A lawyer at a firm I know pasted a client’s full contract into ChatGPT to get a summary. A developer at another company submitted proprietary source code for debugging help. A financial analyst shared internal revenue projections to get a presentation drafted. None of them thought they were doing anything unusual. All of them were sending confidential data to OpenAI’s servers, where it could be used for model training by default.
This isn’t a reason to avoid AI tools for work. It’s a reason to understand exactly what happens to data you submit, which settings change that behavior, and what alternatives exist when the confidentiality requirements genuinely can’t be met through ChatGPT — even with privacy settings enabled.
Quick Answer:
- By default, conversations with ChatGPT on the free and Plus plans are used to train OpenAI’s models — this can be disabled in Settings → Data Controls → Improve the model for everyone → toggle off
- Even with training opt-out enabled, your conversations are still stored by OpenAI for up to 30 days for safety monitoring purposes
- For genuinely sensitive data (client PII, proprietary code, financial projections, health information), the correct tool is either ChatGPT’s Enterprise tier (zero data retention by contract), a self-hosted open-source model, or Microsoft Copilot through a business Microsoft 365 subscription with data processing agreements in place
What OpenAI Actually Does With Your Data
The default behavior differs by plan, and most people are on the wrong plan for work use.
Free plan: Conversations are used to train OpenAI models by default. You can opt out (Settings → Data Controls), but the opt-out only applies to future conversations — it doesn’t retroactively remove data already submitted.
ChatGPT Plus ($20/month): Same default behavior as free. Training opt-out is available in the same location. Conversations stored for 30 days even after opt-out for abuse monitoring.
ChatGPT Team: Data is not used for training by default. Conversations are stored. The Team plan provides a workspace where admins can manage settings across users, but it doesn’t come with a formal data processing agreement (DPA) that would satisfy GDPR, HIPAA, or enterprise security requirements.
ChatGPT Enterprise: Zero data retention — OpenAI contractually commits to not storing or training on submitted data. Comes with a formal DPA. SOC 2 Type 2 compliant. This is the tier where genuinely confidential business data can be submitted with appropriate legal protection.
ChatGPT API: When you access GPT-4o or other models through the API, OpenAI’s API data usage policy is different from the consumer product — API data is not used for training by default, and OpenAI retains it for up to 30 days for abuse detection only. If your company’s developers are building internal tools on the API rather than the consumer interface, the data handling is meaningfully better.
(Sources: OpenAI Privacy Policy, OpenAI API Data Usage Policy, OpenAI Enterprise documentation — all as of July 2026)
Step 1: Enable the Training Opt-Out (Minimum Baseline)
If you’re using the free or Plus plan and haven’t changed any settings, start here.
ChatGPT web → click your profile icon → Settings → Data Controls → “Improve the model for everyone” → toggle off.
In the mobile app: Settings → Data Controls → same toggle.
What this does: your future conversations will not be used to train OpenAI models. What it doesn’t do: it doesn’t prevent OpenAI from storing your conversations (they’re kept for 30 days), and it doesn’t prevent OpenAI staff from reviewing conversations flagged by safety systems.
This is a meaningful improvement over the default but is not sufficient for data that has legal confidentiality requirements — attorney-client privilege, HIPAA-covered health information, financial data subject to SEC regulations, or trade secrets with contractual non-disclosure obligations.
Step 2: Disable Chat History (Reduces Storage Window)
Settings → Data Controls → “Save new chats” → toggle off.
When chat history is disabled, OpenAI states that conversations are retained for 30 days and then permanently deleted. They’re not used for training. This reduces the storage window but doesn’t eliminate it.
The practical implication: for a conversation about a client matter submitted on a Monday, the data sits on OpenAI’s servers until the following month. For most business data, this is acceptable with training opt-out enabled. For data subject to formal legal protection, it’s not.
The Anonymization Approach: Redact Before You Submit
The most practical technique for using ChatGPT with sensitive work data on the free or Plus plan is systematic redaction — replace identifying information with generic placeholders before submitting, then re-identify after you get the output.
For contracts and legal documents: Replace client names with “Party A” and “Party B.” Replace specific dollar amounts with “[AMOUNT]” or round numbers if the exact figure isn’t relevant to the task. Replace dates with “[DATE]” or relative references (“the effective date,” “90 days after signing”). Replace identifying case or matter numbers.
