AI answers can be factually wrong even when they read naturally. Being able to create an image or text and being allowed to publish it are also different matters. This chapter introduces, with concrete examples and steps, what to check before you enter information, before you use an answer, and before you publish.
Chapter goal
Learn to notice errors in AI answers and to check on your own whether the sources, rights, and permission to share are sound. You'll also learn guidelines for when to pause entering information or publishing.
Five Checkpoints When Using AI
Here are five points to keep in mind when using AI for writing or research. How cautious you should be depends on the situation.
Hallucinations: Verify Even Natural-Sounding Answers Against Source Material
Hallucination is when AI gives a plausible-sounding wrong answer. It may write content that contradicts the material you provided, and it may even add a convincing-looking reason or source for the error.[2]
Generating text and verifying facts are separate processes
Language models that generate text predict tokens (units such as characters or parts of words) that continue the context. They often produce correct content, but text that flows naturally and text that is factually correct are different things. Even with services that combine search or calculation, you need to check yourself whether the material the AI referred to or the calculation conditions are correct.[2]
Original memo
The meeting room for the internal study session is only tentatively reserved. The date and speaker have not yet been decided.
Example of an incorrect announcement
"The study session will be held on November 12. The speaker is Tanaka from the Sales Department, and the meeting room has been confirmed."
The date, speaker, and venue are all information that is not in the memo.
Revision based on the memo
"We are planning an internal study session. The date and speaker have not yet been decided. The meeting room is tentatively reserved, and we will share details as soon as they are finalized."
While shaping the text into the form of an announcement, things that haven't been decided get added. This is a familiar example of hallucination.
Four steps to verify an answer
- Pick out what to verify Extract numbers, dates, names, quotations, and phrases that make a commitment, such as "we will."
- Open the source yourself Check that the source actually exists and that the passage says what the answer says. A link alone is not evidence.
- Compare under the same conditions Check that the target audience, region, period, and units match the original, and redo any calculations yourself. Also check that "planned" hasn't changed into "decided."
- Decide what you can use Separate the facts you have verified from points you haven't. Check important decisions with the person in charge or an expert before using them.
Even when multiple AIs agree, that is only a starting point for checking.
They may all be referring to the same false information or source. Don't stop at asking "Please cite sources" or "Review it." Go back to the original material and verify. For specific ways to check, practice in the chapter Verify the basis of an answer.
Copyright: Separate the Conditions for Input, Generation, and Publication
For copyright, consider separately whether you may use someone else's work and whether your generated output is protected. Even if a service's terms say "commercial use allowed," that does not mean you have verified the rights to someone else's work.
① Source material you input
Check who created the work, whether you have permission or a license to use it, and whether any legal exceptions apply. Just because a work can be viewed on the web does not mean you may enter it into AI or reuse it.
② Generated expression
Check whether the output reproduces the creative expression of an existing work. Japan's Agency for Cultural Affairs organizes infringement at the generation and use stages in terms of similarity and reliance.[5]
③ Uses for publication or sale
Uses that are fine during trial stages may require separate checks for publication or distribution. Check the purpose, medium, region, and contract terms together.[5]
Whether you own copyright is a separate check
According to the Agency for Cultural Affairs' approach, the question is examined case by case: how much a person was creatively involved. Long instructions or repeated regeneration alone do not decide whether protection applies. The U.S. Copyright Office's 2025 report also distinguishes between cases where a person determined or creatively edited the expression and cases where a prompt was simply given.[5] [6]
"Similar style" and "reproducing the expression" are different.
Similarity in art style or ideas is not the same as similarity in the creative expression of an existing work. You can't judge this by whether an artist's name was included in the prompt, so compare the original work with the output. Also, even if the generated output is not recognized as copyrighted, that does not mean you are free to use other people's works or trademarks contained in it.[5]
Example: If you use AI to modify an illustration you found online for use in an advertisement, check the usage terms of the original illustration, any original expression remaining in the output, and the terms for using it in ads. Don't assume "AI-modified means it's fine." For difficult cases, consult the rights holder or the person in charge within your organization.
Privacy: What You Should and Should Not Give to AI
Whether you can enter information depends on three things: the type of information, the service and plan you use, and your organization's permission. It's important not to decide based on a single condition, such as "a paid plan is safe" or "it's fine because I hid the name."
