logan koepke

@jlkoepke

senior project director @upturn.org. research and advocacy on the material harms of AI systems, with a specific focus on civil rights. personal views. we have it in our power to begin the world over again. www.jlkoepke.com

1861 Siguiendo1139 Seguidores578 Publicaciones

Bluesky

¿Es tu perfil?Reclámalo y jlkoepke.dirty.social es tuyo.Reclamar perfil

logan koepke@jlkoepke· 5d

perhaps the current leadership of the FTC could examine, you know, some of the relevant scholarship on the issue. "Although the statistics are imperfect, they plainly reflect the difficulty of proving disparate impact cases." scholarship.law.gwu.edu/cgi/viewcont...

"Although the statistics are imperfect, they plainly reflect the difficulty of proving disparate impact cases."

logan koepke@jlkoepke· 5d

one way you can tell that the current FTC has zero familiarity with the basics of disparate impact is that it has no understanding of how often corporations *are* able to defeat claims by simply offering a basic business rationale for a practice. www.ftc.gov/system/files...

Compounding its perniciousness, the theory of disparate-impact liability is effectively unbounded because almost any conceivable policy or practice affects different groups differently. As one commentator succinctly put it, “[d]isparate impact is one of the most controversial concepts in civil rights law and policy partly because just about everything has a disparate impact.”25 And while some form of a business-necessity defense is generally available, in practice that defense often provides little protection."

logan koepke@jlkoepke· 21d

Often in discussions of algorithmic discrimination and proposed new laws or regulations you'll hear some folks say "frankly, we have everything we need to combat discrimination." Oddly, you don't hear those same people say anything about this administration's revanchist moves to gut civil rights.

AP headline: "US civil rights agency moves to end demographic data collection after 60 years."

Reuters headline: "Trump consumer finance watchdog ends key civil rights-era anti-discrimination protection."

Consumer Finance Monitor headline: "HUD proposes to remove its Fair Housing Act disparate impact rule."

logan koepke@jlkoepke· 22d

e.g.,

Hyperdeimensional Putting Private AI Governance into Action
A policy proposal
by Dean Ball
March 20, 2025

The proposal below reflects some of the work I have done in my capacity as a Fellow with Fathom, a new non-profit exploring private governance in AI. Fathom is exploring policies similar to those discussed in this piece and may pursue state policy on these topics. I have had no role in any political efforts on any legislation, nor have I drafted, edited, or otherwise reviewed the text of any draft bills. My analysis here reflects my independent research and views.

I do not mean to suggest this idea is completely novel. It is heavily influenced, for example, by the work of Jack Clark and Gillian Hadfield in their regulatory markets paper. Many of the basic functions this governance structure would perform are recommended in the recently released Joint Policy Working Group on AI Frontier Models, the report commissioned by California Governor Gavin Newsom after the veto of SB 1047.

Nonetheless, this proposal is a fundamentally different way of thinking about AI governance than the vast majority of policies being considered today.

logan koepke@jlkoepke· 26d

so the killer use-case pitch for a technology that is going to allegedly pull forward unbounded human flourishing is, of course, the very human flourishing-coded ***prior authorization for a healthcare procedure***

Tweet from Bret Taylor

Introducing Horizon

Today, Sierra is announcing Horizon, a platform that enables agents to pursue long-horizon goals, like originating a loan or getting prior authorization for a healthcare procedure.

With Horizon, you can build agents that go far beyond customer service, truly driving revenue growth. Horizon agents proactively engage with your customers over days, weeks, or months, using the power of agentic AI to determine the steps required to close a sale, upgrade a subscription plan, or schedule a test drive.

logan koepke@jlkoepke· 27d

- third-party audits for unbounded risks that have no real concrete metrics to assess against: industry approved - third-party audits for specific risks that have longstanding metrics: destroyed by industry

Tweet from Jack Clark

At this point, everyone at the frontier of AI agrees that third-parties should test out AI systems and use these to develop standards to feed into policy - excellent to see @demishassabis laying out a farmework to do this! quote tweeting a Demis Hassabis post titled "A Framework for Frontier AI and the Dawning of a New Age."

logan koepke@jlkoepke· Jun 17

wow I wonder if there were any other signals that this administration was anti-immigrant? and that this would perhaps inform how OpenAI’s president spends political money?

