A titular question mark is often trinitarian, signaling explorations, probing reactions, and enticing contrarians. Here, it also puns a dual trinity: What is ‘intelligence’? Is ‘Homo sapiens intelligence’ a thing? How do HI (human intelligence), AI (artificial intelligence), and HSI (Homo sapiens intelligence) interact, assuming each 'I' is as certain as 'I think; therefore, I am'?
Yes, one can discover answers—intentionally plural—to each of these questions within this issue of HDSR. Some answers are explicitly stated. Others need assembly. Yet others may challenge some readers' HI, whether aided by AI or not.
This issue features three conversations regarding ChatGPT and generative AI. For two of them, Liberty Vittert (HDSR’s Media Feature Editor) and I discussed the questions of intelligence with two of the most prominent and versatile scholars: Professor Andrew W. Lo, Director of the MIT Laboratory for Financial Engineering; and Professor Steven Pinker, my colleague at Harvard’s Department of Psychology. Both have been named in TIME’s The World’s 100 Most Influential People (Pinker in 2004 and Lo in 2012), among many other accolades each has received. Their answers are remarkably consistent in essences, which perhaps should not come as a surprise given their extraordinary intellectual breadth and depth, exemplified by their in-depth studies of human minds and behaviors.
As Lo states in his interview, “for me and for the mathematical models that I developed, intelligence is really captured as an adaptation that provides an advantage for survival” (Lo et al., 2023). Pinker’s response to the same prompt in his interview, “Intelligence is the ability to use knowledge to attain goals,” came with American philosopher and psychologist William James’ humorous illustration of intelligence: the difference between Romeo’s pursuit of Juliet and iron filings’ pursuit of a magnet is that Romeo would find ways to jump over a barrier (e.g., a wall) to touch Juliet’ lips, while the filings would stick to a barrier (e.g., a card) when it is placed between them and the magnet (Pinker et al., 2023). In other words, whether the goal is to live or to love, intelligence is the ability to adapt to achieve the goal, to which we may add that the effectiveness of the adaptation indicates the degrees of intelligence.
Furthermore, neither answer implies that intelligence is unique to humans—a fundamental premise for our contemplation of AI. As Lo emphasized, if we accept his definition of intelligence, “then I would argue that AI is already here in terms of being able to display a certain degree of intelligence. As it evolves, as AI becomes more sophisticated, it will at some point achieve the same or greater level of intelligence as humans” (Lo et al., 2023). Pinker explicated it even more when responding to my naively posed rhetorical question that if his answer would imply that intelligence is not unique to humans, “clearly it can’t be restricted to humans, otherwise it would just be a kind of chauvinist description of one of our traits, and then the concept of artificial intelligence would be an oxymoron” (Pinker et al., 2023).
I managed somehow to bring up the Turing test in a subsequent question concerning a report from Microsoft Research (Bubeck et al. 2023) on testing the general intelligence of GPT-4 in our interview with Pinker. Pinker reminded me that the Turing test has sometimes been considered Alan Turing’s worst idea because “It’s a test of how easily you can fool people, and the answer is ‘easily’” (Pinker et al, 2023). I couldn’t help but laugh, not just because of the humorous answer but also at myself for continually displaying my naivety. Evidently, I’m not very adept (and adapt) at contemplating intelligence, but I will press on. This unprompted self-reflection of our differing levels of intelligence is nevertheless a sign of my HI and self-awareness, one we have yet to detect in AI. More importantly, variations in individual HIs and their resulting life imprints appear to be essential for the impressive performance of generative AI.
Both Lo and Pinker have expressed their pleasant surprises about ChatGPT’s capabilities, a sentiment many of us have echoed at one point or another, whether we consider such capabilities as a sign of intelligence or not. In Pinker’s words, “the big surprise is—and I’ve got to confess that it surprised me—is how much intelligence is sitting implicitly in databases of language if they’re big enough” (Pinker et al., 2023). This brings forth two primary sources of astonishment: an unrecognized form of intelligence and the GPT-like technology that unveils it. The 'implicit intelligence' that astounded Pinker isn't artificial in nature. Any language database is, in essence, a snapshot of human expressions and imprints in the cosmos. It is also more encompassing than the individual human intellect we typically contrast AI with, as seen when comparing AI, like AlphaGo, to world champions.
This concept also seems to transcend the idea of collective intelligence, epitomized by the phrase ‘wisdom of crowds,’ which alludes to improved competencies arising from interactions within a group (Malone & Bernstein, 2022). Collective intelligence isn't static, that is, it does not “sit” there to be discovered, nor does it require a “big enough” group to manifest (to borrow terms from Pinker's quote in the preceding paragraph). In fact, the wisdom of crowds often wanes when the crowd grows too large—just imagine the task of getting 1,000 experts to agree on anything, let alone taking any action.
