Employer-Side AI Screening Is On by Default at LinkedIn, Indeed, and SimplyHired

The American Counsel, September 30, 2026

LinkedIn and Indeed now evaluate job applicants with AI on the employer’s side of the hiring transaction, and the applicant has no setting that turns it off. LinkedIn’s Hiring Assistant evaluates and summarizes applicants by default under every recruiter contract that includes it.[1] Indeed attaches an AI summary and a match score to applications for jobs enrolled in its Smart Screening product,[2] and SimplyHired’s users are bound by Indeed’s terms.[3] On January 1, 2027, California’s automated decisionmaking regulations begin to govern employers that use tools like these to make hiring decisions,[4] and a federal court has held, at the pleading stage, that a vendor whose software screens applicants can be liable as the employer’s agent.[5]

A June 2026 study by the data-removal company Incogni identified the line that separates these tools from the ones the law can leave alone. Of ZipRecruiter and FlexJobs it reported that “[c]andidates have access to on-platform ‘AI’ tools to help them with their applications, but it seems that employers aren’t offered analogous tools.”[6] Of Indeed, LinkedIn, and SimplyHired it reported “tools that enable employers to rate, evaluate, or summarize potential employees’ candidatures using so-called AI-powered systems.”[7] On one side of that line is a tool the applicant chooses to use on her own application. On the other is a tool someone else chooses to use on her. This essay argues that the line should be a legal one and shows how far existing law already draws it. California’s regulations reach the second kind of tool at the moment of decision. Mobley v. Workday makes the vendor of such a tool answerable as the employer’s agent. The Federal Trade Commission’s staff has found that feeding user data into AI by default, with no way to object, was the norm among the large platforms it studied. None of these reaches the training, and the training is where the competitive stakes lie. The platform that receives the most applications trains the screener that employers must use, and the default is what supplies the training data.

The defaults are written into the platforms’ own documents. LinkedIn’s help center for recruiters states that applicant evaluation “is enabled by default at the contract level and applies to all users on the contract,” and that when it is on, “Hiring Assistant will evaluate applicants and provide summaries across all projects.”[8] Indeed’s product page describes the timing of its score: “As soon as a candidate hits ‘Submit,’ the AI parses the data and assigns a Smart Fit Score.”[9] When an employer adds jobs to Smart Screening, “[a]ll your open jobs will be selected automatically,” and the next prompt offers to “auto-enroll future jobs to add all new jobs by default.”[10] SimplyHired’s footer tells the seeker that “SimplyHired is part of the Indeed Site, and by using SimplyHired you agree to the Indeed Terms of Service,” and its privacy link leads to Indeed’s policy, which states that “Employers may use Indeed to generate an AI-powered summary evaluation and a match score based on Employer job criteria.”[11] Every switch in these descriptions belongs to someone other than the applicant. LinkedIn’s belongs to the recruiter’s contract administrator. Indeed’s belongs to the employer. The applicant at Indeed is given an email address to which she may write to “contest any decision generated by said automated processing,” beside a footnote stating that “there are no services on Indeed where Indeed as a data controller conducts automated processing that has a legal or similarly significant effect.”[12]

Most applicants believe a machine reads their application first, and a majority say they do not mind the machines learning from them. Incogni’s survey of 1,000 recent job seekers found that 64 percent feel they write their resumes for algorithms rather than for people, and that 51 percent were at least somewhat comfortable with their data being used to train AI models.[13] Its platform review found that LinkedIn is testing automated interviews it calls “screening calls” and that Indeed processes application data to generate summaries for employers.[14]

I. The Seeker Chooses One Kind of Tool and the Employer Chooses the Other

Candidate-side tools work for the applicant, and the platforms describe them that way. Indeed calls Career Scout “an interactive AI experience designed to help you with your job search and career exploration.”[15] ZipRecruiter introduces “’Phil,’ your AI-powered career advisor,” and explains that “Phil helps job seekers discover new opportunities and stand out to employers.”[16] FlexJobs markets ExpertApply as “a powerful AI job application tool that streamlines the application process” and offers it to paid members.[17] The seeker decides whether to use these tools, what to ask of them, and whether to submit what they produce. A resume that a candidate-side tool has rewritten is still her resume. She can discard it, revise it again, or take it to another site. Incogni found that about one in three respondents had used AI features on professional sites to optimize their profiles and that about a third had used general-purpose language or image models, so a seeker who wants a machine’s help has many to choose from.[18]

The employer-side tools run in the other direction. They take the application as input and return something about the applicant to someone else. LinkedIn describes the model behind Hiring Assistant’s candidate summaries as “[a]n in-house LLM which takes in hiring intent (the list of required and nice-to-have qualifications), candidate LinkedIn profile, candidate resume, screening questions, and recruiting notes as input and generates a response about the qualifications the candidate matches, partially matches, or does not match.”[19] Recruiters “can filter applicants by Relevance, with those identified as top applicants by Hiring Assistant appearing first.”[20] LinkedIn’s product for smaller employers, Hiring Pro, “reviews LinkedIn profiles and resumes against your hiring qualifications to presort top applicants.”[21] The automated interviews Incogni flagged are a Hiring Pro feature in which an “AI interviewer” conducts “an audio or video (if available) screening call,” after which “[h]irers receive a transcript of the candidates’ AI screening interviews, AI-generated summaries and ratings, along with the audio or video recording,” including “a 5-point rating based on your ideal responses.”[22]

Indeed’s terms for Smart Screening describe “an application match score and summary based on your job criteria, your screening and scoring instructions, and application information provided by the Job Seeker.” The employer’s required criteria are converted “into mandatory ‘deal breaker’ screener questions,” and “applications that do not meet those screener question criteria and your other required criteria will not receive an application match score.”[23] An employer that enrolls a job is “requesting Indeed to include an AI-powered summary with each candidate application delivered to you,” and “AI Recruiter is automatically enabled for jobs for which you activate Smart Screening.”[24]

The seeker’s options are confined to the parts of this process that ask her for more. LinkedIn’s AI interviews “are optional for applicants, and they will not be automatically disqualified if they decide not to participate.”[25] Indeed’s AI follow-up questions are “an optional step after you apply.”[26] The evaluation of the application itself has no such option. The summary attaches to “each candidate application,” the relevance ordering applies to everyone who applied, and neither platform offers the applicant a setting that removes her application from the evaluation. LinkedIn’s account of its own data use says as much: “We also process your data for features that help hirers find you such as Hiring Assistant.”[27] The one AI control LinkedIn gives its members, the Data for Generative AI Improvement setting, governs training rather than evaluation. LinkedIn says the setting “does not control the use of your data while you or others are using generative AI features on LinkedIn,” that it is set to “on” by default, and that opting out “does not affect training that has already taken place.”[28]

Incogni’s line sorts tools by who invokes them and for whose benefit. A tool the seeker invokes works for the seeker. A tool the employer invokes works on the seeker. The line also follows the price structure, which I described in an earlier essay on this market: the seeker uses the site free and the employer pays.[29] FlexJobs charges seekers, and seekers get its AI tools. Indeed and ZipRecruiter are free to seekers and charge employers, and employers get the tools that evaluate seekers. LinkedIn sells Hiring Assistant to employers as an add-on to its Recruiter subscription. The AI goes to the side that pays for it.

