The FTC’s 2016 big data report said stripping identifiers would not take an eligibility score outside the FCRA
On January 6, 2016, the FTC said a score crafted for a decision about a particular applicant is a consumer report even with identifiers stripped. Coverage turned on use, which put the compliance question with the business owner.
Looking Back revisits past weeks one at a time, up to the most recent, and asks what it meant for the banks, health systems and public agencies we serve. The series was written in 2026, with the benefit of hindsight, and each article says which week it looks back on.
Two-minute summaryThe dek and the first sentence of each section, quoted from the article
On January 6, 2016, the FTC said a score crafted for a decision about a particular applicant is a consumer report even with identifiers stripped. Coverage turned on use, which put the compliance question with the business owner.
- The week On January 6, 2016, the Federal Trade Commission released a report titled “Big Data: A Tool for Inclusion or Exclusion?
- Coverage: the score, not the name The report started from the FCRA’s definition.
- Notice: what the consumer is told The report reminded users of consumer reports of two notices.
- Fair treatment: treatment and impact The report described two theories.
- The seller’s duty: knowing the customer The fourth part concerns the company that sells data or analytics.
- How the questions reach different organizations A bank or credit union that uses outside scores for underwriting, account opening or pricing faces all four questions.
- Objections, in order The report said its discussion articulated considerations relevant to the Commission’s enforcement authority. [2] It was not a rule.
- Checks a buyer of analytics could run List every decision about an individual that uses an outside score, model or data set.
- With hindsight The CFPB took up the report’s notice footnote in 2022.
The week
On January 6, 2016, the Federal Trade Commission released a report titled “Big Data: A Tool for Inclusion or Exclusion? Understanding the Issues.” [1] The Commission’s press release said the vote to issue it was 4-0, and that Commissioner Maureen Ohlhausen issued a separate statement. [1]
The release described the report’s scope. It said the report looks specifically at big data at the end of its lifecycle: how it is used after being collected and analyzed. [1] It said the report drew on the FTC’s 2014 workshop of the same name and on a Commission seminar on alternative scoring products. [1]
The report itself is dated January 2016. [2] It divided the life cycle of big data into four phases: collection; compilation and consolidation; analysis; and use. [2] It said it focused on the fourth phase. [2] It also said it addressed only the commercial use of big data consisting of consumer information, with a focus on low-income and underserved populations. [2]
The workshop behind it took place on September 15, 2014, according to the report. [2] The report said the workshop had four panels, and that the Commission received sixty-five public comments. [2]
The release summarized the report’s balance. It said the report highlights uses of big data that benefit underserved populations, including access to credit through non-traditional methods and specialized health care. [1] It also listed possible risks, including more individuals mistakenly denied opportunities based on the actions of others and the targeting of vulnerable consumers for fraud. [1]
The legal part of the report named three bodies of law: the Fair Credit Reporting Act (FCRA), federal equal opportunity laws, and Section 5 of the FTC Act. [2] It said the discussion articulated considerations relevant to the Commission’s exercise of its enforcement authority. [2] It added that other federal, state and local laws may apply. [2]
On the FCRA, the report restated the definition of a consumer report. It said the communication must be prepared or provided to others to make an eligibility determination about a particular consumer. [2] It then took up a hypothetical that mattered for anyone buying a model’s output.
