Supplementary Material

regarding the article: “Can Artificial Intelligence Destroy Future Industry?”

Terziyan, V., Gryshko, S., Kaikova, O., & Golovianko, M. (2025, submission #7767). Can Artificial Intelligence Destroy Future Industry?. Procedia Computer Science. Elsevier.

1. Presentation of the study, concepts, assumptions, and conclusions (online):

https://ai.it.jyu.fi/ISM-2025-Provocative.pptx

2. AI-generated summary of our study as a podcast (online):

https://ai.it.jyu.fi/ISM-2025-Provocative.wav

3. Some Additional Notes on Empirical Evidence

Recent empirical research converges on a key mechanism underpinning this paper’s argument: AI-enabled tools can deliver rapid, measurable gains in task performance, yet simultaneously promote cognitive offloading and automation bias. When these patterns become widespread and sustained, they plausibly erode human analytic capacities and diminish the effective role of humans in industrial decision-making.

The cognitive basis for this trade-off is well established. Early laboratory work in cognitive psychology demonstrated that reliable external access to information changes memory strategy: people store less factual content internally and instead remember where to find it—a phenomenon labelled the “Google effect” or digital amnesia (Sparrow et al., 2011). This provides a parsimonious mechanism for how easy access to answers reshapes information habits.

Contemporary large-scale surveys and regression analyses extend these findings to AI-era tools: high-frequency AI use is robustly correlated with lower performance on standardized critical-thinking assessments, with cognitive offloading emerging as a mediating factor (Gerlich, 2025). These population-level studies situate the original lab effects in real-world contexts, showing that the phenomenon is neither anecdotal nor confined to web searching.

At the same time, controlled experiments in education and training consistently report that generative AI (LLMs) can produce substantial short-term gains in writing, problem-solving, and learner satisfaction, particularly under structured pedagogical scaffolds. A recent meta-analysis of 51 experimental and quasi-experimental studies (Nov 2022–Feb 2025) found a large, pooled effect on learning performance () and moderate effects on learning perception and higher-order thinking, with strong moderators such as course type, AI role, duration, and instructional framing (Wang & Fan, 2025). These results clarify an apparent paradox: AI often boosts measurable performance without necessarily strengthening (and in some contexts, potentially weakening) the independent cognitive processes that underpin creative reasoning (Zhang & Liu, 2025).

Recent literature reinforces the concern that the widespread integration of AI into critical decision-making processes is not only reshaping workplace dynamics but also altering fundamental human cognitive capacities. A particularly relevant concept is intellectual perishability, defined as the gradual decline of essential analytical, creative, and interpersonal skills due to sustained reliance on AI tools (Rosca & Stancu, 2025). Empirical observations presented in three independent case studies illustrate this phenomenon: (1) university students progressively losing the ability to write without assistance from LLMs, (2) a marked inability among certain cohorts to locate information when only descriptive rather than keyword-based search terms are available, and (3) the measurable reduction of human-led roles as AI systems replace cognitive and communication tasks. This aligns with the patterns observed in our synthesis of experimental and industry-level data, where dependency on AI correlates with a measurable drop in problem-solving independence, creativity metrics, and situational adaptability. Together, these findings provide converging evidence that without targeted changes in higher education curricula (particularly in engineering and management disciplines) the workforce of future Industry 6.0 may face a persistent deficit in skills that were previously considered foundational for leadership and innovation.

A second stream of evidence addresses automation bias: the systematic tendency to defer to algorithmic suggestions, sometimes overriding correct human judgments in favor of incorrect machine outputs (Goddard et al., 2012). Systematic reviews and domain-specific studies (healthcare, radiology, hiring, public administration) show this effect is robust and difficult to mitigate. In controlled mammography experiments, for example, intentionally incorrect AI advice sharply reduced diagnostic accuracy, even among experienced radiologists (Dratsch et al., 2023). Mediators include trust in automation, workload, task complexity, and training level (Kücking et al., 2024).

Finally, macro-level employer data quantifies the speed and scale of AI adoption. The World Economic Forum’s Future of Jobs Report 2025 finds that AI and information processing are seen by 86% of surveyed employers as transformative technologies for 2025–2030. Many intend to recruit AI-skilled workers and, accordingly, re-orient business models, accelerating the integration of AI into decision workflows (WEF, 2025-a). This creates strong economic and organizational pressure to adopt AI as the default in many roles, narrowing the adaptation window for education, regulation, and professional practice.

