Lesson 13 of 32 · Part One, the principles

The Search for Absolute Truth

Fact-Checking

These machines state a falsehood in exactly the same confident voice they use for a fact, because they have no way of telling the two apart. So build the checking into the asking: demand a source beside every claim, give the model explicit permission to say "I am not sure", ask which of its own statements it is least certain about - and then open the links yourself. Trust is pleasant. Verification is power.

Reach for it whenWhenever a number, a case name, a date or a quotation is about to leave your hands and appear in a memo, an article, a client's inbox or a slide behind you.

How to use The Search for Absolute Truth in ChatGPT

The whole technique rests on one habit here: make it search, and make the link sit next to the claim rather than in a list at the bottom. Ask for your facts as a table - claim, source, link, how sure you are - because a table makes a missing citation visible in a way that prose hides. Then do the part no model can do for you: open every link. When one does not say what was claimed, paste the line back and tell it so; that correction is the most useful message you will send all session. For anything you check often, put "cite a source for every factual claim and tell me when you cannot" into Custom Instructions and stop asking.

  • Turn on web search before asking for any fact you will publish.
  • Ask for a table: claim, source, link, confidence.
  • Open every link - a citation is not evidence until you read it.

From GenAI for Everyone: Prompt Engineering Handbook, First Edition 2025, by Raja Rajeshwari Seshadri and Shreeram Iyer.

The whole technique rests on one habit here: make it search, and make the link sit next to the claim rather than in a list at the bottom. Ask for your facts as a table - claim, source, link, how sure you are - because a table makes a missing citation visible in a way that prose hides. Then do the part no model can do for you: open every link. When one does not say what was claimed, paste the line back and tell it so; that correction is the most useful message you will send all session. For anything you check often, put "cite a source for every factual claim and tell me when you cannot" into Custom Instructions and stop asking.

From the book Use web search for anything recent and ask for links beside every fact. Open the links yourself; if one does not say what was claimed, tell ChatGPT and ask it to correct the answer.
Worth knowingWithout search switched on it will invent citations fluently and without embarrassment - plausible case names, section numbers that do not exist, a real professor paired with a paper they never wrote. With search on it is far better but not safe: it will sometimes cite a page that genuinely exists and genuinely does not contain the claim. Treat its confidence rating as a guess about a guess, useful only for deciding what to check first. And it is obliging: push back on something it got right and it may apologise and withdraw a correct answer, so when you challenge a claim, ask it to defend or correct rather than simply agree.
  • Turn on web search before asking for any fact you will publish.
  • Ask for a table: claim, source, link, confidence.
  • Open every link - a citation is not evidence until you read it.
  • Paste back any link that does not support its claim.
  • When challenging a claim, ask it to defend or correct, not agree.
Ask for the Receipts

Instead of

Tell me cases on consumer rights for defective goods.

ChatGPT version

I am a second-year law student in Kochi preparing a moot-court memo on defective goods under the Consumer Protection Act, 2019. Use web search for this - I will be citing it in a memo that a judge reads.

First, explain the main legal principles on defective goods under the 2019 Act: how defect is defined, what remedies the forums can order, and how liability sits between manufacturer and seller.

Then give me the authorities as a table with these columns, one row per authority:

| Claim it supports | Section or case name | Full citation | Link | How sure are you it exists |

Rules for that table:
- A section is only a row if you can link the text of the Act.
- A case is only a row if you can link a judgment on an official or court-run database. Not a blog, not a question bank.
- In the last column write high, medium or low, and in one phrase say why. If the answer is low, leave the row in and tell me so - I would rather chase a doubtful case than miss a real one.
- Where you cannot find a citation at all, write a row saying so instead of substituting a case you half remember.

Last, list the three claims in your own explanation that a judge is most likely to ask me to support, so I know what to read first.
Open ChatGPT 1,226 characters

Then loop it

  1. Of the three cases you cited, which are you least certain about, and what exactly makes you doubt it - the name, the year, the forum, or the proposition you attached to it?
  2. I opened your second link. The judgment is real but says nothing about the point you attached to it. Correct the row, and tell me whether the proposition survives with a different authority or has to go.
  3. Now search for the text of section 84 of the Act, quote it exactly as printed, and give me the link to the page you took it from.

Why it worksDemanding citations and permission to say "I am not sure" stops invented cases from slipping into her memo - and she still verifies every one.

Triangulate the Statistic

Instead of

How many people are affected by Assam floods every year?

ChatGPT version

I am a journalism student in Guwahati writing an article on floods in Assam. Search the web for this; nothing goes in the piece unsourced.

Build me a five-year table, one row per year, with these columns:

| Year | People affected | Districts affected | Source named | Link | Date on the page |

Then three things underneath it.

