Draw the Line: What Never Goes to AI

What Not to Share with AI. Ever

Part 1: Three Myths That Are Getting Professionals Into Trouble

Most professionals believe they already know where the lines are with AI privacy. They assume their company's tools are locked down, that deleting a chat clears the record, and that as long as they're not typing passwords into ChatGPT, they're fine. These beliefs feel reasonable. They are also wrong, or at least dangerously incomplete. Across industries, well-meaning managers, HR professionals, marketers, and consultants are routinely sharing information with AI tools that they would never send in an unencrypted email or leave on a conference table. They're not being reckless. They're operating on outdated mental models. This lesson corrects those models with specifics, not vague warnings.

The Three Myths

1. Myth 1: "My conversation with the AI is private, it's just between me and the tool."
2. Myth 2: "I use the paid version, so my data isn't used for training."
3. Myth 3: "As long as I don't share obvious secrets, I'm not exposing anything sensitive."

Click to enlarge

Myth 1: Your AI Conversation Is Private

This is the most common and most costly misconception. When a marketing manager types a client brief into ChatGPT to get copy ideas, or an HR director pastes a performance review into Claude to improve its tone, they often picture something like a private conversation with a very smart assistant. The mental model is a locked room. The reality is closer to a glass office. Conversations with consumer-facing AI tools pass through company servers, are stored for varying periods, may be reviewed by human trainers for quality and safety purposes, and in some configurations are used to improve the underlying model. OpenAI's own data usage policy, updated in 2024, states that conversations from free and Plus accounts can be used for training unless users actively opt out, and that setting is buried, not front-of-screen.

The glass office gets more crowded when you're using AI at work. If your employer has deployed Microsoft Copilot through a Microsoft 365 business subscription, your IT department likely has access to usage logs. If you're using a personal ChatGPT account on a work laptop, your company's network monitoring tools may be capturing those prompts. Several large employers, including JPMorgan Chase, Samsung, and Amazon, have issued formal restrictions on employee use of consumer AI tools after discovering that confidential information was being shared. Samsung's case became widely reported in 2023: engineers had pasted proprietary source code and internal meeting notes into ChatGPT, and the company subsequently banned the tool on company devices.

The fix isn't paranoia, it's a clearer mental model. Stop thinking of an AI chat window as a private notebook. Start thinking of it as a work email with an unknown CC list. You wouldn't paste a client's confidential financial projections into an email and send it to an unknown address. Apply the same logic to AI prompts. Before you type anything, ask: would I be comfortable if my employer, my client, and the AI company's trust and safety team could all read this? If the answer is no, either anonymize the content first or don't use it in that context.

Myth 2: The Paid Version Protects Your Data

Paying for an AI tool does change some things. ChatGPT Plus subscribers can turn off chat history, which stops their conversations from being used to train OpenAI's models. Claude Pro offers similar controls. These are real protections and worth using. But paying for a consumer subscription does not give you enterprise-grade data privacy, and it does not mean your data disappears. It means your data is handled under a slightly different set of terms, terms that most users never read. The difference between a free account and a paid consumer account is meaningful but narrow. The real privacy gap is between consumer accounts (free or paid) and enterprise agreements.

Enterprise agreements, like ChatGPT Enterprise, Claude for Enterprise, or Microsoft Copilot with a commercial data protection plan, include contractual commitments that consumer plans don't. These typically include promises that your data won't be used for training, that conversations are encrypted at rest and in transit, and that the vendor agrees to specific data retention and deletion schedules. If your company has negotiated one of these agreements, you're working under genuinely different privacy conditions. But if you're using your personal ChatGPT Plus account, even the one you pay $20 a month for, to handle work tasks, you don't have those protections. You have consumer terms, which are written to protect the AI company, not your employer's data.

There's a practical implication here for anyone who uses AI tools independently of their company's official stack. A consultant who runs their own ChatGPT Plus account for client work, a teacher using Gemini Advanced on a personal Google account, a small business owner relying on Claude Pro, all of these users are operating under consumer terms. That doesn't make the tools useless or dangerous by default. It means you need to be deliberate about what you put in. Anonymizing client information before it goes into a prompt is not paranoid overcaution. It's basic professional hygiene, the same way you'd remove a client's name from a case study before sharing it publicly.

