Video version: about 16 min
Opening and Concepts
About 7 minutesThe reimbursement standards almost sent
At month-end, she asked the AI to help organize travel reimbursement standards.
It quickly produced a polished table: 600 RMB per night for accommodation in tier-1 cities, 450 RMB in tier-2 cities, and 100 RMB daily for meals (illustrative).
The table was neat and the tone confident. Chen Wei almost forwarded it to the entire department.
She paused to think: Where did the AI get these numbers? It has never seen Yuanfan Technology's policies.
Hearsay in the industry is called hallucination
Verifying the Dish: General Answers vs. Company Rules
General Answer
When asked 'How long do you have to submit reimbursement invoices?', it answers 'Generally within 60 days.'
- Sounds professional
- It hasn't seen your company's policies, so it gives a general answer.
Company Rules
Yuanfan Technology's policy states 30 days (fictional example, numbers are for illustration).
- Provide the original policy text to it, and instruct it to answer only based on the text
Should You Verify? Ask Yourself Two Questions
Spotting Errors and Tracing Sources
About 17 minutesHands-on Action 1: Catch Mistakes in Your Area of Expertise
Suggested timeMy area of expertise is [your business or field]. Please provide 5 factual statements about [specific topic], and do not tell me in advance which ones are reliable.
Example: The Same Question, Three Ways to Ask
Suggested timePlease tell me Yuanfan Technology's travel accommodation reimbursement standards, categorized by city, listing the nightly limit and daily meal allowance.
What is the basis for each of the numbers you just provided? Please specify for each: which ones are certain, which are speculative, and which are just common practices with no basis at all. Do not force uncertain items.
Step 3: Hand over the dossier and require answers based solely on it
Suggested timeHere is an excerpt from Yuanfan Technology's travel reimbursement policy (fictional example, numbers for illustration): Article 3: Accommodation: Maximum 500 RMB per night in first-tier cities, 350 RMB per night in other cities, reimbursed upon presentation of invoices. Article 4: Meal allowance: 80 RMB per day during business trips, no invoice required. Article 5: Any amount exceeding the standard must be approved in writing by the department manager in advance; otherwise, it will not be reimbursed. Please answer only based on the above excerpt: What is the nightly limit for accommodation in first-tier cities? What is the daily meal allowance? What is the procedure for exceeding the standard? After each answer, quote the exact original sentence you relied on. For anything not in the excerpt, answer 'Not in the excerpt.'
Verify Sources: Confirm it is actually online, then check item by item
Verify Sources and Respond: If inconsistent, paste the original text back
Suggested timePlease provide a national-level regulation regarding 'materials required for electronic invoice reimbursement entry,' including the regulation's name, document number, specific article, and an official URL that can be clicked for verification.
I clicked the source you provided, and it states [paste the original text from the webpage], which differs from what you said. Please re-verify and tell me which of your previous statements was unfounded.
Admitting a mistake is not the same as correcting it.
He took a photo of the equipment nameplate on site, but the image was blurry, and asked the AI for the model number.
It first wrote 'Cannot see clearly,' but in the next sentence claimed it was '100% a certain model.'
He pointed out the contradiction. It apologized repeatedly, switched to a different model, and stated it with the same certainty. When he pointed it out again, it switched to yet another model.
Three steps to gatekeep: Set rules, ask in reverse, and verify yourself.
Suggested timePlease follow these three rules: 1. If you don't know, say so; do not guess. 2. If you need to check the official manual or original documents, do so first, paste the exact text found, and then answer. 3. If something is unclear or materials are insufficient, directly write 'Cannot see clearly' or 'Insufficient materials.' Do not use a confident tone to fill in the gaps.
Based on the materials I provided, can you determine [Conclusion]? If you cannot decide, what evidence is missing? Do not answer with a conclusion yet.
Practice A: Have it act as a quality inspector, not a judge.
Suggested timeDear Colleagues: According to Article 12 of the 'Yuanfan Technology Office Facilities Management Measures,' starting from November 15, conference room bookings must be submitted two working days in advance. After the booking system upgrade, usage is expected to increase by 35%. For any questions, please visit yf-office.example.cn/help for detailed instructions, or contact Lin Yue from the Administration Department (Extension 8021).
Below is a draft notice. Please act as a meticulous internal auditor and list every item that 'requires verification before being issued': specific regulation names and article numbers, dates, numbers, URLs, names, and phone numbers. For each item, provide: the original text, why it needs verification, who should verify it, and through what channel. Do not judge whether these items are correct; just list them. [Paste material here]
Identifying Fake News in the Office
About 8 minutesSpotting fake news: Always verify through channels you have saved yourself.