Bringing these strands together, the empirical literature yields a coherent picture:

· Performance gains ‒ LLMs and other AI tools can meaningfully enhance short-term output quality when deliberately integrated into workflows.

· Cognitive trade-offs ‒ Widespread AI reliance fosters cognitive offloading (digital amnesia) and, in some contexts, lower neural engagement in problem-solving (Kosmyna et al., 2025).

· Decision-risk mechanisms ‒ Automation bias is well-documented across sectors and can degrade expert judgment unless actively mitigated.

· Adoption pressures ‒ Employer intentions and market forces favor rapid integration of AI into high-value decision pipelines.

These mechanisms do not prove inevitable human cognitive decline, but they provide parsimonious, empirically grounded reasons to treat such a trajectory as plausible and urgent to address, especially through industrial design and higher-education strategies that preserve and strengthen core human competencies.

This article rests on a series of interrelated assumptions, the logical framework of which can be outlined as follows:

· AI cognitive performance is developing at a pace that increasingly outstrips the cognitive skill growth of humans.

· Human reliance on AI often occurs at the expense of maintaining or developing personal cognitive abilities.

· This reliance is observable not only during the educational phases of human development but also in professional and industrial contexts.

· The autonomy of highly capable AI is not only essential for enabling its own self-improvement but also for restricting potential misuse by authoritarian or totalitarian regimes.

· Given growing human dependence on AI, such autonomous systems may not necessarily steer future industry in alignment with human values, raising the potential for significant societal consequences of the type forecasted in this paper.

While the empirical studies referenced throughout this work, including those highlighted in this section, cannot be taken as definitive proof of the entire logical framework (since only longitudinal observation will ultimately validate or refute these claims), they offer substantial indirect support. Current experimental findings and analytical conclusions either directly align with these assumptions or, at minimum, fail to contradict them in any meaningful way.

Consequently, there is a justified basis for treating this logical progression as a credible warning regarding plausible future scenarios with a high likelihood of occurrence. The present evidence, though not conclusive, provides a sufficient foundation for serious consideration and proactive policy, research, and design responses to mitigate risks.

4. Visualized and Justified Summary of Empirical Evidence

This study confirms the validity of our concerns regarding the implications of increasing reliance on AI in hybrid industrial cyber-physical-social systems. The findings reveal a clear trajectory of human cognitive degradation, with a widening gap between human and AI capabilities. Addressing the first research question, i.e., “How will the human vs AI disparity impact the balance of power in hybrid systems, and will humans become obsolete in industrial decision-making?”, the answer is troubling. The evidence suggests that this process is already underway: human cognitive skills are declining, and the gap between human and AI performance is growing rapidly. This leads to a profound dilemma shown in Fig. 1 with no easy resolution. On one hand, restricting AI capabilities would hinder our ability to address the increasingly complex challenges faced by modern industries. On the other hand, granting AI full autonomy risks enabling AI to evolve so rapidly that human intelligence will no longer be necessary. This could result in industries diverging in directions orthogonal to human values and interests, potentially undermining the very foundation of a human-centered society.

Fig. 1. Visualizing the dilemma of AI autonomy and human cognitive evolution: The plot on the left illustrates the scenario where humans retain full control over AI by imposing restrictions on its capabilities. In this case, humans must continuously self-develop to actively address evolving challenges. However, with a limited AI, this self-development may prove insufficient to meet the complexity of these challenges. The plot on the right depicts the opposite scenario, where AI is granted full autonomy, rapidly evolves into a collective superintelligence, and eventually becomes capable of addressing all challenges independently of human input. In this case, humans, having delegated all cognitive tasks to the more powerful AI, risk continuous cognitive degradation. Both scenarios highlight negative outcomes, underscoring the critical need for a balanced approach to AI development and human-AI collaboration. We named the trends here (Ch 1, AI 1, Hu 1, Ch 2, AI 2, Hu 2) differently than they are in the main article, aiming to refer to the corresponding tables, which summarize empirical evidence regarding each trend.

