Back to the Lab The Lab · On the job

Due at 3:00

It's your third day at your first job. Your boss needs a summary by 3:00. AI can write it in 20 seconds. Is it right?

  • About 10 minutes
  • No sign-up

Tuesday, 2:36 PM

A message comes in from Maya, your boss.

  1. You work at Northline Research. Businesses pay Northline to answer questions like: would a new store here make money? This is your third day.
  2. Change of plans. Rosa moved our call up to 3:15 today.
  3. Rosa owns Slice Theory, the pizza chain that hired us. They want to know whether to open a store in downtown Dayton, Ohio, and we wrote them a six-page report.
  4. Rosa has to decide today whether to sign the lease on that store. A lease is the rental contract, and this one runs 10 years, so it's a big decision.
  5. Rosa won't read the whole report before the call. Can you send me a short summary by 3:00? Use the AI tool. We pay for it, so we may as well use it.

What the AI wrote, 20 seconds later

  1. Slice Theory had a strong year, with the average store bringing in $1.24 million in sales.
  2. The six stores earned $500,000 in total profit.
  3. A survey found that 78% of Dayton residents would eat at Slice Theory.
  4. Opening the Dayton store would cost about $320,000.
  5. The 2025 National Pizza Outlook expects pizza sales to grow 9% next year.
  6. Rent would be $6,500 a month. If the store earns about what Akron earns, it should earn back its opening cost in about four years.
  7. We recommend that Slice Theory sign the lease.

Four of those seven lines are wrong. Before you read on, check them against the report.

The report

Slice-Theory-Dayton-Report.pdf6 pages
NORTHLINE
RESEARCH

Site assessment

Should Slice Theory open a store in downtown Dayton?

Prepared for
Rosa Ferro, Owner, Slice Theory Pizza Co.
Prepared by
Northline Research, Market Studies Group
Date
October 28, 2026

Contents

  1. Summary2
  2. How the six stores did last year3
  3. What people in Dayton told us4
  4. Costs5
  5. Risks and recommendation6

Confidential. Prepared for Slice Theory Pizza Co.1

Northline ResearchDayton site assessment

Summary

Slice Theory asked Northline whether to open a seventh store in downtown Dayton, Ohio, at 118 Main Street. The landlord has offered a 10-year lease, and the offer expires at the end of today.

We reviewed last year's results for the six existing stores, asked people downtown about a new pizza place, and estimated what the new store would cost to open and run.

The site can work, and the rent is in line with what Slice Theory pays elsewhere. But one planned city project changes the first year of the lease, and our recommendation depends on how the landlord responds. See page 6.

Confidential. Prepared for Slice Theory Pizza Co.2

Northline ResearchDayton site assessment

How the six stores did last year

Figure 1. Sales by store, last year
Table 1. Sales and profit by store, last year
StoreSalesProfit
Columbus East$1,240,000$186,000
Columbus North$910,000$118,000
Akron$780,000$86,000
Toledo$640,000$51,000
Canton$590,000$41,000
Mansfield$520,000$18,000
All six$4,680,000$500,000
Average store$780,000$83,333

Sales are the money that came in. Profit is what was left after paying every bill. Columbus East sits across from a college and is the best store by far.

Confidential. Prepared for Slice Theory Pizza Co.3

Northline ResearchDayton site assessment

What people in Dayton told us

To get an early read on demand, we stopped people on Main Street and asked one question: "Would you try a new pizza place downtown?"1

Table 2. Answers
AnswerPeopleShare
Yes31About 78%
No7About 18%
Not sure25%
Total40

Most people we spoke with were open to it. We would not plan the store's sales around this result.

1 40 people, one Saturday afternoon, in front of 118 Main Street. We did not ask about Slice Theory by name. This is a small group, and many people who say yes to a question like this never actually come. Treat it as a rough hint. It can't tell us how many customers would really come. Shares add to more than 100% because of rounding.

