Skip to content
Docs
Open Bearly

How AI works

Five short ideas that help you use AI well, from how it writes an answer to when to check it, each with an example to try.

⋯
View as markdown

A chat AI writes one word at a time, picking a likely next word each time. It isn’t looking anything up while it does this.

Pick the word you think comes next.
The capital of France is  

After “The capital of France is”, the next word is almost always “Paris”. After “Our team meeting is on”, Monday, Tuesday, Wednesday, and Thursday are close to even, yet the answer reads just as confidently. A confident sentence isn’t evidence that it’s right.

Because it writes what sounds likely, AI can produce a citation that looks right and doesn’t exist. Try it: three of the five citations below are real.

Is this source real? Choose Real or Made up.
Reference 1
Turing, A. M. (1950). “Computing Machinery and Intelligence.” Mind, 59(236), 433–460.
RealMade up
Or swipe the card.

The real ones are Turing’s “Computing Machinery and Intelligence” (1950), Shannon’s “A Mathematical Theory of Communication” (1948), and Kahneman’s Thinking, Fast and Slow (2011). The other two, a 2019 article in a journal called Review of Machine Reading and a 2008 book called The Paris Lectures on Machine Judgment, were invented to look right. Open a source before you cite it.

Without sources, AI answers from what it learned in training, which can be out of date or wrong. In Bearly, the Web setting decides whether a chat can search the web: open the sliders button in the message box and choose Auto or Off under Web. With Auto, Bearly searches when a question needs current information, and the answer cites numbered sources you can open.

Web is set to Off. Send the question.
Ask about something recent
What did the city council decide last night?
Web

Asked about last night’s council meeting with Web set to Off, the answer can only offer general background. With Auto, Bearly looks it up and links the pages it used, so you can check each claim. See Web, Image & Deep Research for more.

A vague request gets a generic answer. Say who it’s for, how long it should be, and how it should end.

Drag the slider toward Specific.
Length and who it's forTone and what to bringHow it should end
Your prompt
write a welcome email
The answer
Welcome to the team! We're so glad to have you. We look forward to working together and can't wait to see everything you'll accomplish here…

“Write a welcome email” gets a few friendly, empty lines. “Write a 120-word welcome email for three new teammates starting Monday. Friendly. List what to bring on day one. End with who to ask for help.” gets an email you could send after a quick edit.

AI is a fast first draft. Open the source before you rely on any of these:

  • Quotes: find them in the original before you repeat them.
  • Numbers: statistics, prices, and dates come from the source, not the summary.
  • Deadlines: confirm them in the original email, contract, or policy.
  • Legal or safety-critical claims: read the source and ask the person responsible.

To get sources in the first place, use Web Search, Scholar for research papers, or attach the file and ask Bearly to cite the page.

  1. AI predicts the next word, so it sounds sure even when it’s guessing.
  2. It can invent sources that look real. Open them before you cite them.
  3. Set Web to Auto when the answer depends on recent or local facts.
  4. Say who it’s for, how long it should be, and how it should end.
  5. Check quotes, numbers, deadlines, and legal or safety-critical claims.