The line needs one qualification, and the qualification makes it more useful. Incogni wrote that “it seems” ZipRecruiter’s employers are not offered analogous tools, and the hedge was warranted. ZipRecruiter’s annual report describes an “AI-powered marketplace” in which, “[i]mmediately after a job is posted, ZipRecruiter’s matching technology presents the employer with a list of the best potential candidates in the market,” and its revised privacy policy says its AI systems “recommend qualified candidates to employers.”[30] Its terms give employers “functionality to help them identify or filter candidates that have applied.”[31] What I did not find in ZipRecruiter’s documents is a generated summary or score attached to each application. The line therefore sorts tools, and a single platform can have tools on both sides of it.

LinkedIn shows how porous the boundary can be. Its practice interviews are a candidate-side tool in the fullest sense. The seeker takes one for her own preparation, and “LinkedIn is the data controller for practice interviews.” The data they generate “is used by LinkedIn to provide the service and help us improve the AI interview product and experience,” and the AI interview product, as the same page describes it, is the one hirers use to rate applicants.[32] A candidate-side tool whose data improves the employer-side model has crossed the line in the direction that matters.

II. California’s Automated Decisionmaking Rules Reach the Screen and Stop at the Training

California’s regulations on automated decisionmaking technology reach employer-side screening tools, leave candidate-side tools alone, and say nothing about the training that produces the screener. The California Privacy Protection Agency adopted them in July 2025, and the Office of Administrative Law approved them that September. A business already using the technology for a significant decision “must be in compliance with the requirements of this Article no later than January 1, 2027.”[33] The requirements apply to “[a] business that uses ADMT to make a significant decision concerning a consumer.”[34] A significant decision includes one that results in the provision or denial of “employment or independent contracting opportunities or compensation,” and the first employment decision the regulation lists is “[h]iring.”[35] Job applicants have been consumers under the statute since the exemption for applicant and employee data lapsed on January 1, 2023.[36]

The regulations draw Incogni’s line without naming it. The duties attach to a business that uses the technology to make a decision “concerning a consumer.” When a seeker uses Career Scout to tailor her resume, no business is making a decision about her with the tool’s output. When an employer uses a match score to decide whom to interview, a business is. A candidate-side tool falls outside the regulation by construction. An employer-side screening tool falls inside it if it meets the definition of automated decisionmaking technology.

That definition is where the contest will be. The regulation defines the technology as “any technology that processes personal information and uses computation to replace human decisionmaking or substantially replace human decisionmaking,” and it defines substantial replacement as using “the technology’s output to make a decision without human involvement.”[37] Human involvement has three elements. The reviewer must “[k]now how to interpret and use the technology’s output to make the decision,” must “[r]eview and analyze the output of the technology, and any other information that is relevant to make or change the decision,” and must “[h]ave the authority to make or change the decision based on their analysis.”[38]

The platforms have written their disclosures to that definition. Indeed’s privacy policy states that “[u]nless otherwise notified, Indeed does not use AI constituting an AEDT or an ADMT as those tools are defined in applicable laws,” and adds that “all final hiring outcomes are determined by humans.”[39] Its employer terms describe its candidate summaries as “not employment recommendations” and direct that “Employers must review the full application and not base employment decisions solely on summaries.”[40] LinkedIn’s compliance statement says that Hiring Assistant “will not screen out applicants or make hiring decisions,” and that “[r]ecruiters will still need to take action and make final decisions on applicants whether they advance or not.”[41]

The regulation’s test asks whether the output was used to make “a decision” without a human who reviewed the output “and any other information that is relevant,” and it says nothing about finality. In my reading, the applicant ranked two hundredth for a position filled from the top twenty is the case the test was written for. A human read down a list the machine had ordered and stopped well before reaching her. The decision not to interview her was made on the output, because no one reviewed anything else about her. Indeed’s instruction that employers “review the full application” acknowledges that the regulatory line runs through the employer’s conduct, and the platform that supplies the ranked list and hosts the interface in which it is read is better placed than anyone to know whether its customers read past the top of it.

New York City supplies a comparison. Its Local Law 144 defines an automated employment decision tool to include a computational process “that issues simplified output, including a score, classification, or recommendation, that is used to substantially assist or replace discretionary decision making for making employment decisions,” and it bars an employer from using one to screen a candidate without a bias audit conducted within the prior year.[42] Indeed’s help center lists the jobs that cannot be enrolled in Smart Screening, the product that assigns the score. One category is a job whose “location is New York City, due to legal requirements.”[43] Indeed’s AI disclosures state that “[w]e do not consider any of our offerings that are available in New York City to be ‘Automated Employment Decision Tools.’”[44]Both statements can be true at once. The scoring product is not offered in the city whose law expressly counts a score as an automated employment decision tool.

If a platform’s screener is automated decisionmaking technology, the California regulation gives the applicant three things. The first is a pre-use notice, presented “at or before the point when the business collects the consumer’s personal information that the business plans to process using ADMT” and “in the manner in which the business primarily interacts with the consumer.”[45] The notice must explain “the specific purpose” of the technology in more than “generic terms,” “the categories of personal information that affect the output generated by the ADMT,” and “[t]he type of output generated by the ADMT, and how that output is used to make a significant decision.”[46] Indeed’s terms promise that “[e]ach applicable Job Ad will include a notice to Job Seekers that your job uses Indeed to provide AI summaries.”[47] A sentence saying that AI summaries are provided begins to satisfy the first element and does not address the others.

The second is a right to opt out, which the hiring exception will usually remove. An employer need not offer the opt-out for hiring decisions if it uses the technology “solely for the business’s assessment of the consumer’s ability to perform at work” and the technology “works for the business’s purpose and does not unlawfully discriminate based upon protected characteristics.”[48] The third is a right of access. The employer must provide “[i]nformation about the logic of the ADMT,” which “may include the parameters that generated the output as well as the specific output with respect to the consumer,” and must explain “how the business used the output of the ADMT to make a significant decision with respect to the consumer.”[49] LinkedIn’s present answer to an applicant who wants her AI interview rating is that “[c]andidates don’t automatically receive interview ratings, transcripts, or interview summaries” and “can request access by contacting the hirer directly.”[50] From January 2027, for a California applicant to a covered employer whose rating was used without human involvement, the request becomes a right.

Each of those duties belongs to the employer, because the regulation follows the decision and the decision is the employer’s. The platform, which designed the model, trained it, hosts the interface, and holds the data, appears in two places, both concerned with risk assessment. A business that “makes ADMT available to another business” for a significant decision “must provide to the recipient-business all facts available to the business that are necessary for the recipient-business to conduct its own risk assessment,” a duty that applies “only to ADMT trained using personal information.”[51] And a business must itself assess the risk before “[p]rocessing the personal information of consumers, which the business intends to use to train an ADMT for a significant decision concerning a consumer,” where “intends to use” includes a business that “permits others to use” the technology or “is advertising or marketing the use of” it.[52] A platform that trains a screener on applicant data and sells it to employers fits that description in its own marketing. For processing already under way, the assessment is due by December 31, 2027, and a summary goes to the Agency by April 1, 2028.[53]

What the regulation does not give the applicant is any say over the training. The pre-use notice, the opt-out, and the access right each attach to the use of the technology to make a significant decision. None of them mentions training. The applicant is protected, to the extent she is, at the moment her application is scored, and she has no right at the moment her application becomes an input to the next version of the scorer. The platforms fill that silence with their own terms. Indeed’s privacy policy says that “we use personal data to develop, train, build and use statistical models, including artificial intelligence,” and offers no setting to decline.[54] LinkedIn’s setting reaches only the training of models used for content creation. It “does not apply to our training of AI models used for other purposes,” and members who object to other training are directed to a separate objection form.[55]

The closest the regulations come to a rule about pooling is the limit on service providers. A service provider may use a client’s personal information “to build or improve the quality of the services it is providing to the business . . . provided that the service provider or contractor does not use the personal information to perform services on behalf of another person.”[56] The Agency’s own illustration shows how much that permits. An email marketing vendor “could analyze those customers’ interactions with the marketing emails to improve its services and offer those improved services to everyone,” so long as it does not reuse the client’s list for another client.[57] Applied to a screener, the rule lets a platform learn from every employer’s decisions and sell what it learned to every other employer. Indeed does not rely even on that allowance. Its policy says that while it processes application materials for the employer, “Indeed may also independently act as a controller of that same information so we can use the insights to improve our services for you.”[58]LinkedIn tells its enterprise customers that its contracts “[i]n many cases” include “rights to use customer-provided data in connection with providing and improving the services.”[59] The applications come from seekers and the decisions come from employers, and the rule that governs the combination treats the resulting model as the platform’s own. That is where the privacy question becomes a competition question, and Part V returns to it.