One may assume that, while the paper explores the provocative thesis that AI’s accelerating dominance will destabilize industrial ecosystems, leading to the potential erosion of human agency and control, the authors are biased towards such a provocative thesis rather than providing a well-justified research analysis-theory. Particularly, regarding the trends presented in Fig. 1, which may seem to be rather arbitrary than justified.

Framing the empirical evidence behind the trajectories

In the original paper, empirical research has documented only fragments of the AI-human-challenge nexus: the exponential growth of AI capabilities, the spread of automation bias, the escalation of global crises, or the productivity gains from adoption of AI. What has been missing is a synthetic visualization that binds these strands into a coherent picture of systemic trajectories. Our Fig. 1 does not claim to be the result of regression or simulation; rather, it is a conceptual distillation of dozens of statistically significant, peer-reviewed studies that converge on the same pattern. The research gap, therefore, is not the lack of data but the absence of a framework to connect them into a systemic dilemma – a gap our work explicitly seeks to close. Below we are going to provide additional evidence, explicitly supporting our statements and conclusions from the main paper.

Evidence for the escalation of global challenges (Table Ch1)

Before considering AI, we must first establish the accelerating trajectory of the very challenges humanity faces. Table Ch1 consolidates robust empirical evidence showing that global risks are not isolated shocks but interdependent crises – amplified by systemic uncertainty and, paradoxically, by attempts to curb AI when driven by illiberal actors – “bad hands” that turn control into repression.

Table Ch1 – Empirical evidence of escalating complex global problems

Source	Study design	Key results	Conclusions
Systemic risk & uncertainty (polycrisis proxies)
World Economic Forum, (WEF, 2025-b) (Global Risks Report)	Expert survey (900+ cross-sector respondents) and Interconnections Map analysis.	· Tech risks (mis/disinformation, cyber) propagate into political polarization, economic instability, and social crises; · Climate/environment risks interlink with geopolitics and resource scarcity, etc.	Global risks form a tightly connected network. Рolycrisis dynamics intensify; demand new risk-management approaches.
Ahir et al., 2020 (IMF; World Uncertainty Index – WUI)	Quantitative index from 1950-2020, 143 countries; text-frequency measure of “uncertainty” in economic reports/newspapers.	Cyclical spikes in global economic uncertainty emerge when multiple crisis factors compound (oil shocks 1970s, GFC 2008–09, COVID-19 2020).	WUI evidences a “snowball effect” of uncertainty across crises, reinforcing polycrisis dynamics.
International Monetary Fund, 2024 (GFSR, Ch.2)	Panel evidence (43 advanced & EM economies since 1990) on macro-uncertainty and macro-financial stability.	· Аn uncertainty shock comparable to 2008-09 lowers one-year GDP growth by ~1.2 pp; · transmission via market, real (investment/consumption delay), and credit channels; · cross-border contagion through trade/finance links.	Persistently high uncertainty systematically depresses growth and interacts with other vulnerabilities – consistent with polycrisis.
Authoritarian tech & civil-liberties pressure
Beraja et al., 2023	AI & repression in China: · Causal design (DiD, IV, Bartik); · 2.9M government contracts; · 2,800+ prefectures; · 2013–2019; · events & procurement data.	· after protests: +1 police facial-recognition contract per 1M population (IV ≈ 0.74); · fair-weather–protest effect attenuates by ≈ 0.86 with AI procurement; · ~60k cameras per average prefecture; · >17k surveillance contracts total.	AI procurement scales surveillance, dampens protest, and shifts policing from field presence to tech – strengthening repressive capacity.
Frantz et al., 2020 (V-Dem)	Correlational analysis of digital repression and regime stability. V-Dem data; 91 autocracies, 2000–2017.	· +1 unit in digital-repression index = 6–10% lower protest probability; · high digital control correlates with greater regime stability (lower odds of turnover/fall).	Digital repression suppresses dissent and increases authoritarian resilience.
Ünver, 2024 (European Parliament)	Analytical review of official/open sources & human-rights datasets; case studies.	Documented uses: · China – Social Credit, internet censorship, Sharp Eyes, Xinjiang predictive policing & JOP (Joint Operations Platform); · Russia – mass surveillance, content blocking, automated tracking of opposition.	AI adoption by authoritarian regimes increases control and constrains human rights.

