BlurDataGuide · Health records and AI
Is ChatGPT HIPAA compliant? And Claude?
The direct answer first, with links to what OpenAI and Anthropic say. Then the practical part: how to take the patient's identifiers out of a PDF, a scan or a screenshot on your Mac before anything gets pasted.
Short answer
- Not by default. Neither tool is HIPAA compliant just by existing. Protected health information (PHI) belongs only in products covered by a Business Associate Agreement (BAA) with OpenAI or Anthropic, and only in the features that agreement covers.
- ChatGPT: OpenAI offers a BAA for products such as ChatGPT for Healthcare, ChatGPT Enterprise with a Regulated Workspace, ChatGPT for Clinicians and its API with modified data retention. It offers no BAA for ChatGPT Business, and Free, Plus and Pro are not on its list of HIPAA eligible products.
- Claude: Anthropic offers a BAA for HIPAA-ready Claude Enterprise, its API and Claude in Microsoft Foundry. Team plans and the Free, Pro and Max plans can't enable HIPAA.
- No BAA? Don't paste the record. De-identify it first: remove the 18 Safe Harbor identifiers, check the copy, then paste the copy.
As of October 2026, from OpenAI's and Anthropic's own help pages, linked below. General information, not legal advice.
- Made in Europe
- 100% offline
- 14-day money-back guarantee
What a BAA is, and which plans can have one
HIPAA allows a covered entity, such as a healthcare provider or a health plan, to give PHI to a vendor that handles it on its behalf when there is a business associate agreement: a written contract in which the vendor commits to safeguard the data (HHS on business associates). An AI service that receives a patient record is that kind of vendor. So the useful question is not whether the tool is HIPAA compliant, but whether there is a BAA, and whether it covers the plan and the features you use.
ChatGPT (OpenAI)
On OpenAI's list of HIPAA eligible products, available with a BAA:
- ChatGPT for Healthcare
- ChatGPT Enterprise with a Regulated Workspace
- ChatGPT for Clinicians, a workspace for verified US clinicians who can sign their own BAA
- The API with modified data retention: eligible organizations can accept the standard BAA in the API settings
For a sales-managed ChatGPT Enterprise or Edu account, OpenAI says to contact sales.
Not covered: OpenAI offers no BAA for ChatGPT Business, says Health in ChatGPT offers none either, and does not list Free, Plus or Pro. On eligible products, some features still fall outside the BAA.
Sources: HIPAA eligible products · BAA for the API and ChatGPT · Health in ChatGPT
Claude (Anthropic)
Anthropic offers a BAA for its HIPAA-ready services:
- Claude Enterprise, once the organization's Primary Owner turns on HIPAA in the organization settings and accepts the BAA. A standard Enterprise plan is not covered without that step
- The Claude API, after the organization signs a BAA and Anthropic turns it on
- Claude in Microsoft Foundry
Not covered: Team plans and individual plans (Free, Pro and Max) can't enable HIPAA. Inside a covered organization, Anthropic excludes some features, such as Claude Console, Cowork in the cloud and features in beta, and data sent to third parties through connectors.
Sources: HIPAA-ready Enterprise plans · BAA for commercial customers
Both companies make the same point in different words: a BAA is where compliance starts, not where it ends. OpenAI's help center says that accepting a BAA does not, on its own, make your use of the service HIPAA compliant. Plans and coverage change often, so check the linked pages before you rely on this summary.
Why a personal or team plan is not the place for PHI
Most people who search this are on a personal or team plan. These are the reasons to keep patient records out of it.
There is no BAA
For a covered entity or a business associate, giving PHI to a service that hasn't signed a BAA can be a disclosure HIPAA doesn't allow, even if nothing bad ever happens to the data.
Chats are kept
A pasted record sits in your chat history. Depending on the plan and your settings, chats can be kept for a long time and may be used to improve the models.
Deleting doesn't undo it
Once the record has reached someone else's servers, deleting the chat later doesn't change the fact that it was shared.
It's outside your organization's controls
A personal account has none of your organization's access rules, audit logs or retention settings, and your privacy officer can't see what was shared.
The 18 identifiers HIPAA Safe Harbor asks you to remove
HIPAA gives two ways to de-identify health information. In Expert Determination, a qualified expert documents that the risk of identifying someone is very small. Safe Harbor is the checklist: remove 18 kinds of identifiers. Under the Privacy Rule, information de-identified with either method is no longer protected health information. The list applies to the patient and to the patient's relatives, employers and household members:
- Names
- Places smaller than a state: street address, city, county, ZIP code (the first three ZIP digits may stay in some cases)
- Every part of a date tied to the person except the year: birth, admission, discharge, death; and ages over 89
- Phone numbers
- Fax numbers
- Email addresses
- Social Security numbers
- Medical record numbers
- Health plan beneficiary numbers
- Account numbers
- Certificate and license numbers
- Vehicle identifiers and serial numbers, including license plates
- Device identifiers and serial numbers
- Web addresses (URLs)
- IP addresses
- Biometric identifiers, such as finger and voice prints
- Full-face photos and comparable images
- Any other unique number, characteristic or code
There is one more condition: you must not actually know that what is left could identify the patient. The full text is in the HHS de-identification guidance. For how BlurData handles each identifier, see the item-by-item table.
How to de-identify a record on your Mac before you paste it
BlurData reads PDFs, scans, photos and screenshots on your Mac, and the record is never uploaded. This is the workflow for a US medical record.
