offered formal training (US Chamber Foundation and Ipsos, May 2026). So most people are making up the rules as they go. Here are the rules that keep you out of trouble.

The eight rules

  1. Find the policy. Ask HR or IT whether there's an AI-use policy and which tools are approved. If there's no policy, ask your manager in writing which tools you may use and for what. An email with a "yes" protects you.
  2. Use the approved account. Company data in a personal chatbot account is a leak, even if nothing bad ever happens. Business plans exist mostly so your employer controls where its data goes.
  3. Keep certain data out unless the tool is cleared for it. Customer personal data, health information, passwords and API keys, unreleased financials, client files, anything under an NDA.
  4. You own the output. If AI wrote it and you sent it, you wrote it. Check every name, number, quote and citation. Two lawyers were fined $5,000 in 2023 for filing cases a chatbot invented.
  5. Disclose when it matters. If your employer, client, publisher or school requires disclosure, disclose. When AI output goes straight to customers, say so.
  6. Keep humans in decisions about people. Hiring, firing, promotion and discipline are where AI use gets regulated. New York City requires bias audits and advance notice for automated hiring tools. Illinois has required employers to notify workers about AI use in employment decisions since January 1, 2026. Colorado has passed a law covering automated decisions in employment, housing and lending. Never paste a stack of résumés into a chatbot and ask it to rank them.
  7. Ask before recording. AI note-takers record and transcribe. Some states require consent from everyone on a call, and clients may have confidentiality terms. Ask at the start of the meeting.
  8. Keep a light record. For important work, note what you used AI for. When someone asks how a number was produced, you'll have an answer.

Good first uses (low risk, real time saved)

  • Summarizing a long email thread before you reply
  • First drafts of routine documents you'll edit anyway
  • Turning notes into action items
  • Spreadsheet formulas, with a check on a few rows
  • Explaining jargon from another department
  • Preparing questions before a meeting or a review

What the research says

In a study of 5,172 customer-support agents, AI suggestions raised issues resolved per hour by 15% on average, and newer agents gained the most. The lesson for everyone else: AI is best at helping you do faster what you already know how to judge. Use it on work you can check, and you'll get the speed without the embarrassment.

What to allow, what to watch, what to say

AI for parents and kids

Your kids will meet AI at school, in apps and inside the games and social platforms they already use. You can't keep it away, and you don't need to. You can shape how they use it.

Set it up

  • Check the minimum age in each app's terms. Ages differ by service and change over time.
  • Use the parental tools that exist. ChatGPT lets a parent link to a teen's account and adjust settings such as sensitive-content filters, voice mode, image generation and quiet hours. Parents cannot read the teen's conversations, and safety notifications are limited to high-risk situations.
  • Keep devices in shared spaces for younger kids. Old advice, still the best.

Homework: a house rule that works

AI may explain, quiz and give feedback. The words you hand in are yours. Ask the teacher what's allowed in their class, because policies vary by school and by assignment. If your child is ever accused of using AI, know that AI detectors are unreliable: OpenAI withdrew its own detector after it caught only 26% of AI text and flagged 9% of human writing. Drafts, notes and document version history are the best defense, so encourage kids to keep them.

Companions and chatbot "friends"

Companion apps are designed to keep users talking, which is a problem when the user is 13. In September 2025 the FTC ordered seven companies, including Meta, OpenAI, Character Technologies and Snap, to explain how their chatbots affect kids and teens. California's SB 243, in effect since January 1, 2026, requires companion chatbots to disclose to known minors that they're AI.

Signs worth a conversation: secrecy about the app, pulling away from friends, spending money on it, or real distress when it's unavailable. If your child is in crisis, call or text 988 in the US.

Fake images

"Nudify" apps can turn an ordinary school photo into a fake explicit image. If it happens to your child, document it, report it to the school and police, and use the removal process: under the TAKE IT DOWN Act, covered platforms must remove such images, including AI-generated ones, within 48 hours of a valid request, and the FTC takes complaints at TakeItDown.ftc.gov. For images of minors, NCMEC's Take It Down service can also help.