III. A Job Board Screens Like Workday and Procures Like an Employment Agency

Mobley v. Workday holds, at the pleading stage, that a vendor whose software screens applicants for employers can be liable as their agent, and it rejects the claim that Workday was also an employment agency because Workday did not find candidates. The job platforms screen applicants and find them. On Mobley’s reasoning, that exposes them on both theories.

Derek Mobley alleged that he applied to more than a hundred jobs at companies that used Workday’s screening tools and was rejected every time, on one occasion “at 1:50 a.m., less than one hour after he had submitted his application.”[60] The division of labor in his applications is worth noticing. He would “see a job posting on a third-party website (e.g., LinkedIn),” and the link would take him to the Workday platform on the employer’s site. One company showed him the job. Another screened him for it.

In July 2024 Judge Rita Lin of the Northern District of California held that Mobley had plausibly alleged Workday’s liability under Title VII, the Age Discrimination in Employment Act, and the Americans with Disabilities Act as the employers’ agent. “[A] third-party agent may be liable as an employer where the agent has been delegated functions traditionally exercised by an employer,” she wrote.[61] The complaint plausibly alleged such a delegation because it described software that “is not simply implementing in a rote way the criteria that employers set forth, but is instead participating in the decision-making process by recommending some candidates to move forward and rejecting others.”[62] The court declined to treat software differently from a person. Workday’s role “is no less significant because it allegedly happens through artificial intelligence rather than a live human being who is sitting in an office going through resumes manually to decide which to reject,” the court wrote, and “[d]rawing an artificial distinction between software decisionmakers and human decisionmakers would potentially gut anti-discrimination laws in the modern era.”[63] The court also marked the limit of the theory: “many software vendors do not qualify as agents because they have not been delegated responsibility over traditional employment functions.”[64]

The same order rejected a second theory. Title VII and the ADA define an employment agency as “any person regularly undertaking with or without compensation to procure employees for an employer or to procure for employees opportunities to work for an employer.”[65] The court held that Workday was not one. “There are no allegations in the FAC that Workday brings job listings to the attention of those looking for employment,” and “[t]he thrust of the FAC is that Workday allegedly screens applicants using allegedly discriminatory algorithmic tools. This is not sufficient to allege that Workday ‘procures’ employees, i.e., finds candidates for employers.”[66]

In May 2025 the court preliminarily certified a collective of applicants aged forty and over, and in doing so it answered the argument the job platforms now make in their disclosures. The court described Workday’s position as being that “because its AI recommendation system supposedly cannot auto-reject applicants without some degree of participation by the employer, Workday does not ‘recommend.’” The theory of the case, the court responded, was that Workday’s AI “participat[es] in the decision-making process,” which is “a concept that is broader than auto-rejections,” and “[a]n individual who received a ‘low’ CSM score, for example, could qualify as having been denied an employment recommendation.”[67] The collective is defined by scoring. It reaches applicants whose applications were “scored, sorted, ranked, or screened by Workday’s AI,” where “the result of the AI scoring, sorting, ranking, or screening was not a recommendation to hire,” and “that result was communicated to the prospective employer, or the result was an automatic rejection by Workday.”[68] In July 2025 the court held that the collective includes applicants processed by the HiredScore features Workday later added to its products.[69] Workday itself told the court that “1.1 billion applications were rejected using Workday” in the period at issue.[70] The plaintiffs moved for Rule 23 class certification in September 2026, and the motion is pending.[71]

Set the platforms’ disclosures beside these rulings. LinkedIn’s Hiring Assistant “will not screen out applicants or make hiring decisions.” Indeed “does not make employment decisions on behalf of the Employer.” Those statements answer the question Workday tried to make decisive, which was whether the tool itself rejects anyone. The collective order treats a score or ranking communicated to the employer as the relevant act. LinkedIn’s relevance ordering, LinkedIn’s five-point interview rating, and Indeed’s match score are each a ranking or score communicated to the employer. The order came at the preliminary stage, where the question was whether the applicants were similarly situated, and Workday will be able to move to decertify. The court’s account of what counts as participation in a hiring decision does not depend on that posture, however, and nothing in it would spare a score because it was generated on a job board rather than inside an applicant tracking system.

The job platforms also do the thing Workday was held not to do. They bring job listings to people looking for work, which is the service a job board exists to provide, and they increasingly find candidates for employers. Indeed’s Talent Scout “scans millions of searchable profiles, surfaces the strongest matches, and helps draft personalized outreach.”[72] The chief executive of Indeed’s parent, Recruit Holdings, told investors in May 2026 that 70 percent of applications on Indeed now come from recommendations, including messages sent “based on data from employer AI tools such as AI sourcing and AI screening” telling seekers that “[t]his company seems to want you to apply for this position.”[73] LinkedIn’s compliance statement describes its hiring products as focused on “widening the hiring funnel through candidate sourcing and discovery,” and says Hiring Assistant “helps recruiters build a pipeline of individuals in an effort to encourage those individuals to apply for roles.”[74] Indeed’s terms add that its affiliates “Indeed Recruiting Services, Inc. and Indeed Ireland Operations Limited may provide services as an employment agency to their clients pursuant to separate written agreements.”[75] Indeed’s February 2026 privacy policy attributed its AI screening to that affiliate: “Employers may use Indeed Recruiting Services to generate an AI-powered summary evaluation and a generalized match rating of applications received.”[76] The version Indeed published in May 2026 attributes the same function to “Indeed.”[77]

The strongest authority against treating a job board as an employment agency is also the oldest. In 1970 a federal court in San Francisco held that a newspaper publishing help-wanted advertising was not an employment agency. Newspapers, the court wrote, “are not in any other or ordinary sense engaged in the business of procuring employees or employment opportunities any more than they are engaged in the used car business or in the real estate business when they accept and print advertising copy designed to bring buyers and sellers together in those fields,” and the statutory word “regularly” limited the definition to “those engaged to a significant degree in that kind of activity as their profession or business.”[78] Judge Lin relied on the same opinion in Mobley.[79] Brush describes a publisher that accepts and prints what employers submit. A platform that scans profiles, surfaces matches, drafts outreach, and tells a seeker that a company “seems to want you to apply” is engaged, to a significant degree and as its business, in the procurement the newspaper did not undertake. And the agent theory does not require procurement at all. The platforms screen, and on Mobley’s reasoning that is enough.

