/>iAs we have previously detailed here, the latest generation of regulations under the California Consumer Privacy Act (CCPA), drafted by the California Privacy Protection Agency (CPPA), have advanced beyond public comments are closer to becoming final. These include regulations on automated decision-making technology (ADMT), data processing evaluation and risk assessment requirements and cybersecurity audits.
Assessments and Evaluations Overview
The new ADMT notice, opt-out and access and appeal obligations and rights go into immediate effect upon the regulation’s effective date, which follows California Office of Administrative Law (OAL) approval and would either be subject to the quarterly regulatory implementation schedule in the Government Code, or as has been the case with prior CCPA regulations immediately on OAL sign-off. We will not know if the CPPA will again seek a variance from the schedule until they submit the final rulemaking package.
Moving on to evaluations and risk assessments, the draft regulations do propose a phase-in, but only in part. Evaluations must be undertaken beginning on the regulation’s effective date, whereas assessment requirements apply to practices commencing on the effective date, but there is a 24 month period to complete, file certifications and abridged versions, and make available for inspection.
However, since Colorado, which like California, has very detailed requirements for conducting and documenting assessments, and New Hampshire, Oregon, Texas, Montana, Nebraska, and New Jersey already require assessments, Delaware and Minnesota will this summer, and Indiana, Rohde Island and Kentucky will by the new year, query whether the California phase-in is of much use. Out of the 20 state consumer privacy laws, all but Utah and Iowa require assessments.
Further, without at least a cursory assessment, how can you determine if the notice, opt-out and access and appeal rights apply?
So, what is the difference between an evaluation and an assessment?
First, they are required by different provisions. Evaluations are required by Section 7201, and risk assessments by Section 7150.
Next, there is no phase-in of evaluations as with risk assessments.
Risk assessments are much more complex and prescribed, and are at the core of a risk benefit judgment decision, and must be available for inspection and abridged summaries must be filed.
The content of the evaluation, which need not be published or subject to inspection demand outside of discovery, need only address if the process and technology is effective, in other words, materially error free, and if it discriminates against a protected class, in other words free of material bias. As such, they have similarities to assessments under the Colorado AI Act, effective next year but likely to be amended before then, and the recently passed Virginia HB 2094 AI bill that may or may not get signed by Governor Yougkin.
Thus, an evaluation alone won’t help you determine if the ADMT notice, opt-out and access and appeal rights apply, nor meet the risk assessment requirements. While it is a separate analysis, it can be incorporated into assessments assuming a company begins those immediately.
Also, evaluations are not required for selling and processing of sensitive personal information (PI), as are assessments, and assessments are only required for identification processing to the extent AI is trained to do so, whereas any processing for identification is subject to an evaluation. Since CCBA is part of behavioral advertising, which is part of extensive profiling, sharing needs to be addressed in both evaluations and assessments.
Finally, under Section 7201, a business must implement policies, procedures, and training to ensure that the physical or biological identification or profiling works as intended for the business’s proposed use and does not discriminate based on protected classes.
So on to assessments, what activities need to be assessed?
First selling or sharing. All 18 states that require assessments require them for this; though, for the non-California states, the trigger is processing for targeted advertising rather than “sharing”, which is broader than sharing for CCBA, but the California regulations catch up with the new concept of behavioral advertising.
Next, processing of sensitive personal information. The same 18 states require assessments for the processing of sensitive data, with differing definitions. For instance, what is considered children’s personal data differs considerably. Notably, the California draft Regulation amendments would raise the age from 13 to 16, and Florida is under 18. There is also variation in the definition of health data.
Note, while the Nevada and Washington (and potential New York) consumer health laws do not explicitly require assessments, they are practically needed, and Vermont’s data broker law requires initial risk assessments and a process for evaluating and improving the effectiveness of safeguards.
Other Risk Assessment Triggers
Assessments are mandatory before using ADMT to make or assist in making a significant decision, which is “profiling in furtherance of decisions that produce legal or similarly significant effects concerning a consumer.” This is a General Data Protection Regulation (GDPR) and European Data Protection Board (EDPB) inspired provision. The other states that require assessments also have a similar obligation, although the definitions may differ somewhat. In California, “Decisions that produce legal or similarly significant effects concerning a consumer” means decisions that result in the provision or denial of financial or lending services, housing, insurance, education enrollment or opportunity, criminal justice, employment opportunities, healthcare services, or access to essential goods or services. Only California gives guidance on what are essential goods or services, by means a parenthetical “(e.g., groceries, medicine, hygiene products, or fuel). Critics are concerned that this limits algorithmic custom pricing, sometimes derogatively referred to as surveillance pricing, or even AI of consumer behavior to decide where to open or close stores, though the aggregate or de-identified data should suffice for that. There is considerable guidance out of the EU, which can be looked to, though is clearly not binding. The EU approach is quite broad.
Speaking of looking to the EU, beware that the California and Colorado regulations diverge considerably from what is required under GDPR assessments, and keep in mind the material differences between GDPR with its lawful basis and legitimate interest tests and the US laws with opt-out concepts.
Uniquely amongst the states, California proposes the concept of extensive profiling, which covers any:
1) work or educational profiling;
2) public profiling; or
3) behavioral advertising.