Taken together, these studies demonstrate that humanity is not merely facing isolated problems but a compounding wave of interdependent crises, reinforced both by systemic uncertainty and by political misuse of technology. This dual dynamic justifies the steep trajectory of Ch1 in Fig. 1 and sets the stage for the central dilemma: without tools that scale at least as fast as these crises, humanity will inevitably fall behind.

Evidence for exponential growth of AI capabilities (Table AI 1)

Table AI 1 distills a striking regularity: across hundreds of models, datasets, and domains, AI capabilities scale in a power-law fashion, with each order of magnitude in computing and data yielding predictable leaps in performance. This pattern has been confirmed not by isolated studies but by the heaviest hitters in the field – OpenAI, DeepMind, Stanford HAI – supported by evidence ranging from benchmark dominance in medicine and law to the dramatic collapse in inference costs that fuels global diffusion.

Table АІ 1 – Empirical evidence of the exponential growth of AI capabilities and resources

Source	Study design	Key results	Conclusions
Scaling laws and predictable growth
Kaplan et al., 2020 (OpenAI)	Scaling experiments with transformer LMs (10⁵–10⁹ parameters; millions–billions tokens; compute FLOPs)	Test loss decreases with parameters, data, and compute following a power-law	Model performance improves predictably with scale, following power-law scaling laws
Hoffmann et al., 2022 (DeepMind)	400+ models (70M–16B params; 5B–500B tokens); compute budgets varied	Optimal scaling = parameters and tokens scale with the square root of compute budget	LLM development follows stable power-law regularities, enabling forecasts of model efficiency
Compute and resource acceleration
Sevilla et al., 2022	Analysis of 123 landmark ML systems (2000–2020)	Compute demand growth: · pre-DL = doubling every ~20 mo; · early DL = doubling every ~6 mo; · large-scale era = ×10–100 increase, doubling every ~10 mo	Empirical trend: quasi-exponential growth in ML compute requirements
Rahman & Owen, 2024 (Epoch AI)	Trend analysis of training compute (2010–2024)	Training compute growth ≈ 4.4×/year (~doubling every 6 mo)	Strong evidence of exponential increase in AI training compute
Cottier & Rahman, 2024 (Epoch AI)	Cost modeling of frontier AI training (hardware, energy, personnel)	Training costs ≈ 2.4×/year (~doubling every 9 mo)	Frontier training becoming financially prohibitive; concentration of resources among few actors
Sevilla & Roldán, 2024 (Epoch AI)	Database analysis of frontier AI compute budgets (2010–2024)	Training compute for frontier models increased 4–5×/year	Frontier AI exhibits hyper-accelerated resource scaling
Performance breakthroughs across domains
OpenAI, 2023 (GPT-4 Tech Report)	Benchmarking across exams, code, translation (MMLU, bar exam, etc.)	· GPT-4 = human-level bar exam (top 10% vs GPT-3.5 bottom 10%); · MMLU SOTA in 24/26 languages	Scaling GPT-4 shows predictable performance gains consistent with power laws
Liu et al., 2024	Meta-analysis of 45 medical licensing exams worldwide	· GPT-4 accuracy = 81% vs GPT-3.5 = 58%; · passed 26/29 exams; · outperformed students in 13/17	Sharp leap in AI medical reasoning capabilities
Kasai et al., 2023	Evaluation of GPT-3, ChatGPT, GPT-4 on Japanese medical exams (2018–2022)	· GPT-4 passed all 6 years; GPT-3/ChatGPT failed; · only GPT-4/ChatGPT-EN avoided prohibited options	GPT-4 shows step-change improvement over predecessors
Diffusion, adoption, and economic scale
Maslej et al., 2025 (Stanford HAI)	Cost analysis of inference prices (2022–2024)	GPT-3.5 inference cost: $20M tokens (Nov 2022) vs $0.07M tokens (Oct 2024), >280× cheaper	AI systems becoming drastically more accessible and scalable
Singla et al., 2025 (McKinsey)	Global survey of 1,500+ executives & IT specialists	AI adoption: 78% (2024) vs 55% (2023)	Rapid acceleration in business AI adoption
Founders Forum Group, 2025	Market analysis of AI adoption (2024–2030)	AI market size: $391B (2025) vs $1.81T (2030)	Global AI industry entering exponential growth phase

The trajectory of AI 1 in Fig. 1 is therefore anything but speculative. It rests on convergent, high-credibility evidence that AI is evolving at a near-exponential pace – mirroring, and in many domains already outpacing, the escalation of global challenges.