Consider training, storage, deletion, and transmission separately
This is a setting for whether data is used to improve the model. Turning it off does not mean your history disappears or is no longer stored.
The history shown on screen, storage on the service side, and safety checks are all separate things. Even Temporary Chats may be saved in some cases.
Check what is deleted, by when, and whether there are exceptions. Even if something disappears from the screen, storage on the service side or past usage may not be completely removed.
Connected apps, shared links, and sending features each have their own conditions. They can't be judged from AI settings alone.
Separate practice input from real work input
- Confidential company information or unpublished source code
- Personal information of customers or business partners
- Confidential information entrusted to you by others
- Unpublished financial or personnel data
- Detailed circumstances or numbers that can identify an individual or case
- Nonexistent companies, people, or inquiries
- Samples you created that are unrelated to any real person or case
- Public materials whose permitted use you've confirmed
- Your own unpublished writing (after confirming it may be shared)
Passwords, API keys, and private keys never need to be entered, either for writing or for practice. Also, even if you change a name to "Mr. A," someone may be identifiable from the address, case name, date and time, and so on. Hiding a name alone does not make the information anonymous or grant permission to share it.
Check the terms of three services in official documentation
The following applies to personal chat use (checked against each company's official documentation on October 14, 2026). If you use a plan contracted by your company, the conditions and administrative settings differ, so check them individually.
Turning off training use does not remove your history. Different terms apply when you send feedback on a response, for example, and Temporary Chats may also be stored for safety purposes.
Turning off training use does not cancel training that has already begun. Incognito chats are stored even though they don't appear in your history, and safety processing is subject to separate terms.
If you turn off Activity and don't send feedback, later chats won't be used to improve the model. However, responses, safety processing, storage, and human review are handled separately.
"Not used for training" on business plans does not mean data is never stored, is not connected to external services, or that no one looks at it. OpenAI's organizational guide also advises checking how storage and administration are handled for each plan, feature, and connection. After confirming the contract terms, compare them with your organization's rules.
Practice reducing input when replying to inquiries
An inquiry containing a real name, contact details, order number, or an unpublished matter. Even if you hide the name, the other party or the transaction may be identifiable from other information.
"We received an inquiry from a business partner about the delivery date for a product. We have not yet decided on a delivery date. Please draft a polite reply for external use. Do not promise a delivery date."
Order before input: limit the information to what the purpose requires → check how the service, plan, and connected destinations handle it → compare with your organization's rules. If you can't tell, stop entering information and try with nonexistent examples. For details, see Precautions when entering information into AI.
If you have already entered information, first check the scope and conditions under which it can be deleted, and if it is work information, consult your organization's administrator. For example, OpenAI's official explanation of ChatGPT shows that archiving is different from deletion, that there are exceptions to the retention period after deletion, and that separately saved files or the original data in connected destinations are handled separately. Note that simply entering information does not mean it has been used for training or leaked to others.
Dealing with Bias and Impersonation
Bias: Involves Design and Use as Well as Data
AI bias arises not only from the composition of training data but also from human judgment and from the systems in which AI is used by societies and organizations. People with certain attributes may be treated unfairly, or accuracy may differ by language or group.[7]
Example of an application document
An application document says only "a six-month gap in employment," with no reason given.
Assumption you should not add
Deciding that "there's a gap, so motivation must be low" lets a guess that is not in the document enter the evaluation.
How to check
State that "the reason cannot be determined from the documents," then evaluate using the same criteria related to the job. Asking AI for a different perspective alone does not mean bias has been removed.
In situations such as hiring or personnel evaluation that affect how people are treated, decide who is affected and what criteria are used for judgment, and prepare a process where people can review the results.
Deepfakes: Verify Identity Even When Voices or Videos Seem Real
Images, videos, and audio created or altered with AI are also used for impersonation and fraud. The FBI cites as warning examples scams that mimic a family member's voice to request urgent money transfers, and video calls impersonating executives.[3]
- Don't act on the spot Don't proceed based only on a voice or video when there is an urgent money transfer request, a request for credentials such as passwords, or a change to payment details.
- Contact them back yourself Hang up, then call back using a number you already know or the organization's official contact information.
- Stop until you can verify Verify whether the person is who they claim to be, what the request involves, and whether the organization has approved it. If you cannot verify, stop sending money or providing information and consult the person in charge.
