Week 9a: Questions - Automating Innovation with AI?

[jamesallenevans]

Post your (<150 word) question for class about or inspired by the following readings:

“Machine science.” 2010. James Evans and Andrey Rzhetsky. Science 329(5990): 399-400.
“Catching crumbs from the table.” 2000. Ted Chiang. Nature 405(517).
“Machine learning as a tool for hypothesis generation.” 2023. Jens Ludwig, Sendhil Mullainathan. The Quarterly Journal of Economics 139 (2): 751-827.
“AI Expands Scientists’ Impact but Contracts Science’s Focus.” 2025. Hao, Qianyue, Fengli Xu, Yong Li, James Evans.

[aveeshagandhi]

Ted Chiang’s “Catching Crumbs from the Table” raises some super interesting questions about the ethical and societal implications of artificial intelligence surpassing human (behavior and intelligence). How can AI be considered in relation to economic inequality given the story's themes? Are advanced AI systems likely to reinforce existing disparities, or could they be mechanisms for reducing them? In addition, how can researchers and policymakers ensure the equitable distribution of the benefits of AI?

[JaslinAg]

"Innovation and Top Income Inequality", found that innovation by both entrants and incumbents increases top income inequality and that this correlation “at least partly reflects a causal effect” from innovation to top income. That innovation is positively correlated with social mobility. In class, we discussed that we invest in innovation to improve future welfare, so I’m curious about how innovation is affecting broader economic and welfare disparities.

What are the global effects of innovation on income inequality? Does the positive correlation between innovation and social mobility hold when comparing high-income economies like the US to middle- and low-income economies? Is U.S. innovation widening or narrowing the wealth gap between the U.S. and other countries? For example, the COVID-19 vaccine highlights how wealthier nations not only have the resources to innovate but also the ability to gatekeep access. Are there studies examining whether innovation is also a driver of inequality between top economies and less developed economies?

[e-uwatse-12]

The article Catching Crumbs from the Table revisits the question of the role of human scientists in an era where metahumans dominate research, raising several thought-provoking issues. With metahumans relying on digital neural transfer (DNT) to communicate their findings, human scientists struggle to keep up, leaving their journals reduced to second-hand accounts of discoveries they can barely comprehend. This shift prompts reflection on whether human researchers can still make meaningful contributions to science. In response, many have turned to hermeneutics—deciphering metahuman texts, reverse-engineering artifacts, and even remotely analyzing metahuman research facilities.

[grozdanickata]

“Catching crumbs under the table” paints the picture of metahumans that are so intelligent that humans become relatively primitive— an inferior species that is incapable of fully contributing to cutting-edge scientific discoveries.
In this scenario, would the emergence of AI-driven intelligence and metahuman cognition mark a sort of evolutionary event and split not just between humans and metahumans, but between pre-metahuman humans and post-metahuman humans? Would the cultural and cognitive gap between those who integrate with metahumans and those who came before them be so different from the human species as we now know it, that they effectively become distinct species? Does integration of technology into our bodies for enhancement start the creation of a new species, the way that we see ourselves today different from paleolithic man? Given the current speed of AI advancement, is it even possible to stop/ delay this trajectory, or has the process already begun?"

[ypan02]

The article "AI Expands Scientists’ Impact but Contracts Science’s Focus" raises an important philosophical question: Where does scientific innovation truly come from? Is it from asking original questions and discovering new evidence? While AI tools can accelerate finding evidence and analyzing data, they are limited by being trained on existing human-generated data. These tools can’t generate genuinely original questions or uncover new data, as they are based on pre-existing information. Overreliance of researchers on AI tools could lead to a focus on solving existing problems rather than generating new ones, or on analyzing existing data rather than conducting original data collection experiments.

My question is: Is it possible to design AI tools that support original research by generating original questions and helping uncover new data, despite these tools being trained on existing data? If so, what would such a tool look like?

[siyakalra830]

Question on "AI Expands Scientists’ Impact but Contracts Science’s Focus"

AI tools are increasingly integrated into scientific research, enhancing individual productivity and career advancement. Scientists who adopt AI publish more papers, receive more citations, and advance to team leadership roles faster than non-adopters. However, this comes with the unintended consequence of a narrower focus on scientific topics and reduced engagement in follow-on research. Given these trends, how might science policy and AI tool design be modified to encourage exploration of less data-rich, novel research areas and promote broader scientific engagement? Specifically, how can we incentivize AI to catalyze new discoveries rather than simply accelerate progress in well-established domains?

