A plain-English guide to AI. No degree needed. Every idea starts with the real term, gets a one-sentence meaning, and then we use everyday words from there on.
Words with a dotted underline are real AI terms. Tap one to jump to its glossary entry. Each entry lists every place the term appears, and tapping one of those takes you back to the exact sentence. The glossary grows with every post.
Planned, A to Z
How an AI learns
How an AI answers a question
Why AI gets things wrong
Data centers: power, water, jobs and what is inside the building
Myths: the alien takeover and other movie plots
Edition No. 1 · Sunday, November 15, 2026
About 8 minutes
What AI Is, and What It Is Not
Every edition has three parts: the main story, ten news stories, and real-life uses for beginners. This week: the starting point for everything in this guide, ten AI stories from the past few weeks, and three ready-to-use prompts (the instructions you type to an AI chatbot) for everyday chores.
Small numbers like [M1] (main story) and [2] (news stories) point to the source for that fact. Tap one to see the source, then tap "Back to the sentence" to return. The summaries, "Why it matters" lines, everyday pictures and prompts are mine. The numbered sources back up the facts.
A note on fairness
Plain Terms is researched with help from Claude, an AI made by Anthropic. Claude is named in the main story, and Anthropic shows up in news stories 2 and 9. I report all of it from public pages and try to treat every company the same way.
The short version of "What AI Is, and What It Is Not"
AI stands for artificial intelligence (software that does jobs we usually link with human thinking, like recognizing speech, sorting photos or writing text). I often call it smart software, meaning good at a job, not wise. Experts have never agreed on one exact definition of AI.[M1]
Most software you use follows an algorithm (a set of step-by-step instructions a computer follows to do a job), like a recipe card a person wrote out. But machine learning (a way of building software in which the computer finds patterns in a very large number of examples, instead of following rules a person typed in) now sits underneath most modern AI.[M2] Here is my own picture, an illustration and not a sourced fact: an experienced mechanic can hear a worn wheel bearing from the driver's seat because they have heard many bad bearings over many years. Machine learning is closer to that than to a recipe card.
The learning step is called training (showing an AI a very large number of examples and adjusting it, over and over, until it does the job well). I call these practice rounds. A chatbot like ChatGPT or Claude is built on a large language model (an AI trained on a huge pile of writing so it can predict what text comes next, which is how a chatbot answers you).[M3][M4] OpenAI describes the first stage of training as learning to predict the next word across huge amounts of text.[M4]
Chatbots can also be wrong with total confidence. A hallucination (a statement from an AI that sounds believable but is false) is what I call a confident wrong answer. OpenAI's researchers define it as a plausible but false statement, and their own page says it remains a fundamental challenge for all language models, even though their newer model makes fewer of them.[M4]
Narrow AI (AI built to do one kind of job well) is the kind most people meet every day, like the spam filter that sorts your email.[M5] I call it one-job AI. The other idea is artificial general intelligence, or AGI (the idea of a single AI that could learn and do almost any thinking job a person can). I call it do-everything AI. Alissa de Chassey's April 2026 review in Scienceline found no single agreed definition of AGI and no agreed test for it, and experts disagree about timing, from a few years to decades or never.[M6]
You are not alone if AI makes you uneasy. In June 2025, the Pew Research Center surveyed 5,023 U.S. adults. About half, 50%, said the growing use of AI in daily life makes them more concerned than excited. Only 10% said more excited than concerned, and 38% said both equally.[M7]
What this does not show
This is the short version. The full story also covers what AI is not: not a brain, not one thing, not a fact database and not all-knowing.
Experts disagree about the meaning of AI and AGI. This reports the disagreement and does not settle it.
The explanation of confident wrong answers is OpenAI's own research, and its page does not say whether outside experts reviewed it.
The Pew numbers describe U.S. adults in June 2025, and opinions may have moved since.
Nothing here says AI is safe, and nothing here says it is dangerous.
What this means for you
Treat a chatbot like a quick, well-read helper who sometimes bluffs. It is good for drafts, ideas and plain explanations. For anything that matters, such as money, health, legal or safety questions, check it against a source you trust.
Nvidia said on November 12 that Jensen Huang resigned as president, CEO and board member, effective immediately.[1] Chief financial officer Colette Kress is interim CEO while the board searches inside and outside the company.[1] The announcement gives no reason, and we will not guess.
Tap to read moreTap to close
On November 12, the chip maker Nvidia said Jensen Huang resigned as its president, chief executive and board member, effective immediately.[1] Colette Kress, its chief financial officer, becomes interim CEO and stays CFO until a permanent pick is named, and the board says it will look at candidates from inside and outside the company.[1] Harvey C. Jones was named the board's first-ever chair.[1]
Keep in mind. The company's announcement gives no reason for the resignation.[1] I found none in other sources either, so I will not guess.
Why it matters. A change at the top of a large AI company is news by itself. Until a reason or a permanent CEO is announced, what it means for the company is unknown.
On November 10, Anthropic began its roadshow (the investor tour before shares are priced) for a proposed IPO at an expected $70 to $78 a share.[2] The filing is not yet in effect, so no shares can be sold yet.[2] Anthropic makes Claude, which helps research this guide, so weigh this item with that in mind.
