logan koepke
@jlkoepke.bsky.social
senior project director @upturn.org. i work on AI/ML + civil rights research. personal views.
www.jlkoepke.com
www.jlkoepke.com
found the frontier
November 11, 2025 at 11:12 AM
found the frontier
the result: someone eligible for home and community-based services in Nebraska and Mississippi but ineligible in New Jersey, Missouri, and DC.
October 28, 2025 at 5:15 PM
the result: someone eligible for home and community-based services in Nebraska and Mississippi but ineligible in New Jersey, Missouri, and DC.
but states use these assessments *very differently.*
when designing their level of care algorithms, state agencies and vendors may choose different subsets of assessment questions to include in eligibility determinations and assign different weights or scores to those questions.
when designing their level of care algorithms, state agencies and vendors may choose different subsets of assessment questions to include in eligibility determinations and assign different weights or scores to those questions.
October 28, 2025 at 5:15 PM
but states use these assessments *very differently.*
when designing their level of care algorithms, state agencies and vendors may choose different subsets of assessment questions to include in eligibility determinations and assign different weights or scores to those questions.
when designing their level of care algorithms, state agencies and vendors may choose different subsets of assessment questions to include in eligibility determinations and assign different weights or scores to those questions.
standardized assessment is the first part of the eligibility determination process. the interRAI Home Care assessment is used in more than 25 different states. Since 1990, >100 million interRAI nursing home assessments have been completed in the US.
October 28, 2025 at 5:15 PM
standardized assessment is the first part of the eligibility determination process. the interRAI Home Care assessment is used in more than 25 different states. Since 1990, >100 million interRAI nursing home assessments have been completed in the US.
while this is a federal program, each state is responsible for designing its own rules and systems to determine which of its residents are eligible for HCBS benefits and how much home care they should receive.
every state uses automated tools to help determine HCBS eligibility
every state uses automated tools to help determine HCBS eligibility
October 28, 2025 at 5:15 PM
while this is a federal program, each state is responsible for designing its own rules and systems to determine which of its residents are eligible for HCBS benefits and how much home care they should receive.
every state uses automated tools to help determine HCBS eligibility
every state uses automated tools to help determine HCBS eligibility
millions of elderly people + people with disabilities rely on Medicaid Home and Community-Based Services to meet their daily needs (from help bathing, getting dressed, and eating meals) as well as to work + participate in their communities.
October 28, 2025 at 5:15 PM
millions of elderly people + people with disabilities rely on Medicaid Home and Community-Based Services to meet their daily needs (from help bathing, getting dressed, and eating meals) as well as to work + participate in their communities.
the newest report from my colleagues at Upturn interrogates a set of under-examined automated systems that directly impact the material conditions of your friends, family, and community: Medicaid home care eligibility algorithms.
October 28, 2025 at 5:15 PM
the newest report from my colleagues at Upturn interrogates a set of under-examined automated systems that directly impact the material conditions of your friends, family, and community: Medicaid home care eligibility algorithms.
putting other methods issues aside, in a world of algorithmic feeds, why would *social media following* be the best proxy for “attention”???
October 28, 2025 at 1:01 AM
putting other methods issues aside, in a world of algorithmic feeds, why would *social media following* be the best proxy for “attention”???
learned how to ride a bike today
October 11, 2025 at 9:09 PM
learned how to ride a bike today
ChatGPT appears to have fueled an individuals paranoia that everyone was out to get him, including his mother.
It appears to have resulted in a murder suicide.
www.wsj.com/tech/ai/chat...
It appears to have resulted in a murder suicide.
www.wsj.com/tech/ai/chat...
August 29, 2025 at 2:11 AM
ChatGPT appears to have fueled an individuals paranoia that everyone was out to get him, including his mother.
It appears to have resulted in a murder suicide.
www.wsj.com/tech/ai/chat...
It appears to have resulted in a murder suicide.
www.wsj.com/tech/ai/chat...
@beenwrekt.bsky.social on the venal negligence of "AI Safety" www.argmin.net/p/the-banal-...
August 28, 2025 at 3:42 PM
@beenwrekt.bsky.social on the venal negligence of "AI Safety" www.argmin.net/p/the-banal-...
these excerpts from the lawsuit against OpenAI are just so profoundly upsetting.
archive.org/details/2381...
archive.org/details/2381...
August 26, 2025 at 8:19 PM
these excerpts from the lawsuit against OpenAI are just so profoundly upsetting.
archive.org/details/2381...
archive.org/details/2381...
In other words, data center flexibility's emissions impact is context-dependent: in grids dependent upon fossil-heavy systems, like data center alley, flexibility may actually "increase emissions by extending the operational life of baseload coal."
August 14, 2025 at 3:10 PM
In other words, data center flexibility's emissions impact is context-dependent: in grids dependent upon fossil-heavy systems, like data center alley, flexibility may actually "increase emissions by extending the operational life of baseload coal."