California has already written both theories into its employment regulations. Since October 1, 2025, the Civil Rights Council’s rules have defined an employer’s agent to include “any person acting on behalf of an employer, directly or indirectly, to exercise a function traditionally exercised by the employer,” including “applicant recruitment, applicant screening, hiring,” and “including when such activities and decisions are conducted in whole or in part through the use of an automated decision system.”[80] The same rules define an employment agency to include “persons undertaking these services through the use of an automated-decision system.”[81] An automated-decision system is “[a] computational process that makes a decision or facilitates human decision making regarding an employment benefit,” and the listed examples include systems that “[s]creen, evaluate, categorize, and/or recommend applicants,” that screen “resumes for particular terms or patterns,” and that analyze “facial expression, word choice, and/or voice in online interviews.”[82] The words “facilitates human decision making” dispose of the argument that a summary written for a human reader falls outside the definition. Evidence of “anti-bias testing or similar proactive efforts to avoid unlawful discrimination,” or “the lack of evidence,” is relevant to liability and to defenses, and automated-decision system data must be kept for four years.[83]

In June 2026 the Mobley court added a point about geography. Under the California Supreme Court’s decision in Raines, an agent is directly liable for its “own engagement in FEHA-regulated activities on the employer’s behalf,” and so, the court held, “Workday may fairly be held to account under the FEHA for its own California-based conduct” in designing and operating tools that evaluate applications submitted to employers nationwide.[84] At least at the pleading stage, a screening vendor that builds and runs its tools in California can be held to account under California law for how those tools treat applicants elsewhere.

IV. The FTC Found the Default and Named the Moat

The Federal Trade Commission’s staff has already found that feeding user data into AI by default, with no way to object, was the norm among the large platforms it studied, and it tied that default to market power. The finding came in a September 2024 report on nine large social media and video streaming services, studied for 2019 and 2020. No job platform was among them. The report’s finding on AI nonetheless describes the arrangement at LinkedIn and Indeed without translation. “User and non-user information was, by default, ingested into and used by Algorithms, Data Analytics, or AI.” The companies “did not appear to give users the choice to opt-in, nor did Companies appear to seek meaningful consent,” and “with one exception, the Companies did not appear to offer users or non-users the opportunity to opt in or to opt out of the use of their data by Algorithms, Data Analytics, or AI. In other words, the Companies set a default policy of using consumer data for these purposes, and consumers had no real way to object.”[85] The staff recommended that companies “address the lack of access, choice, control, transparency, explainability, and interpretability relating to their use of automated systems.”[86]

Two features of the finding matter here. The first is its object. The finding concerned the default itself: the companies set AI use as the starting condition and gave users no way to object. That is the condition of the applicant who would prefer that a person read her application. On LinkedIn and Indeed, no setting she controls produces that result. The second is that the report connected the default to market structure in its executive summary: “In digital markets, acquiring and maintaining access to significant user data can be a path to achieving market dominance and building competitive moats that lock out rivals and create barriers to market entry.”[87] The body of the report adds that “[d]ata abuse can raise entry barriers and fuel market dominance, and market dominance can, in turn, further enable data abuses and practices that harm consumers in an unvirtuous cycle.”[88]

The Commission is unlikely to carry the finding into enforcement against job platforms, and its own recent conduct suggests why. In December 2025 it reopened and set aside a 2024 consent order against an AI writing service, finding that “the order unduly burdens artificial intelligence (AI) innovation in violation of the Trump Administration’s Artificial Intelligence Executive Order and America’s AI Action Plan.”[89] Its staff’s 2024 warning that it “may be unfair or deceptive for a company to adopt more permissive data practices,” including “using that data for AI training,” and “to only inform consumers of this change through a surreptitious, retroactive amendment to its terms of service or privacy policy” is still posted, but it is a staff statement, and the Commission’s most recent action on AI ran the other way.[90]That is a forecast, and it points to California. The federal report supplies the diagnosis. The state regulations supply the rules, and, as Part II showed, the rules reach the decision and stop short of the training.

V. The Platform With the Most Applications Trains the Screener Employers Must Use

The platform that receives the most applications can train the best screener, and Recruit Holdings, which owns Indeed, has told investors as much. Its chief executive, Hisayuki Idekoba, said in May 2026 that “[i]n HR matching, no matter how much data on the user side you collect, it actually does not mean very much. What matters is whether the company likes that person or not.”[91] For that reason, he said, “over the past two to three years, our AI development investments have focused on developing AI tools for the employer side,” because “what is extremely important for improving matching is having solid data on the company side: data showing that ‘this company hires these kinds of people for this position,’ and being able to collect hundreds of millions of such data points around the world.” He added: “That is why we have been investing in encouraging companies to use more AI tools.”[92]

On Recruit’s own account, then, the employer-side tool is the instrument that collects the data that matters. An employer that advances or passes over a scored application tells the platform what “this company hires.” That data is proprietary in a way that job listings are not. “Of course, the job data that Indeed possesses can also be created by crawling other sites and gathering data from them,” Idekoba said. “However, it is much more difficult to determine which companies truly have strong hiring intent.”[93] The data now routes the applications themselves. Only 30 percent of applications on Indeed come from a seeker’s search. The rest come from recommendations, including messages sent “based on data from employer AI tools such as AI sourcing and AI screening.” “In fact, this accounts for 70%.”[94]Recruit’s annual results add that “by expanding integrations with ATSs, the Company aims to improve matching by bringing outcome data from external ATSs into the Indeed platform.”[95] In the same transcript, the company describes “the virtuous cycle on Indeed,” which “sustains through the utilization of AI tools.”[96]

LinkedIn describes its advantage in terms of the member side. Hiring Assistant is “the only AI hiring assistant powered by the world’s largest professional community,” and “[d]ata collected for the AI tools in LinkedIn Recruiter come from LinkedIn’s economic graph.”[97] The product also learns from recruiters’ decisions. It “learns from your actions and feedback,” including when a recruiter does such things as “[r]eject an applicant and provide written feedback.” LinkedIn states that this feedback “is applied only to the specific role and user” and “does not train a general-purpose AI model,” and it says separately that recruiters’ search inputs are used to improve search ranking and that “[w]e may also use this data in the future to train in-house models.”[98]

The economics of this arrangement have a name. Jens Prüfer and Christoph Schottmüller modeled markets in which a firm’s data about past users lowers its cost of improving quality for future ones, which “gives rise to data-driven indirect network effects,” and showed that “[s]uch markets tip under very mild conditions, moving towards monopoly.”[99] The Commission’s staff made the same point about generative AI in 2023: “generative AI products can get better the more people use them,” and “[a] first mover could secure a significant advantage over its competitors.”[100] The 2023 Merger Guidelines, which remain in effect, observe that “[n]etwork effects can create a tendency toward concentration in platform industries” and that “acquiring data that helps facilitate matching, sorting, or prediction services may enable the platform to weaken rival platforms by denying them that data.”[101] A hiring screener is a matching, sorting, and prediction service.

The strongest objection comes from economists who have examined data advantages and found most of them overstated. Andrei Hagiu and Julian Wright opened a 2020 article with the judgment that “[i]n most instances people grossly overestimate the advantage that data confers.”[102] Their formal work shows how the advantage depends on “the shape of firms’ learning functions, asymmetries between their learning functions, the extent of data accumulation, and customer beliefs.”[103] The Recruit transcript reads like an application of that test. The learning that matters, on Idekoba’s account, comes from employer decisions, which a rival cannot crawl and which accumulate only where applications are submitted and evaluated. The asymmetry is structural, because the platform that receives the applications is the only one that sees what employers do with them. Indeed was used by 83 percent of the seekers Incogni surveyed.[104] An employer that wants those applicants posts where they are, and the screener it is offered there was trained on what earlier employers did with earlier applicants. That is the sense in which employers must use it. The distribution of applicants supplies the compulsion.