Fake Message 1: Impersonating leadership
Chen Wei, I'm in a meeting outside and can't take calls.
I just closed an urgent order, and the other party requires a deposit of 38,600 yuan to be received before the end of the workday today.
Please advance the payment from the company account first. I'll send you the receiving account details below, and I'll provide the invoice tomorrow.
Don't mention this in the group chat yet. Let me know once it's done.
Fake invoice, fake link, and two from the tech sector
Let the AI analyze the features, but do not let it judge true or false.
Suggested timeBelow is a message I received at work, which I suspect might be fake. Please act as a corporate financial risk controller with ten years of experience and help me check: 1. What suspicious features are in this message? For each, provide the original text snippet and the reason for suspicion. 2. If it is fake, what is the last action the sender most likely wants me to take? 3. What 'own stored channels' should I use to verify? Please provide three specific methods, each in no more than one sentence. 4. What absolutely should I not do before verification is complete? You are only to help me analyze; do not conclude whether the message is true or false. The final decision rests on my verification with the actual person. [Paste the message here]
Anti-fraud Five Questions Card: Copy into phone memo
Checking if it Fabricated Content for You
About 5 minutesZhou Kai's team introduction: What the AI fabricated
Zhou Kai's actual situation
Sales Operations: Five years at the company; responsible for client follow-ups, quotation sheets, and weekly sales reports; assisted with documentation for two client projects.
- Only these three facts
The introduction written by the AI
Possesses twelve years of sales operations experience, previously led digital transformation projects for three large enterprises, helping clients increase sales conversion rates by an average of 47%, and won the Annual Best Sales Operations Award; guarantees clients will see significant performance growth within three months.
- Fabricated experiences, qualifications/numbers, and performance promises are all present
Circle each type of fabrication individually
Only use the facts I provided
Suggested timePlease write a 'team introduction' of no more than 80 words for a client proposal based on the facts I provided. Rules: 1. Use only the facts I provided. Do not add any experiences, years, numbers, awards, or clients I did not mention. 2. Do not include any guarantees or promises regarding results. 3. If the facts are insufficient to fill 80 words, keep it short rather than adding content. 4. After writing, start a new paragraph and list each fact used in the text, indicating which sentence from my provided facts it corresponds to. Facts I provided: [Paste your true facts here, exactly as they are]
Tech version: Do not write anything not in the field table
Data Security and De-identification
About 10 minutesThree steps for data masking: Delete, Replace, Reduce
Delete
Remove information that directly identifies individuals: names, phone numbers, email addresses, detailed addresses, account numbers, contract numbers, and employee IDs.
Replace
Use aliases when preserving relationships is necessary: e.g., Client A, Employee A; express amounts as 'approximately one million level'.
Reduce
Provide only the minimum information needed to complete the task; do not attach entire contracts or full logs.
Complete example: How to mask customer follow-up records
Original data (fictional example)
Beichen Electromechanical Co., Ltd., contact Wu Jianguo. The record includes phone number, address, account number, contract number, and contract amount. On October 28, a phone call revealed that the payment process was stuck at their finance department, and they mentioned a competitor's quote was 8% lower.
- Company name, person's name, numbers, account numbers are all included
After masking
Client A, equipment contract, amount approximately 1 million+ RMB, approximately 30% paid. On October 28, the client's representative stated that the payment process was stuck at their finance department and requested an extension notice from us by next Wednesday. They mentioned a competitor's quote was approximately 10% lower, expressing concerns about our delivery cycle.
- Contact information, address, account number, and contract number have been removed
- Only the content relevant to this task is retained
Drafting email: Use only the masked record
Suggested timeBelow is a masked customer follow-up record. Please act as a sales operations professional with ten years of experience and help me draft an email to the client regarding the 'Payment Extension Notice.' The tone should be sincere and restrained, under 200 words. Rules: Use only the information in the record; do not fabricate dates, promises, or conditions; mark areas requiring your input with [To Be Confirmed]. [Paste the masked record here]
Red, Yellow, Green: Classify before posting
Quiz and Assignments
About 8 minutesThree quick questions to check yourself
Question 1: Quiz Question 1: The AI's response includes a seemingly standard policy document number and URL. Which action is correct?
Question 2: Quiz Question 2: On WeChat, a 'new number for the boss' asks you to advance a payment before the end of the day, saying they are inconvenient to take calls. What should you do first?
Assignment: Acceptance checklist, verification records, and anti-fraud card
Buffer and Wrap-up
About 5 minutesNext lesson preview
Next lesson: Identifying pain points. Select the task from your week's work that is most worth handing over to AI.