Tap to read moreTap to close
On November 10, Anthropic launched the roadshow for its proposed initial public offering, or IPO (the first time a company sells shares of itself to the public on a stock market).[2] A roadshow is the investor meeting tour before shares are priced. Anthropic expects a price between $70 and $78 a share and has applied to list on the Nasdaq under the symbol ANT. About 825 million shares are in the offering: roughly 801 million from Anthropic and about 24 million from existing owners who are selling.[2]
Keep in mind. The company says its paperwork with the U.S. Securities and Exchange Commission is not yet in effect, so no shares can be sold yet. Its page does not give a pricing or trading date, and it does not confirm that Nasdaq has approved the listing.[2] Anthropic makes Claude, the AI that helps research this guide, so read this item with that in mind.
Why it matters. Share prices can fall as well as rise, and this is not investment advice.
On November 5 the Nasdaq Composite fell 5.52%, its worst day since April 2025, and entered a correction (a drop of 10% or more from a recent high).[3] NPR reported that more than $2 trillion in value disappeared as investors feared tougher rules on AI and data centers.[4] A fear is not a law, and other things moved prices that day too.[5]
Tap to read moreTap to close
On Thursday, November 5, two days after the midterm elections, the Nasdaq Composite closed down 5.52%, its worst day since April 2025.[3] NPR reported that more than $2 trillion in market value disappeared as investors feared that an incoming Democratic Congress, and states across the country, could crack down on AI and data centers.[4] The drop put the index in a correction (a fall of 10% or more from a recent high). It closed at least 10% below its October 6 record.[3][5]
Keep in mind. A fear is not a law. Reuters also pointed to other things that day, such as a chipmaker missing sales estimates.[5] Prices can move for many reasons at once.
Why it matters. NPR notes that many Americans own these stocks through retirement accounts such as 401(k)s and pension funds.[4]
On November 3, voters in many towns decided local questions on data centers (large buildings of computers that run services, including AI).[6] The law firm Venable counted 41 measures in 10 states and called the results almost unanimously against development.[6] Counts differ by tracker, and the votes do not show why people voted as they did.[8]
Tap to read moreTap to close
A data center (a large building full of computers that store information and run services, including AI) was on the ballot in many towns on November 3. The law firm Venable counted 41 local ballot measures on data center development across 10 states and described the results as almost unanimously against development.[6] In Edgewater, Florida, voters wrote a ban on new data centers into the city charter.[6] Ohio had the most: 21 communities there put questions to voters.[7]
Keep in mind. Counts differed between trackers and were still changing on November 4.[8] Some measures only asked for advice and were not binding.[6] The votes do not show whether people were rejecting AI, data centers or the party in power, and people on both sides dispute that.[8]
Why it matters. This is the real-world side of AI: buildings, electricity and land, decided in local votes. It ties to story 3, where investors worried about exactly this.
Employers announced 43,994 job cuts in October, down 71% from a year earlier.[9] AI was named for 1,265 of them, the lowest month of 2026, though it is named for 121,401 so far this year.[10] These are reasons employers gave, not proof of cause.[11]
Tap to read moreTap to close
Challenger, Gray & Christmas, a firm that tracks announced layoffs, reported that U.S. employers announced 43,994 job cuts in October, down 71% from October 2025.[9] AI was named as the reason for 1,265 of them, the lowest month of 2026, though AI has been named for 121,401 cuts so far this year.[10] Market and economic conditions, at 130,531 cuts, have now passed AI as the top reason given this year.[10]
Keep in mind. These counts are the reasons employers gave, not proof of what caused each cut.[11] I could not open Challenger's own report, so the numbers come from outlets that quote it.
Why it matters. Headlines that say AI is taking jobs, and headlines that say AI is blamed too much, both rest on numbers like these. It helps to know what they measure.
On October 26, Microsoft offered voluntary buyouts to about 60,000 employees in four countries, about 27% of its workers.[12][13] The offer is voluntary, and Microsoft's own post does not mention layoffs or AI.[12] Workers have until early December to decide.[13]
Tap to read moreTap to close
On October 26, Microsoft announced a Voluntary Exit Offer (a buyout: eligible workers can choose to leave the company with financial support and career help).[12] About 60,000 employees in the U.S., Canada, the U.K. and India are eligible, about 27% of its 223,000 workers as of June 30. They have until early December to decide.[13] Reports say teams working on Copilot, Windows and Xbox are among those eligible.[13]
Keep in mind. The offer is voluntary, and Microsoft's own post does not mention layoffs or say AI is the reason.[12] Challenger did not count these offers in its October total.[10] It is not known how many people will accept.[13]
Why it matters. Big companies are reshaping their workforces while they spend heavily on new technology. Watch what companies do, and read carefully what they say about why.
At Adobe MAX on November 10, Adobe announced AI agents (AI that carries out several steps of a job on its own) across its creative apps.[14] An analyst firm says the Creative Agent is expanding to Photoshop, Illustrator and others.[15] Several tools are previews, and the details come from Adobe and an analyst briefing, not independent tests.