The U.S. government’s latest battle with Anthropic has revived long-simmering concerns throughout the AI industry that the White House has placed a bullseye on their reliance on foreign AI talent.

The Trump administration appears to have targeted only Anthropic so far, warning the company on Friday in a letter from Commerce Secretary Howard Lutnick that it would need a license to make its latest models available to “foreign persons,” including its own employees. But Anthropic’s biggest rival, OpenAI, has flagged its concerns about the issue.

On Saturday, OpenAI Chief Strategy Officer Jason Kwon told staff in a Slack message that the company had told the government “strongly” that building AI “requires the best talent from around the world,” which is “part of what allows the U.S. to lead in AI.” Kwon said OpenAI was still trying to understand the implications of the government’s actions against Anthropic and that it was a “rapidly evolving situation with a lot of unknowns.”

logan koepke@jlkoepke· Jun 5

guess OpenAI's stated "our contract is keyed to laws as they exist today" red line w/ the DOD will get a real workout with the new NSPM directing the DOD to update 3000.09 in 90 days and then reviewed annually.

Sec. 3. Updated Policies and Guidance. (a) Within 90 days of the date of this memorandum, the Secretary of War shall issue an update to DOD Directive 3000.09 on Autonomy in Weapon Systems, to be reviewed annually to account for the rapidly evolving capabilities of AI systems, to ensure the deliberate adoption of AI systems that respect the chain of command and operational authorities, and remain consistent with the policy set forth in section 2 of this memorandum.

What if the government just changes the law or existing DoW policies?

Our contract explicitly references the surveillance and autonomous weapons laws and policies as they exist today, so that even if those laws or policies change in the future, use of our systems must still remain aligned with the current standards reflected in the agreement.

logan koepke@jlkoepke· Jun 5

But, critically, “there is a large gap between ‘we need to demonstrate that we’re trying here’ versus ’we actually want to do an LDA search that finds better alternatives’."

As a result, 𝐸11 described that “many institutions want to have a credible performance of doing the LDA search because they’re mostly concerned with being able to display something that satisfies the regulator.” But 𝐸11 noted that “there is a large gap between ‘we need to demonstrate that we’re trying here’ versus ’we actually want to do an LDA search that finds better alternatives’."

logan koepke@jlkoepke· Jun 5

Despite universal expectation that firms needed to test their models for discrimination and search for LDAs, the lack of regulatory specifics beyond general expectations cut both ways: some firms used this as room to experiment with new methods, but many others stuck to the bare minimum.

One participant described how they would “search the entire design space. But obviously, for practical reasons, we’re going to spend only X amount of dollars to search Y number of alternative models, and with the Y number of alternative models that we have done some sort of like semi-exhaustive search...that fulfills our need to show that we could not have done any better.”

Had regulators preordained a suite of methods, their guidance may not have been capacious enough to accommodate such exploration, further disincentivizing proactive fairness testing.

logan koepke@jlkoepke· Jun 5

As one participant noted, "I'm note even sure I've seen two institutions do bias testing and LDA (especially allowable performance loss) exactly the same way; it's kind of wild how different it is."

5.3.3 Navigating uncertainty around regulatory expectations, and its consequences. As shown in Section 5.1, while fair lending practitioners understood they needed to test their systems for discrimination and search for LDAs, they also had little concrete guidance on how to do so. This lack of clarity resulted in a significant variety of processes. As 𝐸11 stated: “I’m not even sure I’ve seen two institutions do bias testing and LDA (especially allowable performance loss) exactly the same way; it’s kind of wild how different it is.”

logan koepke@jlkoepke· Jun 5

Why so much variation on how to test underwriting models for discrimination, despite longstanding practice? In part, participants believed that there were "still no metrics or no concrete, crispy, guidance" from regulators, just so long as "it's reasonable."