[willowzhu]

Question based on readings for Tuesday *as posted on Canvas:

In Creative Destruction and Subjective Well-Being, Akcigit et al analyze the relationship between turnover-driven growth and subjective well being. The authors ask: does higher (per capita) GDP or GDP growth increase happiness? This paper is a first attempt at filling the gap between existing empirical literature on how various measures of subjective well-being (SWB) relate to income or income growth and what exactly drives the growth process, and at how the determinants of growth affect well-being. In the Schumpeterian model, SWB is captured by life satisfaction: the expected discounted valuation of an individual’s future earnings. The authors also use the Cantril Ladder of life to assess individual’s self-valuation of their current and future satisfaction for SWB. Do you think these measures sufficiently capture the nuances of well-being? What made the authors choose these two approaches to measuring well-being? If the Cantril Ladder helps us determine whether people are forward looking or not, and income helps us determine people’s satisfaction (if their happiness is correlated with how much they make), how do we, and do we need to, incorporate other elements to measuring happiness? If this paper is a first attempt at connecting the gap between well being/growth and the drivers of growth/well-being, what else needs to be studied in the next paper to continue filling this gap?

[michelleschukin]

Machine Learning as a Tool for Hypothesis Generation discusses how machine learning can aid in hypothesis generation by identifying patterns that humans might overlook. However, the authors acknowledge that this process is inherently constrained by human input. They state:

"The fact that a sizable share of subjects all name the same discovery suggests that human-algorithm communication has occurred and the procedure is replicable, rather than reflecting some unique spark of creativity. At the same time, the fact that our procedure is not fully automatic implies that it will be shaped and constrained by people."

This raises an important question: does this method truly enhance scientific discovery, or does it risk reinforcing existing biases by anchoring human hypothesis generation to algorithmic insights that are themselves shaped by human-curated data? If human creativity plays a crucial role in curating data and interpreting outputs, does this approach merely amplify existing human biases rather than uncovering novel insights? How might this reliance on human data shape applications in fields such as medicine or finance, and what safeguards must be in place to ensure meaningful, unbiased hypothesis generation?

[jesseli0]

"AI Expands Scientists’ Impact but Contracts Science’s Focus" covers an emerging issue of AI usage in scientific research, namely that while it seems to beneficial impacts to the careers of scientists, it's overall harming the quality of research that scientists output. The researchers advocate for refining AI tools to try and make sure that scientific research does not become too contracted. Given that AI is introducing a harmful positive feedback loop by rewarding researchers for using it and simultaneously narrowing the breadth of scientific output, is it advisable to regulate usage of AI in research until a model is released which can overcome this issue? Will there be long term negative impacts to letting AI usage in research continue in this manner?

[Adrianne-Li]

AI is increasingly used in scientific research, not just as a tool for accelerating discovery but as an agent shaping what kinds of questions get asked. If AI prioritizes problems where data is abundant and easily structured, could this bias scientific exploration away from more complex, less data-rich questions? How might this affect fields like theoretical physics, philosophy, or even exploratory social sciences? Should policymakers and funding agencies intervene to ensure that AI-driven research does not unintentionally narrow the breadth of human inquiry?

[siqi2001]

A Critique

This week’s readings are based on the scenario where computational tools can automate hypothesis generation processes. “Machine Learning as a Tool for Hypothesis Generation” sets the case into stone. “Catching crumbs from the table” meditates on the implications of machine science: what it means for scientists to shift from original research work into hermeneutics. And “AI Expands Scientists' Impact but Contracts Science's Focus” elaborates on one concern for the proliferation of AI-driven research.

I want to introduce a quote from a piece of philosophy I recently encountered, so we can reflect on them together. In an interview, Isabelle Stengers, a Belgian philosopher noted for her philosophy of science, says the following,

“Scalability for Tsing is the aim of any enterprise that aims at working definitions impervious to encounters, contingency, conjunctures, that deals with beings which “do not tell stories.” But to me this is also the very definition presupposed by the idea of an “advancement of knowledge”: the knowledge produced must keep its relevance whatever the circumstances in order to make a clean contrast between before and after. The privilege of scalable questions, able to escape the messiness of situations, has different consequences in the experimental sciences and in the social sciences, but everywhere it concurs to the devastation of the world. But this privilege also directly communicates with the ways sciences align their way of understanding with other institutions that also privilege scalability, namely the industry and the state."