In that hypothetical, a company collects a credit applicant’s zip code and information about her social media and shopping behavior. [2] The company strips her identifying information and sends the application to an analytics firm. [2] The firm scores people with similar characteristics and returns a score, grade or recommendation, knowing it will be used for a credit decision. [2] The report said the Commission would likely regard that analysis as a consumer report. [2]
A footnote went further. It quoted a 2011 FTC staff report saying that information that does not identify a specific consumer does not constitute a consumer report. [2] The report then said: “The Commission does not believe that this statement is accurate.” [2] It said a report crafted for eligibility purposes with reference to a particular consumer, or a set of particular consumers such as applicants, would be a consumer report even with identifying information stripped. [2]
The report drew a contrast with a general report. An analytics firm’s “aggregate credit score” for every zip code, found by a company through a search engine and used to inform its policies, would likely not be a consumer report. [2] The report said the store would still have to disclose the nature of the report on a consumer’s written request, if a denial or higher charge resulted. [2]
On equal opportunity laws, the report described disparate treatment and disparate impact. [2] It said credit decisions based on zip codes may violate the Equal Credit Opportunity Act (ECOA) if they have a disparate impact on a protected class and are not justified by a legitimate business necessity. [2]
On the FTC Act, it said companies must not sell big data analytics products to customers they know or have reason to know will use them for fraudulent purposes. [2] It said Section 5 may also apply where products are sold to customers that use them for discriminatory purposes. [2]
The report closed its legal section with a list it called “Questions for Legal Compliance,” and its research section with four questions for companies. [2] Those asked how representative a data set is, whether a model accounts for biases, how accurate its predictions are, and whether reliance on big data raises ethical or fairness concerns. [2]
The separate statement, printed as an appendix, supported the report. [2] It asked that concerns about inaccurate analysis be weighed against market and competitive forces, and that hypothetical harms be tested with economic reasoning and empirical evidence. [2]
The January 2016 report is easy to file as a policy paper about fairness. [1] For a bank, a health system or an agency that buys analytics, its practical core was narrower. It told buyers how the Commission would decide whether a model’s output is regulated, and what follows when it is. [2]
Our argument is that the report made coverage turn on how a score is used, not on whether a name is attached to it. The Commission said a score crafted for a decision about a particular applicant is a consumer report even when identifiers are stripped. [2] That moves the compliance question from the data team, which controls identifiers, to the business owner, who controls use.
We take the report’s use test in four parts, each a separate question a buyer of analytics has to answer:
- Coverage: is the score a consumer report, which makes the seller a consumer reporting agency and the buyer a user?
- Notice: what the person affected must be told when the score counts against them.
- Fair treatment: whether the inputs or the results treat protected groups differently, or fall on them harder.
- The seller’s duty: what the company that sells the analytics owes, including to people it never meets.
The parts do not overlap. The first decides which statute applies, the second covers communication with the person, the third covers outcomes across groups, and the fourth covers the supply side. After them we look at how the questions reach different organizations, take the objections in order, and end with checks a buyer could run.
Coverage: the score, not the name
What the definition asks
The report started from the FCRA’s definition. A consumer report, it said, is a communication from a consumer reporting agency bearing on a consumer’s personal characteristics or mode of living. [2] It must be used or expected to be used as a factor in establishing the consumer’s eligibility. [2] The report noted that a federal appeals court had called the first part of that test not a very demanding one. [2]
That leaves the second part to do the work. Under the report’s reading, the question is whether the communication is prepared or provided for an eligibility determination about a particular consumer. [2] Eligibility, in the report’s list, covers credit, employment, insurance, housing and other similar decisions. [2] The report’s compliance questions added government benefits. [2]
The stripped application
The hypothetical in the report is worth reading as a system design. An applicant’s data leaves the lender without her name. A vendor scores a cohort that looks like her. The score comes back and decides her application. [2]
On a privacy review, that design can look clean. No direct identifier crosses the boundary, and the vendor never learns who the applicant is. The report said that does not decide the FCRA question. [2] Because the company uses information about the consumer to generate an analysis and then uses that analysis to decide about her, the Commission would likely treat it as a consumer report. [2]
In our reading, the consequence is that de-identification and FCRA coverage are separate controls. De-identification limits what the vendor knows. It does not change what the lender does with the answer. A buyer that relies on stripping identifiers to keep a vendor outside the FCRA is relying on the reading the Commission called inaccurate. [2]
The aggregate report
The report did not say every use of outside analytics is covered. Its counterexample was an aggregate score per zip code, published generally, that a company finds and uses to set policy. [2] The Commission said it would likely not consider that a consumer report, because it does not relate to a particular consumer. [2]
The line, then, runs between analysis crafted for a decision about a person and analysis that informs general policy. A lender that commissions a vendor to score its own applicant pool is on one side. A lender that reads a public market study is on the other. Organizations should check with counsel where a product falls between those poles.
Even the aggregate case had a tail. The report said FCRA Section 615(b) lets a consumer denied credit, or charged more, based on information from a person other than a consumer reporting agency request in writing the nature of that information. [2] The store in the report’s example would have to disclose the nature of the aggregate report on request. [2]
The own-data exception and its limit
The report restated an exception. The FCRA does not apply when companies use data derived from their own relationship with their customers to make decisions about them. [2] In our reading, a bank scoring its own deposit history for its own overdraft decisions is in that territory.
