Start here
How AI actually works
No math required. Here is what's happening when you type a question into ChatGPT, Claude or Gemini, and why you should still check the answer.
Facts checked September 24, 2026
It's predicting words
The chatbots people use today run on large language models. A language model learns by reading an enormous amount of text and practicing one skill: guessing the next word. OpenAI describes this first stage as "predicting the next word in huge amounts of text."1 The federal government's AI standards body, NIST, puts it the same way: these models predict the next word or token, and their answers approximate the patterns in their training data.2
Nobody writes the rules by hand. Anthropic, the company behind Claude, says its models "aren't programmed directly by humans" and are trained on large amounts of data instead.3 That's why AI can write a poem, summarize a contract and explain a recipe. It has seen millions of examples of each.
Think of it this way
Picture the world's most well-read autocomplete. It is very good at producing text that sounds right. Sounding right and being right are usually the same thing, but not always.
Why it makes things up
Because the model is producing likely-sounding text, it can produce false text that sounds just as confident as true text. The industry calls this a hallucination. OpenAI defines hallucinations as "plausible but false statements."1 NIST uses the word "confabulation" and notes the AI presents the error confidently.2
The companies say this openly. Anthropic's help center warns that Claude can show quotes that "look authoritative" but "are not grounded in fact."4 Google's Gemini privacy page says these systems "can hallucinate and present inaccurate information as factual."5
When it went wrong in real life
Federal court, New York, 2023
Mata v. Avianca. A lawyer used ChatGPT for legal research and filed a brief citing court cases that did not exist. The judge found that ChatGPT "fabricated the cited cases" and imposed a $5,000 penalty on the lawyers and their firm.6
Federal appeals court, 2024
Park v. Kim. A lawyer cited a case that doesn't exist and told the court it "was suggested by ChatGPT." The Second Circuit referred her to its grievance panel.7
British Columbia tribunal, February 2024
Moffatt v. Air Canada. The airline's website chatbot gave a passenger wrong information about bereavement fares. Air Canada argued the chatbot was responsible for its own words. The tribunal disagreed and ordered the airline to pay about CA$650 plus interest and fees.8,9
hitechmike's take
Use AI like a sharp intern: great for a first draft, a summary or ideas. For anything that matters (money, health, law, a name or a number) ask it for its source, then open the source yourself.
Does it know what happened today?
On its own, no. A model learns from data collected up to a point in time. AI companies publish that date: Anthropic lists a "reliable knowledge cutoff" for each Claude model,10 and its help center notes Claude may not have the most up-to-date information about current events.4
Many chatbots can now search the web to fill that gap. OpenAI says ChatGPT "can search the web to answer questions with current information and links to relevant sources."11 Google says some Gemini answers are grounded on Search results.5 When you see links in an answer, the AI searched. When you don't, it's answering from what it learned in training.
Who is using it
Americans are using AI more and trusting it warily. In a June 2026 Pew survey, 52% said they were more concerned than excited about AI in daily life. Only 9% were more excited than concerned, and 37% said both equally.15
What it costs in energy
Every answer runs on computers in a data center. Google published a detailed measurement in August 2025: the median Gemini text prompt uses 0.24 watt-hours of electricity and about 0.26 milliliters of water, "about five drops." Google says that's like watching TV for less than nine seconds, and that the energy per prompt fell 33-fold in one year.16
The bigger picture is growth. The International Energy Agency says data centers used around 415 terawatt-hours in 2024, about 1.5% of the world's electricity, and projects that will more than double to around 945 terawatt-hours by 2030.17
Sources for this page
- Why language models hallucinate. OpenAI.
- Artificial Intelligence Risk Management Framework: Generative AI Profile (NIST AI 600-1). National Institute of Standards and Technology.
- Tracing the thoughts of a large language model. Anthropic.
- Claude is providing incorrect or misleading responses. What's going on?. Anthropic Help Center.
- Gemini Apps Privacy Hub. Google Gemini Help.
- Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), Opinion and Order on Sanctions. U.S. District Court, S.D.N.Y. (via CourtListener).
- Park v. Kim, No. 22-2057, 91 F.4th 610. U.S. Court of Appeals, 2nd Circuit (via FindLaw).
- Moffatt v. Air Canada, 2024 BCCRT 149. BC Civil Resolution Tribunal (via CanLII).
- BC Tribunal Confirms Companies Remain Liable for Information Provided by AI Chatbot. American Bar Association, Business Law Today.
- Models overview. Anthropic (Claude Platform docs).
- ChatGPT search. OpenAI Help Center.
- Americans and AI 2026: Chatbots, smart devices and views on impact. Pew Research Center.
- Accelerating the next phase of AI. OpenAI.
- The Gemini app has surpassed 1 billion monthly users. Google.
- Democrats are now more worried than Republicans about AI and its impact on jobs. Pew Research Center.
- Measuring the environmental impact of AI inference. Google Cloud Blog.
- Energy and AI: Executive summary. International Energy Agency.