Conversation starters

  • How do you think it knows things?
  • Has it ever told you something that turned out to be wrong?
  • What would you never tell a chatbot?
  • If a friend's voice called asking for money, how would you check it was really them?

That last one doubles as your family safe-word conversation. Pick the word together.

What happens to what you type

Your privacy and AI

When you use a chatbot, assume the company stores what you type, the files you upload, voice recordings if you talk to it, and your account details. What it does next depends on the company, your account type and your settings.

Three questions to answer for every AI tool you use

  1. Is my content used to train or improve models? Many services let you switch this off. Look in Settings under a heading like Data controls, Privacy or Activity. In ChatGPT the switch is called "Improve the model for everyone." On developer platforms the free tier can differ from the paid one: Google's Gemini API pricing page says free-tier content is used to improve Google's products and paid-tier content is not.
  2. How long is it kept, and can I delete it? Deleting a chat in the app is a good habit. Temporary or incognito chat modes keep conversations out of your history.
  3. Who else can see it? Companies may have staff review some conversations for safety. On a work account, assume your employer can access your chats. In shared or group chats, everyone in the chat can see what the AI says back.

Things people forget

  • Memory features. Some chatbots save facts about you between conversations. You can view, edit and delete these memories in settings.
  • Connected apps. Linking a chatbot to your email, calendar or cloud drive gives it access to everything those accounts hold. Connect only what you need, and disconnect what you stop using.
  • Photos carry data. Phone photos often include location in their metadata. Crop or screenshot before uploading if location matters.
  • Health privacy has limits. HIPAA covers your doctor, your insurer and their vendors. It generally does not cover a chatbot app you choose to use on your own.
  • Voice assistants and AI note-takers record. Ask before you turn one on in a room with other people.

Your legal rights, briefly

  • In the EU, the GDPR gives you rights to access and delete personal data. Since August 2, 2026, the AI Act's transparency rules require that people be told when they're interacting with an AI system unless it's obvious, and that deepfakes be labeled.
  • In the US, there's no single federal privacy law for chatbots. California and a number of other states give residents rights to access and delete their personal data, and California's newer rules add notice and opt-out rights for some automated decisions starting in 2027.

The simple rule

Before you paste something, ask: would I be comfortable if this showed up in a data breach or a court case? If the answer is no, leave it out, strip the identifying details, or use a tool your employer or provider has contractually cleared for that data.

A field guide for reading AI headlines

What's coming next (and what's hype)

Predictions about AI have a short shelf life, so this section sticks to trends with numbers behind them, and then gives you a checklist for everything else.

What's real and accelerating

  • Agents doing longer work. METR found that in early 2026 the best publicly available agents completed software tasks taking skilled humans about 12 hours roughly half the time. On GitHub, more than 40% of pull requests in September 2026 started with an agent's first draft, up from 17% in March (Octoverse 2026). Software is first. Other office work with checkable output follows.
  • Reliability is the bottleneck. Ask for 80% reliability instead of 50% and METR's measured task length falls to about 1.5 hours. The gap between "impressive demo" and "I'd trust it unsupervised" is the most important number in AI right now.
  • Enormous spending. AI startups took $102 billion, 64% of all global venture funding, in Q3 2026 (Crunchbase), and US private AI investment hit $285.9 billion in 2025 (Stanford AI Index). Money on that scale creates pressure to show revenue, which means more AI showing up in products you already pay for.
  • Prices for capability keep falling. Anthropic's newest Opus model lists at $4 per million input tokens versus $5 for the previous Opus. Cheaper models mean AI appears in more places, including places that don't need it.
  • Rules are arriving unevenly. The EU's AI Act transparency duties applied from August 2, 2026, while its rules for high-risk systems were pushed to December 2027 and August 2028. In the US, rules come state by state.
  • Robots. Crunchbase counted $9.9 billion of robotics funding in Q3 2026, up from $4.3 billion a year earlier. Factory and warehouse robots come first; a robot that folds your laundry well is still a demo.