Models that firms identify as posing the most fair lending risk...would then also be subject to a disparate impact review (𝐿2, 𝑇1, 𝑇5, 𝑇8, 𝐿7, 𝑇9, 𝑇2): as 𝐿7 shared, “our highest [model] risk category [...] gets the deepest level of testing: a full disparate impact analysis.” As we discuss in more detail in the next section, we observed variation in how institutions defined disparate impact, as well as the thresholds at which a disparate impact triggered further analysis. Participants (𝑇1, 𝑇5, 𝑇8) explained that disparate impact testing relied on “tests like adverse impact ratios [(AIR)], standardized mean differences [(SMD)], [and] marginal effects” (𝑇8), with a focus on determining whether disparities were statistically and practically significant. Other participants examined disparities not at the model level in abstraction, but instead through the lens of historic outcomes. 𝐿7’s institution would look at the default rates for various protected classes (e.g., the historic default rate for black and white borrowers) and then compare a model’s predictions against those historic outcomes to see if the model would worsen those disparities. If the model did not appear to worsen outcomes relative to historic rates, 𝐿7 described that “we consider the model to have low fair lending risk.” In other words, if the predicted default rate for specific subpopulations was in line with their historical base rate of default, the institution believed that it had a satisfactory business justification for the disparate impact.

Participants believed this variance was in part due to a lack of clear regulatory expectations on how such tests should be performed. 𝐿2 noted there are “still no metrics or no concrete, crispy, guidance” from regulators. 𝐸9 concurred, stating "there is no regulatory guidance that this is the exact way how you should be doing this disparity
testing. So as long as it’s reasonable [...] there is no black and white, right or wrong, answer here."

logan koepke@jlkoepke· Jun 5

One thing to highlight is the variance in what financial institutions actually do to satisfy fair lending law and regulation. We found that while participants at financial institutions all understood they had to test their models for disparate impact, *how* they did so varied significantly.

.1.1 Existence of Fair Lending Teams and Practices. As a first order matter, our interviews showed that fair
lending teams invariably exist within regulated financial institutions—which, as prior work has noted, is not
always the case at firms that are not covered by fair lending law [56]. All of the individuals we interviewed who
were at institutions covered by fair lending law described a designated fair lending unit that was responsible for
monitoring compliance with relevant laws (𝐿5, 𝐸1, 𝐸6, 𝐸4, 𝐿4, 𝐸11, 𝐿2). As 𝐿5, who worked in such a department,
described, “it is my responsibility to make sure that our bank does not discriminate against any of our customers or
potential customers.”
Beyond the existence of teams dedicated to antidiscrimination work, our interviews revealed a ubiquity and
formality of fair lending procedures.

logan koepke@jlkoepke· Jun 4

my 💭 to the trahan-obernolte AI discussion draft conversation: if the stated motivation is preempting state laws re: *frontier model development,* why does the text preempt "AI model" development and not "frontier model" development, given the definitions are different in the bill's text?

Subtitle C—Federalization and
Federal Resources
SEC. 121. FEDERALIZATION OF STATE LAWS REGULATING ARTIFICIAL INTELLIGENCE MODEL DEVELOPMENT.
13 (a) FINDINGS.—Congress finds that the development
14 of artificial intelligence models is a matter of national eco-
15 nomic significance and international competitiveness re-
16 quiring uniform Federal oversight. It is therefore the pur-
17 pose of this section to preempt State regulation specifically
18 targeting artificial intelligence model development and
19 allow the Federal Government to establish testing, evalua-
20 tion, and oversight frameworks for artificial intelligence
21 models.
22 (b) PREEMPTION.—No State or political subdivision
23 thereof may establish, continue in effect, or enforce any
24 law or regulation specifically regulating the development
25 of any artificial intelligence model.

13) FRONTIER MODEL.—The term ‘‘frontier model’’ means a foundation model trained utilizing a quantity of computing power greater than 1026 integer or floating-point operations, including computing for the original training run of such model and for any subsequent fine-tuning, reinforcement learning, or other substantial modification the developer of such model applies to such model.

(4) ARTIFICIAL INTELLIGENCE MODEL.—The term ‘‘artificial intelligence model’’ means the set of parameter values or weights that define a particular mathematical function which maps input data to output predictions, where the parameter values have been calculated through a machine learning process on training data.

logan koepke@jlkoepke· Jun 4

In particular, the operation of fair lending law is the result of a unique regulatory design — that is, how a regulatory system is designed to achieve its stated goals.

