The Answer Depends on Which Tier You Use

The short answer is: it depends on which ChatGPT you're using — and most people don't realize there are effectively two different products with two very different privacy postures.

OpenAI draws a hard line between its business offerings and its consumer tiers. On its enterprise privacy page, the company states plainly: "By default, we do not use your business data for training our models." That commitment covers ChatGPT Enterprise, Business, Healthcare, Edu, Teachers, and the API platform. The Enterprise launch announcement went further, telling customers: "You own and control your business data in ChatGPT Enterprise. We do not train on your business data or conversations."

Free and Plus users get a different deal. Their conversations are stored by default and may be used for model training unless they opt out. And once data enters training, removal is technically impossible — which makes your tier choice and default settings the single most important privacy decision you make.

The concern isn't paranoia, either. In March 2023, a Redis cache misconfiguration exposed roughly 1.2% of Plus subscribers' chat titles and payment information over a nine-hour window. Italian regulators later hit OpenAI with a €15 million GDPR fine. These incidents affected consumer tiers and third-party infrastructure — not the business-tier guarantees OpenAI advertises.

So the practical picture looks like this:

• Business tiers (Enterprise, Business, Healthcare, Edu, API): no training on your data by default, SOC 2 Type 2 audits, AES-256 encryption, and 30-day deletion windows.
• Free and Plus: conversations stored by default and potentially used for training unless you opt out.
• HIPAA coverage: only Enterprise/Healthcare tiers offer a BAA — Free, Plus, and Team do not.
• Data residency is available in 10 regions for business customers, including the U.S. and Canada.

This tier-dependent reality is why the question deserves a real answer, not a shrug. The tool's guarantees matter, but so does the discipline of whoever operates it. It's the same reasoning behind how My AI Call Center handles campaign data: contact lists are reviewed for source and consent before launch, and data is never shared, sold, or used to train shared models.

The next variable — and the one research says matters most — isn't the tier at all. It's what users actually type into the box.

Real Incidents Show the Risk Is Not Theoretical

Data leakage from AI tools isn't a hypothetical worry — it has already happened, repeatedly, in ways that are well documented. Looking at where these incidents actually occurred reveals a consistent pattern: the vulnerabilities live in consumer tiers, third-party integrations, and unmanaged employee usage, not in the core business-tier infrastructure.

In March 2023, a bug in an open-source library exposed sensitive data through a Redis cache, leaking chat titles and payment information for roughly 1.2% of ChatGPT Plus subscribers during a nine-hour window, according to a compliance analysis of documented incidents. The breach hit paying consumer-tier users — not enterprise customers whose data is excluded from training by default.

Then there was Samsung. Engineers pasted proprietary semiconductor source code into ChatGPT to help debug it, and that confidential code entered OpenAI's training data. As compliance experts noted at the time, once data enters training data, removal is technically impossible — which makes prevention the only real remedy.

Regulators have responded. Italian authorities temporarily banned ChatGPT and later fined OpenAI €15 million under GDPR for privacy violations. And in 2025, a share-link indexing incident exposed private ChatGPT conversations through Google search results, showing that even sharing features can become a leakage vector.

The most instructive case may be the Salesloft-Drift supply-chain breach. Trend Micro's investigation found attackers compromised an AI chatbot vendor and cascaded into 700+ organizations — including Cloudflare, Palo Alto Networks, and Zscaler — harvesting OpenAI API credentials among the stolen tokens. Notably, Okta avoided the breach solely because IP allow-listing blocked the stolen token.

The pattern across these incidents is clear:

• Consumer and Plus tiers carry the training-data and retention risk that business tiers exclude
• Third-party integrations and vendors extend your attack surface beyond OpenAI itself
• Unmanaged employee usage — "shadow AI" — now appears in 43% of security incidents, per IBM's Cost of a Data Breach reporting
• What users paste matters: 11% of data entered into ChatGPT contains confidential information

This is why discipline on the input side matters as much as vendor guarantees. My AI Call Center applies the same logic to calling campaigns — list source and consent records are reviewed before launch, and client data is never shared, sold, or used to train shared models. The incidents above came from gaps in governance, not from AI being inherently unsafe — and governance is exactly what a managed, reviewed process provides.

Shadow AI Is the Dominant Leakage Vector

The instinct to block AI tools feels like control — until you see where the data actually goes. Cyberhaven research shows that 32.3% of ChatGPT usage flows through personal accounts that bypass enterprise controls entirely. When organizations prohibit AI without providing governed alternatives, employees don't stop — they migrate to unmanaged channels where no logging, retention policies, or training controls exist.

This shadow AI problem is now measurable at incident scale. IBM's 2025 Cost of a Data Breach Report found that shadow AI appears in 43% of security incidents, more than doubling year over year. Meanwhile, the same Cyberhaven study revealed that 39.7% of AI interactions involve sensitive data — confidential information entering systems the organization cannot see, audit, or delete.