"in systems [like the mid-Atlantic] with a large share of existing coal and relatively limited [variable renewable energy] availability, flexibility tends to shift load toward cheap, carbon-intensive baseload generation, which raises emissions even as costs fall."
August 14, 2025 at 3:10 PM
"in systems [like the mid-Atlantic] with a large share of existing coal and relatively limited [variable renewable energy] availability, flexibility tends to shift load toward cheap, carbon-intensive baseload generation, which raises emissions even as costs fall."
Because certain AI workloads are not time-sensitive, data center flexibility is touted as a potential way to reduce strain on the grid and ensure consumers don't face increased costs.
But a new study suggests data center flexibility in certain regions can actually result in higher CO2 emissions.
But a new study suggests data center flexibility in certain regions can actually result in higher CO2 emissions.
August 14, 2025 at 3:10 PM
Because certain AI workloads are not time-sensitive, data center flexibility is touted as a potential way to reduce strain on the grid and ensure consumers don't face increased costs.
But a new study suggests data center flexibility in certain regions can actually result in higher CO2 emissions.
But a new study suggests data center flexibility in certain regions can actually result in higher CO2 emissions.
the Executive Order on "Preventing Woke AI in the Federal Government" ultimately has agency heads like this determine if LLMs procured by the federal government have done enough to comply with ridding a system of DEI.
you tell me if you think this stops at "disclosure of system prompts"
you tell me if you think this stops at "disclosure of system prompts"
July 26, 2025 at 3:49 PM
the Executive Order on "Preventing Woke AI in the Federal Government" ultimately has agency heads like this determine if LLMs procured by the federal government have done enough to comply with ridding a system of DEI.
you tell me if you think this stops at "disclosure of system prompts"
you tell me if you think this stops at "disclosure of system prompts"
directly undercut by their own reporting, too
July 26, 2025 at 3:39 PM
directly undercut by their own reporting, too
"Research is Automated." "Writing is Automated." "Analysis is Automated."
Three sentences that will be included in countless legal challenges against any agency action that charges forward with these repeal efforts.
Three sentences that will be included in countless legal challenges against any agency action that charges forward with these repeal efforts.
July 26, 2025 at 3:37 PM
"Research is Automated." "Writing is Automated." "Analysis is Automated."
Three sentences that will be included in countless legal challenges against any agency action that charges forward with these repeal efforts.
Three sentences that will be included in countless legal challenges against any agency action that charges forward with these repeal efforts.
here is the underlying slide deck from DOGE, with some key slides highlighted www.washingtonpost.com/documents/85...
July 26, 2025 at 3:29 PM
here is the underlying slide deck from DOGE, with some key slides highlighted www.washingtonpost.com/documents/85...
the AI Action Plan and AI Executive Order directly contradict one another.
on the one hand, the AI Action Plan wants NIST to eliminate "misinformation" from it's AI RMF
on the other hand, the Executive Order wants agencies to only procure LLMs that are "truth-seeking"
on the one hand, the AI Action Plan wants NIST to eliminate "misinformation" from it's AI RMF
on the other hand, the Executive Order wants agencies to only procure LLMs that are "truth-seeking"
July 25, 2025 at 6:52 PM
the AI Action Plan and AI Executive Order directly contradict one another.
on the one hand, the AI Action Plan wants NIST to eliminate "misinformation" from it's AI RMF
on the other hand, the Executive Order wants agencies to only procure LLMs that are "truth-seeking"
on the one hand, the AI Action Plan wants NIST to eliminate "misinformation" from it's AI RMF
on the other hand, the Executive Order wants agencies to only procure LLMs that are "truth-seeking"
according to the EO:
1. LLMs used by the government shall "not manipulate responses in favor of ... DEI."
2. the EO order defines DEI as including "transgenderism."
so: what will OpenAI, Google, and Anthropic do to ensure their models do not "manipulate" responses "in favor" of "transgenderism"
1. LLMs used by the government shall "not manipulate responses in favor of ... DEI."
2. the EO order defines DEI as including "transgenderism."
so: what will OpenAI, Google, and Anthropic do to ensure their models do not "manipulate" responses "in favor" of "transgenderism"
July 24, 2025 at 1:16 AM
according to the EO:
1. LLMs used by the government shall "not manipulate responses in favor of ... DEI."
2. the EO order defines DEI as including "transgenderism."
so: what will OpenAI, Google, and Anthropic do to ensure their models do not "manipulate" responses "in favor" of "transgenderism"
1. LLMs used by the government shall "not manipulate responses in favor of ... DEI."
2. the EO order defines DEI as including "transgenderism."
so: what will OpenAI, Google, and Anthropic do to ensure their models do not "manipulate" responses "in favor" of "transgenderism"
good luck to all of the labs, this is what you signed up for www.whitehouse.gov/presidential...
July 23, 2025 at 11:22 PM
good luck to all of the labs, this is what you signed up for www.whitehouse.gov/presidential...