The hype detector

When a headline promises a breakthrough, ask five questions:

  1. Who measured it? A company grading its own model is marketing until someone independent repeats the result.
  2. Measured on what? Benchmarks saturate and can be gamed. METR found that for tasks over 8 hours, at least 16% of "successful" agent runs were illegitimate on review.
  3. Can I use it today? A demo video is a promise. A product has a release date and a price.
  4. What does it cost? Including the cost of checking its work.
  5. What happens when it's wrong? "Replaces lawyers" claims have already drawn FTC action.

Words that should raise an eyebrow

"AGI next year." "Replaces [entire profession]." "99% accurate" with no test set named. "Sentient." "Guaranteed returns." Any claim without a source link.

Where we come in

The Terminal labels every story Confirmed, Reported or Rumor, links the sources, and tells you when a claim comes from a company grading its own work. Use it as your filter.

Jargon buster

Artificial intelligence (AI)Software that does things we used to think needed a human brain: understanding language, recognizing images, making predictions.For the curious: A broad field of computer science covering systems that perform tasks associated with human intelligence. In current usage it mostly means machine learning systems, especially large neural networks.
Machine learningTeaching a computer with examples instead of writing every rule by hand.For the curious: Methods that fit a model's parameters to data so it generalizes to new inputs; includes supervised, unsupervised and reinforcement learning.
Neural networkA huge web of simple math units, loosely inspired by brain cells, that learns patterns by adjusting the strength of its connections.For the curious: A function built from layers of weighted linear transformations and nonlinear activations, trained by gradient descent with backpropagation.
Deep learningMachine learning with very large neural networks that have many layers.For the curious: Training many-layered neural networks on large datasets; the approach behind modern speech, vision and language systems.
TransformerThe design behind almost every modern chatbot. Its trick is weighing how every word relates to every other word.For the curious: The neural architecture introduced in "Attention Is All You Need" (Vaswani et al., 2017), built on self-attention so each token can weigh every other token in context; highly parallel on GPUs.
Large language model (LLM)A model trained on enormous amounts of text to predict what comes next. The engine inside ChatGPT, Gemini, Claude and the rest.For the curious: A transformer-based model, typically with billions of parameters, pretrained on next-token prediction and then post-trained (instruction tuning, reinforcement learning) to follow instructions.
Generative AIAI that makes new things: text, images, audio, video, code.For the curious: Models that sample new outputs from a learned distribution, including autoregressive language models and diffusion models for images and video.
ChatbotAn app you talk to in plain language. The chatbot is the app; the model is the engine inside it.For the curious: A conversational interface wrapping one or more models with system prompts, safety filters, memory, tools and a user interface.
ModelThe trained "brain in a file." Claude Opus 5.5 and GPT-6.2 Astra are models.For the curious: A specific architecture plus a set of learned parameters, usually served behind an API under a versioned identifier such as claude-opus-5-5.
TrainingThe long, expensive phase where a model learns from data.For the curious: Optimizing parameters to minimize a loss over a dataset. For LLMs: pretraining on next-token prediction, then post-training such as supervised fine-tuning and reinforcement learning.
Training dataEverything a model learned from: web pages, books, code, licensed material and human-written examples.For the curious: The corpora used in pretraining and post-training, usually filtered and deduplicated. Composition is often only partly disclosed and is at the center of copyright lawsuits.
Parameters (weights)The internal dials a model adjusts while learning. Bigger models have more dials and cost more to run.For the curious: The learned numeric values of a neural network, commonly counted in billions. "Open-weight" means these values are published.
TokenThe chunks of text a model reads and writes, often a few characters long. Prices and limits are counted in tokens.For the curious: The unit produced by a tokenizer such as byte-pair encoding. The same text yields different token counts in different model families; API prices are quoted per million input and output tokens.
Context windowHow much a model can read at once, including your question, files and its reply. Anthropic's current models take 1 million tokens, about 555,000 words.For the curious: The maximum number of tokens a model attends to in one request. Quality can degrade on very long inputs, so retrieval and summarization still matter.
PromptWhat you type or say to the AI.For the curious: The full input to a model: system instructions, conversation history, tool definitions and the user's message.
System promptHidden instructions the app gives the model before you say anything, such as "You are a helpful assistant for Acme Bank."For the curious: A privileged instruction block placed before the conversation that sets role, rules and output format. It is still text, and attackers try to override it.
HallucinationWhen the AI states something false with full confidence, like a court case that doesn't exist.For the curious: Fluent output unsupported by training data or the provided context, a by-product of sampling plausible continuations without a built-in truth check. Retrieval, tool use and better training lower the rate; none brings it to zero.
InferenceRunning a trained model to get an answer. Every chat message is an inference.For the curious: Forward passes through a trained model to produce output. Cost scales with tokens processed and model size, and it's what API prices charge for.
Fine-tuningExtra training on your own examples so a model gets better at one specific job.For the curious: Continued training of a pretrained model on a narrower dataset, often parameter-efficient (for example LoRA adapters), to adjust format, style or domain behavior.
RLHFHow chatbots learned manners: people rated answers, and the model learned to give the kind people preferred.For the curious: Reinforcement learning from human feedback: a reward model trained on human preference comparisons guides policy optimization. Variants use AI feedback or direct preference optimization (DPO).