Note however, that whilst behavioral advertising is said to include CCBA, it is broader and is defined as “the targeting of advertising to a consumer based on the consumer’s personal information obtained from the consumer’s activity—both across businesses, distinctly-branded websites, applications, or services, (i.e., CCBA) and within the business’s own distinctly-branded websites, applications, or services.” Significantly, this closes the gap between CCBA and the non-California regulation of targeted advertising, by including entirely 1st party behavioral advertising.
There is a carve out for “nonpersonalized advertising” as defined by CCPA Section .140(t), which means advertising and marketing that is based solely on a consumer’s personal information derived from the consumer’s current interaction with the business with the exception of the consumer’s precise geolocation. But also, note that here the exception is specifically limited to where the PI is not disclosed to third parties (i.e., not a processor or contractor). This has led some to argue that this guts the carve out. However, if personal data was disclosed to a third party, that would likely be a sale, especially given the breadth of the concept of “over valuable consideration” in the eyes of the regulators. So, the approach really is not inconsistent with the current treatment of contextual advertising.
PI to train ADMT or AI
Assessments are also proposed to be required for processing of PI to train ADMT or AI. This is another uniquely California concept, at least under state consumer privacy laws, and the California Chamber of Commerce and others, including some members of the legislature, have argued that, like other aspects of the proposed regulation’s treatment of ADMT, it goes beyond the Agency’s statutory authority. It is interesting to note that one of the topics included in the US House Committee on Energy and Commerce’s request for information to inform federal privacy legislation this week is the role of privacy and consumer protection standards in AI regulation and specifically the impact of state privacy law regulation of ADMT and profiling on US AI leadership. Another topic of focus is “the degree to which US privacy protections are fragmented at the state level and costs associated with fragmentation,” which seems to be inviting a preemption scope debate. So by the time at least this part of the regulation requires action, it may possibly be curtailed by federal law. That said, evaluations and assessments are practically necessary to guide compliance and information governance and to date repeated attempts at federal consumer privacy legislation have been unsuccessful.
Assessment Details
Most state laws do not have any specifics regarding how to conduct or document risk assessments, with the notable exception of Colorado. When it started assessment rulemaking, the Agency stated that it would look to try to create interoperability with Colorado and would also look to the guidance by the EDPB. While both can be seen to have influenced California’s proposed requirements, California adds to these.
Some of the content requirements are factual, such as purposes of processing and categories of PI. Others are more evaluative, such as the quality of the PI and the expected benefits and potential negative impacts of the processing, and how safeguards may mitigate those risks of harm. Nine examples are included in Section 7152(a)(5) to guide analysis.
Section 7152(a)(3) calls for analysis of specific operational elements for the processing.
Focus on Operational Elements
These operational elements are listed here[1] and can be seen as not only getting under the hood of the processing operations but also informing consumer expectations, and the risks and benefit analysis that is the heart of an assessment. Note, in particular, the inquiries into retention and logic, the latter meaning ‘built-in’ assumptions, limitations, and parameters that inform, power or constrain the processing, particularly as concerns ADMT.
Analysis and Conclusions
The assessment must not only document those processing details and the risk / benefit and risk mitigation analysis, but the conclusions and what was approved and/or disapproved.
The draft regulations call for participation by all relevant stakeholders, and they must be specifically named, as must the identification of the person responsible for the analysis and conclusions.
Filing and Certification
California diverges from the other states with respect to reporting requirements. Annually a responsible executive must certify to the CCPA that the business assessed all applicable processing activities, and an abridged assessment must be filed for each processing activity actually initiated. This will make it very apparent which businesses are not conducting assessments.
Further, the draft regulations limit what is required in the abridged assessments to largely factual statements:
- The triggering processing activity;
- The purposes;
- The categories of personal information, including any sensitive categories; and
- The safeguards undertaken.
Note that the risk / benefit analysis summary is not a part of the filing.
Inspection and Constitutional and Privilege Issues
Contrast that with the detailed risk / benefit analysis required by the full assessment, which, like all of the other states that require or will require assessments, is subject to inspection upon request.
This GDPR-inspired approach to showing how you made decisions calls for publication of value judgments, which, as I have opined in an article that is in your materials (see a synopsis here), is likely unconstitutional compelled speech. While the 9th Circuit in the X Corp and NetChoice cases struck down harm assessment and transparency requirements in the context of children’s online safety, the Court distinguished compelling disclosure of subjective opinions about a company’s products and activities from requiring disclosure of merely product facts. There is no 1st Amendment in GDPR-land, so we will have to wait and see if the value judgment elements of assessments can really be compelled for inspection.
Inspections also raise serious questions about attorney-client and work product privilege. Some states specifically provide that inspections of assessments is not a waiver of privilege, and/or that they will be maintained as confidential and/or are not subject to public records access requests. The draft regulations do not; however, the CCPA itself provides that the Act shall not operate to infringe on evidentiary privileges. At any event, consider labeling legal analysis and counsel as such and maintaining them apart from what is maintained for inspection.[2]
[1] Planned method for using personal information; disclosures to the consumer about processing, retention period for each category of personal information, categories of third parties with access to consumers’ personal information, relationship with the consumer, technology to be used in the processing, number of consumers whose personal information will be processed and the logic used.
[2] Note – Obtaining educational materials from Squire Patton Boggs Services Ireland, Limited, or our resellers, does not create an attorney-client relationship with any Squire Patton Boggs entity and should be used under the direction of legal counsel of your choice.
