Aurora Zhang
@aurroz.bsky.social
9/9 Join the AI Policy Opportunities mailing list, shown below. Thanks for the great tutorial!
June 26, 2025 at 1:12 PM
9/9 Join the AI Policy Opportunities mailing list, shown below. Thanks for the great tutorial!
8/n Diversify communication formats, such as using blog posts, twitter threads, strong presentations
Align research timelines with policy needs: maybe sometimes put out early work
Align research timelines with policy needs: maybe sometimes put out early work
June 26, 2025 at 1:06 PM
8/n Diversify communication formats, such as using blog posts, twitter threads, strong presentations
Align research timelines with policy needs: maybe sometimes put out early work
Align research timelines with policy needs: maybe sometimes put out early work
7/n How can research interface better with government? Several lessons:
1. policy change takes time, sustained relationships
2. impactful work often frames problems, not just works within them (but this is <5% of the work)
3. understand audience constraints
4. align research with policy realities
1. policy change takes time, sustained relationships
2. impactful work often frames problems, not just works within them (but this is <5% of the work)
3. understand audience constraints
4. align research with policy realities
June 26, 2025 at 1:04 PM
7/n How can research interface better with government? Several lessons:
1. policy change takes time, sustained relationships
2. impactful work often frames problems, not just works within them (but this is <5% of the work)
3. understand audience constraints
4. align research with policy realities
1. policy change takes time, sustained relationships
2. impactful work often frames problems, not just works within them (but this is <5% of the work)
3. understand audience constraints
4. align research with policy realities
6/6 Recommendations and lessons: encourage virtual/hybrid participation, avoid purely extractive interactions. Build meaningful relationships - don’t always aim for the most senior person. Assess the environment the institution is situated in, assess how detailed/broad the communication should be
June 26, 2025 at 12:59 PM
6/6 Recommendations and lessons: encourage virtual/hybrid participation, avoid purely extractive interactions. Build meaningful relationships - don’t always aim for the most senior person. Assess the environment the institution is situated in, assess how detailed/broad the communication should be
5/n Queer in AI wanted to influence NIST and build community l. Difficulties: choosing strategic vs radical approach, volunteers have variable capacity, invitations to panels and workshops didn’t always translate to being listened to, important policy convos often are inaccessible, informal
June 26, 2025 at 12:51 PM
5/n Queer in AI wanted to influence NIST and build community l. Difficulties: choosing strategic vs radical approach, volunteers have variable capacity, invitations to panels and workshops didn’t always translate to being listened to, important policy convos often are inaccessible, informal
4/n Barriers and difficulties: 1. turbulent bureaucratic processes, time delays and sudden deadlines. 2. Communication about avenues of influence is often obscured- often unclear who to talk to. 3. Civil society interests de-emphasized compared to security, 4. hype over genAI over traditional AI
June 26, 2025 at 12:40 PM
4/n Barriers and difficulties: 1. turbulent bureaucratic processes, time delays and sudden deadlines. 2. Communication about avenues of influence is often obscured- often unclear who to talk to. 3. Civil society interests de-emphasized compared to security, 4. hype over genAI over traditional AI
3/n Background info: NIST is an organization that sets standards, doesn’t have enforcement but has influence, and NIST’s risk management framework is internationally influential.
AISIC: ai safety consortium with over 250 orgs and different working groups tasked with building standards
AISIC: ai safety consortium with over 250 orgs and different working groups tasked with building standards
June 26, 2025 at 12:35 PM
3/n Background info: NIST is an organization that sets standards, doesn’t have enforcement but has influence, and NIST’s risk management framework is internationally influential.
AISIC: ai safety consortium with over 250 orgs and different working groups tasked with building standards
AISIC: ai safety consortium with over 250 orgs and different working groups tasked with building standards
2/n Main focus of the tutorial: (how) can we translate sociotechnical research to policymakers? The tutorial draws on specific experiences with the Biden Harris consortium on AI safety
June 26, 2025 at 12:29 PM
2/n Main focus of the tutorial: (how) can we translate sociotechnical research to policymakers? The tutorial draws on specific experiences with the Biden Harris consortium on AI safety
19/n Thank you for an inspiring and incisive keynote @mjcrockett.bsky.social!