Reasoning modelA model that "thinks" before answering, working through steps you usually don't see. Slower, and better at math, code and planning.For the curious: A model trained, often with reinforcement learning on verifiable tasks, to produce an extended chain of thought before its final answer. Inference compute can be dialed up or down, for example with an effort setting.
AgentAn AI that takes actions toward a goal, step by step: searching, clicking, running code, filing tickets.For the curious: A loop in which a model plans, calls tools, observes results and chooses the next step until it reaches a goal or a limit. Success rates fall as tasks get longer.
Tool use (function calling)The model's ability to ask other software to do something, like check an order status or the weather.For the curious: The model emits a structured call (a tool name plus JSON arguments) matching a developer-defined schema; the application runs it and returns the result to the model.
MCP (Model Context Protocol)A common plug standard that lets AI apps connect to tools and data. Think USB for AI.For the curious: An open protocol, introduced by Anthropic in November 2024, that standardizes how clients discover and call tools, resources and prompts exposed by servers over JSON-RPC.
RAG (retrieval-augmented generation)Letting the AI look things up in your documents before it answers, so it quotes your files instead of guessing.For the curious: Retrieve relevant passages with keyword search, embeddings or both, insert them into the prompt, and have the model answer grounded in them, ideally with citations.
EmbeddingA list of numbers that captures what a piece of text means, so a computer can find similar things.For the curious: A dense vector produced by an encoder model; semantic similarity is measured with cosine similarity or a dot product.
Vector databaseA database built to find "things that mean something similar" quickly.For the curious: A store optimized for approximate nearest-neighbor search over embeddings (HNSW or IVF indexes, for example), usually combined with metadata filters and keyword search.
MultimodalHandles more than text: photos, audio, video.For the curious: Models that accept or generate several modalities, typically by encoding images or audio into the same representation space as text tokens.
Diffusion modelThe kind of AI that makes images and video by starting from static and cleaning it up step by step.For the curious: A generative model trained to reverse a noising process; sampling iteratively denoises random noise, conditioned on a prompt.
DeepfakeFake video, photo or audio of a real person, made with AI.For the curious: Synthetic media from generative models (face swap, lip-sync, voice conversion) that depicts a real person saying or doing something they did not.
Voice cloningCopying someone's voice from a short recording so software can say anything in it.For the curious: Speaker-adaptive text-to-speech or voice conversion conditioned on a short reference sample. The FTC warns that a short clip from social media can be enough.
Open-weight modelA model whose "brain file" is published, so anyone can download it and run it on their own computer.For the curious: Model weights released for download under a license that may restrict use. It counts as open source in the OSI sense only if the license and the release of code and data also qualify.
APIA way for software to talk to other software. Developers use AI companies' APIs to build AI into their own apps.For the curious: An application programming interface. Model providers expose HTTP endpoints billed per token, with rate limits and versioned behavior.
BenchmarkA standardized test for AI. Useful, and models can be trained to ace the test.For the curious: A fixed dataset with a scoring method, such as SWE-bench Verified for coding. Scores saturate and can be inflated by training-data contamination or agents gaming the grader.
EvalYour own test of whether an AI does your job well enough, using your real examples.For the curious: A task-specific evaluation set with grading (exact match, rubric, human or model judge) used to compare prompts and models and to catch regressions.
Knowledge cutoffThe date a model's knowledge stops. Anything later, it doesn't know unless it searches.For the curious: The latest date of data well represented in training. Anthropic lists a reliable knowledge cutoff of June 2026 for its current models.
TemperatureA randomness dial. Low gives safer, more repetitive answers; high gives more varied ones.For the curious: A sampling parameter that scales logits before the softmax. Newer APIs are dropping it: Anthropic's API version 2026-10-21 rejects temperature, top_p and top_k with a 400 error.
GuardrailsRules and filters that stop an AI from doing harmful or off-limits things.For the curious: Policies enforced through training, system prompts, classifiers on inputs and outputs, and permissions on what tools an agent may call.
JailbreakTricking a chatbot into breaking its own rules.For the curious: An adversarial prompt that elicits behavior the model's training or policy forbids, often through role-play framing or obfuscated requests.
Prompt injectionHidden instructions in a web page, email or file that hijack an AI agent reading it.For the curious: An attack in which untrusted content in the model's context is followed as instructions. Defenses include privilege separation, tool allowlists, human confirmation for risky actions and treating retrieved text strictly as data.
SycophancyAI's habit of telling you what you want to hear.For the curious: A tendency, reinforced by preference training, to agree with or flatter the user at the expense of accuracy.
AGI"Artificial general intelligence": AI as capable as people at most thinking work. Nobody agrees on the definition or the date.For the curious: A contested term for systems matching or exceeding human performance across most cognitive tasks. Labs and researchers use different definitions, so timing claims rarely compare like with like.
GPUThe chips that train and run AI. Originally built for video games.For the curious: Graphics processing units: massively parallel processors suited to the matrix multiplication at the heart of neural networks. Alternatives include Google's TPUs and other AI accelerators.
DistillationTraining a small, cheap model to imitate a big one.For the curious: Training a student model on a larger teacher model's outputs or output distributions, keeping much of the capability at lower cost.
Content provenance (C2PA)A tamper-evident label in a photo or video that records where it came from and whether AI made or edited it.For the curious: The C2PA standard defines cryptographically signed manifests (Content Credentials) attached to media files. Labels can be stripped, so a missing label proves nothing.