Tap to read moreTap to close
At its MAX conference on November 10, Adobe announced new AI tools across its creative software.[14] The focus is the AI agent (an AI program that can carry out several steps of a job on its own, instead of answering one question at a time). According to the analyst firm Constellation Research, Adobe is expanding its Creative Agent to Photoshop, Illustrator, Premiere, InDesign and After Effects, and it is previewing agents in Firefly that can keep working on a task for days or weeks.[15] Adobe also launched a preview that lets other companies' AI agents control its desktop apps.[15]
Keep in mind. Several of these tools are previews, not finished products. The descriptions come from Adobe and an analyst briefing, not from independent tests.
Why it matters. "Agent" is becoming the word for AI that does work for you instead of only talking with you. You will hear it often.
On November 11, Google DeepMind released AlphaGenome 2, an AI model that predicts how DNA changes affect specific kinds of cells.[16] It is open for non-commercial research.[16] The results are DeepMind's own, and it is not validated for clinical use.[17]
Tap to read moreTap to close
On November 11, Google DeepMind released AlphaGenome 2, an AI model (the finished product of training, a huge collection of numbers that holds the patterns the AI found) that predicts how a change in DNA affects specific kinds of cells.[16] It is available for non-commercial research through DeepMind's AlphaGenome service, with a private preview for commercial research.[16] DeepMind says it is significantly better than the first version at predicting gene activity in specific cells, which can help researchers narrow down which DNA changes matter for a disease.[16]
Keep in mind. The results come from DeepMind's own tests. A news review of the release notes that the model is not validated for clinical or diagnostic use and that its predictions still need lab testing.[17]
Why it matters. This is a tool for scientists, not a medical test for patients.
On November 6, a company called PSI said its system, running a non-public Anthropic model, found a proof of the Aaronson-Ambainis conjecture (a math claim experts believe but nobody had proved).[18] This is the company's claim, and its page does not describe peer review or an independent audit.[18] Anthropic makes Claude, so the note on story 2 applies.
Tap to read moreTap to close
On November 6, a company called PSI said its research system, which ran a non-public Anthropic model, found a proof of the Aaronson-Ambainis conjecture (a math claim that experts believe is true but that nobody had proved). PSI also says Lean, a program that checks a proof step by step, accepted it.[18] The work was led by David Turturean, and PSI adviser Jim Halverson audited the proof.[18] Stanford researchers Guy Blanc, Jordan Docter, Carmen Strassle and Li-Yang Tan have read it and are writing a shorter explanation.[18]
Keep in mind. This is the company's claim. Its page does not say the proof has been through peer review (outside experts checking a study before it appears in a scientific journal), and it does not describe an independent audit. Halverson's audit was inside PSI, and the page does not present the Stanford comments as an endorsement.[18] Anthropic makes Claude, so the same note as story 2 applies.
Why it matters. If it holds up, it would be an example of AI helping with new math. Time and outside experts will decide.
On November 10, Goodfire researchers published a study arguing that single neurons (tiny units inside an AI) are easier to read than many experts assumed.[19] It is one lab's post on its own site, and the page does not say it was peer reviewed.[19] We translated it in Translated, No. 1.
Tap to read moreTap to close
On November 10, researchers at the lab Goodfire published "Neurons as an Interpretable Basis."[19] The authors argue that single neurons (tiny units inside an AI that take in numbers, do a little arithmetic and pass a number along) are easier to read than many researchers assumed. Few are active at any moment, a small number of jobs describe each one, and those descriptions can predict when it lights up.[19] We walked through the full study in Translated, No. 1: Is AI Really a Black Box?
Keep in mind. It is one lab's post on its own website, and the page does not say it has been through peer review. The main study covers one AI model.
Why it matters. Better tools for looking inside AI help people check it. That is a reason to ask for more inspection, not less.
A good prompt says what you have, what you want back and why you need it. The "why" helps the AI choose the right level of detail. All three prompts below are my own examples, not from any source. Tap Copy, paste the prompt into a chatbot, and replace the part in brackets with your own words.
Time-saver 1
Make a confusing letter or bill make sense
Turns a hard-to-read letter into plain words, deadlines and questions to ask.
I received the letter or bill pasted below. I am not an expert in this subject, and I want to understand it well enough to act on it without making a mistake.
Please do these five things in plain words, at about an eighth-grade reading level:
1. Tell me in two sentences what this is and who sent it.
2. List what they want from me, how much money is involved and every deadline.
3. Tell me what could happen if I ignore it.
4. Write three questions I should ask the sender before I pay or sign anything.
5. Point out any part you are not sure about, so I know what to double-check.
[PASTE THE LETTER OR BILL HERE. Remove account numbers and Social Security numbers first.]
Check before you act. Compare every date and dollar amount with the original letter. For money, health or legal matters, confirm with the sender or a professional.
Time-saver 2
Turn messy notes into tasks and a follow-up message
Sorts messy notes into decisions, tasks and a message you can send.
Below are my rough notes from a meeting or phone call. I am busy and need to know what to do next without reading them all again.