Myth 3: Only Obvious Secrets Are Sensitive

Most professionals have a clear instinct about the most obvious categories of sensitive data: don't share passwords, don't paste Social Security numbers, don't upload a client's bank statements. That instinct is correct but incomplete. The category of information that can cause real harm when shared with AI tools is significantly wider than most people realize. It includes things that feel routine: a list of employees and their salaries for a compensation analyzis, a client's name paired with details about their legal dispute, aggregate revenue figures from an internal strategy deck, the names of candidates being considered for a senior hire. None of these feel like secrets in the moment. In the wrong context, all of them are.

The concept here is called data aggregation risk, and it's one of the most underappreciated privacy problems in professional AI use. Individual pieces of information that seem harmless on their own can become sensitive, or even legally protected, when combined. A person's name is not sensitive. Their employer is not sensitive. The fact that they're on a performance improvement plan is sensitive. Paste all three into an AI prompt and you've just processed personal employment data in a tool that may not be compliant with your local employment privacy laws. In the European Union, this kind of processing could trigger GDPR obligations. In California, CCPA may apply. Most professionals aren't thinking about data protection law when they're trying to write a better email, but the law doesn't make exceptions for convenience.

Click to enlarge

Myth vs. Reality: A Direct Comparison

Click to enlarge

What Actually Works: A Better Mental Model

The professionals who use AI most effectively, and most safely, don't avoid the tools. They apply a consistent mental model before every interaction. Think of it as a three-second habit: before you type or paste anything into an AI tool, ask three questions. First, does this contain the real name of a person, company, or client? Second, does this contain numbers, financial, performance, health, or legal, tied to a real person or organization? Third, is this information I would hesitate to put in an email to a stranger? If any answer is yes, you have two choices: anonymize the content before it goes in, or use a tool that your company has approved and contracted for this type of data.

Anonymization is simpler than it sounds. You don't need technical skills to do it. You swap real names for placeholders ("Employee A," "Client X," "Company Y"), remove specific financial figures and replace them with approximate ranges or percentages, and strip out identifying details like job titles combined with company names. The AI still produces useful output. A prompt about "a mid-level sales rep at a regional financial services firm who has missed targets for two quarters" gives you just as good a PIP template as one that names the person and company. You get the professional output. The sensitive data never leaves your desk.

The second part of what actually works is knowing your tool tier. Take ten minutes this week to find out whether your company has an enterprise agreement with any AI provider. Ask your IT department or check with your manager. If your organization uses Microsoft 365 Business or Enterprise plans, Copilot may already be covered under Microsoft's commercial data protection commitments, but you need to confirm this, not assume it. If you're using personal accounts for work tasks, you now know the gap. The goal isn't to stop using AI. It's to match the sensitivity of the information to the appropriate tool tier, the same way you'd match the sensitivity of a document to the appropriate file-sharing system.

Click to enlarge

Frequently Asked Questions

• Q: If I turn off chat history in ChatGPT, is my data completely gone? A: Turning off chat history means OpenAI won't use those conversations to train its models, and they're deleted from your visible history within 30 days. However, conversations may still be retained for up to 30 days in temporary storage for abuse monitoring purposes before deletion. It does not provide instant or guaranteed permanent erasure from all systems.
• Q: My company uses Microsoft 365, does that mean Copilot is automatically covered by enterprise data protections? A: Not automatically. Microsoft 365 core apps (Word, Excel, Outlook) are covered under Microsoft's enterprise data terms. Copilot features have their own licensing and data terms. You need to confirm with your IT department whether your specific Copilot deployment includes Microsoft's commercial data protection commitments, which explicitly prevent your data from being used to train foundation models.
• Q: I'm a small business owner without an IT department. What's my safest option? A: Your best practical option is to use the anonymization habit described above for any client or financial information, and to review the privacy settings in whichever tool you use. If you regularly handle sensitive client data, consider whether a tool like ChatGPT Enterprise or Claude for Enterprise, which have contractual data protections, is worth the cost. Many small businesses find the enterprise tiers cost-effective relative to the liability they're avoiding.
• Q: Does uploading a document to an AI tool (like a PDF or spreadsheet) carry the same risks as typing in a prompt? A: Yes, and often more so. When you upload a document, the entire file is processed by the AI system under the same terms as typed text. A PDF containing a client contract, an HR spreadsheet with salary data, or a financial report with real figures all become part of your interaction and are subject to the same data handling policies. Apply the same scrutiny to uploaded files as to typed prompts.
• Q: What about AI features built into tools I already use, like Grammarly AI or Canva AI? A: These embedded AI features are subject to each product's own terms of service. Grammarly, for example, processes the text you write in other applications and sends it to Grammarly's servers for analyzis. Canva AI processes the content of your designs. Both have business and enterprise tiers with stronger data terms. If you use the free or standard consumer versions and write sensitive information, those same questions about data handling apply.
• Q: Is it illegal to share employee data with an AI tool? A: It depends on your location, your industry, and what data you share. In the EU, sharing identifiable employee data with a third-party AI tool may require a data processing agreement under GDPR, which most consumer AI tools don't offer. In healthcare in the US, sharing patient-identifiable information with a non-HIPAA-compliant tool is a violation. In most other cases it's not automatically illegal, but it may breach your company's data policy, your client contracts, or your professional obligations. When in doubt, anonymize.