Historical philosophical discourses have consistently explored and debated the idea of a singular, universal intellect. Concepts like Aristotle’s ‘maker intellect,’ Plotinus’s ‘divine intellect,’ Al-Farabi’s and Avicenna’s ‘agent intellect,’ and Averroes’s ‘single intellect’ come to mind. Most of these theories propose superhuman forms of intellect, orchestrating or enlightening the knowledge tiers beneath, akin to those owned by individual humans. The 12th century Andalusia philosopher and polymath Averroes’s single intellect is a much-criticized or even ridiculed departure, for it argues that all human beings share a ‘single mind,’ a collective and transcendent entity. Individual intellects, in his view, merely participate in it, granting humans access to universal truths and knowledge.
I, like many, had thought Averroes’s ‘single mind’ (theory) was most “insane” (Adamson, 2016, Chapter 16). However, the advent of ChatGPT and similar models prompted a reconsideration. This ‘implicit intelligence’ that Pinker and many of us find astonishing seems more akin to a universal intellect than to mere collective wisdom. Whether we actively contribute or simply remain unaware, virtually all of us participate through our individual data, in one form or another, in shaping ChatGPT’s training and, consequently, its abilities. The collective astonishment we feel might imply that this capability exceeds our individual understanding, or perhaps we've just underestimated human intellect’s scope. Either way, it alludes to a higher form—in a platonic sense—of human intelligence that stands above individual human intellect.
Since the term ‘Homo sapiens intelligence’ (HSI) apparently is still unattached (at least according to GPT-4 as of this writing), it might not be an insane idea to adopt it for this higher form of intellect, resonating with Averroes’s ‘single intellect.’ Both concepts rely on individual participation and eschew superhuman notions. The main distinction lies in accessibility: Averroes’ single intellect is accessed via a philosophical imagery, while HSI is tangible, brought to life by large language models.
This tangibility is often credited as bringing us ‘artificial intelligence,’ which is an unfortunate adoption of the term, because it suggests that the intelligence itself is artificially generated rather than artificially accessed. (The term AI could still be saved, as an acronym for ‘accessed intelligence.’) After a vivid portrait of very-high-order statistical patterns via a six-folder correlation, Pinker elucidated this distinction: “The human mind can’t wrap itself around what those very-high-order statistical patterns are, and what the large language models tell us is that implicit in those statistics, there is a lot of knowledge and indeed potential intelligence” (Pinker et al., 2023). Our newfound capability to tap into HSI, in its varied forms, undoubtedly enriches us both individually and collectively.
An insightful interview in this issue with Dr. Seema Iyer, Senior Director of The Hive at USA for UNCHR (the United Nations Refugee Agency), offers explicit instances of how accessing such intelligence aids in expediting legal processes related to refugee statuses and enhances communication and awareness about the refugee crisis (Iyer et al., 2023).
As another example, opinion polling is an endeavor to gauge our collective beliefs and preferences, which are otherwise inaccessible to us as individuals. The once-venerated probabilistic sampling paradigm faces growing challenges, leading Bailey (2023) to remark, “random sampling is, for all practical purposes, dead.” (However, some discussions counter this stance—such as Dempsey (2023), Little (2023), Lohr (2023), and Pescott (2023)—reinforcing the value of evaluating diverse collective opinions.) Given these challenges, it is only natural to speculate if sophisticated large language models (LLMs) might provide enhanced access to such collective insights. The article by Nathan Sanders, Alex Ulinich, and Bruce Schneier (2023) delves into this exact topic, noting that, “Because these LLMs are trained on huge corpora of writing by diverse people captured from across the Internet, they are potentially capable of representing a wide range of beliefs on many policy issues.” The prompt-engineering method they developed is yet another example of the increasing need for human and machine interaction, or more precisely the interaction between HI and HSI enabled by AI.