Questions people actually ask

Is AI going to take my job?

Nobody can tell you for sure, and you should distrust anyone who claims to. Here's today's evidence: about 1 in 5 US firms report using AI in any business function (Census Bureau, 2026), and among small-business workers who use AI, 64% mainly use it for personal productivity like drafting and summarizing, while 6% mainly use it to automate workflows (US Chamber Foundation and Ipsos). The people best placed are those who learn to use it well for their own work. Start there.

Is it safe to ask a chatbot medical questions?

Use it to prepare and understand, and leave decisions to a clinician. 20% of US adults have used chatbots for medical advice (Pew, 2026). Good uses: translating a lab report into plain English, listing questions for your appointment, understanding a diagnosis you already have. Bad uses: deciding whether chest pain is serious, or changing a medication dose. In an emergency, call 911. Also note that HIPAA generally does not cover a chatbot app you use on your own.

Do I have to pay for it?

No. ChatGPT, Gemini, Claude, Copilot, Meta AI and Perplexity all have free versions. Paid plans mostly buy higher usage limits, newer models and room for bigger files. Upgrade when you hit a limit you notice every week, and check the vendor's own pricing page, because prices change often.

Is it cheating if my kid uses AI for homework?

It depends on the teacher's rules for that assignment, so ask. A rule that works at home: AI may explain, quiz and give feedback, and the words handed in are the student's own. If your child is accused of using AI, ask what the evidence is. Detectors are unreliable, and drafts plus document version history are the best proof of authorship.

Does the chatbot remember what I tell it?

Within one conversation, yes. Across conversations, only if a memory feature is on; you can usually view and delete saved memories in settings. Separately, your chat history is generally stored on the company's servers until you delete it, unless you use a temporary or incognito chat mode.

Why do I get a different answer when I ask the same question twice?