**Where the numbers disagree.** For any year where a government figure and a newspaper figure differ, show both, say how far apart they are, and tell me the likely reason - peak-day count against cumulative total for the season, affected against displaced, revenue circles against districts. That distinction is usually the whole story.

**What you could not confirm.** List every cell above you filled from a secondary source, or could not fill at all. Do not interpolate a missing year.

**What to ask for.** The three documents I should be chasing myself - which authority publishes them, and whether they are online.

Keep your own commentary out of it. I want figures I can trace, not a summary.
Open ChatGPT 1,031 characters

Then loop it

  1. Find the official source for the most recent year's figure - the authority's own bulletin, not a paper reporting it - and give me the link and the date printed on it.
  2. Here is my draft paragraph. Go through it claim by claim and mark each one verified, unclear or unsupported, with the link beside the verified ones: [paste paragraph]
  3. Two of my claims came back unsupported. For each, tell me whether a source exists that I have not found, or whether the claim has to come out of the piece.

Why it worksAsking for sources, years and disagreements turns a confident-sounding number into something he can trace and check before it goes to print.

Question the Claim

Instead of

Is it true that eating after 8 pm makes you fat?

ChatGPT version

I am a gym trainer in Amritsar. Several clients are convinced that eating after 8 pm makes you gain weight. Search the web before answering; I need to know where this comes from.

Sort what you find into three labelled sections and nothing else:

**Well established** - what the evidence supports, with the kind of study behind each point (randomised trial, long-term observational study, review of many studies) and a link.
**Still debated** - where researchers genuinely disagree, and what the disagreement is actually about.
**Myth or overstated** - claims that circulate widely with little behind them, and where this particular 8 pm version seems to have come from.

For every point, name the source and say in a phrase how strong it is - how many people, how long, and whether anyone has repeated the finding. Where the honest answer is that the evidence is thin, say the evidence is thin rather than picking a side.

I will pass this on to clients as general information only, not as medical or dietary advice, so flag anything I should not repeat without a qualified person involved.
Open ChatGPT 1,091 characters

Then loop it

  1. Find two recent studies on meal timing and body weight through search, and for each tell me the number of participants, the length, who funded it, and what it actually measured.
  2. Of everything above, which single claim would most embarrass me if a dietitian challenged it in front of a client? Defend it properly or withdraw it.
  3. Now write a short, honest WhatsApp note for clients in Punjabi, saying what is known and what is not, and telling them to see a doctor or dietitian for advice about their own body.

Why it worksSorting facts into established, debated and myth makes the AI admit uncertainty instead of picking a side with false confidence.

Questions

What is The Search for Absolute Truth in prompt engineering?

These machines state a falsehood in exactly the same confident voice they use for a fact, because they have no way of telling the two apart. So build the checking into the asking: demand a source beside every claim, give the model explicit permission to say "I am not sure", ask which of its own statements it is least certain about - and then open the links yourself. Trust is pleasant. Verification is power.

When should I use this technique?

Whenever a number, a case name, a date or a quotation is about to leave your hands and appear in a memo, an article, a client's inbox or a slide behind you.

How do I apply it in ChatGPT?

The whole technique rests on one habit here: make it search, and make the link sit next to the claim rather than in a list at the bottom. Ask for your facts as a table - claim, source, link, how sure you are - because a table makes a missing citation visible in a way that prose hides. Then do the part no model can do for you: open every link. When one does not say what was claimed, paste the line back and tell it so; that correction is the most useful message you will send all session. For anything you check often, put "cite a source for every factual claim and tell me when you cannot" into Custom Instructions and stop asking.

What are the limits of this technique in ChatGPT?

Without search switched on it will invent citations fluently and without embarrassment - plausible case names, section numbers that do not exist, a real professor paired with a paper they never wrote. With search on it is far better but not safe: it will sometimes cite a page that genuinely exists and genuinely does not contain the claim. Treat its confidence rating as a guess about a guess, useful only for deciding what to check first. And it is obliging: push back on something it got right and it may apologise and withdraw a correct answer, so when you challenge a claim, ask it to defend or correct rather than simply agree.

How do I improve the answer after the first try?

Never stop at the first factual answer. Read it and underline every number, name, date and quote - these are where errors hide. Then loop: "Give the source for each of these," "Which of these are you least sure about?", "Search the web and confirm this." Gemini's double-check button and web search in all three apps help a great deal, but a source link is not proof until you open it. The final check belongs to you: read the original source yourself before you rely on it.

Loop it

Never stop at the first factual answer. Read it and underline every number, name, date and quote - these are where errors hide. Then loop: "Give the source for each of these," "Which of these are you least sure about?", "Search the web and confirm this." Gemini's double-check button and web search in all three apps help a great deal, but a source link is not proof until you open it. The final check belongs to you: read the original source yourself before you rely on it.

This lesson on the other tools

The technique is the same; the way you write it is not. Claude, Gemini.Perplexity.