This is why Incogni’s line matters for competition as well as privacy. A candidate-side tool builds no moat. Its output leaves with the seeker, who can use any model she likes, and the resume it produces can be submitted anywhere. An employer-side tool builds one, because its output stays with the platform and returns to it as data about what the employer did next. The default is what fills the moat. A screener that ran only when the applicant chose would learn from fewer decisions, and the learning would be the applicants’ to grant. A screener that is on by default learns from every application that reaches an employer who has left it on, and the applicant’s only way to keep her application out of the next model is to keep her application to herself.

The 51 percent of seekers who told Incogni they were comfortable with their data training AI answered the question privacy law asks. That question does not reach the arrangement. In The Interference Fallacy I argued that privacy law asks about consent and antitrust asks about price, and that both miss standing power: power that is asymmetric, arbitrary, and present whether or not it is exercised in a given case. A screener trained on most of a labor market’s applications and embedded in the marketplace where most applications are made is power of that kind. It ranks each applicant for an employer she has never met, on criteria learned from employers she will never meet, and each ranking adds to the data that keeps it ahead.

VI. Four Rules Would Make the Line Enforceable

Incogni’s line can become a legal one through four changes, each within the reach of an existing institution. These are proposals, and I state them as such.

First, the California Privacy Protection Agency should say, in its pending work on employee data, that human involvement is measured applicant by applicant. The Agency closed a preliminary comment period on that subject in May 2026.[105] A rule stating that a business that uses a ranked or scored list to decide which applications a person will read has used automated decisionmaking technology as to every applicant no person read would apply the existing definition as written. It would also end the argument that a final human decision somewhere in the process takes the whole screen outside the regulation.

Second, notice should reach training. The regulation already treats training a screener on consumers’ data as a risk to be assessed, and it requires the platform to give employers the facts they need to assess it. The applicant whose data is the input should receive at least what the employer receives, together with a way to decline the training that does not require her to stop applying for work. The Commission staff’s recommendation that companies address the lack of “choice” and “control” over automated systems applies without modification.[106]

Third, the access right should be delivered by the party that holds the output. The applicant’s score, summary, and rating sit on the platform’s servers, and the regulation already obliges a service provider to help answer an access request, “including by providing the business with the consumer’s personal information it has in its possession.”[107] A platform that showed each applicant her own score in the account through which she applied would discharge the employer’s obligation at the lowest cost, and it would make the score contestable by the only person with a reason to contest it.

Fourth, competition review should treat applicant decision data the way the Merger Guidelines already treat data that facilitates “matching, sorting, or prediction.” Recruit has said it intends to bring outcome data from external applicant tracking systems into Indeed. An acquisition of a screening vendor or an applicant tracking system by the largest job platform should be examined for what it adds to that flow, and an integration that conditions access to a platform’s applicants on the return of employers’ decisions to the platform’s model deserves the scrutiny the Guidelines reserve for denying rivals data.

The line Incogni drew separates a tool the applicant chooses from a tool chosen for her by the party that pays. California’s rules will take hold of the second kind at the moment of decision on January 1, 2027. The training, and the market position it builds, will remain outside them until the law recognizes that the applicant whose application teaches the screener has a claim on what it learns.

---

[1] LinkedIn, Enable or Disable Applicant Evaluation for Hiring Assistant, Recruiter Help, https://www.linkedin.com/help/recruiter/answer/a7460381 (last visited Sept. 25, 2026) (noting that “Hiring Assistant is available as an add-on for LinkedIn Recruiter” and that “[a]s an admin, you can turn this setting on or off”); see also Press Release, LinkedIn, Hiring Assistant, LinkedIn’s First AI Agent for Recruiters, To Launch Globally in English (Sept. 3, 2025), https://news.linkedin.com/2025/hiring-assistant-globally-available (announcing that Hiring Assistant “will be globally available in English by the end of September”).

[2] Indeed, Terms of Service § C.8 (last updated July 17, 2026), https://www.indeed.com/legal (stating that an employer using Smart Screening requests “an AI-powered summary with each candidate application delivered to you” and an “AI-powered match score” for applications that meet its required criteria).

[3] SimplyHired, https://www.simplyhired.com/ (last visited Sept. 25, 2026) (quoting the site footer); Indeed, Terms of Service, supra note 2 (stating that the Terms apply to “the following Indeed websites, which are part of the Site: workopolis.com, simplyhired.com, and resume.com”).

[4] Cal. Code Regs. tit. 11, § 7200(a)-(b) (2026) (applying the automated decisionmaking requirements to “[a] business that uses ADMT to make a significant decision concerning a consumer” and requiring existing uses to comply “no later than January 1, 2027”).

[5] Mobley v. Workday, Inc., 740 F. Supp. 3d 796, 806-08 (N.D. Cal. 2024) (holding that the complaint plausibly alleged that Workday is an agent of its client-employers and thus an “employer” under Title VII, the ADEA, and the ADA).

[6] Incogni, Are Job-Search Platforms Exploiting Job Seekers for Their Personal Data? 11, 14 (June 3, 2026), https://blog.incogni.com/are-job-search-platforms-exploiting-job-seekers-for-their-personal-data/ (reporting the finding for ZipRecruiter and FlexJobs in a study that surveyed 1,000 Americans from May 5 to 8, 2026, and reviewed nine platforms); see J.R. Howell, Two-Sided Market Structure Creates an Antitrust Blind Spot for Non-Paying Users, The American Counsel (2026) (analyzing the study’s finding that eight of nine platforms sell user data); J.R. Howell, Glassdoor Sells to the Employers Its Reviewers Are Warning You About, The American Counsel (2026) (drawing on the study’s usage figures).

[7] Id. at 12-13 (describing tools at Indeed, LinkedIn, and SimplyHired that the platforms are “developing or rolling out”).

[8] LinkedIn, Enable or Disable Applicant Evaluation, supra note 1 (stating also that “Hiring Assistant in LinkedIn Recruiter can evaluate applicants and summarize how their qualifications match job requirements”).

[9] Indeed, Indeed Smart Screening, https://www.indeed.com/employers/solutions/smart-screening (last visited Sept. 25, 2026) (stating that “Indeed Smart Screening screens and scores candidates based on your criteria”).

[10] Indeed, How To Enroll Jobs in Smart Screening, https://www.indeed.com/help/employers/articles/how-to-enroll-jobs-in-advanced-screening (last visited Sept. 25, 2026) (explaining that Smart Screening “can review resumes, score applicants, ask follow-up questions, and more”).

[11] SimplyHired, supra note 3 (quoting the site footer); Indeed, Privacy Policy (last updated Sept. 22, 2026), https://hrtechprivacy.com/brands/indeed (describing, under the heading “How we use Artificial Intelligence or ‘AI,’” the summary evaluation and match score available to employers).

[12] Indeed, Privacy Policy, supra note 11 (describing, under the heading “Automated Processing of Your Personal Data,” automated processing used “to match Job Seekers with Employers effectively” and “to send particular Job Seeker applications to an Employer”).

