Score Breakdown by Category
How OpenAI's terms rate across my evaluation categories.
As the most widely-used AI service, OpenAI's terms affect millions of users. While they've improved transparency over time, questions remain about training data opt-out complexity, frequent policy changes, and the scope of rights you grant when using ChatGPT.
ChatGPT is usable for most people, but the right settings and habits change a lot depending on who you are. Pick the column that matches you. Menu paths reflect the ChatGPT web and app settings as of early 2026; OpenAI relabels these screens periodically, so confirm the exact wording in your own account.
Improve the model for everyone (also labeled Chat History & Training) to off. On free and Plus accounts, OpenAI's documented default is that this is on, so you have to change it yourself.Delete. OpenAI's stated policy describes removing deleted chats from its systems within roughly 30 days, subject to exceptions; confirm the current figure in its help docs.How OpenAI's terms rate across my evaluation categories.
These provisions require careful attention for OpenAI users.
OpenAI does offer a training opt-out, but as I read the consumer terms, the default runs the other way: on free and Plus accounts, content you provide can be used to improve the models unless you change the setting. The control sits in Data Controls and has been relabeled more than once.
The terms reserve a broad right to update themselves, and in my framework's view the notice mechanism leans toward the company: material changes can take effect through posting rather than a direct, individually acknowledged notice. Read the change-of-terms clause so you know what counts as notice to you.
As I read the terms, the usage rules are broad and the suspension and termination power is discretionary. That combination means an account can be restricted on the company's judgment of a violation, which is worth understanding before you rely on continued access.
The terms disclaim accuracy and place the risk of relying on outputs on the user. As I read this, you should treat outputs as a draft to verify, not as a reliable statement of fact, and assume the contract gives you little recourse if an output is wrong.
The detail behind the score: how to actually turn training off, where the consumer and enterprise privacy gap sits, which use cases carry real risk, how OpenAI compares to its two closest rivals, and what never belongs in a prompt. Each block is collapsed; open what is relevant to you.
On consumer ChatGPT, the relevant toggle has been relabeled more than once. As of early 2026 the path is:
Settings › Data Controls › Chat History & Training (or Improve the model for everyone) › set to Off
API and business tiers are different. OpenAI states that data sent through the API, and through ChatGPT Enterprise and Team, is not used to train models by default. That is the opposite default from a free or Plus personal account, which is the gap most people miss.
The biggest misunderstanding I see is people assuming "I pay for Plus, so my data is protected." Plus is still a consumer plan. As I read OpenAI's published terms, the meaningful protections sit in the business tiers. The table below summarizes my reading of those documents; confirm the current wording before you rely on any row.
| Dimension | Free / Plus (consumer) | Team / Enterprise / API (business) |
|---|---|---|
| Used for training by default | Yes, unless you opt out | No |
| Data processing agreement (DPA) | No | Yes, available |
| Admin controls / retention settings | Limited | Yes |
| Contractual confidentiality commitments | Standard ToS only | Stronger, negotiable at Enterprise |
| Suitable for regulated / client data | Generally no | Possibly, with the right terms in place |
If you handle client files, patient data, or anything covered by a confidentiality duty, the consumer plan is usually the wrong tool regardless of how careful you are with individual prompts.
Not every use of ChatGPT carries the same exposure. Here is how I rank common ones by how much caution they call for:
Public-facing writing, idea generation, summarizing content you already have rights to. Little sensitive data involved.
Watch for secrets in code (API keys, credentials) and license terms on generated snippets. Strip secrets before pasting.
Strategy notes and internal memos may be confidential or trade secret. Use a business tier and avoid the consumer app.
Pasting customer names, addresses, or account details into a consumer tool can breach your own privacy promises and policies.
Account numbers, tax data, and financial records deserve a tool with the right contractual protections, not a personal login.
Health data can trigger HIPAA or state health-privacy rules. The consumer ChatGPT is generally not an appropriate place for it.