Yet, as Sanders et al. (2023) indicate, we're merely at the dawn of understanding these interactions. In his interview, Lo exercised the same caution: “At this point, when it comes to financial advice, I would be extremely cautious about trusting advice from ChatGPT” (Lo et al., 2023). Despite this, he remains hopeful, foreseeing a future within our lifetimes where AI tools will play pivotal roles in critical human decision-making. I share his optimism, sensing vast reservoirs of HSI yet to be uncovered, which could accelerate advancements in what we traditionally term artificial general intelligence (AGI) and vice versa. Our individual intelligence is also likely evolving because of the need to adapt to new kinds of intellectual intercourse, such as prompt engineering—skillfully manipulating a Chatbot to get what we want surely is a perfect illustration of the description of intelligence, as stated earlier. In other words, the pursuit of decision-making AGI is becoming more promising because of the increased interactions and mutual reinforcements of multiple forms of intelligence. In essence, the quest for decision-centric AGI looks increasingly feasible due to the growing interplay and mutual enhancement of various intelligence forms.
A part of human intelligence is our ability to make independent judgments, and to be skeptical when information or evidence is unclear. The ever-increased calls for a host of requirements for machine learning algorithms, such as fairness, explainability, interpretability, and transparency, are mostly driven by the black-box nature of these algorithms, inducing our skepticism or even fear. Few of us would feel comfortable to have our high-stake—such as financial and medical—decisions be made by a person or an algorithm that we do not have some degree of trust in (granted, it could be a blind trust). The nuanced findings by Sanders et al. (2023) about the reliability of using chatbots to gather political opinions remind us that the issue of trustworthy AI and data science should always be at the core of our pursuit of enhanced human intelligence and capability.
The two “crisis” articles in this this issue also remind us that the concerns for untrustworthy AI or data science are widespread, but so are the efforts to address them. The article by Stephen Ruberg, Sandeep Menon, and Charmaine Demanuele (2023) discuss serious concerns with clinical predictive algorithms (CPAs) for diagnoses and prognoses, including the disturbing finding that “a recent review of 232 diagnostic or prognostic algorithms developed in the rush of the COVID-19 pandemic deemed none of them fit for clinical use.” Ruberg et al. propose a more holistic approach for optimizing the development of CPAs, and a fit-for-purpose, sequential clinical trial approach for evaluating CPAs, thereby helping to resolve the credibility crisis of CPAs. “With these two recommendations met,” the authors hope, “the benefits and risks of a CPA will be clearer, resulting in greater credibility, transparency, and utility” (Ruberg et al., 2023).
HDSR’s new column on Reinforcing Reproducibility and Replicability features an article by Kim A. Weeden in this issue that brings us to the field of sociology, which apparently has been slow, compared to other social science disciplines, “in its recognition of a replication crisis, adoption of scientific transparency practices, and participation in the large and interdisciplinary literature on reproducibility” (2023). The existing efforts have been mostly ‘bottom up’ and piecemeal in nature, without systematical effort by or even endorsement from leading professional societies or top journals. Weeden attributes this state of the fair to the fragmentation of sociology, which straddles between science and humanity, and with both quantitative and qualitative researchers, who may have very different views towards practices such as data sharing because of different cultures and constraints.
For example, qualitative studies may rely on “nonverbal cues from subjects and the embedded experiences of researchers that cannot be captured in transcripts, field notes, or other data products.” Consequently, “Sharing qualitative data may thus not only lead to ‘failed’ replications, but to misinterpretation of data if they are taken out of context” (Weeden, 2023). Reading Weeden’s column certainly makes me appreciate more the complexity of ensuring trust in data science. For example, Weeden raises the issue of trust from a different angle: “will data sharing undermine the necessary trust that develops between researchers and subjects?” Indeed, without the trust from the data subjects, we will face a vital issue for all data science, that is, the lack of reliable data due to the refusal from the subjects and the selective biases resulting from trustworthiness-induced response behaviors.
This very issue is rather prevalent, especially for data collected by government agencies. Without the public’s trust of the privacy of their data, collecting reliable data would be a very daunting and costly task, even if it is possible at all. This is a key reason that the U.S. Census Bureau embarks on the journey of implementing differential privacy (DP) protection for its 2020 Census, a complex task since DP calls for injecting noise into data before releasing, a process itself will reduce the data utility as a trade-off for enhanced privacy protection. The reduced utility can also erode the trust in data science findings, from the data users and the public alike. The article by Cory McCartan, Tyler Simko, and Kosuke Imai (2023), “Making Differential Privacy Work for Census Data Users,” continues the investigations and debates about these complex trade-off issues as documented in HDSR’s special issue on Differential Privacy for the 2020 U.S. Census: Can We Make Data Both Private and Useful? McCartan et al.’s (2023) article reports the authors’ effort and difficulties in utilizing the DP protected census data and makes recommendations to U.S. Census Bureau for better future disseminations of data with DP protection, with the recommended changes to be “essential for ensuring transparency of privacy measures and reproducibility of scientific research built on census data.”