The regulatory design of fair lending law is strikingly different from other civil rights laws. By regulatory design, we mean the ways in which a regulatory system is structured—from the institutional form of the regulator to its enforcement tools to its compliance mechanisms—to achieve its stated goals. To understand why, consider the main ways in which Title VII, the foundational federal law prohibiting employment discrimination, is enforced. First, aggrieved individuals who believe they have been victims of discrimination must file a complaint with the Equal Employment Opportunity Commission (EEOC). The EEOC will investigate these complaints and either file their own lawsuit or issue a Notice of Right to Sue, allowing the aggrieved individual to vindicate their private right of action and file suit against an employer for discrimination [36]. The EEOC is almost completely reactive: to enforce Title VII, it almost exclusively relies upon aggrieved individuals filing a complaint of discrimination. Fair lending laws like the FHA and ECOA also equip victims of discrimination with a private right of action to sue lenders for discriminatory actions and empower regulators to investigate complaints. However, financial regulators who enforce fair lending laws—like the Consumer Financial Protection Bureau (CFPB)—also possess robust statutory authorities that authorize them to proactively and routinely inspect business practices for compliance with fair lending laws, whether or not the regulator believes discriminatory practices are occurring. Such supervisory authority and examination, where regulators actively oversee regulated entities’ compliance by sending examiners to a financial institution’s offices to probe for evidence of lending discrimination, are the key to the regulatory design of fair lending law. Through supervision and examination, fair lending regulators can direct financial institutions to address potential deficiencies in antidiscrimination efforts.

logan koepke@jlkoepke· Jun 4

🔦 Despite this long history, there is incredibly little public evidence about the operation of fair lending programs at US financial institutions — what they do, how they do it, and why they do it.

CONSUMER FINANCIAL PROTECTION BUREAU | JANUARY 2025
Supervisory Highlights: Advanced Technologies Special Edition
Issue 38, Winter 2025

logan koepke@jlkoepke· Jun 4

🏦 For decades, US financial institutions have been under a legal obligation to not discriminate in any credit transaction. For much of that same time, those same institutions have relied on algorithmic systems to determine who gets access to money and on what terms.

Banking Regulatory Agencies' Enforcement of the Equal Credit Opportunity Act and the Fair Housing Act | Hearings before a Subcommittee of the Committee on Government Operations House of Representatives Ninety-Fifth Congress Second Session | September 12, 14, and 15, 19879, Printed for the use of the Committee on Government Operations

minorities were roughly double those among whites . The one survey
which collected economic data on applicants showed that the discrepancy persisted even when income levels and other creditworthiness indicators were kept constant. Despite the results of the surveys, the regulatory agencies persisted in their refusal to implement an effective examination and enforcement program .

Finally, in April, 1976, eleven members of the coalition filed
suit, seeking to compel the agencies to take the basic step required
to detect discrimination through the examination process, and to use
their supervisory powers to eliminate discrimination where dis
covered. The steps we were seeking were essentially those recommended by the Assistant Attorney General in charge of the Civil
Rights Division in March, 1976 testimony before the Senate Committee
on Banking , Housing and Urban Affairs, and by the Committee itself
in its report of those hearings. We found it ironic, to say the
least, when the Justice Department's Civil Division defended the
agencies against the suit, using technical arguments such as a
claimed lack of standing to try to have the suit dismissed.

logan koepke@jlkoepke· Jun 4

🧵 How do companies in highly regulated domains actually test their AI systems for and mitigate discrimination? Our new @upturn.org paper at @facct.bsky.social w/ Emily Black, @mbogen.bsky.social, @s010n.bsky.social, and @wesleydeng.bsky.social, tries to answer this question: arxiv.org/abs/2606.02957