[13] Incogni, supra note 6, at 7-8 (reporting that “[o]ver 60% of recent job-seekers feel as if they write their resumes for algorithms rather than people” and that “over 50%” of respondents were “at least somewhat comfortable” with “their data being processed and used for the training of ‘AI’ models,” while “[o]nly a quarter” were at least somewhat uncomfortable).

[14] Id. at 12-13 (describing LinkedIn’s automated interviews as “currently in early testing for LinkedIn Hiring Pro users” and quoting Indeed’s statement that “Indeed Recruiting Services processes personal data including application data to generate … summaries”). The Indeed language Incogni quoted comes from the February 2026 version of Indeed’s privacy policy. See infra notes 76-77.

[15] Indeed, Terms of Service, supra note 2, § C.6 (describing “Indeed Career Scout (Beta)”).

[16] ZipRecruiter, Inc., Annual Report (Form 10-K), at Item 1 (Feb. 25, 2026), https://www.sec.gov/Archives/edgar/data/1617553/000161755326000016/zip-20251231.htm (describing ZipRecruiter’s job seeker products).

[17] FlexJobs, ExpertApply, https://www.flexjobs.com/expert-apply (last visited Sept. 25, 2026) (describing the tool); FlexJobs, How FlexJobs Works, https://www.flexjobs.com/how-flexjobs-works (last visited Sept. 25, 2026) (stating that “paid members get exclusive access to . . . AI-powered tools like ExpertApply”).

[18] Incogni, supra note 6, at 7-8 (reporting that “1 in 3 respondents used so-called AI features provided by professional and networking sites to optimize their profiles” and that “[a]round a third of respondents” used language or image-generation tools).

[19] LinkedIn, Hiring Assistant in Recruiter FAQ, Recruiter Help, https://www.linkedin.com/help/recruiter/answer/a7109304 (last visited Sept. 25, 2026) (identifying the models used for intake and search, messaging, and “Summarization of candidates”).

[20] LinkedIn, Review Applicants with Hiring Assistant, Recruiter Help, https://www.linkedin.com/help/recruiter/answer/a7109476 (last visited Sept. 25, 2026) (adding that “[t]he candidate profile card shows a brief summary of the candidate’s qualifications as they relate to your project and hiring needs”).

[21] LinkedIn, Hiring Pro, https://business.linkedin.com/hire/hiring-pro (last visited Sept. 25, 2026) (describing the product’s applicant sorting and “instant candidate summaries”).

[22] LinkedIn, AI Interviews in LinkedIn Hiring Pro, LinkedIn Help, https://www.linkedin.com/help/linkedin/answer/a8330369 (last visited Sept. 25, 2026) (describing the feature as “an early-stage LinkedIn Hiring Pro Jobs feature” and explaining that each response is rated on “[i]ts match to the hirer’s ideal response”).

[23] Indeed, Terms of Service, supra note 2, § C.8 (setting out the “Indeed Smart Screening (Beta)” terms for employers).

[24] Id. (adding that “you can turn off AI Recruiter or Smart Screening at any time in your Employer Dashboard”).

[25] LinkedIn, AI Interviews, supra note 22.

[26] Indeed, About AI Recruiter Questions, Indeed Help, https://www.indeed.com/help/job-seekers/articles/42787723284749-about-ai-recruiter-questions (last visited Sept. 25, 2026) (adding that “[i]f you are asked to answer with audio or video, you may switch to text if needed”).

[27] LinkedIn, LinkedIn and Generative AI (GAI) FAQs, LinkedIn Help, https://www.linkedin.com/help/linkedin/answer/a5538339 (last visited Sept. 25, 2026) (explaining that shared resume data “may be used as an input so that they can find you and see how you match their job qualifications”); see also LinkedIn Help, https://www.linkedin.com/help/linkedin/answer/a1327213 (last visited Sept. 25, 2026) (stating that after an application “the Share resume data with recruiters toggle will be turned on”).

[28] LinkedIn, Control Whether LinkedIn Uses Your Data To Improve Generative AI (GAI) Models That Are Used for Content Creation on LinkedIn, LinkedIn Help, https://www.linkedin.com/help/linkedin/answer/a6278444 (last visited Sept. 25, 2026) (stating that the setting “only controls the use of your data for training the models that may power those features”); Blake Lawit, Updates to LinkedIn’s Terms of Service, LinkedIn (Sept. 18, 2024), https://www.linkedin.com/blog/member/trust-and-safety/updates-to-our-terms-of-service-2024 (announcing that “[w]hen it comes to using members’ data for generative AI training, we offer an opt-out setting”).

[29] Howell, Two-Sided Market Structure, supra note 6 (describing the job seeker as the subsidy side of a two-sided platform); ZipRecruiter 10-K, supra note 16, at 10, 23 (stating that “[o]ur marketplace is free to job seekers”).

[30] ZipRecruiter 10-K, supra note 16, at Item 1 (describing the matching technology); ZipRecruiter, Global Privacy Policy § 5.C (effective Sept. 30, 2026), https://www.ziprecruiter.global/en/privacy (stating also that ZipRecruiter may use personal data “to help train, optimize, and power our AI systems” and that “[o]ur AI does not make independent, automated choices to hire or reject anyone”).

[31] ZipRecruiter, Global Terms of Use § 7.A (effective Aug. 11, 2026), https://www.ziprecruiter.global/en/terms (describing employer screening functions); see also id. § 8.F (stating that “ZipRecruiter makes no hiring recommendations or decisions about any applicant”).

[32] LinkedIn, AI Interviews, supra note 22 (distinguishing practice interviews, for which LinkedIn is controller, from hirer interviews, for which “[h]irers are the data controller” and “LinkedIn is a data processor on behalf of the hirer”).

[33] Cal. Code Regs. tit. 11, § 7200(b) (2026) (setting the compliance date for existing uses); Cal. Privacy Prot. Agency, Rulemaking Page for CCPA Updates, Cybersecurity Audits, Risk Assessments, Automated Decisionmaking Technology, and Insurance, https://cppa.ca.gov/regulations/ccpa_updates.html (last visited Sept. 25, 2026) (reporting that the Board adopted the regulations “[o]n July 24, 2025,” that they “were approved by the Office of Administrative Law” on “September 22, 2025,” and that their effective date is “January 1, 2026”).

[34] Cal. Code Regs. tit. 11, § 7200(a) (2026) (defining the scope of Article 11).

[35] Id. § 7001(ddd), (ddd)(4)(A) (defining “significant decision” and its employment component).

[36] Cal. Civ. Code § 1798.145(m)(4) (West 2026) (providing that the applicant and employee exemption “shall become inoperative on January 1, 2023”).

[37] Cal. Code Regs. tit. 11, § 7001(e), (e)(1) (2026) (defining “automated decisionmaking technology”).

[38] Id. § 7001(e)(1)(A)-(C) (setting the three requirements of human involvement).

[39] Indeed, Privacy Policy, supra note 11 (stating these propositions under the heading “How we use Artificial Intelligence or ‘AI’”).

[40] Indeed, Terms of Service, supra note 2, § B.5 (addressing “Candidate Summaries and Display”).

[41] LinkedIn, Hiring Assistant FAQ, supra note 19 (answering “How does Hiring Assistant comply with applicable privacy, data protection, and AI/ADT laws, regulations, executive orders, and guidance?”).

[42] N.Y.C. Admin. Code § 20-870 (2026) (defining “automated employment decision tool”); id. § 20-871(a) (prohibiting an employer or employment agency from using such a tool to screen a candidate unless the tool has been the subject of a bias audit conducted no more than one year before its use and a summary of the results has been made public).