An unnatural appearance can be a clue, but the absence of unnatural signs does not mean something is genuine. For images and videos you see on social media, check who posted them, when, and the surrounding context before sharing. For measures within companies, you may also refer to our guide to AI agent security.
Laws and Rules: Check Your Region, Purpose, and Role
Beyond laws specific to AI, check rules related to what you're doing, such as copyright, personal information, and consumer protection. The starting point is which region, for what purpose, and whether you are involved as a provider or a user of AI.
The following is based on official materials from each country and region as of October 16, 2026. To determine whether it applies to your use, check the original text at the linked sources or ask your internal department in charge.
EU: Application dates differ by use
The AI Act generally applies from August 2, 2026. Prohibited practices and certain others start to apply earlier. Some high-risk uses, such as employment and education, apply from December 2, 2027, and high-risk AI built into regulated products applies from August 2, 2028, so the dates are staggered.
There are also conditions and exceptions for labeling obligations for AI-generated content, depending on its scope and each party's role.[8]
United States: Existing rules also apply to AI
The FTC explains that privacy commitments made to users must be kept even for AI services, and that being AI does not exempt a company from existing laws.
Check which laws and contracts apply to the work you use or provide.[9]
China: Public-facing services and labeling
The Interim Measures for the Management of Generative AI Services apply, for example, when services are offered to the general public within China. The same provisions do not necessarily apply to research that is not offered publicly or to internal use.
The Measures for Labeling AI-Generated Synthetic Content took effect on September 1, 2025. Check the conditions that apply to providers and to users who publish generated content.[10]
Japan: Promotion and risk management
The AI Promotion Act was fully enforced on September 1, 2025. The Cabinet Office shows both the promotion of AI use and responses to its risks. The AI Basic Plan was revised on July 14, 2026.
Even when the government encourages use, that does not mean a specific use has been approved. Also check rights, contracts, and organizational rules.[1]
Checklist Before Using and Before Publishing
Check your own work
- Before entering information: limit it to what is necessary, and check how the service handles it and whether it's okay to share. Use nonexistent examples for practice.
- Before using an answer: compare numbers, dates, quotations, and promises against the original sources, and leave any points you can't verify marked as unconfirmed.
- Before publishing: check the rights to the materials and generated output, labeling conditions, and the recipients and scope of publication.
- Important decisions: don't decide based on AI answers alone. Consult the person in charge or an expert.
Decide procedures within your organization
- Decide which services, plans, and connected destinations are allowed, and what information may be entered.
- Decide who confirms publication or sending, and which operations require approval.
- Prepare a process for reviewing bias and errors and correcting decisions.
- Decide on contacts for problems and procedures for suspending use or investigating.
Summary of this chapter
Before entering information, before using an answer, and before publishing, check how information is handled, whether it matches the original sources, rights, bias in judgment, and whether the other party is who they claim to be. Verification is an important role for users, but responsibility does not always rest solely with users. Depending on the contract and specific actions, the responsibilities of AI providers and organizations may also be considered.[5] [9]
References
- Cabinet Office, Japan: Overview of the AI Promotion Act, AI Basic Plan.
- NIST: Generative AI Risk Management Profile (2024, English PDF). Covers erroneous generated content and approaches to verifying and managing it.
- FBI/IC3: Warning about financial fraud using generative AI (December 3, 2024, English).
- Official materials for the privacy section (checked October 14, 2026): OpenAI's data controls, Anthropic's model improvement settings, Google's Gemini privacy explanation. Individual materials on storage, deletion, and temporary chats can be found at the relevant points in the text.
- Agency for Cultural Affairs, Japan: Concepts on AI and copyright (March 2024, PDF). Distinguishes development and training, generation and use, and the copyrightability of generated output.
- U.S. Copyright Office: Publication of report on copyright in AI-generated works (January 29, 2025, English).
- NIST: AI bias can come from sources other than data (2022, English).
- European Commission: Scope and application timeline of the AI Act (English).
- FTC: AI companies uphold your privacy and confidentiality commitments (2024, English).
- Cyberspace Administration of China: Interim Measures for the Management of Generative AI Services, Measures for Labeling AI-Generated Synthetic Content (Chinese).
In the next chapter, we explain the latest AI trends. When using drafts at work, also see the email review steps.