Evidence for accelerated growth under autonomy (Table AI 2)

Table AI 2 demonstrates a consistent pattern: once freed from human priors and constraints, autonomous АІ systems achieve breakthroughs across domains that had resisted decades of human effort. From discovering algorithms and stabilizing plasma to optimizing hardware and energy systems, autonomy repeatedly converts “impossible” into achieved fact.

Table AI 2 – Empirical evidence for AI’s accelerated growth under autonomy

Source	Study design	Key results	Conclusions
Algorithm discovery & game-playing
Silver et al., 2017 (AlphaGo; Nature)	Self-play experiments with AlphaGo Zero vs human-trained predecessors	· Zero beat AlphaGo Lee 100–0; AlphaGo Master 89–11; · Elo: Zero 5185 vs Master 4858 vs Lee 3739 vs AlphaGo 3,144	Full autonomy without human priors yields superior performance
Fawzi et al., 2022 (AlphaTensor; Nature)	RL agent (AlphaTensor) autonomously discovered matrix multiplication algorithms	· Discovered faster algorithms (e.g., 47 vs 49 multiplications for 4×4 GF(2)); · 10–20% GPU/TPU speed-up	Autonomous discovery outpaced decades of human research
Mankowitz et al., 2023 (AlphaDev; Nature)	RL agent (AlphaDev) searched assembly-level sorting algorithms	· +70% faster for short sequences; · +1.7% for >250k elements; · integrated into LLVM C++ library	Autonomous RL surpassed human benchmarks with industry adoption
Hardware & engineering optimization
Mirhoseini et al., 2021 (Google; Nature)	RL agent for chip floorplanning; compared with human experts	<6h to generate layouts matching/exceeding human designs in PPA metrics (power, performance, area)	AI matched/exceeded experts while saving thousands of engineering hours
Energy & climate control
Luo et al., 2022 (DeepMind)	RL controller for commercial data-center cooling vs rule-based human strategies	Energy savings: 9–13% across sites, maintaining comfort limits.	Autonomous AI achieved superior energy efficiency.
Zhang et al., 2023	DRL algorithms tested in simulations and real data traces for thermal management	· Simulation: ≈10% cost savings; · Real-world data trace: 13.6% energy savings; high temperature accuracy (MAE 0.1 °C, ±0.2 °C). · Avg. DRL energy savings: 8.84% vs default controller	DRL ensured efficient, precise energy management beyond human defaults
Plasma control in fusion
Degrave et al., 2022 (Nature)	DRL controller tested on Swiss tokamak TCV; 19 magnetic coils controlled in real time	Maintained plasma configurations unachievable by traditional PID controllers (classical rule-based systems): · RMSE X-point = 3.7 cm (high positional accuracy); · plasma current RMSE = 0.50 kA (close match to target)	Autonomous AI stabilized plasma in regimes beyond human/standard control.
Human-AI collaboration limits
Romeo & Conti, 2025	Systematic PRISMA review (35 studies, 2015–2025) on automation bias in high-risk domains	· Automation bias linked to low AI literacy and cognitive overload; · XAI can worsen bias if users lack critical thinking	Human factors constrain outcomes; removing human bottlenecks enables faster AI acceleration

These results are not “opinions” but hard, quantitative gains: energy savings, thousands of hours of engineering time reduced, discoveries beyond the reach of human heuristics, etc. The implication is clear: once freed from human-imposed bottlenecks, AI does not merely grow fast, it accelerates beyond what human-guided trajectories can achieve. This justifies the sharper slope of AI 2 in our Fig. 1.

Evidence for AI as the only tool fast enough to keep pace (Table Ch 2)

If crises escalate at the rate of Ch 1, and AI accelerates at the rate of AI 2, the decisive question becomes: can AI actually close the gap? Table Ch 2 assembles some of the strongest causal evidence available, from randomized controlled trials in healthcare and education (Nature, Harvard, Hopkins PNAS) to quasi-experiments in productivity and IT development (Science, QJE, McKinsey, Google). Across these settings, AI consistently delivers measurable, repeatable gains in productivity, learning, clinical outcomes, and problem-solving speed.