The report then limited it. If an unaffiliated firm regularly evaluates companies’ own data and provides the evaluations back for eligibility determinations, the firm would likely be acting as a consumer reporting agency. [2] Each company would likely be a user of consumer reports. [2]
That limit matters for modern architecture. When a lender sends its own transaction data to a hosted scoring service and receives scores back, the data may be first-party. The report’s analysis turns on who performs the evaluation and for what purpose. [2]
Who decides the use
If coverage turns on use, the person who decides the use owns the question. In a large organization that is a product or credit policy owner, not the data engineer who strips the identifiers. The two may never meet.
A model inventory helps close that gap only if it records use. An entry that lists a vendor score as “fraud screening” tells a reviewer little if the same score also gates account opening. The report’s test asks what the output is used or expected to be used for. [2] The inventory has to answer that question per use, not per model.
Use also drifts. A score bought for marketing can migrate into an approval rule because it is already in the data warehouse. In our reading of the report’s use test, that migration can change the legal status of the vendor relationship without any change in the data. [2] A change-control step for new uses of existing scores is the control that catches it.
Disclaimers
The report described two data broker cases. In one, it said, the FTC alleged that Spokeo marketed consumer profiles for use by human resources departments in hiring, and a consent decree required $800,000 in civil penalties. [2] In the other, it said, the Commission alleged that Instant Checkmate advertised employment and tenant screening uses, and obtained $550,000 in civil penalties. [2]
Both companies, the report said, had disclaimers saying they were not consumer reporting agencies and that users could not use their data for eligibility purposes. [2] It said the disclaimers were not effective in insulating the companies from FTC enforcement. [2]
For a buyer, the point runs both ways. A vendor’s terms that forbid eligibility use do not settle whether the vendor is a consumer reporting agency. A buyer’s own use can also create obligations the vendor’s paperwork denies. The contract is evidence. It is not the answer.
Notice: what the consumer is told
Adverse action and risk-based pricing
The report reminded users of consumer reports of two notices. A company that uses consumer report information to deny credit, insurance, employment, housing or certain other covered benefits must provide an adverse action notice. [2] A company that charges more for credit or insurance based on consumer report information must provide a risk-based pricing notice. [2]
The report said both notices exist to let consumers check their consumer reports and correct inaccuracies. [2] It described two FTC cases on risk-based pricing. In a 2013 action against Time Warner Cable over deposits it required based on consumer reports, it said, the consent order required $1.9 million in civil penalties. [2] In a 2015 action against Sprint over customers placed in a program for those with lower credit scores and charged an extra monthly fee, it said, the consent order required a $2.95 million penalty. [2]
Those cases involved a cable company and a wireless carrier. The report used them to show that pricing decisions carry notice duties as well as denials. [2] A deposit or a fee set by a model is a pricing decision in the report’s terms.
The Section 615(b) request
The second notice path concerns information that does not come from a consumer reporting agency. Under Section 615(b), as the report described it, the consumer may request in writing the nature of the information that led to a denial or higher charge. [2]
That duty is easy to miss in system design. A decision engine that pulls an outside data set into an eligibility rule needs a way to answer a later written request about that data. Someone has to know which outside sources fed which decision. Without a record of which inputs reached a given decision, the request cannot be answered accurately.
Reasons from a model
For creditors, ECOA and Regulation B require a statement of specific reasons for adverse action. The report flagged that requirement in a footnote. [2] It said identifying the specific reasons may be challenging when those reasons implicate big data analytics, and pointed to 12 CFR 1002.9. [2]
The report’s list of legal compliance questions repeated the point. It asked whether a creditor using big data analytics in a credit transaction was complying with the requirement to provide statements of specific reasons for adverse action. [2]
In our reading, the footnote put an engineering constraint on model choice. A model whose drivers cannot be stated for an individual decision makes that requirement hard to meet. The challenge is not the notice template. It is whether the institution can state why the model reached its answer.
Pre-adverse action in employment
Employment uses add a step before the decision. The report’s compliance questions said entities that use consumer report information for employment purposes must give a pre-adverse action notice, and thereafter an adverse action notice. [2] They must also certify a permissible purpose and certify they will not use the information in violation of equal opportunity laws. [2]
For an automated hiring screen, the pre-adverse action step means the screen cannot be the final word in the same instant. There has to be a pause between the result and the rejection, with a notice in between. A system that rejects automatically at the screening stage has no place for that pause.
Explaining the model to the analyst first
The report’s research section made a related point about people inside the company. It said the complexity of the data and statistical models can make it difficult for analysts to fully understand and explain the underlying model or its results. [2] If the analysts cannot explain the result, the notice writer cannot either.
