Can clinic staff paste patient information into ChatGPT? The safe default is no: do not enter identifiable patient information into an unapproved public AI account. Use patient information only in a product and workflow your organization has specifically reviewed and approved for that purpose.
That does not mean every AI tool works the same way or that every use is automatically unsafe. Product versions, account types, features, integrations, and data practices can differ. A tool being useful for general drafting does not, by itself, make it appropriate for patient information. Your practice should establish what is approved before staff use AI in clinical or administrative workflows.
Why patient information in an AI prompt needs care
A prompt may contain protected health information (PHI) even when it is written as an everyday request. A clinical note, appointment detail, patient message, or description of a case can identify someone directly or through the surrounding details. Information entered into an outside tool may also be handled under the vendor’s product terms and technical setup, which your practice should understand before use.
Removing a patient’s name alone does not necessarily make a clinical note de-identified. Dates, locations, rare circumstances, and information about relatives may still point to a particular person. A prompt that seems anonymous to the person writing it may be recognizable to someone with additional context.
For that reason, staff should avoid copying real patient records into a general-purpose tool to “test” it, ask for a summary, or improve a message. Use fictional examples or information that has been appropriately de-identified under your organization’s process when practicing with a tool that is not approved to handle PHI.
Is regular ChatGPT HIPAA compliant?
There is no safe blanket answer based only on the product name. Do not assume a consumer account, a healthcare-oriented offering, an API, or a particular feature has the same privacy and security terms as another option. The relevant question is whether the exact product, account, configuration, and intended workflow have been reviewed and approved by your organization for the information being used.
A business associate agreement (BAA) is an important checkpoint when a vendor handles PHI on behalf of a healthcare organization. But a BAA is not a blanket approval for every product or feature from that vendor. The practice should confirm which services and integrations are covered, how data is handled, and whether its own safeguards and risk review support the proposed use. Direct questions to the people responsible for privacy, security, and compliance at your organization.
Turning off model training does not, on its own, establish that a tool is safe for PHI. Training settings address only one part of the picture. Your review should also consider such matters as data retention, who can access the information, connected services, account controls, and the specific terms that apply to the product and workflow.
What to check before staff use an AI tool
A written healthcare AI policy can help staff make consistent choices instead of guessing at the point of use. Before allowing an AI tool in a workflow involving patient information, have the appropriate organizational leads review the tool and define acceptable uses. Consider questions such as:
- Which exact product and account are approved? Identify the approved version, features, integrations, and user accounts. Avoid vague guidance that simply says “AI is allowed.”
- Can the tool receive the information involved? Confirm whether the intended use includes PHI, and whether the vendor relationship, agreement, and safeguards cover that use.
- What happens to prompts and outputs? Review applicable data handling, retention, access, and training settings rather than relying on assumptions or a single setting.
- Who may use it, and for what? Set boundaries for tasks such as drafting, summarizing, or organizing information. Make clear which tasks require human review and which are not permitted.
- How will staff get help? Give employees a clear contact and reporting path for questions, mistakes, or unexpected tool behavior.
These checks are especially important when a workflow involves clinical notes, patient communication, or sensitive behavioral-health information. Some records may have additional confidentiality considerations, so the organization should review them according to its own policies and requirements.
Can AI help draft patient messages or summarize notes?
Possibly, but the answer depends on the tool and the workflow your organization has approved. Staff should not paste identifiable information into an unapproved tool and assume that editing the result afterward removes the risk. Even a draft request can disclose patient details at the point it is entered.
If a practice approves an AI workflow, staff should follow its instructions about what information may be used, who reviews the output, and how the final message or summary is checked. AI-generated text may be incomplete or inaccurate, so it should not be treated as a substitute for appropriate clinical judgment or the practice’s established review process.
For help building broader privacy safeguards into everyday technology use, see our guidance on protecting patient data and our overview of HIPAA compliance for medical practices.
What if an employee pasted PHI into a public AI tool?
Respond promptly and follow your organization’s incident procedure. Staff should not try to handle the situation quietly or decide on their own that it is harmless. A practical first response is:
- Stop further use. Do not enter more patient information or continue the same task in that tool.
- Notify the designated privacy or security lead promptly. Use the practice’s established reporting channel and share what happened.
- Preserve relevant details. Record the tool and account used, the general type of information involved, and when the entry occurred, following internal instructions.
- Follow the organization’s incident process. Let the appropriate team assess the circumstances and determine next steps. Do not assume that deleting a prompt resolves the issue.
Clear reporting helps the organization assess what happened and respond through its established process. Practices can make this easier by including AI-related mistakes in staff guidance and making the reporting contact easy to find. For related security considerations, see our article on clinic cybersecurity.
Frequently asked questions
Does removing a patient’s name make a note safe to paste?
Not necessarily. Dates, locations, rare circumstances, and details about relatives can still identify a patient. Use only information that your organization has appropriately de-identified or approved for that specific tool and purpose.
Does turning off model training make it safe to enter patient information?
No single setting answers the full privacy and security question. Training preferences do not replace review of the exact product, data handling, agreements, integrations, and organizational safeguards.
Can staff use AI to draft a patient message?
Only through a workflow approved by the practice. Do not include identifiable patient details in an unapproved tool, and follow the organization’s requirements for reviewing any AI-generated message before use.
What should staff do if they are unsure whether a detail identifies a patient?
Do not enter it into the tool while uncertain. Ask the practice’s designated privacy, security, or compliance contact which information and tools are approved.
Make AI use a reviewed part of practice policy
Clinic staff should not paste identifiable patient information into an unapproved public AI account. Approval should apply to the specific product and workflow, with clear guidance on information use, vendor terms, staff responsibilities, and incident reporting. Zevonix provides healthcare-focused IT, cybersecurity, and compliance support for practices in St. Augustine and St. Johns County. Contact the team to discuss how AI use can fit into your practice’s technology and privacy processes.