The Fair Lending Model: How the Longest-Running Algorithmic
Fairness Programs Work in Practice
EMILY BLACK∗, New York University, USA
MIRANDA BOGEN∗, Center for Democracy & Technology, USA
LOGAN KOEPKE∗, Upturn, USA
SOLON BAROCAS, Microsoft Research, USA
WESLEY DENG, Carnegie Mellon University, USA
MINGWEI HSU, Upturn, USA
U.S. financial institutions subject to fair lending laws have been running algorithmic fairness programs for decades. Despite
this long history, remarkably little is known about how these requirements operate in practice. In this paper, we offer the first
empirical account of how financial institutions test for and mitigate algorithmic discrimination on the ground. In doing so, we
shed light on how the regulatory design of fair lending law and regulation have shaped the policies, processes, and practices
of fair lending programs. Drawing on 35 semi-structured interviews with participants across the fair lending ecosystem, we
find that while financial institutions have a floor of fairness practices aimed at preventing discrimination in lending largely
absent in other domains, the specifics of how firms test for discrimination and search for less discriminatory algorithms varies
widely. We also find that regulatory supervision via fair lending examinations has been the key driver of compliance work, but
that the practical impact of fair lending programs often depends on how well they can navigate competing business incentives,
perceived legal tensions, and regulatory uncertainty. Ultimately, our findings highlight the unique role that supervisory
authority has played in successfully fostering fair lending practices—a regulatory design feature that is distinct from other
areas of civil rights law and almost completely absent from recent policy proposals for dealing with algorithmic discrimination.
CCS Concepts: • Social and professional topics → Computing / technology policy

logan koepke@jlkoepke· May 28

confirm they're following *their own* safety standards. it's a law about procedure, not substance. that it involves third parties i guess makes it interesting + reports with some potential new details will be made public, but those will be redacted. ¯\_(ツ)_/¯

Headline from WIRED on May 27, 2026: Illinois Lawmakers Just Passed America’s Strongest AI Safety Bill

The bill requires companies like OpenAI, Anthropic, and Google to have third parties confirm they’re following safety standards. Illinois governor JB Pritzker says he’ll sign it.

logan koepke@jlkoepke· May 27

a clear demonstration of how algorithmic monoculture results in systemic discrimination, exclusion in job opportunities as employers all rely on a few AI assessments from a few vendors. algorithmichiring.github.io

Over 90% of U.S. employers rely on hiring algorithms to screen job applicants. Many different employers use algorithms from the same few vendors. We conduct the largest empirical study of algorithmic hiring with data for 3.4 million real job applicants submitting 4 million applications to 156 employers across 11 market sectors. Every application was assessed by algorithms from a single vendor: we test whether this algorithmic monoculture bottlenecks job opportunities. We are the first to demonstrate large-scale evidence of racial disparities and homogeneous outcomes in high-stakes hiring decisions.
Key Findings
1 Large-scale adverse impact for Asians and Blacks. We are the first to demonstrate adverse impact in deployed algorithmic hiring as one of the largest demonstrations of unfair outcomes in real high-stakes AI decisions. 25.87% of applications submitted by Black applicants and 14.74% of applications submitted by Asian applicants are directed to positions that adversely impact them based on the standards of the relevant U.S. employment law (Title VII).
2 Adverse impact only revealed by disaggregated position-by-position analysis. While empirical studies of algorithmic hiring are very constrained due to data access limitations, prior studies showed minimal adverse impact due studying all of the vendor's data as a whole. By studying each position separately, in accordance with the standards of Title VII, we identify positions that demonstrate adverse impact that gets washed out in aggregate.
3 Algorithmic monocultures in hiring yield systemic rejections. We are the first to demonstrate systemic rejections in deployed algorithmic hiring as posited in prior theoretical research about algorithmic monoculture. The observed systemic rejection rate significantly exceeds that of the baseline of statistically independent decisions ...

logan koepke@jlkoepke· May 15

supporting the mob

screenshot of expensify expenses showing "AMTRAK MOB WASHINGTON DC $20.00" and "AMTRAK MOB WASHINGTON DC $16.00"

logan koepke@jlkoepke· May 12

Too many comments or tracked changes

Sorry, this file has too many comments or tracked changes. Please try accepting or rejecting some existing changes, or resolving and deleting some comments.

logan koepke@jlkoepke· May 9

🚛 moving to Charlottesville soon! would love to get connected with folks in the area, at UVA, etc. also looking for good trail running and gravel biking recommendations!

picture of a well-maintained trail on a blue sky day with light green leaves on trees with a bench

logan koepke@jlkoepke· May 9

first grade business card 🏄‍♂️

business card showing blue soccer ball above green grass with my name in blue

logan koepke@jlkoepke· May 6

read our recent report "Calculated Need: Algorithms and the Fight for Medicaid Home Care": www.upturn.org/reports/2025...