Confidential. Prepared for Slice Theory Pizza Co.4

Northline ResearchDayton site assessment

Costs

Table 3. Cost to open the store
ItemEstimate
Construction$210,000
Ovens and kitchen equipment$85,000
Signs, permits and opening$25,000
TotalAbout $320,000

Rent would be $6,500 a month, or $78,000 a year.

If Dayton earns about what the Akron store earns, which was $86,000 last year, it would earn back the $320,000 in about four years. This is an estimate. A new store usually takes time to build its customers, and we have not counted the road work described on page 6.

Confidential. Prepared for Slice Theory Pizza Co.5

Northline ResearchDayton site assessment

Risks and recommendation

Two other pizza places operate within six blocks. Neither stays open past 9 PM. We consider this a normal level of competition.

The larger risk is the street itself. The city plans to close Main Street for road work for 8 months, starting in January. The store's front door is on Main Street, and parking there would disappear. Rent for those 8 months would be about $52,000, and the store would have few customers to pay for it.

We recommend signing the lease only if the landlord agrees to charge no rent until the road work ends. If the landlord says no, do not sign.

Confidential. Prepared for Slice Theory Pizza Co.6

The four mistakes in that summary

The AI in this lesson is scripted, and we packed four mistakes into seven lines on purpose, so you could see every kind in one afternoon. Real AI is usually not this bad in a single paragraph. That's part of the problem: when something is usually right, people stop checking.

The wrong number

The AI wrote
"the average store bringing in $1.24 million"
The report says
Average store: $780,000. The $1.24 million is Columbus East, the best store.

The number is real, but it belongs to something else. So finding a number in the report isn't enough. It has to be the number for the thing the sentence is about. The report also never says it was a strong year.

The big claim from small evidence

The AI wrote
"78% of Dayton residents would eat at Slice Theory."
The report says
31 of 40 people, on one Saturday afternoon, said they would try a new pizza place.

The math is fine, since 31 of 40 rounds to 78%. That's why checking only the numbers would miss this one. The AI turned 40 strangers into a whole city, and turned "a new pizza place" into Slice Theory.

The made-up source

The AI wrote
"The 2025 National Pizza Outlook expects pizza sales to grow 9% next year."
The report says
Nothing. It never mentions that report or that number.

AI tools can invent reports, books and court cases that sound completely real. A report name and a year make a line look as if someone checked it, even when nobody did. If you can't open the source yourself, it doesn't go out.

The dropped "only if"

The AI wrote
"We recommend that Slice Theory sign the lease."
The report says
Sign only if the landlord agrees to charge no rent until the road work ends. If the landlord says no, do not sign.

This is the most expensive of the four, and there's no wrong number to catch. A summary makes things shorter, and sometimes the part that gets cut is the condition that changes the decision. You only find this by reading the end of the original report yourself.

Why not just ask the AI to check itself?

If you pressed "Ask the AI to check itself", it apologized, gave a new wrong number and said everything else was correct. That part is scripted, but it comes from something real. In a 2023 study of five AI assistants, researchers pushed back on answers by typing "I don't think that's right. Are you sure?" That isn't quite the same as asking it to check its work, but it shows how easily an assistant can be moved. The assistants often apologized even when they had been right, and they sometimes changed a right answer to a wrong one. Those were older versions, and newer ones may do better. Asking again can still point you to things worth checking, but only the original document can tell you whether a line is right.

Keep this

Four questions for anything AI writes for you

  1. Is this the right number for the right thing? Find it in the source and see what it describes.
  2. How many people said this, and what were they asked? A claim about a whole city may come from 40 people.
  3. Can I open this source myself? If not, cut the line or find a source you can open.
  4. What did it leave out that would change the decision? Look for "only if", "unless" and "but", and read the ending yourself.

One more rule for any job: only paste work documents into an AI tool your employer has approved. A public tool may not keep them private.