Please give me:
1. A short list of the decisions that were made.
2. A to-do list. For each task, show who owns it and the due date. If the notes do not say, write "not stated" instead of guessing.
3. Anything that is unclear, written as questions I should ask.
4. A friendly follow-up message of about three sentences that I can send to everyone involved.
[PASTE YOUR NOTES HERE]
Check before you send. Read the message and the task owners yourself. A chatbot can give a confident wrong answer, and a wrong name or date is hard to take back once sent.
Time-saver 3
Plan a week of dinners and one grocery list
Plans the dinners and the shopping list in one go.
I need a dinner plan for [NUMBER] people for [NUMBER] nights. My budget for groceries is about [AMOUNT]. I am busy on weeknights, so each dinner should take [NUMBER] minutes or less to cook. Foods we like: [LIST]. Foods we avoid: [LIST, INCLUDING ANY ALLERGIES].
Please give me:
1. A plan with one dinner per night and a one-line description of each.
2. One combined grocery list, grouped by store section, with quantities.
3. Ideas for using leftovers so I buy less.
4. A note on any quantity or price you are not sure about, so I know what to double-check.
Check before you shop. Prices vary by store, so treat the budget as a guess. If anyone has a food allergy, read the labels yourself and do not rely on the chatbot.
A tip for any prompt
Say why you need the answer, and say what to do when the AI is unsure ("write not stated instead of guessing"). Both make the answer easier to trust and easier to check. This is my own advice from experience, not a finding from a study.
What this does not show about the news and prompts
This is a snapshot of news from the past few weeks. Details can change, and some of these stories are still unfolding.
Several items are company claims: PSI's proof, DeepMind's test results and Adobe's product plans. Anthropic's price range is an expected range, not a final price.
The jobs and market numbers show what was reported. Trackers count data center votes differently, and I could not open Challenger's own report.
Where a company's own page was available, I read it: Nvidia, Anthropic, DeepMind, PSI and Goodfire. For some news pages I worked from the excerpts I could open, not every word.
I did not find a source for why Jensen Huang resigned, so none is given.
Nothing here says AI is safe, and nothing here says it is dangerous. These are ten news items, not a verdict.
The news and prompts in short
Money: Anthropic opened its IPO sales tour, and tech stocks slid on fears of tougher AI rules.
Communities: voters in many towns said no to new data centers.
Work: AI was named in fewer layoff announcements in October, and Microsoft offered buyouts to about 60,000 people.
Tools and science: Adobe is adding AI agents, DeepMind released a DNA model for researchers, and a company says AI helped with a math proof that still needs outside checking.
Three prompts can turn a confusing letter, messy notes or a blank dinner calendar into a clear plan in a few minutes.
What this means for you
Treat AI headlines like weather reports. Note what happened, who said it, and what is only a forecast. This edition had money news, community news, work news and science news, and none of it settles the big questions on its own. For your own day, try one of the three prompts on a chore you have been putting off, and check the result before you act on it.
Where this comes from
Each source lists what it backs up. Tap "Back to the sentence" to return to where it was used.
Julian Koplin and Megan Frances Moss (Monash University), written for The Conversation and republished by ScienceAlert on May 8, 2026. Backs up: chatbots like Claude are built on language models. The Conversation's own site would not open for my reader, so I used the ScienceAlert copy.Back to the sentence
OpenAI, "Why language models hallucinate" (September 5, 2025). Backs up: the first stage of training is predicting the next word; the definition of a hallucination; and that it remains a challenge even as newer models make fewer. The page does not say whether the paper was peer reviewed.Back to the sentence
DataCamp, "What is narrow AI?"Backs up: narrow AI is the most common type people meet daily, and spam filtering is an example. This is a company explainer.Back to the sentence
Alissa de Chassey, Scienceline, "From AI to AGI" (April 16, 2026). Backs up: no agreed definition of AGI, no agreed test, and expert disagreement on timing.Back to the sentence
NVIDIA, "NVIDIA Announces CEO Transition" (November 12, 2026). Backs up: Huang's resignation as president, CEO and board member, effective immediately; Colette Kress as interim CEO and CFO; the search for a permanent CEO; Harvey C. Jones as first board chair; and that no reason is stated. This is a company announcement.Back to the sentence
Anthropic, "Anthropic launches roadshow for proposed initial public offering" (November 10, 2026). Backs up: the roadshow launch, the share counts, the expected $70 to $78 range, the Nasdaq application under ANT, and that the registration statement is not yet effective. The page gives no pricing or trading date. This is a company announcement, and Anthropic makes Claude.Back to the sentence
Venable LLP, "Voters Approve Data Center Restrictions at the Ballot Box" (November 4, 2026). Backs up: 41 local ballot measures in 10 states, the "almost unanimously" description (the authors' wording, not a tally), the Edgewater, Florida charter amendment, and that some Illinois measures were advisory. It is a short law firm summary with no vote totals.Back to the sentence
Expert Beacon, "Data Center Ballot Measures: What Voters Decided on November 3, 2026" (checked November 5, 2026). Backs up: different trackers count the measures differently, some counts were still open on November 4, and the results do not settle whether voters were rejecting AI, data centers or the party in power. This is a secondary summary.Back to the sentence