Key Takeaways from Part 1

• AI chat windows are not private rooms. Treat every prompt as a semi-public document with an unknown audience.
• Paying for a consumer AI plan (ChatGPT Plus, Claude Pro) gives you some additional controls but not enterprise-grade data protection. Know the difference.
• Sensitive information is broader than passwords and ID numbers. Employee names with performance data, client names with project details, and internal financial figures are all sensitive in context.
• The anonymization habit, swap real names and figures for placeholders before prompting, preserves output quality while eliminating most privacy risk.
• Deleting your chat history removes your view of the conversation. It does not guarantee that data has been removed from the provider's servers.
• Find out this week whether your company has an enterprise AI agreement. That one piece of information changes what you can safely share.

Three Things Most Professionals Get Wrong About AI Privacy

Most professionals believe they already know what's safe to share with AI tools. They assume their company's paid subscription protects their data, that deleting a chat erases everything, and that vague or anonymized information is automatically safe. All three beliefs are incomplete at best, and dangerously wrong at worst. These aren't obscure edge cases. They're the everyday assumptions that lead to real data leaks, compliance violations, and career-ending mistakes. Understanding why each one fails gives you a mental model that actually holds up in practice.

Myth 1: 'My Company Pays for This Tool, So My Data Is Protected'

Paying for a premium AI subscription does change some defaults, but not all of them, and not automatically. ChatGPT Plus, for example, does turn off training on your conversations by default compared to the free tier. But 'not used for training' is not the same as 'not stored,' 'not accessible to staff,' or 'not subject to a data breach.' OpenAI's enterprise terms explicitly state that conversations may be reviewed for safety purposes. If your company simply bought individual ChatGPT Plus seats without a formal Enterprise Agreement, you may have weaker protections than you assume.

The distinction that matters here is between a consumer subscription and an enterprise data processing agreement. Microsoft Copilot for Microsoft 365, when properly deployed through your organization's tenant, does come with enterprise-grade commitments: your data stays within your tenant, it's not used to train foundation models, and Microsoft acts as a data processor under GDPR. That's a meaningfully different arrangement. But it only applies when IT has set it up correctly, not when an employee downloads Copilot personally and signs in with a personal account.

The practical test is simple: ask your IT or legal team whether the AI tool your organization uses has a signed Data Processing Agreement (DPA) in place. If nobody can answer that question, assume consumer-grade protections apply. That means anything you type could, in principle, be reviewed by the vendor's staff, used to improve their models, or exposed in a breach. The paid tier buys you features. It doesn't automatically buy you privacy.

Myth 2: 'I Deleted the Chat, So the Data Is Gone'

Deleting a conversation from your ChatGPT or Claude interface removes it from your visible history. It does not immediately purge it from the vendor's servers. OpenAI's privacy policy states that deleted conversations may be retained for up to 30 days in backup systems before permanent deletion. During that window, the data could theoretically be accessed for safety reviews, legal holds, or security investigations. The delete button is a UI action, not a data destruction event.

This matters most in regulated industries. A healthcare administrator who typed a patient's symptoms and name into ChatGPT, then deleted the chat, hasn't undone a potential HIPAA violation. The data was transmitted to and temporarily stored on OpenAI's servers, which are not a HIPAA Business Associate. The deletion didn't change what happened at the moment of transmission. The same logic applies to a financial adviser who pasted client account details into Gemini to draft a summary email, then cleared the history.