According to Stengers, both people who subscribe to the "advancement of knowledge" and institutions (industry and state) want our answers to questions to be scalable such that they could apply to various scenarios without asking each of them to tell their stories. What we lose from our fixation on scalability is “a mode of thought which endeavors to activate what might be possible against the safety of probability.” This is one of Stengers’ criticisms against the knowledge economy. Importantly, here she doesn’t equate the knowledge economy to science, nor does she think science has to serve capitalism. Such a disassociation, for her, gives science a chance to reinvent itself, to create a narrative that disentangles it from its current impatient, anxious economy.

When we think about the relationship between machine and human, can we understand machine as reinforcing the knowledge economy? If so, can human researchers start to pay (more) attention to non-scalable, situated knowledge and knowledge that concerns what might be possible instead of the one that ensures the safety of probability?

[sijia23333]

The paper "Machine Learning as a Tool for Hypothesis Generation" proposes a novel semi-automated procedure that combines machine learning algorithms with human interpretation to discover patterns in data that might be missed by traditional research methods. In their judicial decision-making application, they found that two previously unidentified facial features - "well-groomed" and "heavy-faced" - influence judge detention decisions, even after controlling for other variables.

This approach seems promising, but I wonder about its ethical dimensions. Since the procedure identified physical appearance factors that may lead to discrimination, how should we balance the value of this knowledge against potential harms? If judges learn they're biased toward well-groomed defendants, they might overcorrect in their decisions. Or they might rationalize their biases. How could we ensure that insights from this methodology lead to more equitable outcomes rather than reinforcing existing biases? What ethical frameworks should guide the deployment of such hypothesis-generating tools in consequential domains like criminal justice?

[tHEMORAN02]

Catching crumbs from the table is radically different from the other readings we have had so far. It imagines a future where neo humans dominate all scientific discoveries and only publish their findings in a way incomprehensible to humans. I think the more likely parallel thing we will probably find in our future is that of ai agents or maybe some kind of thinking homunculus. I do think that the role of future scientist being finding areas that met humans overlook is a cool one though.
My question is pretty simple, do meta humans necessarily have a blind spot or is that just an invention by the author to give us hope

[dnlchen-uc]

Hao, Qianyue, Xu, Li, and Evans find that AI can be used to increase the frequency of research publications but may reduce diversity within research topics. This presents an interesting dilemma when evaluated using frameworks we learned in prior weeks. In accordance with assembly theory, would less "components" to build off of result in lower complexity innovations due to less interdisciplinary interactions and fewer novel ideas? Or would lowering the entry barrier to academia incentivize more diverse viewpoints even if a larger proportion of research is conducted in data-rich areas? Hence, should the adoption of AI in academia be viewed as a driver of democratization of research (similar to how automation can increase employment in fields such as aviation) or as a limiting factor for research diversity?

[chrislowzhengxi]

In "AI Expands Scientists’ Impact but Contracts Science’s Focus," the article points out that AI models work best with large, high-quality datasets, which is why fields like medicine and materials science—where there’s plenty of data—see more AI-driven research. But this also means AI tends to focus on existing problems rather than exploring new or underdeveloped areas. As a result, scientific research becomes more concentrated, covering fewer topics than traditional research methods.

Another issue is that AI-generated research seems to have lower engagement, meaning scientists aren’t building on each other’s AI-driven work as much. Instead of pushing the boundaries of knowledge, AI is often used to refine and optimize what's already known, leading to more of the same rather than big breakthroughs. At the same time, many researchers are chasing recognition and citations, which could make them less willing to take risks on completely new ideas.

Since AI is here to stay, how can we make sure it contributes to meaningful discoveries rather than just refining old ones? How can we use AI in a way that actually expands scientific exploration rather than narrowing it?

For example, tools like ChatGPT’s "Deep Research" function are designed to enhance research—but do they actually help uncover new insights, or just organize existing knowledge more efficiently? Could AI be designed to suggest new research directions or highlight gaps in current knowledge?