Deep learning IRR by MURARI AMBATI

Research question

ma91541 — 2025-09-26 16:03:14 UTC

To what extent will the use of deep learning negatively impact human development in society?

Lens

ma91541 — 2025-09-26 16:05:22 UTC

Scientific

Introduction & Context

ma91541 — 2025-09-26 16:05:49 UTC

BP #1: Deep learning's complexity makes transparency and interpretability difficult.

ma91541 — 2025-10-01 02:56:45 UTC

BP #2: Deep learning reduces human oversight in high-stakes decision making.

ma91541 — 2025-10-01 02:56:56 UTC

BP #3: Deep learning creates societal uncontrollable technologies.

ma91541 — 2025-10-01 02:59:25 UTC

Thesis

ma91541 — 2025-10-01 03:33:49 UTC

From a scientific lens, the use of deep learning will negatively impact human development in society to a significant extent because the increasing complexity and optimization of these models creates systems that surpass human interpretability, thereby limiting transparency, reducing human oversight, and risking dependence on technologies that are not fully understood or controlled by users.

Interpretability

ma91541 — 2025-10-01 03:40:17 UTC

"Although these machines can surpass humans in terms of data processing, they lack accountability and interpretability and hence are referred to as 'black boxes'" (Xiong et al. 2023)

"Adverasial attacks...demonstrate that deep models can make unpredictable decisions and that interpretability remains a challenge." (Szegedy et al. 2014; Goodfellow et al., 2025; Su et al., 2019)" <-- supporting evidence/backup.

Interpretability (Additional)

ma91541 — 2025-10-01 03:44:29 UTC

Comments on "Deep learning model-assisted detection of kidney stones on computed tomography" with "accuracy of this automated model" is not "sufficient" which he highlights that "promising technology" is Deep learning.

Complexity

ma91541 — 2025-10-01 03:46:55 UTC

Wachter et al. (2017) on Counterfactual Explanations (4K Citations) stating that "explaining the functionality of complex algorithmic decision-making systems and their rationale in specific cases is a technically challenging problem." Backs the idea that deep learning models are so complex that specialized methods are required to interpret their decisions.

Proposed Solution (Failed)

ma91541 — 2025-10-01 03:48:54 UTC

"Towards A Rigorous Science of Interpretable Machine Learning" by Velez and Kim (7K Citations) showcases that deep models (current at the time) failed to meet even the minimal standards of transparency. Thus, reinforces that deep learning inherently reduces interpretability because of the structural complexity.

AI Oversight & Human Mistakes

ma91541 — 2025-10-01 03:54:13 UTC

From the article "AI Oversight & Human Mistakes: Evidence from Centre Court" that umpires increased "the rate at which they called balls in, producing a shift from making Type II errors to Type I errors." Demonstrates real-world scenario which showcases reduced human oversight as the umpires relied on Hawk-Eye to make calls, which means they did not independently verify each decision. The shift from Type II to Type I errors shows that humans adjust their behavior on AI outputs which introduces new errors.

Effects of machine learning errors on human decision-making

ma91541 — 2025-10-01 03:57:16 UTC

By Matzen and team, deliberates that the "results emphasize the importance of considering human cognition" showcasing that "accuracy led to increasing participant accuracy and decreasing RT" which highlights that ML outputs can reduce human vigilance and oversight. It points to over-reliance on AI for they assumed it was correct which mirrors real-world situations where they could over-trust AI in critical contexts.

Uncovering the dark side of AI-based decision-making

ma91541 — 2025-10-01 03:59:54 UTC

Written by Papagiannidis and team, relevant towards the thesis because pulls on how Ai trading bots affect relationship b/w traders and AI developers. With relevance to how AI fails "to predict accurately...energy prices will rise." Human oversight is required in these stakes, which shows that AI can replace aspects, but not fully take over human oversight in the moment.

Navigating artificial general intelligence development

ma91541 — 2025-10-01 04:05:01 UTC

Written by Raman and team, the article (Nature) highlights the necessity to align AGI with responsible development to prevent risks. AGI "need[s] a pluralistic AGI framework" which to rid of "uncontrollable emergence[s]." Not deep learning specific but focuses on broader AGI which showcases that AI in general is a problem.

Model-Reuse Attacks on Deep Learning Systems

ma91541 — 2025-10-01 04:07:44 UTC

Proposed by Yujie Ji and team, where he mentions that “Many of today’s machine learning (ML) systems are built by reusing an array of, often pre-trained, primitive models…most of such models are contributed and maintained by untrusted sources.” Relevant in the context because it shows that society relies on components outside human control which introduces systemic risks.

Uncertainty Quantification-Based Robust Deep Learning for Power System Applications

ma91541 — 2025-10-01 04:09:30 UTC

Proposed by Liang and team, highlights that “Existing deep learning models are lack of robustness under distribution shift scenarios, which greatly limits their real-world application.” Highlights that deep learning systems CAN behave unpredictably when conditions change which makes them partially uncontrollable in society. “The experimental result demonstrates that the deep ensemble (DE) model…outperformed the other methods, followed by Bayesian neural network (BNN).” Highlights that deep learning models' reliability is not guaranteed and depends heavily on architecture choice.

Survey on Explainable AI

ma91541 — 2025-10-01 04:13:17 UTC

Abusitta highlights that through a comprehensive survey that the success of "most AI-powered applications" are unable to "explain decisions." Relative to explainability needing to be present for a trustworthy system so individuals understand the decision-making, and so that other corporations and figures endorse it for our future growth.

s

ma91541 — 2026-01-13 02:15:27 UTC