That ordering matters for governance. A model validation team that can describe the principal drivers of a model’s decisions is the precondition for a notice. Explanation for regulators, auditors and consumers starts with explanation for the institution’s own staff.
Notices as a design input
Notice duties can be treated as a downstream compliance step. The report’s structure suggests treating them as a requirement on the model and the data pipeline. A model owner should be able to produce, for any declined applicant, the principal reasons and the outside sources used.
That has costs. It favors simpler models or explanation methods that can be checked. It requires keeping inputs and model versions for each decision. It requires a process that ties a consumer’s request to a specific decision record. Each is a build item, with an owner and a budget.
Fair treatment: treatment and impact
Disparate treatment
The report described two theories. Disparate treatment occurs when a creditor or employer treats an applicant differently based on a protected characteristic. [2] The report said a lender cannot refuse to lend to single persons, or offer them less favorable terms than married persons, even if big data analytics show single persons are less likely to repay. [2]
That example matters because analytics can make a prohibited distinction look like a finding. The report’s answer was that the finding does not make the distinction lawful. [2] The control for disparate treatment is a review of model inputs and decision rules against prohibited bases. It also requires a check that proxies have not been added to restore a removed variable.
Disparate impact and the zip code
The report said disparate impact occurs when facially neutral policies have a disproportionate adverse effect on a protected class. [2] The exception it described was a legitimate business need that cannot reasonably be achieved by means with less disparate impact. [2] It said credit decisions based on zip codes may violate ECOA if they have a disparate impact on a protected class and are not justified by a legitimate business necessity. [2]
It added that even where a business necessity is shown, a less discriminatory alternative may still mean a violation. [2] A footnote said zip codes can also raise concerns of redlining, and cited a CFPB and Justice Department action against Hudson City Savings Bank. [2]
For model owners, the disparate impact analysis described in the report implied three steps. Measure outcomes by group. Document the business need for each variable that drives a disparity. Look for an alternative that serves the need with less disparity. The report said big data can also facilitate the identification of disparate impact. [2]
Proxies
The report warned that results may affect sets of individuals differently because models may use variables that turn out to operate no differently than proxies for protected classes. [2] It said researchers may also lack the information needed to tell whether their results have such effects. [2]
The report’s hiring example made the point concrete. It said one company found that employees who live closer to their jobs stay longer. [2] It said another company excluded that factor from its hiring algorithm because of concerns about racial discrimination, since neighborhoods can differ in racial composition. [2] The variable predicted retention. The second company judged the proxy risk outweighed the predictive value.
Advertising and discouragement
The report spent time on targeted marketing. It said an advertisement to a particular community for a credit offer open to all is, in general, unlikely by itself to violate ECOA, absent disparate treatment or an unjustified disparate impact in subsequent lending. [2] It still told companies to proceed with caution. [2]
It pointed to Regulation B’s ban on statements that would discourage, on a prohibited basis, a reasonable person from applying. [2] It noted that for prescreened solicitations Regulation B requires creditors to keep records of the solicitations and the criteria used to select recipients. [2] It also said the Justice Department had in some cases cited a creditor’s advertising choices as evidence of discrimination. [2]
For a marketing team, the recordkeeping point is concrete. If an audience model selects who receives a credit offer, the selection criteria are records. A model that cannot report its criteria makes that recordkeeping harder.
Accuracy as a fairness question
The report treated accuracy as part of fairness. It said big data is good at detecting correlations but does not explain which correlations are meaningful. [2] Its example was Google Flu Trends, which it said appeared accurate at first and generated highly inaccurate estimates over time. [2]
It gave a lending example. Rental payment history can expand access to credit, it said, but consumers have the right to withhold rent when a landlord fails to provide heat or basic sanitation. [2] A model that reads withheld rent as default would punish a tenant for exercising a legal right.
It then contrasted low and high stakes. A wrongly targeted baby formula ad costs a consumer little, the report said. [2] Its higher-stakes hypothetical was a fraud tool that flags people who do not use social media as risky, so that a wireless company demands extra documents. [2] The report said those people may never know why they were turned away, or be able to correct the information. [2]
The research questions
The report’s research section put four questions to companies, about representativeness, bias, accuracy and fairness. [2] It described Hurricane Sandy, which it said generated more than twenty million tweets, with the greatest number from Manhattan rather than harder-hit areas. [2] It described Boston’s Street Bump app, whose team recognized the data was likely not fully representative of road conditions. [2]
These examples come from city services and public data. They apply directly to agencies that use app or sensor data to allocate services. They also apply to health systems that use patient portal activity as a signal, since portal use is not evenly spread across patient populations.