June 26, 2025 at 12:06 PM
19/n Thank you for an inspiring and incisive keynote @mjcrockett.bsky.social!
18/n Interventions: 1. Better Images of AI tries to replace standard images of AI (oftentimes a disembodied big brain) with better images.
2. Realize yourself fully. Helen De Cruz: write boldly
3. Join various refuges, such as the AI workers inquiry, AAUP, anti-autocracy handbook
2. Realize yourself fully. Helen De Cruz: write boldly
3. Join various refuges, such as the AI workers inquiry, AAUP, anti-autocracy handbook
June 26, 2025 at 12:01 PM
18/n Interventions: 1. Better Images of AI tries to replace standard images of AI (oftentimes a disembodied big brain) with better images.
2. Realize yourself fully. Helen De Cruz: write boldly
3. Join various refuges, such as the AI workers inquiry, AAUP, anti-autocracy handbook
2. Realize yourself fully. Helen De Cruz: write boldly
3. Join various refuges, such as the AI workers inquiry, AAUP, anti-autocracy handbook
17/n Solnit: we are in a crisis of narratives. Crockett calls for us to "resist the mythology of the singular knower": the archetype of the lone genius.
A counternarrative comes from feminist social epistemology: Sandra Harding, Donna Haraway, Ann Fausto Starling, and others.
A counternarrative comes from feminist social epistemology: Sandra Harding, Donna Haraway, Ann Fausto Starling, and others.
June 26, 2025 at 11:55 AM
17/n Solnit: we are in a crisis of narratives. Crockett calls for us to "resist the mythology of the singular knower": the archetype of the lone genius.
A counternarrative comes from feminist social epistemology: Sandra Harding, Donna Haraway, Ann Fausto Starling, and others.
A counternarrative comes from feminist social epistemology: Sandra Harding, Donna Haraway, Ann Fausto Starling, and others.
16/n Crockett introduces the term "hermeneutical shackling": monopolistic AI generates epistemic monoculture, stifling imagination and imagination. Hao et al. 2024 shows that research using AI often tends to cluster around particular types of questions
June 26, 2025 at 11:52 AM
16/n Crockett introduces the term "hermeneutical shackling": monopolistic AI generates epistemic monoculture, stifling imagination and imagination. Hao et al. 2024 shows that research using AI often tends to cluster around particular types of questions
15/n c.f. aspirational harms of generative AI (Fazelpour & Magnani 2025) does this constrain the imaginations of AI users by showing them a single picture of what is possible?
June 26, 2025 at 11:50 AM
15/n c.f. aspirational harms of generative AI (Fazelpour & Magnani 2025) does this constrain the imaginations of AI users by showing them a single picture of what is possible?
14/n But this may cause a hegemonic monoculture to emerge. We used to ask questions like "tell me a bedtime story" and "my friend is mad, did I do something wrong?" to our entire community, and now we get it from a single source.
June 26, 2025 at 11:48 AM
14/n But this may cause a hegemonic monoculture to emerge. We used to ask questions like "tell me a bedtime story" and "my friend is mad, did I do something wrong?" to our entire community, and now we get it from a single source.
13/n Zooming out to how this connects to big tech:
The stated goal of chatGPT is to "deeply understand you", be "more than a chatbot," it is an "expert, tutor, advisor, muse, collaborator, translator, entertainer, companion, analyzer". It will be "the way we interact with everything"
The stated goal of chatGPT is to "deeply understand you", be "more than a chatbot," it is an "expert, tutor, advisor, muse, collaborator, translator, entertainer, companion, analyzer". It will be "the way we interact with everything"
June 26, 2025 at 11:44 AM
13/n Zooming out to how this connects to big tech:
The stated goal of chatGPT is to "deeply understand you", be "more than a chatbot," it is an "expert, tutor, advisor, muse, collaborator, translator, entertainer, companion, analyzer". It will be "the way we interact with everything"
The stated goal of chatGPT is to "deeply understand you", be "more than a chatbot," it is an "expert, tutor, advisor, muse, collaborator, translator, entertainer, companion, analyzer". It will be "the way we interact with everything"
12/n Second example: can AI be better scientists than people? Some say that AI is better at scientific reasoning than people, but existing benchmarks are disconnected from participants' personal motivations for studying a question and only measure participants working alone, not collaboratively
June 26, 2025 at 11:42 AM
12/n Second example: can AI be better scientists than people? Some say that AI is better at scientific reasoning than people, but existing benchmarks are disconnected from participants' personal motivations for studying a question and only measure participants working alone, not collaboratively
11/n But this is a DEAD empathy - it is contextualized, done completely over a computer, does not have any face-to-face interaction.