[43] Indeed, How To Enroll Jobs in Smart Screening, supra note 10 (listing the jobs that “cannot be enrolled”).

[44] Indeed, AI and AEDT FAQ (last updated Aug. 25, 2026), https://www.indeed.com/legal/http-indeed-com-legal-aiandaedtfaq (stating also that “AI may be used to generate summaries of candidates’ skills, including whether they match employer criteria”).

[45] Cal. Code Regs. tit. 11, § 7220(b)(2)-(3) (2026) (setting the timing and manner of the Pre-use Notice).

[46] Id. § 7220(c)(1), (c)(5)(A)-(B) (setting the contents of the Pre-use Notice and barring a purpose stated “in generic terms, such as ‘to make a significant decision’”).

[47] Indeed, Terms of Service, supra note 2, § C.8.

[48] Cal. Code Regs. tit. 11, § 7221(b)(2) (2026) (excepting “admission, acceptance, or hiring decisions” from the opt-out on the stated conditions); see also id. § 7221(b)(1) (excepting any significant decision for which the business offers an appeal “to a human reviewer who has the authority to overturn the decision”).

[49] Id. § 7222(b)(2)-(3) (setting the contents of a response to a request to access ADMT).

[50] LinkedIn, AI Interviews, supra note 22.

[51] Cal. Code Regs. tit. 11, § 7153(a)-(b) (2026) (imposing the information duty on a business that makes ADMT available to another business).

[52] Id. § 7150(b)(6) (listing training as a processing activity that requires a risk assessment and defining “intends to use”).

[53] Id. §§ 7155(b), 7157(a)(1) (setting the deadline for risk assessments of processing that began before and continues after the regulations’ effective date, and the date for submission to the Agency). The approved text left the effective date in § 7155(b) as a placeholder for the Office of Administrative Law to supply. The Agency’s consolidated regulations state it as January 1, 2026.

[54] Indeed, Privacy Policy, supra note 11 (stating that Indeed uses personal data to build models “which allows Job Seekers and Employers to quickly and easily find one another and determine if they are a good match”).

[55] LinkedIn, Control Whether LinkedIn Uses Your Data, supra note 28 (limiting the setting to content-generating models and excluding “models used to personalize your LinkedIn experience or models we use for security, safety, or anti-abuse purposes”); LinkedIn, GAI FAQs, supranote 27 (directing objections to the training of other models to the “LinkedIn Data Processing Objection Form”).

[56] Cal. Code Regs. tit. 11, § 7050(a)(3) (2026) (permitting internal use to improve services “even if this business purpose is not specified in the written contract”).

[57] Id. § 7050(a)(3)(A) (illustrating the limit with an email marketing service provider that “cannot use the original email list to send marketing emails on behalf of another business”).

[58] Indeed, Privacy Policy, supra note 11 (stating, under the heading “Who is Responsible for Your Information,” that “[f]or certain processing, such as when the employer sets questions like screener questions, Indeed can be a processor of your application materials”).

[59] LinkedIn, LinkedIn Talent Solutions AI Compliance FAQ, Recruiter Help, https://www.linkedin.com/help/recruiter/answer/a6215067 (last visited Sept. 25, 2026) (answering whether “data provided by our customers” will “be used by LinkedIn to improve the product/service”).

[60] Mobley, 740 F. Supp. 3d at 802-03, 811 (describing the allegations that Mobley applied through third-party job sites and was rejected “for every one of the 100-plus applications that he submitted to companies using Workday’s platform”).

[61] Id. at 806 (holding that the complaint plausibly alleged agency liability).

[62] Id. at 806-07 (describing the alleged delegation of screening functions).

[63] Id. at 807 (rejecting a distinction between automated and human agents).

[64] Id. at 808 (offering spreadsheet and email vendors as examples of software providers that are not agents).

[65] Id. (quoting 42 U.S.C. §§ 2000e(c), 12111(7)).

[66] Id. at 808-09 (dismissing the employment agency theory without leave to amend and noting that “[t]here are no factual allegations supporting that Workday recruits or solicits candidates”).

[67] Mobley v. Workday, Inc., No. 23-cv-00770-RFL, 2025 WL 1424347, slip op. at 10 (N.D. Cal. May 16, 2025) (ECF No. 128) (granting preliminary certification of an ADEA collective).

[68] Id., slip op. at 17-18 (defining the collective).

[69] Order re HiredScore Dispute at 2, Mobley v. Workday, Inc., No. 23-cv-00770-RFL (N.D. Cal. July 29, 2025), ECF No. 158 (holding that “[t]he scope of the collective, at the preliminary certification stage, includes individuals whose applications were scored, sorted, ranked, or screened using HiredScore AI features”).

[70] Mobley, 2025 WL 1424347, slip op. at 19 (quoting Workday’s submission and rejecting it as a ground for “withholding notice”).

[71] Pls.’ Mot. for Class Certification, Mobley v. Workday, Inc., No. 23-cv-00770-RFL (N.D. Cal. Sept. 14, 2026), ECF No. 418 (seeking certification of a class of applicants subject to “Workday’s AI-based recruitment . . . tools, including HiredScore/Spotlight, Fetch, and Candidate Skills Match”). The court earlier declined to certify an interlocutory appeal of its ruling that applicants may bring disparate impact claims under the ADEA. Order Denying Certification for Interlocutory Appeal at 2-3, Mobley v. Workday, Inc., No. 23-cv-00770-RFL (N.D. Cal. July 2, 2026), ECF No. 376 (reasoning that “at this stage of the litigation, it no longer appears that an immediate appeal would materially advance the ultimate termination of the litigation”).

[72] Press Release, Indeed, Indeed Introduces New Suite of Hiring Products: Career Scout, Talent Scout, Premium Sponsored Jobs and Indeed Connect (Sept. 10, 2025), https://www.indeed.com/news/releases/indeed-introduces-new-suite-of-hiring-products-career-scout-talent-scout-premium-sponsored-jobs-and-indeed-connect (describing Talent Scout and stating that “Advanced Screening scores applicants against each employer’s unique criteria and uses conversational AI to gather nuanced candidate information”).

[73] Recruit Holdings Co., Transcript of Earnings Results Call and Sell-side Analyst Follow-up for Q4 FY2025, at 4 (May 15, 2026), https://file.recruit-holdings.com/files/en/Recruit_202603Q4_call-transcript_en.pdf (reporting remarks of Hisayuki Idekoba that “only 30% of applications come from” search). The transcript is an edited English translation of a call conducted in Japanese and is “provided for the convenience of investors only.” Id. at 1.

[74] LinkedIn, Hiring Assistant FAQ, supra note 19 (answering the compliance question quoted supra note 41).

[75] Indeed, Terms of Service, supra note 2, § D.18 (describing “Indeed’s Role” and stating that these services “are distinct from the Site, but may involve such affiliates using the Site on their clients’ behalf”).

[76] Indeed, Privacy Policy (last updated Feb. 20, 2026), archived at https://web.archive.org/web/20260228165147/https://hrtechprivacy.com/brands/indeed (stating also that “Indeed Recruiting Services processes personal data including application data to generate these summaries for the purpose of assisting Employers with their recruiting and hiring” and that “Indeed Recruiting Services uses AI for those screening purposes on behalf of the Employer”).

[77] Indeed, Privacy Policy (last updated May 13, 2026), archived at https://web.archive.org/web/20260601154429/https://hrtechprivacy.com/brands/indeed (attributing the summary and screening functions to “Indeed”); see also Indeed, Privacy Policy, supra note 11 (stating that “Indeed processes personal data including application data to generate these summaries and scores for the purpose of assisting Employers with their recruiting and hiring”).