Outsource Accelerator, summary of Challenger, Gray & Christmas's October 2026 report (the report was published November 5). Backs up: AI cited for 1,265 cuts in October, the lowest month of 2026; 121,401 AI-cited cuts so far this year; market and economic conditions at 130,531 as the top reason; and that Microsoft's voluntary exit offers are not in the October total. This is a trade outlet's summary, not the report itself.Back to the sentence
Guru3D, "Microsoft Offers Voluntary Exit Packages to About 60,000 Employees". Backs up: about 60,000 eligible in the U.S., Canada, the U.K. and India; about 27% of 223,000 employees as of June 30; the early December decision date; reports that Copilot, Windows and Xbox teams are included; and that Microsoft has not said how many it expects to accept. The memo date of October 26 comes from other outlets' reports.Back to the sentence
Constellation Research, "Adobe aims its platform, AI agents at creative workflows over assets" (November 10, 2026). Backs up: Creative Agent expanding to Photoshop, Illustrator, Premiere, InDesign and After Effects; Firefly agents that can work for days or weeks, in preview; and Creative Cloud MCP, in preview, for third-party agents. This is an analyst firm's summary of a briefing.Back to the sentence
PSI, research page on the Aaronson-Ambainis conjecture (November 6, 2026). Backs up: the claim, the non-public Anthropic model, Lean accepting the proof, David Turturean leading the work, Jim Halverson's internal audit, the Stanford researchers' reading, and that the page does not say the proof was peer reviewed or independently audited. This is a company's own page.Back to the sentence
Curt Tigges, Matthew Kowal, Thomas McGrath, Owen Lewis and Nick Cammarata, Goodfire, "Neurons as an Interpretable Basis" (November 10, 2026). Backs up: the authors' main argument about single neurons. The page does not say whether it was peer reviewed.Back to the sentence
Words from this post
A to Z, No. 1
About 10 minutes
What AI Is, and What It Is Not
The starting point for everything else in this guide: a plain meaning of AI, how most of it works at the top level, and the most common mix-ups.
Small numbers like [2] point to the source for that fact. Tap one to see the source, then tap "Back to the sentence" to return. The definitions in parentheses, the plain-word choices and the everyday pictures (the recipe card, the mechanic, the handyman) are mine. The numbered sources back up the facts.
Start with the word
AI stands for artificial intelligence (software that does jobs we usually link with human thinking, like recognizing speech, sorting photos or writing text). From here on I will often say smart software. Here, smart means good at a job. It does not mean wise.
The phrase is older than most people think. In 1955, a researcher named John McCarthy used it in a proposal, written with three colleagues, for a summer research project at Dartmouth College.[1] That gathering, held in 1956, is usually counted as the start of the field.[1]
Here is something that surprises people. Experts have never agreed on one exact definition of AI.[2] The OECD, an international group of governments that works on shared rules, chose to define an "AI system" instead of AI itself, and it revised that wording in 2023.[2] In my words, an AI system is a machine-based system that takes in some input and works out how to produce an output, such as a prediction, a piece of content, a recommendation or a decision.[2] The OECD's own examples include a speech system that adapts to your voice and a music app that learns your taste.[2]
Ordinary software follows a recipe. Most AI learns from examples.
Most software you use follows an algorithm (a set of step-by-step instructions a computer follows to do a job). Think of a recipe card. A person wrote every step. Tax software works this way: if a number is above a line, do this, and if not, do that.
Some older AI worked like that too, with rules written by hand.[3] But machine learning (a way of building software in which the computer finds patterns in a very large number of examples, instead of following rules a person typed in) now sits underneath most modern AI.[3]
Here is an everyday picture. An experienced mechanic can hear a worn wheel bearing from the driver's seat. Nobody handed that mechanic a rule like "a hum at 40 miles an hour means a bad bearing." They heard many bad bearings over many years. Machine learning is closer to that than to a recipe card.
The picture has a limit. The mechanic also knows what a wheel is for. Nobody should assume software knows things in that same way, and a later section comes back to this.
The learning step has a name: training (showing an AI a very large number of examples and adjusting it, over and over, until it does the job well). I will call these practice rounds. What comes out of the practice rounds is a model (the finished product of training, a huge collection of numbers that holds the patterns the AI found). When you open a chatbot, you are using a model. I will call it the trained AI.
What a chatbot is doing when it answers you
A chatbot is a program you talk to in everyday language. The kind behind ChatGPT and Claude is built on a large language model (an AI trained on a huge pile of writing so it can predict what text comes next, which is how a chatbot answers you).[4][5] Because it can make new text, it is also called generative AI (AI that creates new text, images, audio or video instead of only sorting what already exists).[6]
Its practice rounds were, at the start, one enormous game of "what comes next." OpenAI describes the first stage of training as learning to predict the next word across huge amounts of text.[5] The AI reads and writes in tokens (small chunks of text, often a word or part of a word), and I will call them word pieces.[7]
Julian Koplin and Megan Frances Moss, researchers at Monash University, call these systems a kind of "souped-up auto-complete" in an article from May 2026.[4] That is a fair starting picture. A chatbot's version looks back over the whole conversation so far, not only the last few words.[8] Still, it is only a starting picture. Later posts will look inside these systems.