The report suggested it may be worthwhile to have human oversight of data and algorithms where big data tools are used for important decisions, such as those implicating health, credit and employment. [2] It framed that as a research consideration, not a legal requirement. [2]
The seller’s duty: knowing the customer
Promises and omissions
The fourth part concerns the company that sells data or analytics. Under Section 5, the report said, a company that violates a material promise, whether about sharing data, offering choices or safeguarding information, will likely be engaged in a deceptive practice. [2] A failure to disclose material information may also violate Section 5. [2]
Its example was CompuCredit. The report said the FTC’s complaint alleged a credit card marketer failed to disclose that, based on a behavioral scoring model, cash advances or certain transactions would reduce credit lines. [2] The transactions listed included marriage counseling, bars and nightclubs, pawn shops and massage parlors. [2]
For a lender, that example ties model behavior to disclosure. If a model reacts to a category of spending, the customer-facing terms may need to say so. The model’s feature list becomes a disclosure question.
Security
The report said companies that maintain big data on consumers should take care to reasonably secure it. [2] It tied the level of care to the amount and sensitivity of the data, the size and complexity of the company, and the cost of available security measures. [2] It said a company holding Social Security numbers or medical information should have particularly strong security compared with one holding names only. [2]
Selling to the wrong buyer
The sharpest statement in this part was about customers. In one case, the report said, the FTC’s complaint alleged a company sold payday loan applicants’ information to non-lender third parties, and one used it to withdraw money from consumers’ accounts without authorization. [2] It described ChoicePoint, where it said the Commission alleged the company sold personal information of more than 163,000 consumers to identity thieves posing as legitimate subscribers. [2]
From those cases, the report said companies must not sell big data analytics products to customers they know or have reason to know will use them for fraudulent purposes. [2] Its compliance questions added that if a company has reason to believe its data will be used for discriminatory purposes, it should ask more specific questions about how the data will be used. [2]
That is a know-your-customer duty for data sellers. It runs to the people in the data set, who have no contract with either party.
The buyer’s side of the same duty
The report addressed sellers, but the duty has a mirror. A seller asked to judge how its customers use data will ask those customers questions. A buyer that cannot say what it will use a product for, or that gives a vague answer, makes the seller’s diligence fail. Buyers should expect the questions and have the answers in writing.
The answers also bind the buyer. A stated use that later expands, without telling the vendor, can put both parties outside what they agreed. Keeping the stated use current is part of the buyer’s own control set.
Who the FTC Act reaches
Sector precision matters here. A report footnote listed exceptions to the FTC’s Section 5 authority, including banks and savings and loan institutions, federal credit unions, common carriers and air carriers. [2] In our reading, a bank or federal credit union buying analytics falls within those exceptions. Its fair lending and consumer protection supervision comes from other agencies.
The report also said several agencies enforce ECOA and Regulation B, naming the FTC, the Justice Department and the CFPB among others. [2] In our reading, a non-bank analytics vendor can fall under the FTC’s FCRA authority even where its bank customer is outside the FTC’s Section 5 reach. The report said the FTC has authority to enforce compliance with the FCRA. [2]
How the questions reach different organizations
A bank or credit union
A bank or credit union that uses outside scores for underwriting, account opening or pricing faces all four questions. Coverage decides whether the vendor relationship carries FCRA user duties. Notice decides what the adverse action letter can say. Fair treatment governs the variables. The seller’s duty mostly sits with the vendor, but the bank’s diligence is how it learns whether the vendor meets it.
For a credit union, the own-data exception has a specific edge. In our reading, shared service arrangements that evaluate pooled member data and return scores resemble the report’s description of an unaffiliated firm evaluating companies’ own data. [2] The report said only a fact-specific analysis can settle whether a practice is subject to the FCRA, so whether that makes the service a consumer reporting agency is a question for counsel. [2]
A fintech lender or bank partner
A fintech that builds on non-traditional data is the report’s central example. The report described predictive products that may use a consumer’s zip code, social media usage or shopping history to judge creditworthiness. [2] It said the standards for FCRA applicability are the same. [2]
Where a fintech originates through a bank partner, the questions split across two parties. The bank is the creditor for notice duties. The fintech or its vendor may be the party that knows the model’s drivers. Contracts have to move the reasons, records and data lineage to whoever signs the notice.