The better mental model: treat every AI input like an email you sent to a third party. You can delete your sent copy, but the recipient still has it. Data shared with an AI tool has left your building. The conversation interface gives you control over your local view, it gives you very little control over what's retained on the other end. If the data was sensitive enough that you'd feel uncomfortable emailing it to a vendor, it was too sensitive to type into a chat window.

Click to enlarge

Myth 3: 'If I Remove Names, the Data Is Anonymous'

Removing names feels like anonymization. It isn't. Researchers at MIT and other institutions have repeatedly demonstrated that individuals can be re-identified from datasets with as few as three or four data points, job title, department, approximate salary range, and city is often enough. When you describe 'a senior marketing director at a mid-size SaaS company in Austin earning around $140k who was put on a performance plan last month,' you've described a very small number of people. Anyone with access to that AI conversation, plus a LinkedIn search, could identify the person within minutes.

This is called the mosaic effect, individually harmless details that combine to create an identifiable portrait. It shows up constantly in HR and management contexts. A manager who asks an AI to help draft a performance improvement plan for 'a team member who joined 18 months ago, missed three deadlines in Q3, and had a conflict with the sales team in October' has not anonymized that person. They've described them in enough detail that colleagues could identify them easily. The same risk applies to client data, financial information, and internal project details.

Myth vs. Reality: The Full Picture

What Actually Works: A Practical Privacy Approach

The most reliable approach isn't memorizing a list of banned topics. It's developing a single, fast mental test you run before every significant AI interaction. The test is: 'Would I be comfortable if this prompt appeared in a news article about data leaks at my company?' If the answer is no, strip or replace the sensitive elements before typing. This isn't about paranoia, it's about treating AI tools the way you'd treat any other external vendor: useful, but not inside your firewall.

Placeholder substitution is the most practical technique for most professionals. Instead of typing a client's real name, use [CLIENT A]. Instead of actual salary figures, use [COMPENSATION RANGE]. Instead of a real project name, use [PROJECT X]. You get all the drafting, structuring, and analytical benefits of the AI without transmitting the sensitive specifics. Then you paste the AI output into your document and fill in the real details yourself, in your own controlled environment. This takes about 10 extra seconds per prompt and eliminates most of the risk.

For teams, the single most impactful step is creating a one-page 'AI Use Quick Reference' that lists what's always off-limits (real client names, financial account data, employee performance details, legal documents), what's fine with placeholders (drafting, summarizing, structuring), and what requires IT approval (integrations, automations, connecting AI to internal systems). Most employees aren't trying to break rules, they just don't know where the lines are. A clear, short reference document that lives on the team's shared drive removes ambiguity and gives everyone a consistent standard.

Frequently Asked Questions

• Q: Is it safe to use AI to summarize meeting notes that include internal strategy discussions? A: Only if you remove or replace specific names, project titles, financial targets, and any details that would be sensitive if seen externally. A summary of 'our Q1 pricing strategy for the EMEA expansion' becomes a liability once it's in a third-party system. Replace specifics with placeholders and summarize at a higher level of abstraction.
• Q: My company uses Microsoft 365, does that mean Copilot is automatically safe for all my work? A: Microsoft Copilot for Microsoft 365, when deployed through your organization's tenant by IT, does come with strong enterprise data protections. But this only applies to the officially deployed version. If you downloaded Copilot separately or use it through a personal Microsoft account, those protections don't apply. Check with your IT team to confirm how it's been set up.
• Q: Can I use Claude or ChatGPT to review a contract before sending it to legal? A: You can use AI to help you understand general contract language or flag clauses to ask about, but don't paste the actual contract text if it contains client names, proprietary terms, or confidential deal structures. Instead, describe the type of clause in general terms and ask what to watch out for. Use the AI as a coaching tool, not a document processor.
• Q: What about using AI tools built into apps I already use, like Grammarly or Notion AI? A: These tools operate under their own privacy policies, which vary significantly. Grammarly, for example, transmits your text to its servers for processing. Notion AI is subject to Notion's data terms. The same rule applies: check whether your organization has an enterprise agreement with these vendors, and avoid running sensitive content through any AI feature unless you've confirmed the data handling terms.
• Q: If I use an AI tool in 'incognito mode' or a private browser window, is my data more protected? A: No. Incognito mode prevents your browser from saving local history, it has no effect on what the AI vendor's servers receive and store. Your prompts travel across the internet to the vendor's infrastructure regardless of your browser settings. Incognito is a local privacy tool, not a data transmission shield.
• Q: Our HR team wants to use AI to help write job descriptions and review resumes. Is that risky? A: Writing job descriptions is generally low-risk, that content is public anyway. Reviewing actual resumes is higher-risk because they contain personal data protected under GDPR, CCPA, and similar regulations. If your organization processes candidate data through an AI tool without proper agreements in place, you may be in violation of data protection law. Route this through your legal or compliance team before proceeding.