The importance of data sharing, data transparency, and data accessibility is also the main emphasis of the Effective Policy Learning column article, “Continuing Implementation of the Foundations for Evidence-Based Policymaking Act of 2018: Who Is Using the Data?” by Nancy Potok (2023), the co-editor for the column. After a succinct overview of the key provisions and goals of the Evidence Act, the article focuses on meeting a key requirement by the act: “that agencies engage the public in using public data assets of the agency and encourage collaboration by publishing on the website of the agency on a regular basis (but not less than annually) information on the usage of such assets by non-Government users” (Potok, 2023). The article discusses the importance of collecting and publishing such usage data with respect to seven aspects, from public engagement and transparency to accountability and oversight, with the overarching aims to ensuring the act’s goals of “enhancing evidence-based policymaking, improving government efficiency, and increasing transparency while safeguarding privacy and security” (Potok, 2023).
Like all articles listed in this section, achieving desirable goals takes much effort and ingenuity. In that regarding, Potok’s article serves as an introduction to the Democratizing Data project, as well as to a forthcoming HDSR special issue that will take an in-depth look at search and discovery tools provided by this project and how they facilitate evidence building, as well as current use cases and potential future innovations. Together, the four articles highlighted in this section (Potok, 2023; Ruberg et al., 2023; Sanders et al., 2023; and Weeden, 2023) provide a snapshot of how academia, government, and industry make sector-wise as well as collective efforts to ensure the trustworthiness of data, data science, and ultimately the findings and policies derived from them.
A fifth article, “Understanding and Fostering Regional AI Ecosystems: A Case Study in Maine” by Hamit Hamutcu, Usama Fayyad, and Michael Pollastri (2023) in this issue of HDSR further demonstrates the combined power when academia, government, and industry work together to tackle a wide-ranging set of issues, from growing AI talent pipeline (academia) to addressing regulatory challenges (government) and developing predictive tools (industry). That is HIs and HSI in full action, which in turn would supply better and richer training data for GPT-n or whatever the next AI surprise might be.
However, no matter how advanced or surprising future AI technologies might be, their full potential and safe use will require human interaction and intervention. Preparing future digital-age citizens and talents is therefore paramount to the healthy evolution of the AI enterprise and data science ecosystem (Meng, 2019). The earlier such a preparation, the more effective it would be.
The Minding the Future column of HDSR is devoted to pre-college education. To get a snapshot of how high school students and teachers react to the arrival of ChatGPT, the column’s editor, Nicole Lazar, recruited three teachers and seven students from her local high school, and reported the findings in “Perils and Opportunities of ChatGPT: A High School Perspective” (Lazar et al., 2023). Because ChatGPT has been the talk of the town, or rather the world, perhaps no findings would truly surprise most of us. Nevertheless, I never thought about ChatGPT as a “rubber duck debugger” and I smiled when I read the same emphasis by the teachers and students that ChatGPT, as a technology, is neither good nor bad. Such seemingly obvious emphases are at the core of preparing our young minds to develop intellectual mutuality for appreciating and navigating the increasingly more nuanced world, where making informed judgment and considered decisions and taking responsibility of ones’ actions is essential for world peace and for sustaining the Homo sapiens.
My smile widened as I continued in issue 5.4 with a machine learning–based literature review conducted by Koby Mike, Benny Kimelfeld, and Orit Hazzan (2023), studying “The Birth of a New Discipline: Data Science Education.” The study clustered over 1,000 articles into a framework of 26 clusters, which are further categorized into five superclusters: (a) curriculum, (b) pedagogy, (c) STEM skills, (d) domain adaptation, and (e) social aspects. Collectively, they display how HSI acts upon itself for enhancement via perhaps the most effective means we humans ever created for the enhancement—that is, education—and how AI helps to reveal and provide access to this enhancement. The first two superclusters cover broadly what to teach and how to teach, and the last three address the ecosystematic need of data science: substantive applications and technical skills; and societal impact, interactions, and interventions, respectively. Studies as such are never complete, especially with the rapidly evolving nature of data science and AI. Nevertheless, there is so much to learn from and reflect upon this one study, and indeed it would take an even longer editorial to document my thoughts prompted by reading it.
But I must end this editorial in order to catch the last ship to the digital platform for launching HDSR’s 21st issue, an auspicious number to signal HDSR’s transition into adulthood. At the risk of appearing to transition from statistician to numerologist, I will end with a quote by Dr. Seema Iyer in her interview (Iyer et al., 2023), where ‘21st’ is also a nod to maturity. She speaks of mutual data science skills—skills challenging to impart, because they demand instructors themselves to possess sufficient experience and appreciation. That is precisely why we need to tape into our HSI and learn more from each other. Only then can we collectively elevate both the maturity of data science and data science education.