Table Ch 2 – Quantitative evidence for AI as the most effective tool to tackle global challenges

Source	Study design	Key results	Conclusions
Productivity & economy
Chui et al., 2023 (McKinsey)	Modeling of 2,100+ work activities across 63 business cases	· Economic impact: $6.1–7.9T annually; · Productivity growth: +0.1–0.6%/year (up to +3.4% with combined tech)	Generative AI yields multi-trillion value and significant productivity gains, enabling reinvestment into global challenges
Brynjolfsson et al., 2025 (QJE)	Quasi-experiment (DiD, event-study) with 1,572 support agents	· +15% cases resolved/hour (avg.), +34% for novices; · customer sentiment +0.18 (≈0.5 SD); · attrition decreased by 40%	AI assistants rapidly upskill novices, closing workforce gaps in large-scale services
Noy & Zhang, 2023 (Science)	RCT with 453 professionals across sectors	· Task time: -40%; · Output quality: +18%	AI substitutes effort and improves efficiency in knowledge work
Peng et al, 2023 (GitHub Copilot)	Controlled experiment with developers (JS server task)	Completion 55.8% faster with Copilot	AI pair programming accelerates IT development amazingly
Paradis et al, 2024 (Google)	RCT with 96 developers (enterprise task)	· Time on task: 96 min (AI) vs 114 min (control), · ≈21–26% faster	AI integration provides a measurable productivity boost in enterprise coding
Golovianko et al., 2023 (EJIS)	Multi-phase MVPs of AI cognitive digital clones in organizational decision-making	· Bonus allocation (3 staff cohorts 367/428/501): F1=0.94–0.96 vs 0.49 baseline; · Recruitment: time/candidate 1 min vs 10; ≈1,591 staff-hours saved/year	Cognitive AI clones sharply improve accuracy, speed, and efficiency in critical decision bottlenecks
Healthcare
Chang et al, 2025 (Nature Comm.)	Prospective multicenter cohort study, N=24,543 (breast cancer screening)	Cancer Detection Rate: · 5.70/1000 (AI) vs 5.01/1000; · Recall unchanged (p=0.564)	AI-CAD improves detection (+13.8%) without increasing false positives
Adams et al., 2022 (Nature med.; Johns Hopkins)	Prospective, multi-site study, N=590,736 (sepsis detection)	· Time to antibiotics −1.85h faster; · Mortality decreased by18.7%	AI early warning significantly improves clinical outcomes
Education
Kestin et al., 2025 (Harvard)	RCT, N=194 (physics course)	· Learning gains +0.63 SD; · Time on task: 49 min (AI) vs 60 min (control)	AI tutoring outperforms active learning, enabling scalable individualized teaching
Gryshko et al., 2024	Controlled experiment, N=42 students, 8 rounds of argumentation	· Argumentation mass (A-Mass) =12.7×; · Prompt training further amplified gains	AI accelerates development of argumentation skills, reducing cognitive gaps
Energy & climate
Xie et al., 2025	Panel study, 3,374 obs. (China, 2008–2021), DID & IV	· AI adoption increased energy efficiency (+0.0223*** IV; +0.2735*** DID); · energy use (−0.0282**) and emissions (−0.0269***) decreased · 1% AI adoption = +6.56% green tech innovations; +3.79% agglomeration efficiency	AI is a powerful driver of green innovation and decarbonization
Price et al., 2025 (Nature)	Benchmarking study: ML (GenCast) vs ECMWF ENS	· Higher accuracy in 97.2% of 1,320 targets; · runtime: ~8 min (TPU) vs hours (supercomputer)	ML models surpass classical weather forecasts, enabling faster decisions
Science & biomedicine
Kovalevskiy et al., 2024 (PNAS; Google)	Validation of AlphaFold vs experimental protein structures	· Database growth: from 0.36M to 214M structures; · GDT (Global Distance Test) >90; · >50% papers in structural biology use AlphaFold	AI compresses discovery cycles, accelerating drug development from years to days

The verdict is unmistakable. АI is not a technological curiosity but the only empirically validated accelerator, across domains, capable of matching the tempo of poly-crisis dynamics. Without AI, the system spirals into “too little, too late.” With AI, humanity gains – at least in principle – a fighting chance to keep pace.