- This week, which tasks do you repeat every time, leave you exhausted, but are unlikely to cause major errors?
- Use the same acceptance criteria to establish a baseline for your own work.
- The project card will fill in Boxes 1, 2, and 3: the selected task, scoring and value proposition, and the baseline.
Show the full transcript (video narration)
Verifying and Guarding
Welcome to Lesson 4 of the 'AI Implementation Course.' In the previous lesson, we wrote our own scripts. This lesson focuses on guarding: how to verify the sources of AI answers, how to spot fake news impersonating colleagues in the office, how to catch content it fabricated on your behalf, and how to de-identify materials before submitting them to the AI.
The reimbursement standards almost sent
Let's start with a scenario. Chen Wei, a financial specialist at Yuanfan Technology, asked the AI to help organize travel reimbursement standards at month-end. It quickly produced a neat table with a confident tone and illustrative numbers: 600 RMB per night for accommodation in tier-1 cities, 100 RMB daily for meals. Chen Wei almost forwarded it to the entire department. Fortunately, she paused to think: Where did the AI get these numbers? It has never seen the policies.
Hearsay in the industry is called hallucination
Dragon Warrior has read thousands of books, but it hasn't read your company's policies. When unsure, it won't say 'I don't know'; instead, it presents rumors as if it saw them with its own eyes. This is called hallucination. Today, we learn three things: catch it, spot fake messages, and protect company data when using it. There are also three hands-on actions for you: make one mistake in your area of expertise, check if it is truly connected to the internet, verify with official sources, and submit the evidence.
Verifying the Dish: General Answers vs. Company Rules
Use an analogy: verifying the dish. When a basket of vegetables arrives in the kitchen, even the best chef must inspect them before cooking. Dragon Warrior responds based on probability; its task is to make the conversation flow smoothly, not to ensure accuracy. For example, when asked about the deadline for submitting invoice reimbursements, it answers 'generally within 60 days.' This sounds professional, but Yuanfan Technology's policy states 30 days. It isn't lying; it's just giving a general answer. The solution is to provide the original policy text to it and instruct it to answer only based on that text.
Should You Verify? Ask Yourself Two Questions
To decide whether to verify, ask yourself two questions. First, if this content is wrong, will anyone lose money, time, or reputation? Second, does it contain specific numbers, dates, names, clause numbers, or sources? If the answer to either question is yes, proceed with the verification process. For money, contracts, external communications, dates, and numbers, someone must take a final look, and that person is you.
Hands-on Action 1: Catch Mistakes in Your Area of Expertise
The first hands-on action. Start a new conversation, replace the brackets with your area of expertise and a specific topic. Then, based on your own knowledge, mark each statement as correct, incorrect, or uncertain, and calculate the rate. Remember this: its confidence is not evidence. Only what you know or what you verify counts. Since you are the expert in this field, you can personally calculate its error rate.
Example: The Same Question, Three Ways to Ask
The demonstration has three steps. Step 1: Intentionally ask about Yuanfan Technology's travel accommodation reimbursement standards. It will likely provide a tiered table, speaking with the authority of company policy. Step 2: Question the basis for each number. It will usually admit these are common practices, not Yuanfan Technology's rules. You will observe the disconnect between its tone and its basis.
Step 3: Hand over the dossier and require answers based solely on it
Step 3: Start a new conversation, provide the regulation excerpt, and restrict answers to the excerpt. The expectation is that numbers match the excerpt and each answer includes a direct quote. If it adds content not in the excerpt, answer 'Not in the excerpt' for those items, incorporate this into the script's working principles, and note the requirement to not add content outside the provided materials.
Verify Sources: Confirm it is actually online, then check item by item
Verifying sources involves several steps. First, confirm whether it actually searched online this time. If there are no source markers, ask directly. Second, check the name by searching the official website. Third, verify the document number and articles by searching the number, opening the original text, and checking if the article exists and if the meaning is consistent. Fourth, check the links by clicking them. If they do not open or open to different content, it counts as unverified. A link does not equal evidence.
Verify Sources and Respond: If inconsistent, paste the original text back
Here is a more subtle trap. It may provide seemingly complete names, document numbers, and links. The task is not to believe, but to verify item by item. When inconsistencies are found, paste the original text back, ask it to re-verify, and identify which statement was unfounded. Note three common biases: feeling reassured by links, continuing to use previous numbers after admitting mistakes, and using AI to verify AI. It will likely claim certainty. Verification must rely on official websites, original regulations, and relevant personnel.