• Blocking AI pushes usage outside visibility, expanding the attack surface
• Personal accounts lack SSO enforcement, centralized logging, and retention controls
• Once data enters training pipelines, removal is technically impossible
• Governance — not prohibition — is the only control that keeps data inside managed workflows

My AI Call Center applies this principle to outbound calling: every campaign runs against approved, permissioned, or reviewed lists only, with consent records verified before launch. Data is never shared, sold, or used to train shared models — the same discipline the research identifies as the missing layer in most AI governance programs.

How My AI Call Center Applies These Controls to Outbound Campaigns

The research points to one clear conclusion: governance gaps, not AI tools alone, cause most data leakage. That is exactly the gap a managed calling service is built to close.

Consider what the data shows. According to Cyberhaven's AI adoption research, 39.7% of AI interactions involve sensitive data, and 32.3% of ChatGPT usage flows through personal accounts that bypass enterprise controls entirely. Meanwhile, IBM's breach data shows shadow AI now appears in 43% of security incidents — more than doubling year over year.

My AI Call Center structures every campaign to eliminate that unmanaged usage. Because it is a done-for-you managed service rather than self-serve software, there is no shadow AI problem: your team never pastes contact data into a personal chatbot, and every campaign runs inside reviewed, documented controls.

List and consent review comes first, before anything launches. The research is blunt about why this matters — compliance analysts note that 11% of data pasted into ChatGPT contains confidential information, and once data enters training, removal is technically impossible. Our process applies the same discipline upstream:

• List source and consent records are checked before any campaign launches — approved, permissioned, or reviewed lists only.
• Bought lists without clear permission records are flagged, and in most cases declined.
• Your data is never shared or sold, and never used to train shared models.
• AI disclosure runs on every call, and opt-outs are logged and honored immediately.
• Clinic campaigns follow HIPAA-compliant communication standards.

That last point matters more than many realize. Healthcare guidance on AI tools warns that organizations should assume any PHI entered into non-enterprise AI products violates HIPAA — a risk that disappears when communication standards are set and enforced before launch rather than left to individual staff judgment.

The broader lesson from the research is that provider promises alone are not enough. Even with OpenAI's enterprise-grade commitments like AES-256 encryption and SOC 2 Type 2 audits, Trend Micro's breach analysis advises organizations to implement their own protective controls rather than trusting vendors blindly. Consent verification and list discipline are the privacy foundation those controls rest on — and they are the first thing we review, before you spend anything.

If you want to know whether your list will support a compliant campaign, the first campaign review is free. Plan My Campaign walks through your goal, list source, and consent records — and we tell you plainly if the list will not support the campaign.

Practical Checklist: What to Verify Before Any AI Tool Touches Your Data

Every data-leak incident traced in this article shares one trait: someone skipped a verification step. Before any AI tool touches your data, run this checklist — it takes less than an hour and closes the gaps behind most documented breaches.

1. Confirm the tier and its training policy. OpenAI states plainly that business tiers — Enterprise, Business, Healthcare, Edu, and the API platform — do not train on your data by default. Free and Plus accounts are different: conversations are stored and may be used for training unless you opt out, and once data enters training, removal is technically impossible. Ask the vendor for the training policy in writing, tied to the specific tier you are buying.

2. Verify a BAA exists if you handle PHI. Only ChatGPT Enterprise and Healthcare offer a HIPAA BAA — Free, Plus, and Team do not, and compliance experts advise healthcare organizations to assume any PHI entered into non-Enterprise ChatGPT violates HIPAA. If a vendor cannot produce a signed BAA, treat the tool as off-limits for patient data. This is why My AI Call Center applies HIPAA-compliant communication standards to every clinic campaign rather than relying on a general policy.

3. Enforce SSO and centralized logging to eliminate shadow AI. Cyberhaven's research found that 32.3% of ChatGPT usage flows through personal accounts that bypass enterprise controls entirely, and IBM's breach study found shadow AI involved in 43% of security incidents. Blocking tools backfires — employees simply move to unmanaged alternatives. Single sign-on with logging keeps usage visible.

4. Audit third-party integrations for credential exposure. The Salesloft-Drift breach cascaded to 700+ organizations, with attackers harvesting OpenAI API credentials among stolen tokens. Okta avoided damage solely because IP allow-listing blocked the stolen token. Rotate credentials, allow-list where possible, and inventory every integration touching your AI accounts.

5. Retain ownership of deletion timelines. OpenAI commits to removing deleted conversations from systems within 30 days, and offers zero data retention on eligible API endpoints — but only if you configure them. Get the retention and deletion terms in your contract, not in a FAQ.

The pattern behind the checklist:

• Know exactly which tier you are on and what it trains on.
• Demand a BAA before any regulated data moves.
• Make all AI usage visible through SSO and logs.
• Treat integrations as your credentials, not the vendor's problem.
• Control deletion timelines in writing before launch.

The same discipline applies to who handles your contact data. My AI Call Center reviews list source and consent records before any campaign launches, declines bought lists without clear permission records, and never shares, sells, or trains shared models on client data — the same pre-launch verification this checklist asks you to run on any AI tool.