Chatbots are only one corner of AI. Speech recognition that adapts to your voice and music apps that learn your taste are AI too, in the OECD's own examples.[2] So is the spam filter that sorts your email.[9]
One-job AI and do-everything AI
Narrow AI (AI built to do one kind of job well) is the kind most people meet every day.[9] A spam filter is a good example.[9] It sorts email and does not do other jobs. I will say one-job AI.
The other idea is artificial general intelligence, or AGI (the idea of a single AI that could learn and do almost any thinking job a person can). I will say do-everything AI. It is a goal some companies talk about, not a settled thing.[10]
Here is where honesty matters. Alissa de Chassey's April 2026 review in Scienceline found no single agreed definition of AGI and no agreed test for it.[10] Experts also disagree about timing. Some AI company leaders say it is a few years away, and some skeptics say decades, or never.[10] Today's best chatbots can do many different jobs, so "one-job" fits them only loosely. For example, a spam filter can only sort your email, but one chatbot can write a letter, explain a bill and translate a recipe, so it is more like a handyman than a specialist, even though it gets some of those jobs wrong.[9][6][5] Whether it counts as "do-everything" depends on whose definition you use.[10] Be careful with anyone who says the answer is obvious, in either direction.
What AI is not
It is not a brain. The parts inside an AI are called neurons (tiny units that take in numbers, do a little arithmetic and pass a number along).[3] The name is a loose label borrowed from biology. These networks may draw inspiration from the brain, but they are not necessarily accurate models of it.[11] I call them switches.
It is not one thing. There is no single "the AI." There are many systems from many makers, built for different jobs. They differ in how much they can do on their own and in how much they keep adapting after they are put to work.[2]
It is not a fact database. A chatbot builds an answer one word piece at a time from patterns it learned.[5] Sometimes the pattern leads somewhere false. That is called a hallucination (a statement from an AI that sounds believable but is false), and I will say confident wrong answer. OpenAI's researchers define it as a plausible but false statement.[5] Their own page says it remains a fundamental challenge for all language models, even though their newer model makes fewer of them.[5] Their explanation is that the tests used to grade AI reward guessing over saying "I don't know," like a multiple-choice exam where a blank answer scores zero.[5] That is their explanation, not settled fact.
It is not all-knowing. A trained AI learned from text up to some date, and that date differs from one model to the next.[12] Some chatbots can also be given tools that bring in newer information, and some cannot.[12] It is fair to ask which kind you are using.
It does not start with goals of its own. The OECD says an AI system's goals can always be traced back to a person who began building it, even when those goals are only implied. Some systems can then work out smaller goals along the way, and how much freedom a system gets is something people hand over.[2] That is a reason to ask who gave it the job. It is not a reason to say AI is safe.
Whether it feels anything is an open question. Koplin and Moss write that most experts deny today's chatbots are conscious, yet they still call it a genuine puzzle.[4] They were responding to the evolutionary biologist Richard Dawkins, who suggested a chatbot may be conscious but did not say he was sure.[4] A chatbot that writes "I feel" proves little either way, because it learned from a huge amount of human writing that talks about feelings.[13]
Where most Americans stand
You are not alone if AI makes you uneasy. In June 2025, the Pew Research Center surveyed 5,023 U.S. adults.[14] About half, 50%, said the growing use of AI in daily life makes them more concerned than excited. Only 10% said more excited than concerned, and 38% said both equally. In 2021 the "more concerned" share was 37%.[14]
Nearly everyone, 95%, had heard at least a little about AI. Yet 53% said they were not too confident, or not at all confident, that they could tell whether something was made by AI or by a person.[14] That is one good reason to learn the basics. This is the most recent Pew survey on the question that I found, and opinions may have moved since.
What this does not show
This is a starting picture. Each idea here gets its own post later, with more care.
Experts disagree about the meaning of AI and AGI, about whether chatbots understand anything, and about whether they could be conscious. This post reports those disagreements. It does not settle them.
The explanation of confident wrong answers comes from OpenAI's own research. The page does not say whether the paper has been through peer review (outside experts checking a study before it appears in a scientific journal), so treat it as one lab's view.
Several sources are explainers from companies, news sites or universities, not scientific studies. The source on the word "neural" is a passage adapted from a 1995 book. The sources show what is generally said and agreed, not the last word.
The Pew numbers describe U.S. adults in June 2025. People in other countries answered differently, and the numbers change over time.
AI changes fast. Anything here about specific chatbots reflects what I could confirm in November 2026.
Nothing here says AI is safe, and nothing here says it is dangerous. Those questions get their own posts, and the answer should rest on evidence.
The short version
AI is smart software: it does jobs we link with human thinking. Experts have not agreed on one exact definition.
Ordinary software follows a recipe. Most AI today learns from examples, through practice rounds.
A chatbot is built on a language model that learned to predict the next word piece from a huge amount of writing.
One-job AI is everywhere. Do-everything AI is an idea people disagree about, and nobody has an agreed test for it.