“Those are the kind of skills that I think of as a 21st century skill: to take, again, a very broad view of what data might even look like, and creating some algorithms and creating some rules and assumptions so that you can convert that information that's just coming at us into some type of sensemaking.” (Iyer et al., 2023)
Xiao-Li Meng has no financial or non-financial disclosures to share for this editorial.
Adamson, P. (2016). Philosophy in the Islamic World: A history of philosophy without any gaps. Oxford University Press.
Bailey, M. A. (2023). A new paradigm for polling. Harvard Data Science Review, 5(3). https://doi.org/10.1162/99608f92.9898eede
Dempsey, W. (2023). Paradigm lost? Paradigm regained? Comment on “A New Paradigm for Polling” by Michael A. Bailey. Harvard Data Science Review, 5(3). https://doi.org/10.1162/99608f92.49b9b14a
Hamutcu, H., Fayyad, U., & Pollastri, M. (2023). Understanding and fostering regional artificial intelligence ecosystems: A case study in Maine. Harvard Data Science Review, 5(4). https://doi.org/10.1162/99608f92.dc2a1d1b
Iyer, S., Meng, X.-L., & Vittert, L. (2023). Close to refuge: Integrating AI and human insights for intervention and prevention: A conversation with Seema Iyer. Harvard Data Science Review, 5(4). https://doi.org/10.1162/99608f92.1a58d824
Lazar, N., Byrns, J., Crowe, D., McGinty, M., Abraham, A., Guo, M., Mann, M., Narayanan, P., Roberts, L., Sidore, B., & Wager, M. (2023). Perils and opportunities of ChatGPT: A high school perspective. Harvard Data Science Review, 5(4). https://doi.org/10.1162/99608f92.9f0adc39
Little, R. J. (2023). The “Law of Large Populations” does not herald a paradigm shift in survey sampling. Harvard Data Science Review, 5(3). https://doi.org/10.1162/99608f92.6b049957
Lo, A. W., Meng, X.-L., & Vittert, L. (2023). From financial markets to ChatGPT: A conversation with Andrew Lo. Harvard Data Science Review, 5(4). https://doi.org/10.1162/99608f92.49328257
Lohr, S. L. (2023). Assuming a nonresponse model does not make it true. Harvard Data Science Review, 5(3). https://doi.org/10.1162/99608f92.2b901b7f
Malone, T. W., & Bernstein, M. S. (Eds.). (2022). Handbook of collective intelligence. The MIT Press.
Meng, X.-L. (2019). Data science: An artificial ecosystem. Harvard Data Science Review, 1(1). https://doi.org/10.1162/99608f92.ba20f892
Mike, K., Kimelfeld, B., & Hazzan, O. (2023). The birth of a new discipline: Data science education. Harvard Data Science Review, 5(4). https://doi.org/10.1162/99608f92.280afe66
Pescott, O. L. (2023). Seek a paradigm and distrust it? Statistical arguments and the representation of uncertainty. Harvard Data Science Review, 5(3). https://doi.org/10.1162/99608f92.a02188d0
Pinker, S., Meng, X.-L., & Vittert, L. (2023). The intelligence and rationality of AI and humans: A conversation with Steven Pinker. Harvard Data Science Review, 5(4). https://doi.org/10.1162/99608f92.c37d3572
Potok, N. (2023). Continuing implementation of the Foundations for Evidence-Based Policymaking Act of 2018: Who is using the data? Harvard Data Science Review, 5(4). https://doi.org/10.1162/99608f92.c0219d0e
Ruberg, S., Menon, S., & Demanuele, C. (2023). Resolving the credibility crisis: Recommendations for improving predictive algorithms for clinical utility. Harvard Data Science Review, 5(4). https://doi.org/10.1162/99608f92.c1292c54
Sanders, N. E., Ulinich, A., & Schneier, B. (2023). Demonstrations of the potential of AI-based political issue polling. Harvard Data Science Review, 5(4). https://doi.org/10.1162/99608f92.1d3cf75d
Weeden, K. A. (2023). Crisis? What crisis? Sociology’s slow progress toward scientific transparency. Harvard Data Science Review, 5(4). https://doi.org/10.1162/99608f92.151c41e3
©2023 Xiao-Li Meng. This editorial is licensed under a Creative Commons Attribution (CC BY 4.0) International license, except where otherwise indicated with respect to particular material included in the editorial.