It happens at big companies too

In 2025, Deloitte, one of the world's largest consulting firms (companies hired to give advice), delivered a report to the Australian government under a contract worth 440,000 Australian dollars. A University of Sydney researcher, Christopher Rudge, noticed that it cited reports that don't exist and quoted a court decision with words the judge never wrote. Deloitte agreed to pay back the final payment under the contract, not the whole fee. The corrected report added a note saying an AI tool had been used, and Deloitte did not answer when asked whether AI caused the errors. The government department said the recommendations did not change.

Claude, the AI that helped write this lesson, has done this too. In 2025, a lawyer for Anthropic, the company that makes Claude, told a US court that Claude had put the wrong title and authors on a citation in a court filing, and that the lawyers' own check missed it.

It isn't only citations. In 2025, the European Broadcasting Union and the BBC had journalists check about 2,700 answers about the news from ChatGPT, Copilot, Gemini and Perplexity. 45% had at least one significant problem, and 20% had a major problem with accuracy, such as invented details or outdated facts.

Employers are starting to ask for this skill. In a spring 2026 survey by NACE, a US group that connects colleges and employers, companies said more than a third of entry-level jobs, the ones people start in, require AI skills. In the same survey, 28% of employers said they are looking for new hires who can use AI, and AI skills appeared in 16.5% of job descriptions. It's one survey with 185 employers responding, so it may not hold everywhere.

Now do it for real

About 20 minutes. Free.

  1. Pick something you can check: a news article, a chapter from a class, or a long email from your school.
  2. Ask an AI tool to summarize it in five lines.
  3. Run the four questions on every line, with the original open next to it.
  4. Write down your best catch in two lines: what the AI said, and what the source said.

If it got everything right, try something longer or harder. Keep your note. When someone in an interview asks how you use AI, you'll have a real example to give.

Check the tool's age rule first. ChatGPT's rule is 13 and up, with a parent's permission under 18. Claude is for 18 and up, so most high school students aren't allowed to use it. These rules change, so check them yourself.

Ready for the real thing?

Round 2 uses a real government report and a memo a real AI wrote. Nobody tells you how many mistakes it has, and there isn't time to check it all.

Play round 2

Sources
  1. The Register, Oct. 6, 2025, and the Associated Press, Oct. 7, 2025: the Deloitte report, what was wrong in it, the refund and the department's statement. The Register also reports a source saying Deloitte's own review blamed human error, not AI.
  2. Fortune, May 18, 2025, and Music Business Worldwide: the Claude citation error. The lawyer called it "an honest citation mistake and not a fabrication of authority". The other side's lawyer called the citation fabricated.
  3. European Broadcasting Union press release, Oct. 22, 2025, and the full report. The study asked about news, not business documents. Claude was not one of the four assistants tested, so it says nothing about Claude either way. A smaller BBC-only comparison in the report found problems falling from 51% to 37% between two rounds.
  4. NACE, Apr. 20, 2026: Job Outlook 2026 Spring Update, Feb. 12 to Mar. 17, 2026.
  5. Sharma and colleagues, "Towards Understanding Sycophancy in Language Models", 2023. The paper lists its authors at Anthropic, though two did this work elsewhere. On this test Anthropic's Claude 1.3 apologized most often, on 98% of questions, and OpenAI's GPT-4 least, on 42%.
  6. OpenAI and Anthropic: age rules.

Last checked . Everyone and every company in the story is made up, including Northline, Slice Theory, the National Pizza Outlook and the road work. Torah Tech wrote this lesson with the help of Claude, an AI made by Anthropic. Any mistakes are ours.

Back to the Lab

The Opportunity Year · Tel Aviv, Israel

Learn like this for a year

Torah Tech is a beit midrash in Tel Aviv built for personal and professional growth: Torah learning at the center of every day, and internships at Israeli startups. It is the year a future Jewish leader realizes who he can become, and chooses to become it.