MRN 00482913
1840 Larkspur Lane, Springfield, IL 62704
Daughter Ana Ortega, (217) 555-0143
A1c 7.9%. Metformin 1000 mg twice daily.
MRN
Daughter ,
A1c 7.9%. Metformin 1000 mg twice daily.
Redact takes the identifiers out and leaves the clinical content. Pseudonymize would put labels such as [PERSON_1] and [ADDRESS_1] in their place. (Sample data.)
Open the record in BlurData
Drag in the PDF from the patient portal or the EHR, a scan, a phone photo of a paper page, or a screenshot. Long charts are fine: every page is analyzed, and export stays locked until the analysis is complete.
Choose the US Medical Records preset
Install it with one click from the free preset library. It looks for patient names, full dates in common US formats, medical record numbers, Social Security numbers, street addresses, ZIP codes, emails and phone numbers, and leaves lab values such as 12.4 or 1,234.56 alone.
If the record has web links, IP addresses, account numbers, license plates, barcodes or QR codes, switch those detections on too. For health plan, license or device numbers, add a pattern once and it is found every time.
Redact, or Pseudonymize
Redact removes the text from the file. Pseudonymize puts a label in its place, such as [PERSON_1] or [ADDRESS_1], the same on every page, so the AI can still tell the patient from a relative. Either way, diagnoses, medications, lab values and notes stay readable.
Check every page
Every detection is highlighted before anything is saved. Remove what should stay, double-click a missed word to add it on this page or everywhere, and draw a box over faces, signatures, handwriting and wristband barcodes. Read the free-text notes line by line: that is where identifiers hide.
Export the copy
Redact saves a new PDF with the text removed, or images with the pixels rewritten. Pseudonymize can also save as Markdown for AI, a clean text file that is easy to paste into a chat.
With Pseudonymize, Save key gives you a small CSV file that turns the labels back into the real values. Keep it on your Mac, apart from the copy, or don't save it at all.
Check the copy, then paste
Run the exported PDF through the free checker below, then search the copy for the patient's last name, record number and date of birth. Nothing should come up. Now the copy is ready for ChatGPT or Claude.
Free check before you paste. Our redaction checker looks for text left under black boxes and in the file's properties, right in your browser. Your file is never uploaded.
Check a redacted PDFWhat a tool can't do for you
BlurData finds and removes identifiers on your Mac. That is a large part of de-identification, not all of it.
It doesn't make you compliant
HIPAA compliance is your organization's policies, agreements, training and records. No app provides it, BlurData included. Ask your privacy officer whether de-identified copies may go to an AI tool at all: many organizations have their own rules.
Safe Harbor needs a person
The rule asks that you don't know of anything left that could identify the patient. Only someone who reads the result can judge that: a rare condition, an unusual job and a local event mentioned together can point to one person.
Free text hides identifiers
Notes say things like "her husband Tom drove her in", "works at the county fire station" or "turned 93 last week". Partial dates such as "March 2024" don't look like dates to a pattern. Read the narrative yourself.
Names are the weak spot
Unusual spellings, nicknames, names that are also ordinary words (Hope, Rose, Will), names split across lines and handwriting on scans can be missed, and a faded scan can be misread. Search the copy for every name you know is in the record.
Frequently asked questions
Is ChatGPT HIPAA compliant?
Not by itself, and not on every plan. As of October 2026, OpenAI offers a Business Associate Agreement (BAA) for specific products: ChatGPT for Healthcare, ChatGPT Enterprise with a Regulated Workspace, ChatGPT for Clinicians and its API with modified data retention. It offers no BAA for ChatGPT Business, and Free, Plus and Pro are not on its list of HIPAA eligible products. Without a BAA, keep PHI out of ChatGPT.
Is Claude HIPAA compliant?
Only in a HIPAA-ready setup. As of October 2026, Anthropic offers a BAA for Claude Enterprise once the organization's Primary Owner turns on HIPAA and accepts the agreement, for the Claude API and for Claude in Microsoft Foundry, with some features excluded. Team plans and the Free, Pro and Max plans can't enable HIPAA.
Can I paste a de-identified record into ChatGPT or Claude?
Under the HIPAA Privacy Rule, health information de-identified with the Safe Harbor or the Expert Determination method is no longer PHI, so a properly de-identified copy is not the same question as the original record. Two cautions: Safe Harbor needs all 18 identifiers gone and someone who checks the result, and your organization may have its own rules on AI tools. Ask your privacy officer.
What about ChatGPT Health?
OpenAI says Health in ChatGPT comes with no BAA and is not meant for clinical work or for covered entities. For HIPAA-regulated work, OpenAI points to ChatGPT for Healthcare and ChatGPT for Clinicians instead.
I'm a patient. Can I ask ChatGPT or Claude about my own records?
HIPAA sets rules for healthcare providers, health plans and the businesses that work for them. It generally doesn't govern what patients do with their own records. Your privacy is still worth protecting: taking out your name, date of birth, record numbers and address first takes a few minutes, and the AI can still explain the results.
Does BlurData make me HIPAA compliant? Does it upload the record?
No to both. BlurData holds no HIPAA certification, and no app can make you compliant. It finds and removes identifiers: reading, detection and export happen on your Mac, and the record is never uploaded. Whether a copy meets the Safe Harbor standard is your decision, or your privacy officer's.
This page is general information, not legal advice. Plan names and BAA coverage are summarized from OpenAI's and Anthropic's own help pages as of October 2026 and can change. Check them, and your privacy officer, before you use any AI tool with patient data.