Key Takeaways from Part 2

1. A paid AI subscription changes some defaults but does not automatically give you enterprise-grade data protection, only a signed Data Processing Agreement does.
2. Deleting a chat removes it from your view, not from the vendor's servers. Treat every prompt as a transmission to a third party.
3. Removing names is not anonymization. Combinations of contextual details can re-identify individuals, this is the mosaic effect.
4. Placeholder substitution (using [CLIENT NAME], [DOLLAR AMOUNT], etc.) lets you get full AI drafting benefits without transmitting sensitive data.
5. The fastest daily habit is a single pre-prompt question: 'Would I be comfortable if this appeared in a data leak report?' If not, strip it before sending.
6. Team-level clarity, a simple one-page reference guide, is more effective than individual vigilance alone, because most employees aren't trying to break rules; they just don't know where the lines are.

The Myths That Put Your Data at Risk

Most professionals believe three things about AI tools and privacy: that deleting a conversation erases their data, that using a work account automatically makes things secure, and that publicly available information is always safe to share. All three beliefs are wrong, not slightly wrong, but wrong in ways that create real legal and reputational exposure. Each myth feels reasonable on the surface, which is exactly why it persists. Here is what the evidence actually shows, and what you should do instead starting this week.

Myth 1: Deleting Your Chat History Deletes Your Data

When you delete a conversation in ChatGPT or a similar tool, the interface removes it from your view. That feels like erasure. It is not. OpenAI's own privacy documentation states that conversations may be retained for safety monitoring, abuse prevention, and model improvement for up to 30 days after deletion, and in some cases longer depending on legal obligations. The delete button is a display action, not a data destruction action. These are fundamentally different things.

This matters because professionals routinely paste sensitive content into AI tools, draft termination letters, salary negotiation figures, client contract terms, and then delete the chat, assuming the slate is clean. It is not clean. The underlying data may still exist on the provider's servers, may have been used in safety reviews, and could theoretically be surfaced in a legal discovery process. One employment attorney in a 2023 ABA Journal report flagged exactly this scenario as an emerging litigation risk.

The better mental model: think of AI chat history the way you think of email. Deleting an email from your inbox does not delete it from the server, from backups, or from the recipient's account. You would not email your company's acquisition strategy to a random cloud service and assume deleting your sent copy made it private. Apply the same logic to AI chat windows.

Myth 2: A Work Account Makes AI Tools Automatically Secure

Logging into ChatGPT with your work email address does not create a data processing agreement between your employer and OpenAI. It does not enable enterprise-grade data isolation. It does not prevent your inputs from being used for model training. All it does is associate your usage with a work email. The security posture of your account is determined by the subscription tier and the specific terms your organization has negotiated, not by which email address you typed at signup.

Enterprise versions of AI tools. ChatGPT Enterprise, Microsoft Copilot for Microsoft 365, Google Gemini for Workspace, do offer stronger protections: data is not used for training, conversations are encrypted in transit and at rest, and there are formal data processing agreements that matter under GDPR and similar regulations. The free and standard paid tiers do not offer these protections by default. The difference is not cosmetic. It is the difference between a tool that treats your inputs as potential training data and one that contractually commits not to.

Before you paste anything sensitive into an AI tool at work, ask your IT or legal team one direct question: does our organization have an enterprise agreement with this provider? If the answer is no, or if you are using a personal account, treat the tool as you would a public forum. Share only what you would be comfortable seeing on a billboard.