But in reality, people are interested in being understood by someone who has been through the experience they've been through. This is something AI is not capable of.
But in reality, people are interested in being understood by someone who has been through the experience they've been through. This is something AI is not capable of.
June 26, 2025 at 11:36 AM
11/n But this is a DEAD empathy - it is contextualized, done completely over a computer, does not have any face-to-face interaction.
But in reality, people are interested in being understood by someone who has been through the experience they've been through. This is something AI is not capable of.
But in reality, people are interested in being understood by someone who has been through the experience they've been through. This is something AI is not capable of.
10/n First example: AI is thought of as being empathic. In a lab setting, we study this by having some participant describe a complex social situation they are dealing with. We have AI and humans generate responses, and often we find that AI responses are rated to be more empathetic
June 26, 2025 at 11:33 AM
10/n First example: AI is thought of as being empathic. In a lab setting, we study this by having some participant describe a complex social situation they are dealing with. We have AI and humans generate responses, and often we find that AI responses are rated to be more empathetic
9/n In AI, researchers often ask, "can AI outperform humans on some cognitive process" and measure this cognitive process in a DEAD way on AI. Then we often make general claims about the general cognitive process in the world
June 26, 2025 at 11:31 AM
9/n In AI, researchers often ask, "can AI outperform humans on some cognitive process" and measure this cognitive process in a DEAD way on AI. Then we often make general claims about the general cognitive process in the world
8/n The problem is that we use these DEAD cognitive metrics to generalize to phenomena in the larger world. "We mistake DEAD generosity with generosity in the real world"
June 26, 2025 at 11:29 AM
8/n The problem is that we use these DEAD cognitive metrics to generalize to phenomena in the larger world. "We mistake DEAD generosity with generosity in the real world"
7/n DEAD cognition is Decontextualized, Engineered, Anonymized, Disembodied.
One example is using the Dictator Game to measure generosity in a lab setting, which has limited social or relational context, stripped of identity, completely computerized
One example is using the Dictator Game to measure generosity in a lab setting, which has limited social or relational context, stripped of identity, completely computerized
June 26, 2025 at 11:27 AM
7/n DEAD cognition is Decontextualized, Engineered, Anonymized, Disembodied.
One example is using the Dictator Game to measure generosity in a lab setting, which has limited social or relational context, stripped of identity, completely computerized
One example is using the Dictator Game to measure generosity in a lab setting, which has limited social or relational context, stripped of identity, completely computerized
6/n Researchers now often design their studies about what is possible on MTurk as a result of publishing incentives. Due to constraints of the mechanized laboratory, these methodologies are "DEAD"ening cognition.
June 26, 2025 at 11:26 AM
6/n Researchers now often design their studies about what is possible on MTurk as a result of publishing incentives. Due to constraints of the mechanized laboratory, these methodologies are "DEAD"ening cognition.
5/n Where does this pessimism come from? Early cognitive scientists conceptualized the mind as a computing machine, and this metaphor is now ubiquitous. But this metaphor is not inevitable - it ws a choice to emphasize this aspect of the mind as opposed to affective factors, history, culture
June 26, 2025 at 11:22 AM
5/n Where does this pessimism come from? Early cognitive scientists conceptualized the mind as a computing machine, and this metaphor is now ubiquitous. But this metaphor is not inevitable - it ws a choice to emphasize this aspect of the mind as opposed to affective factors, history, culture