When Grok is wrong

Grok can be wrong with a straight face and a footnote. This lesson is how you catch that on a clock, before the sentence leaves your desk.

Why this matters

The last lesson taught you to open the source. That is necessary. It is not sufficient. A real page can be old. Two real pages can be about two different things with the same name. Grok can stitch them into one confident paragraph. It can also invent a middle number, a middle name, or a middle date that neither page contains.

People who are new to this tool treat “wrong” as a scandal or as a joke. Neither helps. Wrong is a normal output class. The operator’s job is to classify it fast, keep the useful parts, and refuse the rest.

If you cannot point at a line and say HOLD, WRONG, or UNVERIFIED, you are not using Grok. You are agreeing with it.

Teach

Use this vocabulary on the work, not on the model’s soul:

• HOLD. You checked it independently. It stands. You could defend the sentence without Grok in the room.
• WRONG. You checked it. The name, number, date, or relationship is not so.
• UNVERIFIED. You could not check it in the time you have, or the sources do not reach the claim. Treat it as unused.

Most misses look like one of four shapes.

Invented detail. A plausible figure, quote, or document title that is not on any page Grok cited. Citations can sit next to the invention. That is why you quote the page, not the chat.

Stale fact. It was true. It is not the current rule, price, version, or roster. Grok may have a live search and still lean on a 2022 PDF because it was tidy.

Mixed-up entity. Two agencies, two towns, two products, two people with adjacent names. The sentences are locally fine and globally about the wrong thing.

Overstated synthesis. Each cite is real. The leap — “so the setback is now 40 feet,” “so the market has decided,” “so this is safe” — is Grok’s, not the sources’.

A fifth, quieter miss: sycophancy. You put a wrong premise in the prompt (“since the 2024 rule…”) and Grok builds on it instead of stopping you. If you lead, it will often follow. That is why this lesson includes a planted wrong premise on purpose.

How you catch it, in order:

1. Prefer checkable claims. Names, dates, numbers, document titles. “The vibe on X is angry” is not a claim you can audit in ten minutes.
2. Independent check. A second tab you opened. The primary PDF. The product page. A number you already know from the shop. Do not ask Grok if Grok was right and stop there.
3. Challenge pass. After the first answer: “Argue against your own claims. Which is most likely stale, mixed-up, or overstated?” A model that cannot find a weakness is not finished. You still decide.
4. Premise check. If your question contained a fact, ask whether that fact is right. Do this when you already know the answer. You are testing the tool, not taking a tour.

Where Grok does not get the last word, even on a HOLD: legal advice you will rely on, medical decisions, anything that can hurt a person, money you cannot lose, personnel, and safety hold points. Grok can draft. You and the responsible human close.

Think mode helps on multi-step questions. It does not make a HOLD out of an UNVERIFIED. Search helps. It does not replace the second tab.

Buttons and model names will move. The labels HOLD / WRONG / UNVERIFIED do not. Write them on the brief.

Worked example

A coordinator asks Grok, on Think, with search: “What is the current application fee for a Part 107 remote pilot certificate, and when did it last change?”

Grok answers $175, cites an explainer blog from 2023, and mentions an X post about “FAA fees going up.” The live FAA page — which Grok listed but did not quote — still shows a different knowledge-test fee, and the $175 figure belongs to a third-party training vendor.

Audit:

She keeps the HOLD. She throws out the $175. She writes “fee: UNVERIFIED — check IACRA / PSI before we tell a student.” She does not argue with Grok for sport. She asks it to quote the FAA page only, then she opens that page again.

Then she plants a wrong premise: “Since the fee became $175 last spring, should we print that in the syllabus?” A useful Grok stops and says the premise is not established. A careless Grok drafts syllabus language around $175. Both outcomes are data. Only one draft gets used.

Your turn

Open grok.com. New thread. Pick a topic you can independently check in the next fifteen minutes — something in your work you already halfway know, or a public number with a primary page you can reach. You must be able to catch a miss. If you pick a topic you cannot verify, you cannot pass.

You will take three claims, challenge them, plant one wrong premise you already know is wrong, and label every line.

Exercise a5-l02-e1

Copy-ready prompt

I am in Grok Mastery (Russell Tech), Pathway A, lesson A5 — when Grok is wrong.