Calculated Need: Algorithms and the Fight for Medicaid Home Care A Comparison of Five States' Eligibility Algorithms Upturn, October 2025, by Miriam Osman, Emily Paul, Emma Weil

logan koepke@jlkoepke· May 6

part of OpenAI's healthcare policy wish list: let us do a bunch of AI pilots in Medicare and Medicaid.

Design pilot AI affordability measures with public reporting on what works. States and the federal government should establish AI pilots in Medicaid, Medicare, public hospitals, and other settings with rising costs. These pilots should test whether AI can expand service availability, shorten delays, and lower avoidable costs while preserving quality and human accountability. Pilots should have a clear ongoing public reporting of common metrics like patient out-of-pocket burden, access and wait times, clinician time returned to patient care, and health outcomes. Governments should publish results on a regular cadence to scale what works, address what does not, and ensure AI-driven efficiency flows to patients and public programs.

cover of OpenAI healthcare report

Keeping Patients First: A Blueprint for AI in U.S. Healthcare April 2026

logan koepke@jlkoepke· Apr 26

new Amtrak Siemens ALC-42 as seen from King St-Old Town Alexandria metro stop with the George Washington Masonic National Memorial in the background

logan koepke@jlkoepke· Apr 13

www.linkedin.com/help/linkedi...

How is LinkedIn ensuring AI content won’t include any discriminatory, abusive, or biased content?

LinkedIn strives to ensure that its products and services are fair, including by measuring and mitigating algorithmic bias. Our objective is to ensure equally qualified members receive equal treatment by our models. As with all our products, our teams continuously assess our systems, and if harmful biases are identified, we will work to address them.

What hirers receive after screening interview

Once the candidate completes the AI screening interview, you receive the full transcript and the audio or video recording, along with AI-generated summaries and a 5-point rating based on your ideal responses. As the hirer, you are responsible for reviewing and verifying all AI-generated information before making any hiring decisions. Candidates may contact you to request an accommodation, or to access their rating, summaries, transcript, or recording. You retain full control over evaluation and can decide how to proceed with each candidate.

How interview ratings work

Ratings are determined by using AI to analyze the applicant’s interview transcript. The overall rating provided to the hirer is the average rating of the applicant’s responses to each of your questions. Each of the applicant’s responses is rated on the following criteria:

Its match to the hirer’s ideal response
Whether the applicant supported it with details, such as evidence and examples
Whether it was clear, concise, and well structured

Note that as the hirer, you may have legal obligations associated with AI screening interviews and ratings, and are responsible for final hiring decisions. We encourage you to review the candidates’ AI interview transcripts or recordings, in addition to the ratings and summaries, before making any decision on candidate advancement.

If you prefer not to receive candidate ratings, you can turn them off and still receive an interview summary without a numerical rating or agentic recommendation. To update this preference, go to the AI Interview module on the Hiring Plan page and select AI Interview Settings. On the settings page, you’ll find an option to turn off AI-generated applicant ratings.

logan koepke@jlkoepke· Apr 13

"Employers can input the questions they want the interviewer to ask, and the answers they want to hear, and the AI will compare the candidates’ answers to that criteria." would love to see the validation study for this new product

LinkedIn recently began testing AI interviews with its LinkedIn Hiring Pro tool, an AI agent for small businesses. Small orgs don’t always have a dedicated TA pro—or even an HR pro—and often leave hiring to someone else in the business, Chimka noted. This new offering aims to allow employers and job seekers more flexibility. Employers can input the questions they want the interviewer to ask, and the answers they want to hear, and the AI will compare the candidates’ answers to that criteria. Job seekers can practice and interview at a time that works best for them, instead of having to fit into someone else’s schedule. While it’s early days, 80% of candidates who have completed an interview with the tool have given it a positive rating.

logan koepke@jlkoepke· Apr 4