[78] Brush v. S.F. Newspaper Printing Co., 315 F. Supp. 577, 580 (N.D. Cal. 1970) (emphasis omitted) (holding that a newspaper publishing help-wanted advertising is not an employment agency under Title VII), aff’d, 469 F.2d 89 (9th Cir. 1972).

[79] Mobley, 740 F. Supp. 3d at 809 (quoting Brush, 315 F. Supp. at 580).

[80] Cal. Code Regs. tit. 2, § 11008(a) (2025) (defining “agent” and providing that “[a]n agent of an employer is also an ‘employer’ for purposes of the Act”); Press Release, Cal. C.R. Dep’t, Civil Rights Council Secures Approval for Regulations To Protect Against Employment Discrimination Related to Artificial Intelligence (June 30, 2025), https://calcivilrights.ca.gov/2025/06/30/civil-rights-council-secures-approval-for-regulations-to-protect-against-employment-discrimination-related-to-artificial-intelligence/ (announcing that the regulations “are set to go into effect on October 1, 2025”).

[81] Cal. Code Regs. tit. 2, § 11008(g) (2025) (defining “employment agency”).

[82] Id. § 11008.1(a), (a)(1)(A)(iv), (C), (D) (defining “Automated-Decision System” and listing examples).

[83] Id. § 11009(f) (making evidence of anti-bias testing relevant “to any such claim or available defense”); id. § 11013(c) (requiring preservation of “automated-decision system data” for four years).

[84] Order Granting in Part and Denying in Part Motion to Dismiss at 4, 6, Mobley v. Workday, Inc., No. 23-cv-00770-RFL (N.D. Cal. June 22, 2026), ECF No. 360 (quoting Raines v. U.S. Healthworks Med. Grp., 534 P.3d 40, 53 (Cal. 2023), and holding that the complaint “adequately alleges material participation in the allegedly unlawful discriminatory employment decisions by an actor . . . in California”).

[85] Fed. Trade Comm’n, A Look Behind the Screens: Examining the Data Practices of Social Media and Video Streaming Services 59-60 (Sept. 2024), https://www.ftc.gov/system/files/ftc_gov/pdf/Social-Media-6b-Report-9-11-2024.pdf (identifying default ingestion and the absence of opt-in or opt-out as trends among the companies studied); see id. at i, vi (describing the orders issued to nine social media and video streaming services and the 2019 to 2020 period studied); see also id. at 69 (finding that “Companies implemented a default policy of using consumers’ Personal Information, without giving consumers the opportunity to opt in or to opt out”).

[86] Id. at 82 (adding that “there was limited or no ability to opt in to or opt out of such uses”).

[87] Id. at vi (summarizing the competition implications).

[88] Id. at 79 (discussing the relationship between data practices and market power).

[89] Press Release, Fed. Trade Comm’n, FTC Reopens and Sets Aside Rytr Final Order in Response to the Trump Administration’s AI Action Plan (Dec. 22, 2025), https://www.ftc.gov/news-events/news/press-releases/2025/12/ftc-reopens-sets-aside-rytr-final-order-response-trump-administrations-ai-action-plan (announcing the set-aside).

[90] Staff in the Off. of Tech. & the Div. of Priv. & Identity Prot., Fed. Trade Comm’n, AI (and Other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive, Tech. Blog (Feb. 13, 2024), https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/02/ai-other-companies-quietly-changing-your-terms-service-could-be-unfair-or-deceptive (stating also that “[a] business that collects user data based on one set of privacy commitments cannot then unilaterally renege on those commitments after collecting users’ data”).

[91] Recruit Transcript, supra note 73, at 3 (reporting remarks of Hisayuki Idekoba).

[92] Id. at 3-4 (reporting remarks of Hisayuki Idekoba).

[93] Id. at 5 (reporting remarks of Hisayuki Idekoba).

[94] Id. at 4 (reporting remarks of Hisayuki Idekoba).

[95] Recruit Holdings Co., Consolidated Financial Results for FY2025, at 9 (May 15, 2026), https://file.recruit-holdings.com/files/en/Recruit_202603Q4_earnings_en.pdf (stating also that “[t]he Company aims to leverage the vast amount of data available in each service combined with AI and machine learning technology” and citing “the 3.5 million employers that use Indeed each year to hire” and “over 665 million verified job seeker profiles created on Indeed”).

[96] Recruit Transcript, supra note 73, at 24 (describing how “business clients . . . further utilize our sourcing products, Sponsored Jobs, and the AI features on Indeed,” and concluding that “[t]his creates a virtuous cycle”).

[97] LinkedIn, Hiring Assistant FAQ, supra note 19 (answering “What is Hiring Assistant?” and “What data is Hiring Assistant trained on?”).

[98] Id. (answering “How does Hiring Assistant learn from my actions and feedback?,” “What happens to the feedback I provide Hiring Assistant?,” and “How does Hiring Assistant process input and output data?”).

[99] Jens Prüfer & Christoph Schottmüller, Competing with Big Data, 69 J. Indus. Econ. 967, 967 (2021) (concluding also that “[m]arket tipping can be avoided if competitors share their user information”).

[100] Staff in the Bureau of Competition & Off. of Tech., Fed. Trade Comm’n, Generative AI Raises Competition Concerns, Tech. Blog (June 29, 2023), https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2023/06/generative-ai-raises-competition-concerns (observing also that “companies’ control over data may also create barriers to entry or expansion that prevent fair competition from fully flourishing”).

[101] U.S. Dep’t of Just. & Fed. Trade Comm’n, Merger Guidelines 24-25 (Dec. 18, 2023), https://www.ftc.gov/system/files/ftc_gov/pdf/2023_merger_guidelines_final_12.18.2023.pdf (discussing platform competition under Guideline 9); Memorandum from Andrew N. Ferguson, Chairman, Fed. Trade Comm’n, to Staff (Feb. 18, 2025), https://www.ftc.gov/system/files/ftc_gov/pdf/ferguson-memo-re-merger-guidelines.pdf (stating that the 2023 Guidelines “are in effect and are the framework for this agency’s merger-review analysis”).

[102] Andrei Hagiu & Julian Wright, When Data Creates Competitive Advantage, Harv. Bus. Rev., Jan.-Feb. 2020, https://hbr.org/2020/01/when-data-creates-competitive-advantage (arguing that most claimed data advantages are overstated).

[103] Andrei Hagiu & Julian Wright, Data-Enabled Learning, Network Effects, and Competitive Advantage, 54 RAND J. Econ. 638, 638 (2023) (modeling the conditions under which data-enabled learning confers competitive advantage).

[104] Incogni, supra note 6, at 3 (reporting platform usage in the past five years among respondents).

[105] Cal. Privacy Prot. Agency, Laws & Regulations, https://cppa.ca.gov/regulations/ (last visited Sept. 25, 2026) (noting, under “Employee Data,” that “[t]he preliminary comment period closed on May 20, 2026”).

[106] Fed. Trade Comm’n, A Look Behind the Screens, supra note 85, at 82 (recommending that companies address the lack of “access, choice, control, transparency, explainability, and interpretability relating to their use of automated systems”).

[107] Cal. Code Regs. tit. 11, § 7222(i) (2026) (requiring a service provider or contractor to “provide assistance to the business in responding to a verifiable consumer request to access ADMT”).