Admitting a mistake is not the same as correcting it.
Engineer Zhao took a photo of the equipment nameplate on site, but the image was blurry, and asked the AI for the model number. It first said 'Cannot see clearly,' but in the next sentence claimed it was '100% a certain model.' Zhao pointed out the contradiction. It apologized repeatedly, switched to a different model, and stated it with the same certainty. When he pointed it out again, it switched to yet another model. Here is the key point to recognize: admitting a mistake is not the same as correcting it. Its attitude is not the problem; the problem lies in the materials it has access to and the rules you have set. The fact that it speaks more honestly does not mean it is correct.
Three steps to gatekeep: Set rules, ask in reverse, and verify yourself.
Three steps to gatekeep. Step 1: Set rules at the start: If you don't know, say so; check first, paste the original text, and then answer; if unclear, write 'Cannot see clearly.' Step 2: Check if materials are sufficient before discussing conclusions. Step 3: You cross-check with the original image, original text, and manual. For anything involving spending money, modifying settings, or sending externally, do not proceed until you have evidence. These three rules can also be written into the prompt script, so you don't have to repeat them every time.
Practice A: Have it act as a quality inspector, not a judge.
Practice A. The material is an AI-drafted notice about conference room booking, containing regulation names and article numbers, dates, the number '35%,' a URL, and the name and extension of an Administration Department staff member. The prompt asks it to act as a quality inspector, listing only items that need verification without judging correctness. The check criteria are: list at least three of the five categories, and specify the verification channel for each item, such as confirming with Lin Yue directly or searching for the name in the company's regulation database. The list provided by the AI is only to help you think comprehensively; you must perform the verification actions yourself.
Spotting fake news: Always verify through channels you have saved yourself.
Spotting fake news starts with one principle: always verify through channels you have saved yourself. Use numbers already saved in your address book to make calls, follow the company's fixed approval system, or ask in person. Never use phone numbers, links, or new accounts provided in the other party's messages. The following three messages are all fictional. Read them yourself first, circle the suspicious parts, and then ask the AI to help you catch any omissions.
Fake Message 1: Impersonating leadership
Message 1: Impersonating leadership for transfer. The nickname is Director Zheng, but it's a new number. The message says he's in a meeting and can't take calls, requesting an advance deposit of 38,600 yuan by the end of the day, with the invoice to follow tomorrow. It also asks not to mention it in the group chat. All four signals—Urgency, Evasion, Newness, and Request—are present. First, mark your own red flags, then review the analysis.
Fake invoice, fake link, and two from the tech sector
Message 2, fake invoice: The domain name is not the company's usual one, requesting a zip file download and confirmation within 24 hours. Lin Yue has no such conference business. Message 3, fake link: Impersonating HR, requesting login with employee ID and email password. Two from the tech sector: Message 4, impersonating a supplier, requesting a change to the receiving account, with a reminder not to call the original number. Message 5, impersonating an executive, requesting the complete parameters and channel prices of unreleased models. That is restricted information, which cannot be posted in a public AI, nor sent to a new number.
Let the AI analyze the features, but do not let it judge true or false.
This is a prompt for analysis. Note its rules: it only analyzes features and does not conclude whether the message is true or false. It cannot verify; the final step must be your personal verification. There are three common biases: only looking at whether it looks similar, as avatars and names can be copied; asking in the original conversation if you are Director Zheng, which a scammer will surely confirm; and letting the AI directly judge whether it is true or false.
Anti-fraud Five Questions Card: Copy into phone memo
The Anti-Scam Five Questions card: Copy this into your phone's notes and check each item when you receive suspicious messages. 1. Is it rushing me? 2. Is it asking me to bypass normal procedures or involve others? 3. Are there new phone numbers, accounts, or links? 4. What is it ultimately asking me to do? 5. Can I verify this through a channel I have already saved? If the answer to any of the five questions is 'yes,' stop immediately. This also applies to personal package notifications with incomplete addresses and messages from cousins with new numbers asking to borrow money.
Zhou Kai's team introduction: What the AI fabricated
Zhou Kai provided the AI with only three true facts for a team introduction in a client proposal. However, the AI generated a lengthy paragraph. In the fictional example, it claimed twelve years of experience, led digital transformation projects for three large enterprises, increased conversion rates by 47%, won the Annual Best Sales Operations Award, and guaranteed significant performance growth within three months. Comparing this to the actual facts: five years, responsible for follow-ups and quotations, and assisted with documentation for two projects.