Example transformation:
- Original: “This agreement between Acme Corporation and Johnson & Smith LLP, dated March 15, 2026, for $2.4 million…”
- Redacted: “This agreement between Company A and Firm B, dated [DATE], for [AMOUNT]…”
The AI output (summary, drafting, analysis) refers to “Company A” and “Firm B” — you mentally re-insert the real names when using the result.
For code: Replace proprietary function names, database schema names, and API endpoints with generic equivalents. A function called calculateAcmeRevenueShare() becomes calculateRevenueShare(). Database table names like acme_customer_transactions become customer_transactions. The logic you’re debugging or asking about is identical; the proprietary naming is gone.
For financial data: If you need help structuring a presentation or analysis, work with percentage changes rather than absolute figures. “Revenue grew 23% YoY with margins expanding 4 points” contains the same analytical structure as the actual numbers but reveals nothing about scale.
[COMMON TRAP] Partial redaction creates a false sense of security. Replacing a client’s name while leaving their industry, deal size, jurisdiction, and counterparty type in the text may still allow identification by someone who sees the conversation. Either redact comprehensively or use a more privacy-appropriate tool for that specific task.
ChatGPT Enterprise: When You Need a Real Solution
For organizations that use AI regularly with genuinely confidential data — law firms, healthcare providers, financial services, companies with significant IP — ChatGPT Enterprise is the appropriate tier, not a workaround.
What Enterprise provides:
- Zero data retention: OpenAI contractually commits to not storing submitted data beyond the immediate session
- No training: submitted data is never used for model training, by contract
- Data Processing Agreement: a formal legal agreement that satisfies GDPR requirements and provides contractual basis for data processing
- SOC 2 Type 2 compliance
- Admin controls for managing user access, conversation policies, and data governance across the organization
Pricing is negotiated per organization — OpenAI doesn’t publish Enterprise pricing. For most small businesses, the Enterprise tier is cost-prohibitive. For organizations where a single data breach or confidentiality violation would be more costly than the Enterprise subscription, it’s worth evaluating.
Local AI: The Zero-Data-Transmission Alternative
If the data is sensitive enough that you’re uncomfortable with any external transmission — regardless of OpenAI’s policies — running a local AI model on your own hardware eliminates the problem entirely. No data leaves your machine.
Tools like Ollama make running capable open-source models (Llama 3.1, Mistral, Qwen) locally straightforward on modern hardware. A laptop with 16GB of RAM can run models sufficient for document summarization, drafting assistance, code review, and most business writing tasks. A desktop with a GPU handles larger models that approach GPT-4o capability for many tasks.
The tradeoff: local models require setup, consume local hardware resources, and are generally less capable than GPT-4o for complex reasoning tasks. For routine document work, the capability gap is smaller than most people expect.
The full setup process for running local AI on a PC — including which models are appropriate for different tasks and how to configure Ollama — is covered in how to run local AI on your PC and why it matters for data privacy.
Microsoft Copilot: The Enterprise Alternative Already in Your Stack
For organizations using Microsoft 365 Business or Enterprise, Microsoft Copilot (GPT-4 integrated into Word, Excel, Outlook, and Teams) operates under Microsoft’s commercial data processing terms rather than OpenAI’s consumer terms.
Under the Microsoft Products and Services Data Processing Agreement (which applies to all Microsoft 365 commercial subscriptions), Microsoft commits to:
- Not training AI models on your tenant’s data
- Processing data only to provide the subscribed service
- GDPR compliance through standard contractual clauses
If your organization already pays for Microsoft 365 Business Premium or Enterprise, Copilot features are either included or available as an add-on, and the data governance framework is already in place. Using Copilot within Word to summarize a confidential document is meaningfully different from pasting that document into ChatGPT.com, because the legal and contractual data protection framework is already established through the existing Microsoft relationship.
Practical Workflow: Categorizing What You Can and Can’t Submit
The useful mental model is a three-tier classification for work tasks:
Safe to submit to ChatGPT (free/Plus with opt-out):
- Drafting from scratch with no existing confidential information (“write a template NDA” vs. “review this specific NDA”)
- Research questions about general topics, industry trends, legal frameworks
- Formatting and editing generic documents that don’t contain confidential content
- Brainstorming and ideation where the ideas themselves aren’t confidential
- Learning and skill development
Submit with redaction:
- Document summarization where the structure and terms matter but specific parties, amounts, and dates can be replaced
- Code debugging where the logic matters but proprietary naming can be genericized
- Data analysis where percentage relationships or anonymized examples convey the same analytical question
Don’t submit to ChatGPT (free/Plus) — use Enterprise, local, or Copilot:
- Documents containing client PII that you’re legally obligated to protect
- Source code containing proprietary algorithms or API keys
- Financial projections or M&A information subject to securities regulations
- Healthcare information subject to HIPAA
- Information covered by NDAs where external transmission would constitute a breach
- Anything where a data incident would create legal liability
[PRO TIP] Create a one-page internal policy document for your team or for yourself that lists which categories of information can go to which AI tool. Written policies create awareness and serve as evidence of due diligence if a data handling question ever comes up. “We had a clear policy that client PII goes to our Copilot environment, not ChatGPT” is a much better position than “we didn’t have a policy.”