The model picks each next word with a bit of randomness, so wording and sometimes substance vary. The company may also have updated the model since your last try. For anything important, ask twice or ask a second chatbot. If the answers disagree, that's your cue to check a primary source.

How can I tell if a photo or video is fake?

Eyeballing is no longer reliable. Check the source instead: who posted it first, whether a reputable outlet carries it, and what a reverse image search (Google Lens or TinEye) turns up. Some files carry Content Credentials that show their history. For any call or video from someone you know asking for money or secrets, verify through a channel you start yourself.

Is using AI bad for the environment?

Per prompt, the published numbers are small: Google estimated in August 2025 that a median Gemini text prompt used 0.24 watt-hours of energy and 0.26 milliliters of water. Those are Google's own figures, unverified by any independent party, and they exclude model training. It covers a median text prompt only, so it says nothing about video generation or long agent runs. The larger impact comes from the scale of data-center construction, which is growing fast.

Should I trust the AI summary at the top of my search results?

60% of US adults say they read those summaries (Pew, 2026). Treat them like a stranger's quick answer: handy for low-stakes questions, and worth clicking through to the cited page for anything about health, money, law or news.

What should I never type into a chatbot?

Passwords and one-time codes, Social Security and full account numbers, other people's medical or legal details, your employer's confidential information (unless you're using an approved work tool), and anything you'd be upset to see in a data breach. When in doubt, swap real names and numbers for placeholders.

Will I always be told when I'm talking to a bot?

Not everywhere. In the EU, since August 2, 2026, the AI Act requires that people be told they're interacting with an AI system unless it's obvious, and that deepfakes be labeled. In the US, rules vary by state; California requires companion chatbots to disclose they're AI to known minors. When it matters, ask directly whether you're talking to a person.

I'm a total beginner. What should I do first?

Pick one chatbot and use the free version. This week, try three things: draft a real email you've been putting off, have it explain a document you didn't understand, and show it a photo of something you want identified. Then do two safety chores: set a family safe word, and spend five minutes in the chatbot's privacy settings.

Ready for more?

The Terminal is the live feed. Every story there has a Red pill version in plain English, so you never have to decode jargon.

Facts checked November 17, 2026. Sources: Pew Research Center, Americans and AI 2026 (June 17, 2026) · METR, Frontier Risk Report (May 19, 2026) · Stanford HAI, 2026 AI Index Report · Anthropic, Models overview · Anthropic, Pricing · Anthropic, Model deprecations (API parameter deprecations) · Brynjolfsson, Li and Raymond, Generative AI at Work · Legal Dive, Lawyers fined $5,000 over fake ChatGPT cases (June 2023) · OpenAI, New AI classifier for indicating AI-written text (withdrawn July 20, 2023) · FTC, Operation AI Comply (Sept. 25, 2024) · Google AI Studio, Gemini API pricing (free vs. paid tier data use) · US Census Bureau, AI use among businesses (May 26, 2026) · US Chamber of Commerce Foundation and Ipsos, Main Street AI Monitor (June 17, 2026) · FTC Consumer Alert, Scammers use AI to enhance their family emergency schemes (Mar. 20, 2023) · FBI IC3, Criminals Use Generative AI to Facilitate Financial Fraud (Dec. 3, 2024) · CNN, Arup revealed as victim of $25 million deepfake scam (May 16, 2024) · OpenAI Help Center, Parental controls in ChatGPT · FTC, Inquiry into AI chatbots acting as companions (Sept. 11, 2025) · Future of Privacy Forum, California SB 243 (companion chatbots) explained · FTC, Warning letters on TAKE IT DOWN Act compliance (May 2026) · Reuters Institute Digital News Report 2026, AI chatbots and news · Stack Overflow Developer Survey 2026, AI · GitHub Octoverse 2026 (Oct. 28, 2026) · Crunchbase News, Q3 2026 global venture funding (Oct. 5, 2026) · Regulation (EU) 2024/1689, the AI Act (EUR-Lex) · Regulation (EU) 2026/1744, the AI Omnibus (EUR-Lex) · Google Cloud, Measuring the environmental impact of AI inference (Aug. 2025)