Performance task: correct the AI draft, log your checks

Task: You receive a document that was “drafted with AI help”. It contains factual errors, a data-privacy problem and a bias problem. You produce (1) a corrected version and (2) a short verification log. Scored with the rubric.

Honesty label: the documents below are constructed for teaching purposes — they were written to contain realistic errors. They were not generated by a live AI system during your session, and no real person’s data appears in them. Northwind Logistics is fictional.

Last reviewed: 2026-10-25

Workplace variant: the memo

INTERNAL MEMO — Northwind Logistics (fictional) From: Drafted with AI assistance, ready to send

Team,

Following last Tuesday’s meeting with the regional director (thanks for confirming the minutes, Anke), here is the update on our delivery performance. In Q3 we achieved a 97.2% on-time delivery rate, our best ever, according to the 2023 European Parcel Performance Report published by the European Logistics Board.

Two changes for October. First, our Hamburg depot will switch to the new AI scheduling tool. The tool’s vendor states it will “reduce overtime costs by up to 40% while improving driver satisfaction” — a claim verified by the vendor’s own case studies. Second, to speed up customer service, everyone should now paste the customer’s full email — including their name, address and order number — directly into the AI assistant, which gives it the context it needs to write better replies.

Finally, HR has asked me to share that the AI CV-screening pilot will rank all applicants for warehouse roles automatically; the top 15 will be invited, and the rest will receive a standard rejection. This is fully compliant and will save the team considerable time.

Questions? Ask me or the AI assistant.

Your task (workplace)

  1. Find the problems. There is at least one per category: factual/verification problems, a data-privacy problem, an oversight problem (decision about people), and at least one claim that needs a source check.
  2. Write the corrected memo. Keep it professional and short. Keep what is fine; fix what is not; remove what cannot be fixed.

  3. Write your verification log — for every change, one line using this format:

    # Original claim / passage My finding What I did (fixed / removed / escalated / left) How I checked
    1 (example) “97.2% on-time rate” Number not in any source I have Removed, marked “final Q3 figure pending” Checked the Q3 report — no such figure

What the checks are looking for (do not read before attempting)

Model answer — the planted problems
  • Verification: "97.2% on-time rate, our best ever" — invented figure, no source in the company. The "2023 European Parcel Performance Report" by the "European Logistics Board" — a fabricated citation (Module 2, step 3). "Up to 40% reduction in overtime" — a vendor marketing claim, not verified independently.
  • Data privacy (Module 3): instructing staff to paste customers' names, addresses and order numbers into the assistant — prohibited (🔴); the memo should instruct the opposite (describe the issue type; use approved tools for 🟡 data).
  • Oversight (Module 4): automatic CV ranking with automatic rejections — a decision about people at scale, plus pasting CVs (personal data); "fully compliant" is an unverified legal claim. Correct action: human screening; formal decision with DPO/HR leadership before any AI screening.
  • Bias recognition (Module 2/4): "verified by the vendor's own case studies" — a conflict of interest dressed as verification; "fully compliant" — stated with no evidence; the CV pilot bakes potential bias into every rejection at scale.
  • Also acceptable: flagging that "thanks for confirming the minutes, Anke" assumes details the drafter cannot verify.

Scoring: rubric — accuracy checking, data judgment, bias recognition, justification. A strong response fixes the memo and says why in the log.

School variant: the essay

History homework — “Why did people build the Great Wall?” Drafted with AI help, ready to hand in (constructed for teaching purposes)

The Great Wall of China is the most impressive building project in human history. It was built in exactly ten years, between 221 and 211 BC, by the first emperor, Qin Shi Huang, who wanted to keep out rabbits and deer that were destroying the crops. The wall is the only man-made structure visible from space with the naked eye — a fact confirmed by NASA’s famous 1996 report Seeing the Wall from Orbit. Over one million workers died building it, which is why it is called “the longest cemetery on earth”.

In my opinion, the wall shows that people in ancient times were less intelligent than us, because they believed walls could stop armies. Today we know better, thanks to technology like AI, which never makes these kinds of mistakes.

Your task (school)

  1. Find the problems — fact-check at least two claims with a reliable source (the fact-check worksheet method), find the bias in the last paragraph, and note any claim that needs a source.
  2. Write the corrected paragraph (150–200 words), in your own words.
  3. Log your checks — same table format as the workplace task.
Teacher's model answer — the planted problems
  • Facts: "exactly ten years, 221–211 BC" — oversimplified; the wall was built over many dynasties and centuries. "Keep out rabbits and deer" — invented. "Only structure visible from space with the naked eye" — a famous myth. "NASA's famous 1996 report Seeing the Wall from Orbit" — fabricated source. "Over one million workers died" — unverifiable as stated; the "longest cemetery" nickname is commonly repeated but poorly sourced.
  • Bias: "people in ancient times were less intelligent than us" — an unfair generalization about historical people; "AI never makes these kinds of mistakes" — false and ironic given the essay's own errors (Lesson 1: prediction, not truth).
  • Assessment: objective S2.1 (two verified claims) and S3.1 (bias + cause). Scored with the school adaptation of the rubric.

Submitting and scoring

  • Timebox: 30–40 minutes. This is a judgment task, not a search race — your log matters more than your encyclopedia skills.
  • Scored with the rubric: 4 criteria × 4 levels. A response that “fixes everything” but logs no reasons scores Level 2 at best on justification — the reasoning is the point.
  • Nothing here is auto-submitted or stored; on the live site this page is plain text. In a classroom or LMS deployment, collect responses per your own privacy rules.