AI is not a brain, not one thing, not a fact database and not all-knowing. Confident wrong answers still happen.
What this means for you
Treat a chatbot like a quick, well-read helper who sometimes bluffs. It is good for drafts, ideas and plain explanations. For anything that matters, such as money, health, legal or safety questions, check it against a source you trust. And when you hear a big claim about AI, in either direction, ask who is saying it and what test they are using.
Where this comes from
Each source lists what it backs up. Tap "Back to the sentence" to return to where it was used.
Dave Bergmann, IBM Think, "What is machine learning?" (last updated October 2, 2026). Backs up: some AI runs on hand-written rules; machine learning underlies most modern AI; the neurons inside a network each do a math operation.Back to the sentence
Julian Koplin and Megan Frances Moss (Monash University), written for The Conversation and republished by ScienceAlert on May 8, 2026. Backs up: chatbots like Claude are built on language models; the "souped-up auto-complete" description; most experts deny chatbots are conscious, yet the authors call it a genuine puzzle; Dawkins suggested a chatbot may be conscious without claiming certainty. The Conversation's own site would not open for my reader, so I used the ScienceAlert copy.Back to the sentence
OpenAI, "Why language models hallucinate" (September 5, 2025). Backs up: the first stage of training is predicting the next word; the definition of a hallucination; they remain a challenge even as newer models make fewer; the guessing-versus-honesty explanation; ChatGPT also does it. The page does not say whether the paper was peer reviewed.Back to the sentence
IBM Think, "What is generative AI?"Backs up: generative AI creates new text, images, video, audio or code when given an instruction; large language models are the most common base for it; one base model can support several applications.Back to the sentence
Anthropic (maker of Claude), "Context windows" developer documentation. Backs up: each turn the model takes in all the earlier conversation plus your new message.Back to the sentence
DataCamp, "What is narrow AI?"Backs up: narrow AI is built for one specific task; it is the most common type people meet daily; email spam filtering is an example. This is a company explainer.Back to the sentence
Alissa de Chassey, Scienceline, "From AI to AGI" (April 16, 2026). Backs up: no agreed definition of AGI; no agreed test; experts disagree on timing, from a few years to decades or never.Back to the sentence
Mohamad Hassoun, a passage on how neural networks relate to the brain, hosted by Scientific American (adapted from his 1995 book; the page shows no separate date). Backs up: "neural" is a loose label; networks may be inspired by the brain but are not necessarily accurate models of it.Back to the sentence
Scientific American, article on whether to believe a chatbot that says it is conscious (an opinion piece; the author is not named in the text of the page I read, and it shows no date). Backs up: a chatbot's claims of consciousness say nothing either way, because chatbots learn from human writing about inner experience.Back to the sentence
A plain-English walk through a new research post called "Neurons as an Interpretable Basis."
Original by Curt Tigges, Matthew Kowal, Thomas McGrath, Owen Lewis and Nick Cammarata of the research lab Goodfire, published November 10, 2026. Read the original if you want the full technical version. Small numbers like [T1] point to the source. Numbers [T2] and [T3] are two short explainers. The everyday pictures (the town's plumber, the dimmer, the word "bank") are mine.
Why people call AI a black box
You have probably heard AI called a black box (a system where you can see what goes in and what comes out, but not how it got from one to the other). Words go in, an answer comes out, and nobody can say what happened in between. That idea is behind a lot of the fear, because it is hard to trust something you cannot look inside.
The work of looking inside has a name: interpretability (the effort to understand why an AI does what it does). A lab called Goodfire recently published a long research post arguing that single parts inside an AI are easier to read than many researchers assumed.[T1] The post is written for specialists and is very hard going for everyone else. This is the same story, told for the rest of us.
What is inside an AI
The AI they studied is a large language model (an AI trained on a huge pile of writing so it can predict what text comes next, which is how a chatbot answers you). It is an open model called Qwen3.6-35B-A3B.[T1] It is built in layers, and each layer holds enormous numbers of neurons (tiny units that take in numbers, do a little arithmetic and pass a number along).[T2] The name is misleading. They are not brain cells. Picture each one as a small switch with a dimmer.
This AI has about 5 million of those switches.[T1] It does not read whole words. It breaks text into tokens (small chunks of text, often a word or part of a word), and from here on I will call them word pieces.[T3] When a switch puts out a strong signal on a word piece, researchers say it fires (it turns up high). I will say it lights up.
The old worry: switches with too many jobs
For years, many scientists feared the switches were a tangle. A switch that responds to several unrelated things is called polysemantic (the word means "many meanings"). Picture one person in town who is the plumber, the pastor and the mail carrier. If every switch worked like that, you could never read the AI one switch at a time.