A broker-dealer or wealth manager
Investment firms meet the report mainly through employment. The report’s compliance questions said entities using consumer report information for employment purposes must certify a permissible purpose, certify compliance with equal opportunity laws, and give pre-adverse action and adverse action notices. [2] A firm that buys an analytics product to screen candidates or advisers should run it through those duties.
A health system or ambulatory network
Health systems use outside data in hiring, in patient financing and in deciding whom to reach for programs. Hiring tools bring the FCRA employment duties. Patient payment plans that extend credit may bring ECOA questions. Outreach models bring the report’s representativeness question, because a model trained on one population may miss another.
A public agency
The report’s own scope was commercial use. [2] Its compliance list still named government benefits among eligibility decisions. [2] An agency that buys a third-party risk score for benefits screening should ask whether the vendor’s product is sold as a consumer report and what notice the vendor’s terms assume. Agencies also face their own program rules and, for federal systems, the Privacy Act; those are outside the report.
An analytics vendor
The vendor carries the heaviest version of the questions. If its products are used for eligibility decisions about particular consumers, the report said it may be a consumer reporting agency with accuracy, notice, access and correction duties. [2] The report’s list put those as reasonable procedures for maximum possible accuracy, notices to users, consumer access and correction of inaccuracies. [2]
What the contract should carry
Across these organizations, the contract with an analytics vendor is where the four questions become obligations. It should state the permitted uses and whether either party treats the product as a consumer report. It should give the buyer the factors a model uses, in enough detail to state principal reasons. It should require notice before the vendor changes model drivers or data sources. It should require the vendor to keep the records a consumer’s written request would need. These terms do not settle the legal question. They make the facts available to whoever has to answer it.
Objections, in order
1. The report is not a rule
The report said its discussion articulated considerations relevant to the Commission’s enforcement authority. [2] It was not a rule. The objection is accurate. The answer is that the report stated how the agency that enforces the FCRA against non-bank sellers would read the statute. A buyer that chooses a different reading should know it is doing so, and why.
2. Markets punish inaccurate models
The separate statement made this argument. It said businesses have strong incentives to seek accurate information about consumers, and that competitors with better analysis will fill gaps. [2] That is a real constraint on inaccuracy that costs a firm money.
It does not reach every harm. An accurate model can still use a prohibited basis. An accurate score can still trigger notice duties. A seller can be accurate and still sell to a fraudulent buyer. The four questions survive the market argument because three of them are not about accuracy at all.
3. Stripping identifiers protects the consumer
It protects the consumer against some risks, such as the vendor re-using her data. The report did not say otherwise. Its point was narrower: de-identification does not take an eligibility analysis outside the FCRA. [2] Both controls can be worth having. Neither substitutes for the other.
4. The duties fall on the vendor
Some do. The user duties, adverse action and risk-based pricing notices among them, fall on the company that makes the decision. [2] ECOA’s reasons requirement falls on the creditor. [2] A buyer cannot contract those away.
5. Human review fixes the model
The report suggested human oversight may be worthwhile for important decisions. [2] Oversight helps when the reviewer can see the drivers and has authority to overrule. It helps less when the reviewer sees only a score and a recommendation. A review step needs the same explanation the notice needs, or it becomes a signature on the model’s output.
6. The guidance moved after 2016
It did, as the next section shows. Some interpretive documents were issued and then withdrawn. A rule on de-identified data was proposed and then withdrawn. A 2026 rule changed Regulation B on disparate impact. The FCRA’s definition and the specific-reasons requirement were not removed by those steps. Controls built on statute and regulation text hold up better than controls built on one agency’s guidance.
Checks a buyer of analytics could run
Map decisions to scores
List every decision about an individual that uses an outside score, model or data set. For each, record whether it is an eligibility decision in the report’s sense: credit, employment, insurance, housing or a government benefit. [2]
Ask the coverage question per use
For each eligibility use, ask whether the output is crafted for a decision about a particular applicant. If it is, assume the FCRA question is live even if the data was de-identified, and get counsel’s view. [2]
Check the notices
Confirm that each declining or pricing decision can produce the required notice. For credit, confirm the institution can state specific principal reasons drawn from the factors the model actually used. Test this on real declined files, not templates.