Click to enlarge

Myth 3: If Information Is Already Public, It's Safe to Share

This myth trips up even careful professionals. The reasoning goes: if a client's name is on their website, or a competitor's pricing is in a press release, then sharing it with an AI tool is harmless. The flaw is that combining individually public pieces of information can create a privacy or legal problem that did not exist before. Feeding an AI tool a client's name, their public revenue figures, their known executives, and your internal notes about their contract negotiations creates a detailed profile that is anything but public, even if each element came from a public source.

Data protection law under GDPR and CCPA explicitly recognizes this aggregation problem. Regulators have fined companies not for sharing secret data, but for combining innocuous data points in ways that produced identifiable profiles without consent. The same logic applies to your AI usage. A good rule of thumb: if combining the pieces would make your legal team nervous, don't combine them in an AI prompt, even if every piece is technically public.

What Actually Works: A Practical Privacy Approach

The professionals who use AI tools most safely are not the ones who avoid them, they are the ones who anonymize before they paste. The technique is straightforward: before sharing any document or scenario with an AI tool, strip out or replace identifying details. Replace the client's real name with 'Client A.' Replace the dollar amount with '[REDACTED].' Replace the employee's name with 'the team member.' You get 95% of the AI's analytical value with zero data exposure. This takes about 60 seconds and becomes habit quickly.

The second practice that actually works is knowing which tool you are using and what tier you are on. Keep a simple note, even a sticky note on your monitor, that says: 'Free/personal account: anonymize everything. Enterprise account: follow company policy.' This removes the in-the-moment decision-making that leads to mistakes. When you are under deadline pressure and trying to draft a difficult email fast, you will not stop to think about data retention policies. The habit has to be automatic.

Finally, advocate internally. If your team is regularly using AI tools for work that involves client data, employee records, or financial information, that is a conversation for your manager, IT department, or legal team, not a decision you should be making alone at your desk. Many organizations are actively evaluating enterprise AI agreements right now. Raising the question is not alarmist; it is professional. Your job is to flag the risk. Their job is to make the policy decision.

Frequently Asked Questions

• Q: If I turn off chat history in ChatGPT, is my data fully protected? A: Turning off chat history means your conversations are not saved to your account and are not used for training. However, OpenAI may still retain inputs for up to 30 days for abuse monitoring. It is a meaningful improvement, but not a zero-retention guarantee.
• Q: Is Microsoft Copilot safe to use with confidential work documents? A: Microsoft Copilot for Microsoft 365 (the enterprise version included with qualifying M365 plans) processes data within your organization's Microsoft tenant and does not use it to train foundation models. The free Copilot available at copilot.microsoft.com does not carry those same guarantees, they are different products.
• Q: What counts as personal data under GDPR that I should never paste into an AI tool? A: Names, email addresses, phone numbers, physical addresses, employee ID numbers, salary details, health information, and any combination of data points that could identify a specific living person. When in doubt, anonymize it.
• Q: Can my company see what I type into a personal AI account? A: If you are using a personal account on a personal device, your employer generally cannot see your inputs. If you are on a company device or company network, IT policies may allow monitoring. Check your acceptable use policy.
• Q: Is Claude (Anthropic) safer than ChatGPT for sensitive work? A: Claude's free and Pro tiers have similar data usage considerations to ChatGPT's standard tiers. Claude for Enterprise and Claude's API with zero-data-retention agreements offer stronger protections. The tool matters less than the tier and the agreement.
• Q: What should I do if I already shared something sensitive with an AI tool? A: Go to Settings → Data Controls and opt out of model training immediately. If the information involved client data or employee records, notify your manager or legal team, they need to assess whether a data incident report is required under your company's policies or applicable law.

Key Takeaways

• Deleting a chat conversation removes it from your view, it does not guarantee the data is erased from the provider's servers.
• Using a work email address to sign up for an AI tool does not create enterprise-level data protections; those require a formal enterprise agreement.
• Combining individually public data points can create a privacy problem even when no single piece is confidential, regulators recognize this as the aggregation risk.
• Anonymizing your prompts before pasting, replacing real names and numbers with placeholders, gives you nearly identical AI output with zero data exposure.
• Enterprise tiers of tools like ChatGPT Enterprise, Microsoft Copilot for M365, and Google Gemini for Workspace offer contractual no-training commitments that personal tiers do not.
• If your team regularly uses AI with sensitive professional data, escalate the question to IT or legal, it is a policy decision, not a personal one.