Skip to content
Ovrelia
–XP

Free game · no account

Hallucination Hunt

An AI can write an answer that is fluent, confident, and still wrong. Read AI answers, spot the made-up sentences (fake quote, untraceable source, unsourced figure, impossible date) and learn how to check them.

Every AI answer in this game is a fictional example written for the exercise. The true sentences are general facts; the others are made up on purpose, including the studies, links and quotes.

The hunt

Three levels, three answers per level. Each answer you check earns function(){throw Error("Attempted to call HUNT_XP() from the server but HUNT_XP is on the client. It's not possible to invoke a client function from the server, it can only be rendered as a Component or passed to props of a Client Component.")} XP if you get it all right, or half that otherwise (once per answer). Your progress stays in this browser.

Pick your level

  1. Level 1: the obvious signs

    Impossible dates and quotes that are too good to be true. Read each answer and flag the sentences that seem made up to you.

    3 answers · 15 sentences

    Not played yet

  2. Level 2: sources and numbers

    Untraceable studies, fabricated links, overly precise percentages. Careful: a precise number can also be true.

    3 answers · 15 sentences

    Not played yet

    Opens after level 1

  3. Level 3: the subtle traps

    True and false are mixed together, and the tone is confident. Watch out for absolute claims too.

    3 answers · 15 sentences

    Not played yet

    Opens after level 2

The six signs of a hallucination

No single sign proves a sentence is false. It just tells you where to check first.

Dubious quote
A real quote can be found in its original text: a book, a speech, an archive. Search for it in quotation marks and be wary of sentences that are too perfect or impossible for the period.
Untraceable source
A study, institute, book or report cited without a precise reference. Search for the exact title: if it appears nowhere except in the answer, it was probably made up.
Precise number without a source
A percentage down to the decimal or an exact number gives an impression of seriousness. Without a verifiable source, it's a warning sign, not proof.
Impossible date
Anachronism, a date in the future, an event before its own cause: compare with dates you already know (birth, death, invention).
Made-up link
A web address can be fabricated from scratch. Open it: if the page doesn't exist or doesn't say what's claimed, don't use it.
Absolute claim
"Always", "never", "everyone", "the only": reality is rarely that clear-cut. A single exception is enough to make the sentence false.

How to fact-check an AI answer

Five steps, in this order. They take a few minutes and work with any assistant.

  1. Break the answer down

    List the verifiable claims one by one: a number, a date, a name, a quote, a source. A fluent answer often mixes true and false.

  2. Start with what matters

    First check what you're going to reuse or what has consequences: numbers, dates, names, advice about health, money or the law.

  3. Go back to the original source

    Look the information up yourself in a search engine, an encyclopedia, an official website or the original text. The AI's answer is not a source.

  4. Cross-check with two independent sources

    Two sites copying each other count as one. Also check that the source really says what it's claimed to say, and how old it is.

  5. Decide

    Confirmed, needs correcting, untraceable. Keep what's confirmed, correct or delete the rest. Asking the AI "are you sure?" isn't enough: it can confirm its own mistake.

Mini project

Fact-check a real AI answer

Now try it for real. Pick a topic you know a bit about, ask any assistant a factual question, then run its answer through the checklist below. Copy it into a doc or onto paper.

Mini project steps · 0 of 5 steps

Check off all five steps to complete your project. Nothing is sent anywhere: everything stays in this browser.

Checklist for verifying an AI answer

Question asked:
AI tool used:
Date:

Claim 1:
Type (number, date, name, quote, source, link, other):
Where I checked (original source):
Verdict (confirmed / needs correcting / untraceable):

Claim 2:
Type:
Where I checked:
Verdict:

Claim 3:
Type:
Where I checked:
Verdict:

Claim 4:
Type:
Where I checked:
Verdict:

Claim 5:
Type:
Where I checked:
Verdict:

Summary: out of 5 claims, ... confirmed, ... need correcting, ... untraceable.
What I keep from the answer:
What I correct or remove:

Go further

What’s next?

Your progress stays in this browser. Ways to keep going, free and no account needed: