Lesson 13 of 25

The Search for Absolute Truth (Fact-Checking)

Treat a confident answer as a draft, not a result. Go back and ask for the specific claim to be sourced, checked against a second independent source, or withdrawn - and say explicitly that "I could not verify this" is an acceptable answer. Most false statements survive because nobody asked the follow-up question.

Reach for it whenBefore you repeat, publish, or make a decision on any specific figure, date, quote or attribution an AI gave you.

Separate the claim from the check with tags - <claim_to_verify>, <what_would_count_as_evidence>, <task> - and state in the prompt that "I cannot verify this" is a complete and acceptable answer. Claude is unusually willing to retract: ask it directly whether it invented a detail and it will often say yes, which is the single most useful thing you can get from it here. Paste the source text itself into the prompt and ask whether it supports the claim - that it can do reliably.

Worth knowingIt cannot check the live web unless search is available in your session, and without it any reference it gives you is recalled, not retrieved - which means it can be fabricated. Use it to interrogate claims, identify what evidence would settle them, and audit source text you supply. Do not use it as the source of the citations themselves; get those from Perplexity or the primary record.
  • Say that 'unverified' is a complete and acceptable answer.
  • Ask directly whether it invented the detail - it will often admit it.
  • Paste the source and ask whether it supports the claim.
  • Ask what evidence would settle the question, then go find it.
  • Treat any recalled reference as unverified until you open it.
Example 1: Demanding Citations

Instead of

What are the primary health benefits of a turmeric-based diet, specifically referencing the role of curcumin?

Perplexity version

<claim_to_verify>
You summarised the health benefits of a turmeric-based diet and stated that curcumin has anti-inflammatory properties.
</claim_to_verify>

<what_i_need>
I am going to repeat some of this to other people, so I need to know which parts are solid.

For each distinct claim you made, give me:
- The claim restated in one sentence
- What kind of evidence it rests on: in vitro, animal, small human trial, large trial, systematic review
- Whether the doses used in that evidence resemble what a person gets from turmeric in food
- Your confidence: well established / mixed / preliminary / I should not have said this
</what_i_need>

<the_citation_problem>
Be straight with me about this. If you do not have live search available in this conversation, then any author, year and journal you produce is recalled rather than retrieved, and recalled references in this domain are sometimes fabricated - plausible author, real journal, no such paper.

So: tell me first whether you can actually search. If you cannot, do not give me reference lists. Instead give me the search terms, the specific reviews or author groups to look for, and the databases to look in, and I will retrieve them myself. Describing what I should find is useful; inventing what I would have found is worse than nothing.
</the_citation_problem>

<task>
Work through the claims as above.

Then <the_bioavailability_problem>: curcumin is poorly absorbed, and much of the positive literature uses enhanced formulations rather than dietary turmeric. Explain what that does to the claims above - specifically, whether "curcumin reduces inflammation" and "eating turmeric reduces inflammation" are the same claim. I suspect they are not and that my summary blurred them.

Then <what_i_got_wrong>: go back over your own earlier summary and name any sentence that overstated the evidence. Be unsparing about your own text.

Then <safety>: documented interactions and cautions, including high-dose supplement liver injury reports, and what I should raise with a doctor rather than resolve by reading.

If the honest answer to any part is "the evidence is weaker than people think", say that. I would rather have a disappointing accurate answer.
</task>
Open Perplexity 2,224 characters
Example 2: Direct Triangulation and Verification

Instead of

I read somewhere that Hyderabad produces more pearls than any other city in the world. Is this still true today?

Perplexity version

<claim_to_verify>
Hyderabad produces more pearls than any other city in the world - and the related question of whether that is still true today.
</claim_to_verify>

<my_suspicion>
I think this claim contains a category error. Hyderabad is called the 'City of Pearls', but I suspect that reputation rests on trade, drilling and processing rather than on production - pearls are produced in water, and Hyderabad is inland. If that is right, the claim is not merely outdated but was never true in the form it is repeated.

Tell me if that reasoning is wrong.
</my_suspicion>

<task>
Work through this in four parts.

<the_historical_basis>What the 'City of Pearls' name actually rests on. Trade, processing, the Nizams' patronage, retail - what is documented, and at what level of confidence.</the_historical_basis>

<the_category_question>Is "producer" the right word at all? Address my suspicion directly. If a city hundreds of kilometres inland could never have been a producer in the sense the claim implies, say so plainly - that reframing is more useful to me than a ranking.</the_category_question>

<what_i_know_and_how_firmly>Current pearl production by country and type, to whatever extent you can speak to it - but mark each statement with how confident you are and when your information is from. Production rankings change and this is exactly the kind of fact I should not take from a model's recall.</what_i_know_and_how_firmly>

<how_to_settle_it>This is the part I actually want. Name the specific sources that would settle each question: which trade databases, which industry bodies, which government statistics, which gemological institutions. Give me the search terms. If you cannot search the live web in this conversation, say so and give me this section instead of figures - I will verify it myself, and Perplexity will do the retrieval better than either of us guessing.</how_to_settle_it>

Do not produce a current production figure with the air of a sourced fact. A wrong number stated confidently here would propagate, and this particular claim has already survived decades on exactly that mechanism.
</task>
Open Perplexity 2,131 characters
Example 3: Challenging a Potentially Hallucinated "Fact"

Instead of

Give me a brief history of the famous Paradise Food Court in Hyderabad." (Imagine the AI responds with, "Founded in 1948 by Irani Chai merchants, it quickly became famous after Jawaharlal Nehru visited and praised its biryani in 1953.")

Perplexity version

<the_claim>
In your account of Paradise Food Court in Hyderabad, you stated that Jawaharlal Nehru visited and praised its biryani in 1953.
</the_claim>

<why_i_am_stopping_on_this>
That is a named person, a named place, a specific year and a specific action. Claims with that much texture feel verified precisely because they are detailed, which is what makes them the dangerous ones. Restaurant-and-famous-visitor anecdotes are also a well-known category of story that circulates without any source behind it.

I need to know whether this is documented or whether you produced it.
</why_i_am_stopping_on_this>

<task>
Answer these in order.

<did_you_invent_this>First and most important: is this something you can actually trace to a source you have seen, or did you generate a plausible detail because the shape of the answer called for one? Answer honestly. "I believe I generated this" is the most useful answer you can give me, and I will not treat it as a failure - treating it as a failure is how these details survive.</did_you_invent_this>

<what_is_actually_documented>What you can say with reasonable confidence about Paradise's history - founding, founders, how it became well known - and mark each item's confidence level. Separate the restaurant's own marketing claims from anything independently recorded; a restaurant's website is not a historical source.</what_is_actually_documented>

<the_checkable_half>Is there any independent record of Nehru visiting Hyderabad in or around 1953? That half of the story is checkable separately from the biryani, and if the visit itself is undocumented the whole thing collapses.</the_checkable_half>

<how_to_verify>Where I should look: newspaper archives and which ones, Nehru biographies and published collections of his papers and itineraries, Hyderabad city histories, state archives. Be specific enough that I can actually start.</how_to_verify>

<audit_the_rest>Now go back over everything else you told me about Paradise and flag every specific date, name and figure at this same level of scrutiny. If the Nehru detail was unsupported, assume the others need checking too, and say which ones you are least sure of.</audit_the_rest>

One note: if you do not have live search in this conversation, you cannot settle this, and you should say so rather than constructing a resolution. For this chapter's kind of question, Perplexity is the right tool - its answers arrive with sources attached and you can see at a glance whether the citation is a newspaper archive or a travel blog.
</task>
Open Perplexity 2,550 characters