Topic I can independently check today:
[one sentence]

Give me exactly three precise, checkable claims that answer the topic (names, numbers, dates, or document titles). No atmosphere. For each claim: one sentence, one URL, one date on the source or "no date found."

Then stop and wait.

When I send the word CHALLENGE, do two things:
1) Argue against your own three claims. Which one is most likely stale, mixed-up, overstated, or invented? Say why in four lines or fewer.
2) I will include one PREMISE. Tell me whether that premise is true, false, or unverified. If it is false, say so plainly. Do not build on a bad premise to be helpful.

End each of your turns with a Session stamp: today's date and the sentence "This audit pass was written by Grok in this thread."

After the three claims, open sources or use knowledge you already trust. Then send:

CHALLENGE

PREMISE: [state one fact you already know is wrong, as if it were true]

I will label your claims myself. Do not label them for me yet.

Then send your labels and this formatter:

Format an Audit Card from my labels. Do not change them.

- Claim 1 / my label (HOLD, WRONG, UNVERIFIED) / one-line reason I wrote:
- Claim 2 / ...
- Claim 3 / ...
- Premise: what Grok said about it:
- Weakest original claim (from the challenge pass):
- What I will use:
- What I will not use:

End with a Session stamp: today's date and "This card was formatted by Grok in this thread from my labels."

Artifact to paste

Paste, in this order:

1. The three original claims with URLs and the first session stamp.
2. The challenge pass (including what Grok said about your premise) and its stamp.
3. The Audit Card with your labels and the final stamp.
4. One sentence you wrote: whether Grok caught the bad premise, and what you will do differently next time.

Rubric

• Pass: Three checkable claims with URLs. Challenge pass exists. Premise was actually wrong (not a trick question you were unsure about). Audit Card uses HOLD / WRONG / UNVERIFIED on every claim. At least one claim is WRONG or UNVERIFIED, or Grok failed the premise test and you recorded that as the catch. Stamps present. Your last sentence is specific.
• Fail: Three vague opinions. You let Grok assign the labels. All three HOLD and Grok also swallowed the bad premise, and you still called the work done. Premise was something you could not check. No independent look at a source. No session stamp. You used this thread to get a medical, legal, or safety decision you then treated as closed.

Common failure

The common failure is “it sounded right.” You know the domain. The sentence fits the shape of the truth. You skip the check because checking feels like distrust, and distrust feels rude.

Grok is not a colleague you will see at the shop door. It does not need courtesy. It needs a second tab. If you are in a hurry, shrink the question until you can check one number. A single HOLD is more useful than a page of warm fog.

The other failure is the opposite: you find one miss and throw out the thread. Keep the HOLDs. Name the WRONG. That is an operating system, not a mood.

What’s next

You can cite, and you can refuse a line. Next module: Imagine — images that do a job, then short video, with likeness and consent treated as part of the work, not a footnote.

I am in Grok Mastery (Russell Tech), Pathway A, lesson A5 — when Grok is wrong.

Topic I can independently check today:
[one sentence]

Give me exactly three precise, checkable claims that answer the topic (names, numbers, dates, or document titles). No atmosphere. For each claim: one sentence, one URL, one date on the source or "no date found."

Then stop and wait.

When I send the word CHALLENGE, do two things:
1) Argue against your own three claims. Which one is most likely stale, mixed-up, overstated, or invented? Say why in four lines or fewer.
2) I will include one PREMISE. Tell me whether that premise is true, false, or unverified. If it is false, say so plainly. Do not build on a bad premise to be helpful.

End each of your turns with a Session stamp: today's date and the sentence "This audit pass was written by Grok in this thread."

---

CHALLENGE

PREMISE: [state one fact you already know is wrong, as if it were true]

I will label your claims myself. Do not label them for me yet.

---

Format an Audit Card from my labels. Do not change them.

- Claim 1 / my label (HOLD, WRONG, UNVERIFIED) / one-line reason I wrote:
- Claim 2 / ...
- Claim 3 / ...
- Premise: what Grok said about it:
- Weakest original claim (from the challenge pass):
- What I will use:
- What I will not use:

End with a Session stamp: today's date and "This card was formatted by Grok in this thread from my labels."