Keep the decision record
Store, per decision, the model version, the inputs and the outside sources. Without that record, a written request about the nature of outside information cannot be answered. [2]
Measure outcomes by group
Run outcome analysis by protected group where law and data allow. Document the business reason for drivers of any disparity and the alternatives considered. [2]
Review marketing models
Treat audience selection for credit offers as part of the lending process. Keep the criteria used for prescreened solicitations, as the report noted Regulation B requires. [2]
Control new uses
Add a change-control step when an existing score is used for a new decision. A marketing score that moves into an approval rule can change the legal analysis without any change in the data. [2]
Diligence the seller
Ask a vendor whether it treats its product as a consumer report, what accuracy procedures it runs, and how it vets its own customers. The report expected sellers to know the purposes for which customers use their data. [2]
Check the representativeness of the data
Before using app, sensor or portal data to allocate services, ask who is missing from it. The report’s Hurricane Sandy and Street Bump examples are the template. [2]
With hindsight
The CFPB took up the report’s notice footnote in 2022. On May 26, 2022, it released Consumer Financial Protection Circular 2022-03 on adverse action notices for credit decisions based on complex algorithms. [3] It was published in the Federal Register on June 14, 2022. [3]
The circular answered its question with “Yes.” [3] It said ECOA and Regulation B do not permit creditors to use complex algorithms when doing so means they cannot provide the specific and accurate reasons for adverse actions. [3] A footnote said creditors relying on post-hoc explanation methods must still be able to validate the accuracy of those approximations. [3]
On December 13, 2024, the CFPB published a proposed rule to amend Regulation V, which implements the FCRA. [4] It said the proposal would implement the FCRA’s definitions of consumer report and consumer reporting agency, among other provisions. [4] Comments were due by March 3, 2025. [4]
The proposal set out three alternative versions of a provision on de-identified information. [4] Under the first, de-identification would not be relevant to whether a communication is a consumer report. [4] Under the second, it would not be relevant if the information is still linked or linkable to a consumer. [4]
The third listed conditions, any of which would make de-identification irrelevant. One was that the information is used to inform a business decision about a particular consumer. [4] In our reading, that condition restated the use test the FTC had applied in its 2016 report. [2] [4]
The proposal did not become a rule. On May 15, 2025, the CFPB withdrew it. [5] The withdrawal said the proposal was not aligned with the Bureau’s current interpretation of the FCRA, which it said it was in the process of revising. [5] It said that if the Bureau determines a rule is necessary, it will propose a new rule and seek comment. [5]
Before that, the CFPB had withdrawn a long list of guidance documents, applicable as of May 12, 2025. [6] The list included Circular 2022-03 and Circular 2023-03 on adverse action notices. [6] It also included Circular 2024-06 on background dossiers and algorithmic scores used in employment decisions. [6]
The notice said the withdrawal was not necessarily final, and that the Bureau would continue reviewing the documents. [6] It said the withdrawn guidance should not be enforced or otherwise relied upon by the Bureau while the review is ongoing. [6]
The regulation the circulars interpreted remained. As checked on November 16, 2026, the eCFR text of 12 CFR 1002.9(b)(2) still says a statement of reasons for adverse action must be specific and indicate the principal reasons. [8] It still says a statement that an applicant failed to achieve a qualifying score on the creditor’s scoring system is insufficient. [8]
The report’s disparate impact discussion fared differently. On April 22, 2026, the CFPB published a final rule amending Regulation B, effective July 21, 2026. [7] The Bureau said the rule provides that ECOA does not authorize disparate-impact liability, which it called the effects test. [7] The rule also further defined discouragement. [7]
As checked on November 16, 2026, the eCFR text of 12 CFR 1002.6(a) says the Act does not provide that the “effects test” applies for determining whether there is discrimination in violation of the Act. [8] In our reading, the zip-code passage in the 2016 report rested on that test under ECOA. [2]
In our view, what held up from 2016 is the coverage analysis and the notice duty. The FCRA question still turns on the statute’s definition, which none of these withdrawals amended. The specific-reasons requirement is still in Regulation B. [8]
What did not hold up, as a matter of CFPB rules, is the report’s account of ECOA disparate impact. The 2026 rule speaks to ECOA and Regulation B. [7] The report also named other laws, including the Fair Housing Act and Title VII. [2] Organizations should ask counsel how each applies to them today.
Spotted an error? Tell us and we will check it and correct the article where needed.