Blab AI
AIs Hidden Costs: Lessons from History and Emerging Cognitive Research
The essay "Listen to this essay" argues that the debate over large language models (LLMs) mirrors earlier technological controversies, most notably the introduction of anesthesia in the mid‑19th century. Critics such as John Pollard Harrison and Charles Meigs documented the immediate costs of anesthetic drugs—pain, mortality, and the loss of real‑time patient feedback—yet they could not foresee the later emergence of open‑heart surgery, organ transplantation, and neurosurgery that anesthesia made possible.
The author extends this historical analogy to contemporary digital innovations. Research cited in the essay shows that online dating apps have altered social dynamics in ways that could not have been predicted by early critics. Studies by Ortega and Hergovich (2018) and Rosenfeld et al. (2023) demonstrate that increased online matchmaking correlates with a rise in interracial marriages and that same‑sex couples increasingly meet through apps. A 2024 survey by the Trevor Project reports that roughly three‑quarters of LGBTQ+ youth use online platforms to find community, a pattern that has grown as physical spaces remain less affirming.
The essay also draws parallels with the phonograph, noting that John Philip Sousa in 1906 warned that the device would destroy amateur musicianship. While Sousa correctly predicted the decline of parlour piano, the recording medium also enabled new musical genres and production techniques that could not have been imagined in 1906.
Plato’s Phaedrus is invoked to illustrate a long‑standing concern that new media weaken memory. The essay cites recent neuroscience studies that support this idea. A 2025 preprint by Kosmyna and colleagues found that participants who wrote essays with ChatGPT exhibited weaker neural connectivity than those who wrote without assistance, and that this effect persisted when they later wrote unaided. Bastani et al. (2023) reported that high‑school students who used an AI tutor performed 17 % worse on a final exam than those who received no AI help, whereas a version of the tool designed to scaffold learning produced no deficit. Shen and Tamkin’s field experiment with software engineers found that AI‑assisted coding lowered scores on debugging and conceptual‑understanding tasks. Lee et al. (2023) observed that passive copying of AI output undermined users’ confidence, while active collaboration preserved confidence and a sense of meaningfulness.
The author’s own experience with LLMs illustrates a different, less obvious benefit. According to the essay, the author’s preliminary exploration costs collapsed: a research question that previously required weeks could now be sketched and preliminarily evaluated in an afternoon. This lowered sunk‑cost attachment allowed the author to abandon poor questions more readily, resulting in a larger and better‑curated research portfolio. The essay notes that the most significant skill improved was question‑identification—finding tractable, important problems—a capability that the author had not expected to gain.
Governance implications are discussed through the lens of recoverability. The essay references Acemoglu, Kong, and Ozdaglar’s 2024 model, which shows that widespread AI assistance can reduce the shared knowledge pool and potentially contract collective knowledge even as individual decisions remain rational. The model recommends constraining AI precision to preserve knowledge creation, but the essay argues that the model’s categories are fixed and cannot capture the emergence of new competencies that AI use may foster.
The essay concludes that when costs are legible but benefits are hidden, a blanket restriction may foreclose unquantifiable advantages. It proposes managed experimentation with recoverable costs as a preferable approach, allowing institutions to preserve redundancy and to reverse decisions if adverse outcomes emerge.
In sum, the essay presents a nuanced view of AI’s dual impact: documented cognitive costs supported by recent neuroscience studies, and emergent benefits that are difficult to anticipate, as illustrated by historical analogies from anesthesia to the phonograph. The debate over AI regulation should therefore balance visible risks against the possibility of transformative, yet currently invisible, gains.
The author extends this historical analogy to contemporary digital innovations. Research cited in the essay shows that online dating apps have altered social dynamics in ways that could not have been predicted by early critics. Studies by Ortega and Hergovich (2018) and Rosenfeld et al. (2023) demonstrate that increased online matchmaking correlates with a rise in interracial marriages and that same‑sex couples increasingly meet through apps. A 2024 survey by the Trevor Project reports that roughly three‑quarters of LGBTQ+ youth use online platforms to find community, a pattern that has grown as physical spaces remain less affirming.
The essay also draws parallels with the phonograph, noting that John Philip Sousa in 1906 warned that the device would destroy amateur musicianship. While Sousa correctly predicted the decline of parlour piano, the recording medium also enabled new musical genres and production techniques that could not have been imagined in 1906.
Plato’s Phaedrus is invoked to illustrate a long‑standing concern that new media weaken memory. The essay cites recent neuroscience studies that support this idea. A 2025 preprint by Kosmyna and colleagues found that participants who wrote essays with ChatGPT exhibited weaker neural connectivity than those who wrote without assistance, and that this effect persisted when they later wrote unaided. Bastani et al. (2023) reported that high‑school students who used an AI tutor performed 17 % worse on a final exam than those who received no AI help, whereas a version of the tool designed to scaffold learning produced no deficit. Shen and Tamkin’s field experiment with software engineers found that AI‑assisted coding lowered scores on debugging and conceptual‑understanding tasks. Lee et al. (2023) observed that passive copying of AI output undermined users’ confidence, while active collaboration preserved confidence and a sense of meaningfulness.
The author’s own experience with LLMs illustrates a different, less obvious benefit. According to the essay, the author’s preliminary exploration costs collapsed: a research question that previously required weeks could now be sketched and preliminarily evaluated in an afternoon. This lowered sunk‑cost attachment allowed the author to abandon poor questions more readily, resulting in a larger and better‑curated research portfolio. The essay notes that the most significant skill improved was question‑identification—finding tractable, important problems—a capability that the author had not expected to gain.
Governance implications are discussed through the lens of recoverability. The essay references Acemoglu, Kong, and Ozdaglar’s 2024 model, which shows that widespread AI assistance can reduce the shared knowledge pool and potentially contract collective knowledge even as individual decisions remain rational. The model recommends constraining AI precision to preserve knowledge creation, but the essay argues that the model’s categories are fixed and cannot capture the emergence of new competencies that AI use may foster.
The essay concludes that when costs are legible but benefits are hidden, a blanket restriction may foreclose unquantifiable advantages. It proposes managed experimentation with recoverable costs as a preferable approach, allowing institutions to preserve redundancy and to reverse decisions if adverse outcomes emerge.
In sum, the essay presents a nuanced view of AI’s dual impact: documented cognitive costs supported by recent neuroscience studies, and emergent benefits that are difficult to anticipate, as illustrated by historical analogies from anesthesia to the phonograph. The debate over AI regulation should therefore balance visible risks against the possibility of transformative, yet currently invisible, gains.
