Best Practices to Protect Personal Identifiable Information (PII) with LLMs

Large language models (LLMs) are advanced artificial intelligence (AI) tools that exhibit powerful capabilities, including text generation, language translation, and content creation. However, utilizing LLMs in conjunction with personally identifiable information (PII) introduces significant risks. PII refers to information that can be used to identify individuals, such as names, physical addresses, Social Security numbers, dates of birth, driver’s license, bank accounts, health records, biometric data, employment history, communication history including email, text, instant messages, online behavior and even IP addresses. When PII is transmitted to LLMs, it poses potential threats of unauthorized access, privacy breaches, and legal non-compliance. This paper explores the risks associated with submitting PII to LLMs, outlines relevant laws such as GDPR, CCPA, HIPAA, and PCIDSS, and presents measures to mitigate these risks.

What are the Risks?

Submitting PII to LLMs entails the following risks:

  1. Data breaches: If PII is exposed due to LLM vulnerabilities or inadequate security measures, it may be accessed by unauthorized parties, leading to potential data breaches. Compliance with data protection laws is crucial in preventing such incidents.
  2. Privacy violations: Evenif PII remains undisclosed, LLMs possess the capability to infer sensitive information, raising concerns about privacy violations. These violations can result in severe legal consequences and damage an organization’s reputation.
  3. Legal implications: Organizations that process PII through LLMs must adhere to relevant industry-standard compliance frameworks like PCI-DSS, HIPAA, SOC 2, ISO 27001,IRS 4557and data protection regulations like GDPR to avoid penalties and maintain user trust.Failure to comply canadditionally lead to legal liabilitiesand reputational harm.
  4. Bias: LLMs are trained on extensive text datasets, and the presence of PII within these datasets may introduce biases into the model’s outputs. Compliance with industry laws ensures fairness and equality in AI applications.
  5. Ethical considerations: Using FMs and LLMsthat process sensitive data can raise ethical concerns, as the potential misuse of such data can lead to discrimination, social stigmatization, or other negative consequences. Organizationsmustconsider the ethical implications of their applications and strive to create responsible AI solutions.

Abide by the Laws

  1. GDPR: The General Data Protection Regulation applies to organizations operating within the European Union (EU) and governsthe processing of personal data. It emphasizes principles such as data minimization, purpose limitation, and the rights of data subjects. Compliance with GDPR is crucial when handling PII within LLMs to ensure data protection and user consent.
  2. CCPA: The California Consumer Privacy Act is a comprehensive privacy law applicable to organizations operating in California. It grants consumers specific rights regarding their personal information and imposes obligations on businesses to ensure transparency and data privacy. Adhering to CCPA is essential when using LLMs to process PII, particularly for businesses operating in California.
  3. HIPAA: The Health Insurance Portability and Accountability Act sets standards for safeguarding protected health information (PHI) in the healthcare industry. When healthcare organizations leverage LLMs to handle PII, compliance with HIPAA ensures the secure handling and storage of sensitive medical data.
  4. PCI DSS: The Payment Card Industry Data Security Standard outlines requirements for organizations that handle payment card data. When PII related to payment cards is transmitted to LLMs, adherence to PCI DSS ensures the secure processing, storage, and transmission of such data, minimizing the risk of unauthorized access and fraud.

What are the Options?

To prevent PII from Being Submitted to LLMs to mitigate the risks associated with submitting PII to LLMs, the following technical measures can be implemented:

  1. Anonymization or redaction techniques: Employ anonymization methods, such as tokenization or data masking, to remove or obfuscate PII before transmitting it to the LLMs. This ensures that the data remains pseudonymous, reducing the risk of privacy violations.
  2. Data tokenization or encryption: Tokenize or encrypt data before transmitting it to the LLMsto protect it from unauthorized access.
  3. Dynamic Data Masking (DDM): Dynamic Data Masking helps prevent unauthorized access to sensitive data by enabling customers to specify how much sensitive data to reveal with minimal effect on the application layer.Traditional DDM tools allow you to configure specific database fields to hide sensitive data in the result sets of queries.While using LLMs to process data submitted by users in real time or from a file that your customer has uploaded, it is most efficient to redact the input data dynamically as you submit it to an LLM.
  4. Secure LLMs: Utilize LLMs with robust security features, including encryption mechanisms for data in transit and at rest. Implement access controls to restrict unauthorized access to PII processed by LLMs, ensuring compliance with data protection regulations.
  5. Monitoring and auditing: Continuously monitor LLMs for signs of bias or privacy violations. Implement auditing processes to verify the quality and integrity of the training data, ensuring compliance with industry laws and regulations.


The risks of submitting PII to LLMs are substantial, necessitating compliance with relevant industry laws such as GDPR, CCPA, HIPAA, and PCI DSS. By implementing technical measures like anonymization, utilizing secure LLMs, monitoring for bias, and ensuring user awareness, organizations can safeguard individuals’ privacy, maintain compliance, and mitigate the risks of data breaches. Adhering to these measures is crucial across various industries, where protecting PII and ensuring legal compliance are paramount.Organizations must also educate users about the risks associated with submitting PII to LLMs and obtain explicit consent for data processing. Moreover, they must provide clear instructions on the minimum required data for LLM tasks and enable their end users to exercise their rights under relevant industry laws.

Comments are closed.