These three are among the most widely used consumer AI services. On the terms and privacy dimensions I score, my reading is that they are not equal. My ToS Watchdog scores, which reflect my opinion under my published methodology: Anthropic 72 (B), OpenAI 55 (C+), Google Gemini 48 (C). Each rating is my evaluative read of the company's own published terms and privacy policy, not a statement of objective fact.
| Dimension | OpenAI (ChatGPT) | Anthropic (Claude) | Google (Gemini) |
|---|---|---|---|
| Fairness score | 55 / C+ | 72 / B | 48 / C |
| Consumer training default (per published policy) | On unless you opt out | States consumer chats are not used to train by default | On unless you opt out |
| Opt-out clarity (my read) | Moderate, relabeled often | Strong | Moderate, in account settings |
| Account / ecosystem linkage (my read) | Narrower | Narrowest | Tied to a broader account, per its policy |
| My fit for privacy-first users | Acceptable | Strongest | Weakest of the three, in my view |
Scores reflect my reading of each company's published terms and privacy policies at the time of review. Companies revise these documents, so treat the comparison as a snapshot. See the full Anthropic review and Gemini review for the detail behind each grade.
The cleanest privacy strategy is to keep sensitive data out of the prompt in the first place. Before you hit send, scan for anything on this list:
On free and Plus consumer accounts, the default is that your content can be used to improve the models unless you opt out in Data Controls. Business tiers and the API state the opposite default: not used for training. So the answer depends entirely on which plan you are on and whether you flipped the setting.
Opting out limits future use; it does not automatically erase what was already processed. Separately, you can delete individual chats, which OpenAI says are removed from its systems within about 30 days. A court-ordered litigation hold can override routine deletion, so deletion is best treated as a goal, not a guarantee.
The consumer app generally is not the right tool for regulated or confidential data. If you need it for business, use Enterprise, Team, or the API, get a data processing agreement in place, and confirm how retention and deletion are handled before you rely on it.
Yes. The terms let OpenAI suspend or terminate access for Usage Policy violations at its discretion, and a warning is not guaranteed. Export your data through Data Controls before you lose access, and if you are banned, ask support which specific policy provision was cited and keep that record.
Under OpenAI's terms, as between you and OpenAI, you are generally assigned its rights in the output, subject to the terms and the law. Copyright protection for purely AI-generated material is uncertain in the United States, and others can receive similar outputs, so do not assume exclusive rights over what you generate.
Anthropic scores higher on my fairness index (72 vs 55), largely because it states it does not use consumer chat data to train models by default and its opt-out posture is clearer. If training-data exposure is your main concern, that difference is meaningful. The full Anthropic review is linked above.
In my framework's view, yes, in principle. When an AI company is in active litigation, a preservation order can require it to retain certain records, which can cut against a routine deletion request. The practical takeaway is that, while any such hold is in effect, some data you might expect to be deleted may need to be retained. That is a reason to be conservative about what you put in, not a reason to panic. Confirm the current status of any specific litigation before relying on it.
"My honest read: OpenAI's terms are middle of the pack. They are not predatory, but the consumer default of training on your chats, the frequent and quietly relabeled changes, and the broad termination discretion mean you carry more of the risk than the marketing suggests. The single highest-value move for almost everyone is not a setting, it is discipline about what you paste. Turn training off, yes, but assume anything you type could be retained, reviewed, or pulled into litigation, and act accordingly. For regulated or client data, a consumer plan is the wrong tool no matter how careful you are."
This page is attorney-supervised informational content, not legal advice, and does not create an attorney-client relationship. Terms and privacy policies change; verify the current language in your own account and OpenAI's published documents before relying on anything here. For a question specific to your situation, you can email me the facts and any documents.
Below is how I read OpenAI's current Terms of Use and related published policies on the points that matter most: data retention, content ownership, and training. Verify the current language in OpenAI's own documents and your own account before relying on any of it.
As I read the privacy terms, a deletion request reduces future exposure but does not guarantee erasure of everything already processed, and a preservation order in pending litigation can require retention of data you asked to delete. Treat deletion as a goal, not a guarantee.
OpenAI's Terms of Use address content ownership, API usage, and acceptable use. OpenAI states that it may update these terms over time, so confirm the current version applies to you.
OpenAI provides a training opt-out for consumer ChatGPT in Settings › Data Controls, and states that data sent through the API and through its business tiers is not used to train its models by default. Confirm the exact wording in your own account and the current published terms.
Need help understanding your rights or resolving a platform dispute?
Analysis