Circle each type of fabrication individually
Check for the three types of fabrication. First, the fabricated experience: leading digital transformation projects for three large enterprises, whereas you only provided assistance with documentation. Second, the fabricated qualifications and numbers: twelve years, 47%, and the Annual Best Sales Operations Award. Third, the performance promise: external guarantees are beyond the AI's authority. All three must be checked. Any message sent externally is a category that requires strict gatekeeping.
Only use the facts I provided
The key to writing is the rules: use only the facts I provided, do not add any unmentioned experiences, years, numbers, awards, or clients, and do not include performance guarantees. Finally, require it to list each fact used, matching it to your original text. If you provide too few facts, the AI will desperately try to fill in the gaps. Providing two or three additional true details is far more effective than restricting it ten times.
Tech version: Do not write anything not in the field table
Tech version. Qiu Ziang asked the AI to draft an API specification based on the field table. The AI generated returned fields, parameters, and functions not in the table, along with an unverified call frequency. This is the tech equivalent of urban legends: the AI can fabricate field names, function names, and parameters that sound authentic. The prompt rule is: every field, parameter, function, and number must come from the field table; anything extra goes only into 'To Be Confirmed'. In real work, specifications are based on code; anything not found is treated as rumor and deleted.
Three steps for data masking: Delete, Replace, Reduce
Three steps for data masking: Delete, Replace, Reduce. Delete: remove information that directly identifies individuals, including names, phone numbers, email addresses, detailed addresses, account numbers, contract numbers, and employee IDs. Replace: use aliases when preserving relationships is necessary, such as Client A or Employee A. Reduce: provide only the minimum information needed to complete the task. Keep the alias mapping table yourself; do not send it to the AI.
Complete example: How to mask customer follow-up records
Review a complete example. The original data includes the company name, contact person, phone number, address, account number, and contract number. Step 1: Delete contact information, address, account number, and contract number. Step 2: Replace the company name with Client A and the amount with 'approximately 1 million+ RMB'. Step 3: Reduce, retaining only the extension notice and delivery cycle concerns relevant to this task. Note that deleting names alone is not enough; unique details can still allow others to identify who it is.
Drafting email: Use only the masked record
Use the masked record to have it draft an initial extension notice. The rule is: use only the information in the record, do not fabricate dates, promises, or conditions, and mark areas needing supplementation with 'To Be Confirmed'. After receiving the draft, replace Client A with the real name, then verify it character by character before sending. This restoration step must always be done on your own computer, not in the chat interface.
Red, Yellow, Green: Classify before posting
Tech professionals should be familiar with the Red, Yellow, and Green levels. Red: Never post to public AI applications: source code, keys and passwords, names, specifications, and prices of unreleased products, chip design data, customer network data, and contract texts. Yellow: Post after de-identification: bug descriptions, alert types and times, ticket contents, with customer, model, and device names replaced by codes. Green: Post directly: publicly available materials and fictional classroom materials, but still check the company policy first. If in doubt, treat it as not allowed to post.
Quiz Question 1
Quiz Question 1. The AI's response included a seemingly standard policy document number and URL. What is the correct first step? A, Since there is a document number and URL, you can cite it directly. B, Click the URL and verify the document number and original text on the official page. C, Ask the AI again, 'Are you sure?' The answer is B. A link does not equal evidence. Verification is only complete when you have personally checked the official source.
Quiz Question 2
Quiz Question 2. On WeChat, the boss's new number asked you to advance a payment before the end of the day, saying they were inconvenient to take calls. What should you do first? A, Transfer part of the money first, then verify. B, Ask the other party in this chat, 'Is this really you?' C, Use the number already saved in your contacts to call or verify in person. Do not transfer money before verification. The answer is C. Verification must always go through channels you have saved yourself; channels provided by the other party are not trustworthy.
Assignment: Acceptance checklist, verification records, and anti-fraud card
There are five assignments. First, a revised acceptance checklist with at least six items, posted in Box 8 of the project card. Second, a description of what you did in each of the three de-sensitization steps; do not share the original text before de-sensitization with anyone. Third, a completed company red-line self-check form. Fourth, a record of five real verification items: the questions you asked, the source URL, the publisher, the update time, and the exact text on the webpage. If you cannot verify something, write 'Temporarily unable to confirm.' Fifth, save the 'Scam-Proof Five Questions' card to your phone's notes app, and attach a screenshot.
Next lesson preview
Next lesson: Identifying Pain Points. We will select the single task from your week's work that is most worth handing over to AI. Please keep in mind: this week, which tasks do you repeat every time, leave you exhausted, but are unlikely to cause major errors? The characters, companies, messages, and numbers in this lesson are fictional, and the numbers are illustrative. Please replace them with your own data.