The Browser Extension Risk
A significant and underappreciated data leakage vector: browser extensions that “enhance” ChatGPT or that access your browser’s text content can intercept what you type into ChatGPT before it’s encrypted and transmitted. Extensions with access to all page content can read everything you type.
Before using ChatGPT for any work purpose, audit which browser extensions have “Read and change all your data on websites you visit” permissions. Remove any extension you don’t recognize or no longer use. For sensitive work sessions, use a browser profile with no extensions installed.
For a detailed look at which privacy settings across browsers and tools actually reduce data exposure, how to protect your privacy online in 2026 covers browser-level privacy configuration in detail.
Comparing Your AI Options for Work Data
| Tool | Training opt-out | Data retention | Legal DPA | Best for |
|---|---|---|---|---|
| ChatGPT Free/Plus | Manual (off by default) | 30 days | No | Generic tasks, redacted content |
| ChatGPT Team | On by default | Yes (stored) | No | Small teams, non-regulated data |
| ChatGPT Enterprise | Always | Zero | Yes | Regulated industries, client data |
| Microsoft Copilot (M365 Business) | On by default | Per Microsoft DPA | Yes | Organizations already on M365 |
| Local AI (Ollama) | N/A (no transmission) | None | N/A | Maximum privacy, offline use |
| ChatGPT API | On by default | 30 days | Available | Developer-built internal tools |
For a direct comparison of how ChatGPT handles different task types versus Claude and Gemini — including which performs better for document work, coding assistance, and research — ChatGPT vs Claude vs Gemini in 2026 covers the capability differences alongside the data handling differences.
FAQ
Does deleting a ChatGPT conversation remove it from OpenAI’s servers? Deleting from your chat history removes it from your view, but OpenAI retains deleted conversations for up to 30 days for safety purposes before permanent deletion. If training opt-out is enabled, deleted conversations aren’t used for training during that retention period.
Is ChatGPT HIPAA compliant? ChatGPT free and Plus are not HIPAA compliant. ChatGPT Enterprise can be made HIPAA compliant through a Business Associate Agreement (BAA) with OpenAI, which is available for Enterprise customers. Without a BAA in place, submitting any Protected Health Information (PHI) to ChatGPT violates HIPAA regardless of which plan you’re on.
Can my employer see what I type into ChatGPT? If you’re using ChatGPT on a company-managed device or through a company account, your IT department may have visibility into your browsing activity. Personal ChatGPT accounts on personal devices aren’t visible to employers, but company policies may still prohibit using external AI tools with work data regardless of which device you use.
Does incognito mode make ChatGPT more private? No. Incognito mode prevents your browser from storing local history, but it doesn’t affect how ChatGPT or OpenAI processes your data. The same data transmission and storage behavior applies regardless of whether you’re in incognito mode.
What should I do if I already submitted confidential data to ChatGPT? Enable the training opt-out immediately if it wasn’t already on. Delete the specific conversation from your history. If the data was subject to legal confidentiality requirements (attorney-client privilege, HIPAA, NDA), consult with the relevant legal or compliance team about disclosure obligations — some jurisdictions and contracts require notification of unauthorized data transmissions.
Conclusion
ChatGPT is genuinely useful for work, and the data risk is manageable with the right approach. The training opt-out and chat history disable settings are the minimum baseline for any work use. Systematic redaction extends what you can safely submit on the free and Plus plans. For data with formal confidentiality obligations — client data, health information, regulated financial data — ChatGPT Enterprise, Microsoft Copilot under a commercial agreement, or a local AI model are the appropriate tools. The decision tree isn’t complicated: identify what category the data falls in, match it to the appropriate tool, and make that matching process explicit enough that it becomes habitual rather than a case-by-case judgment call