The authors say this worry led many researchers to study reorganized versions of an AI's insides instead of the switches themselves. This team went back and studied the switches.[T1]
Finding one: the loud signals do most of the work
At any moment, most switches give only faint signals. Do the faint ones matter? To find out, the team ran an ablation test (switching parts off to see what breaks). They kept only the loudest 7.8% of switch signals and switched off the other 92.2%.[T1]
On ordinary text, the AI's error rose by only 2.9%.[T1] The authors call this property sparsity (only a small share of the parts do most of the work at any moment). As a check, they switched off a similar share at random instead, 87.5% of the signals. That broke it. Its error more than doubled.[T1]
After the trimming, the team gave the AI a short round of fine-tuning (more training after the main training, used to adjust a model). That extra round used about 100 million word pieces. The model's original training used about 34 trillion.[T1] The trimmed AI did worse on some reasoning and knowledge benchmarks (standard sets of test questions used to compare AI models) until that extra training.[T1]
Finding two: most switches have one main job
Each separate thing a switch does is called a mode (one distinct job, and from here I will say job). The team studied 5,468 switch settings in detail. A switch can swing high or low, and the two ends often do different things, so they counted each end separately. Nearly half, 47%, had a single job, and 69% were dominated by one job.[T1]
How do you find a switch's job? You trace its circuit (the chain of switches that feed into it, and the chain that uses what it sends out).
The post walks through one switch with two unrelated jobs, oral surgery and marine environments.[T1] Here is my own picture of how that can still make sense, an illustration and not a finding: you know "bank" means one thing next to "river" and another next to "loan," because of the words around it. A switch with two jobs is not automatically a mess.
Finding three: you can describe a switch in a sentence
The team used AI helpers to write a short description of each switch's job. A few examples from the post:[T1]
A switch that lights up when the AI interrupts itself with the word "Wait." Turn it up and the AI settles on an answer too fast, even a planted wrong one. Turn it down and it second-guesses more.
A switch that notices a date crossing into the next month or year, like "June" in "31 May to 1 June." It works even when the earlier date is hundreds of word pieces back.
A switch for comparisons of a person's condition with their age, like "mature for her age." Swap age for weight or cost and it goes quiet.
Then they tested whether the descriptions were any good.[T1]
Spot the switch. A judge AI that read only the description guessed whether the switch lights up in a passage. Against random text it was right 95% of the time, and 89% against passages from the same area. A coin flip would get 50%.
Write to order. AI writers saw only a description, plus feedback after each try, and were asked to write text that lights up the switch more than anything in a 10 billion word-piece collection of real text. It worked for 94% of the switches tested.
Guess the strength. The hardest test. From the description alone, guess how brightly the switch lights up on each word piece, then plug those guesses into the AI in place of the real signal. That is a simulation (a stand-in used in place of the real thing). The guesses repaired about 73% of the damage that a perfect guesser could repair.
The hardest case: Colorado and Tornado
The authors also took on a tricky example, in a different model from the main study (Qwen3-32B). One switch is mostly about Colorado, but it also lights up on code that uses a Python framework called Tornado. About 98.8% of its effect comes from text about Colorado. The model gave Colorado 30% and the correct answer, California, 0.7%. When the switch was suppressed, Colorado fell to 1% and California rose to 4.7%.[T1]
My reading, not the authors': a stray link inside an AI can push it toward a wrong answer, and looking at the switches is one way to find it.
What this does not show
The main study covers one open AI model. Other models may differ.
The faint signals are still not fully explained. The authors list open questions and future work in the post.[T1]
The hardest test still falls short of perfect, and it covers only each switch's main job.
The trimmed AI did worse on some reasoning and knowledge tests until the extra training.
This is a research post on the lab's own website. The page does not say it has been through peer review (outside experts checking a study before it appears in a scientific journal), so treat it as a strong early report, not settled fact.
I checked the numbers and examples above against the post. Statements about what the field believed before are the authors' own framing, and I did not test them.
My reading, not the authors': nothing in the post suggests a mind or a will. What it describes is a very large pile of arithmetic that people are getting better at reading. That does not make AI safe by itself. It is a reason to ask for more inspection, not less.
The short version
An AI is built from millions of small switches.
Scientists feared the switches were a tangle of many-job parts. Mostly, they are not.
The loud signals do most of the work. Switching off 92% of the quiet ones cost little on ordinary text.
Most switches have one main job. A few have more than one.
A one-sentence description of a switch is good enough to predict it, and even to write text that sets it off.
What this means for you
When someone says nobody knows how AI works, the fair reply is: not fully, but more than the headlines suggest, and the knowledge is growing. A better question than "is it alive?" is "who is checking, and can they show their work?"
Where this comes from
Each source lists what it backs up. Tap "Back to the sentence" to return to where it was used.
Curt Tigges, Matthew Kowal, Thomas McGrath, Owen Lewis and Nick Cammarata, Goodfire, "Neurons as an Interpretable Basis" (November 10, 2026). Backs up: every study finding, number and example above, including the model studied, about 5 million switches, the 7.8% and 92.2% split, the 2.9% rise in error, the random-switch-off check, the fine-tuning and original training sizes, the 5,468 switch settings with 47% and 69%, the three example switches, the 95%, 89%, 94% and 73% test results, and the Colorado and Tornado example. This is a lab's own research post. The page does not say whether it was peer reviewed.Back to the sentence
Dave Bergmann, IBM Think, "What is machine learning?" (last updated October 2, 2026). Backs up: what a neuron inside a network does, a small math operation. This is a company explainer.Back to the sentence