Natural Language Processing (NLP) teams need to stay up to date with the latest data protection regulations when training their models with user data. Not only do customers continue to demand privacy, but according to the GDPR they must give positive consent for their personally identifiable information (PII) to be used in training NLP models. Additionally, data must be free of any unnecessary personal information, detailed records kept of where personal data is stored, and measures are needed to ensure that machine learning models cannot leak data or enable reverse-engineering outputs or weights (a common concern with any language model) when in production.
Data Protection Regulations to Watch for in 2022: A Guide for NLP Teams
While many companies are aware of the requirements for the GDPR, CCPA, POPIA (South Africa), and LGDP (Brazil), it’s difficult to keep track of every upcoming data protection regulation and the regular changes to the current regulations. Even figuring out whether they even affect one’s business can be painful. Do they apply to a region’s citizens worldwide, like the GDPR and Thailand’s PDPA? Do they only apply to businesses that surpass a threshold of data collected from residents of a state, like the CPRA? Have they been amended to remove any such threshold and only apply to data subjects physically located in the region in question, like Japan’s APPI amendment?
Here’s what to keep an eye out for 2022:
- California Privacy Rights Act (United States)
- Act on the Protection of Personal Information Amendments (Japan)
- Shenzhen Special Economic Zone Data Regulation (China)
- Personal Data Protection Act (Thailand)
California & the CPRA
The California Privacy Rights Act (CPRA), that takes effect on January 1st, 2023, will apply to information collected after January 1st, 2022. A few highlights from The Complete Guide to California Privacy Rights Act (CPRA) on how the CPRA enhances the California Consumer Privacy Act (CCPA) are:
- New categories of sensitive data were added, including: SSN, driver’s license, state ID, passport number, login credentials (e.g., passwords), precise geolocation, racial/ethnic origin, religious beliefs, union membership, content of messages, genetic and biometric data, medical data, sex life, and sexual orientation. Consumers must have a clear option to limit the use or opt out of having their sensitive information be used by a business.
- Consent should be “a freely given, specific, informed and unambiguous indication of the consumer’s intent.”
- Consumers must have the option to not have their personal information shared with third parties (e.g., advertisers) and must have the ability to opt out of having that data be shared.
- Consumers must have the option to opt out of automated decision-making. Automated decision-making will have to be explainable.
- Consumers must be informed about the length of retention of their personal data and have the right to request deletion or correction of their data.
- Data minimization becomes a requirement, with businesses only being allowed to collect the necessary personal data and only hold onto it for the necessary amount of time.
- Like the CCPA, the CPRA applies to businesses who: 1) had a gross annual revenue of $25 million or more in the previous year; 2) buy, receive, or sell personal information of 100,000+ California residents, households, or devices; or 3) derive 50% or more of their annual revenue from selling or sharing California residents’ personal information (as described in California’s Office of the Attorney General’s site).
Japan & the APPI
Amendments to the Act on the Protection of Personal Information (APPI) in Japan are taking effect on April 1st, 2022. These amendments include:
- The need to explicitly mention the purposes for which personal data will be used, along with the need to provide detailed information if personal data is shared with third parties or with parties outside of Japan.
- Information on how the data subject can exercise rights over their “retained personal data.”
- Purposes of use also have to be specified for pseudonymized information that is linked to the original data.
- Now applies to every Personal Information Controller in Japan (see also Japan – Data Protection Overview).
Shenzhen Local Data Regulation
- Data processors need to implement de-identification or anonymization measures for personal data, sensitive personal data, and important data which is specified by the state.
- Data subjects can partially withdraw consent (i.e., for specific steps of the data processing, storage, and transfer pipeline).
- The personal information of minors (13 years old or less) is considered sensitive personal data. No data profiling or ‘recommendation personalization’ is allowed for minors.
- This regulation establishes standards for data trading platforms, which requires the protection of personal data, among other information.
- Applies to the “data activities of natural persons, legal persons, and unincorporated organizations in the special zone” (see China: Draft Data Regulations of the Shenzhen Special Economic Zone).
Note that the Shanghai Data Regulation is in addition to China’s Personal Information Protection Law (PIPL). Other regional data protection laws are also popping up and companies processing personal data in the region should be on the alert.
The Personal Data Protection Act (PDPA) will go into effect on June 1st, 2022. Since the PDPA took quite a bit of inspiration from the GDPR, we’ll highlight their key differences:
- Special categories of personal data are not defined in the PDPA (whereas under the GDPR there are 7 special categories, the processing of which is prohibited except under special circumstances). The PDPA does, however, require explicit consent to collect “personal data pertaining to racial, ethnic origin, political opinions, cult, religious or philosophical beliefs, sexual behaviour, criminal records, health data, disability, trade union information, genetic data, biometric data, or of any data which may affect the data subject in the same manner, as prescribed by the PDPC.”
- The PDPA gives data subjects the right to request that their personal data be anonymised.
- No mention of whether data subjects need to be informed of automated decision-making and profiling.
- Applies to “all organisations that collect, use or disclose personal data in Thailand or of Thai residents, regardless of whether they are formed or recognised under Thai law, and whether they are resident or have a business presence in Thailand (see Data Protection Laws of the World).
What these data protection regulations mean for NLP teams
There’s a common thread of ‘consent’ across most data protection regulations: do data subjects know for which purposes you’re using their personal data, how long you’re storing their personal data, and with whom you’re sharing the data? But where they often differ are with regards to the rights of the data subject to request, correct, or have their data deleted, how consent should be obtained, how explicit and fragmented that consent should be, and how opting out should work.
If you’re looking for help with consent management, we’ve heard great things about Dataships.
Another recurring trend among data protection regulations is data minimization: businesses should only collect the personal data they need, and only store it while it’s actively needed – no more hoarding information to figure out a purpose afterwards.
The safest thing for organizations to do has been to follow the GDPR as the baseline for all data, but with different types of information being categorized as sensitive or not (and personally identifiable or not), it becomes draining for NLP teams to keep track of data protection regulations on top of the latest research in NLP & Machine Learning. This is what we live and breathe at Private AI, so NLP teams can take dealing with personally identifiable information off their very full plates and focus on their core products.