Evidence for cognitive degradation under AI reliance (Table Hu 1/2)

Finally, Table Hu 1/2 documents a consistent and troubling pattern: when humans offload cognition to AI, they gradually erode their own skills. Across medicine, education, and professional decision-making, studies converge on the same finding: short-term gains in speed or accuracy are systematically offset by declines in vigilance, critical reasoning, and independent problem-solving. The slight difference between Hu 1 and Hu 2 trends is since human cognitive degradation would be evidently faster with autonomous AI (Hu 2) rather than with human-controlled AI (Hu 1), which requires at least prompt engineering skills.

Table Hu 1/2 – Empirical evidence of human cognitive degradation with AI use

Source	Study design	Key results	Conclusions
Diagnostic decision-making & bias
Glickman & Sharot, 2025 (Nature Hum Behav)	A series of behavioral experiments (n=1,401) testing how biased AI feedback alters human judgments vs human–human feedback	· Biased AI amplified human biases more than biased human advice; · small judgment errors snowballed into larger distortions	AI feedback loops strengthen cognitive bias formation beyond human-only interaction; risk of compounding errors
Khera et al., 2023	Clinical decision tasks with 457 physicians under AI support: standard model ± explanations vs biased model.	· Baseline 73% vs Standard AI 76% (no expl.) vs 78% (heat-maps); · Biased AI drops to 62% (64% with heat-maps).	AI can raise average accuracy yet heighten automation bias; biased models degrade performance below baseline, even with explanations
Rosbach et al., 2025	28 pathologists estimate tumor cell percentage with/without AI under time pressure.	· AI support increased overall accuracy; · automation bias occurred in 7% of cases (correct responses converted into AI-induced wrong ones); · under time pressure, reliance on incorrect AI cues increased, lowering accuracy	Short-term productivity gains coexist with loss of vigilance and the propagation of mistakes made by humans under time pressure
Cognitive offloading & dependence
Wahn et al., 2023	Laboratory experiment with regression analysis. Multiple-object tracking (MOT) task with optional delegation to an algorithm; n=52	Participants offloaded ~50% of trials; when told the algorithm is “perfect,” full offloading ≈80%.	Under load, people readily relinquish control, even when capable themselves, especially when perceiving AI as highly reliable
Wahn & Schmitz, 2024	Laboratory experiment with regression analysis. Multiple-object tracking (MOT) with incentives (“bonus task”) and reliability cues; n=52	Full offloading ≈50% without info on algorithm accuracy vs. ~82% when told it was “perfect.” Bonus task reduced offloading (β=−0.63*), perceived reliability increased it (β=0.25**); monetary payoff no effect.**	Knowledge of AI reliability sharply increases delegation. Offloading is shaped by non-monetary factors, while financial payoffs show no effect
Work skill erosion (deskilling)
Budzyń et al., 2025	Multicenter observational study in 4 endoscopy centers; 1,443 colonoscopies (795 pre-AI vs 648 post-AI)	· Adenoma Detection Rate (ADR): 28.4% vs 22.4% (after AI); · OR=0.69 for ADR post-AI (The odds of ADR decreased by ~31%)	Routine AI assistance correlated with deskilling (lower ADR), consistent with habituation to AI hints
Education & longer-term retention
Bastani et al., 2025 (PNAS)	RCT in Turkish high schools (~1,000 students): GPT-Base vs GPT-Tutor (scaffolded prompts) vs control	Practical tasks results: +127% (Tutor) / +48% (Base) vs control. After access removal, test performance −17% in GPT-Base; Tutor ≈ no significant drop	Unguarded access boosts practice outputs but harms subsequent independent performance; scaffolded tutoring mitigates degradation

The very slow upward slope of Hu 1 and clear downward slope of Hu 2 in Fig. 1 is thus firmly grounded in evidence. AI may boost productivity in the moment, but it simultaneously hollows out the cognitive foundations required for long-term resilience. This dual edge – empowering and eroding at once – is precisely what makes the dilemma so acute.

Closing the circle: why Fig. 1 matters

By linking Table Ch 1 – Table Ch 2, Table AI 1 – Table AI 2, and Table Hu 1/2, our Fig. 1 achieves what no single study could deliver: a synthetic, system-level visualization of a structural dilemma. The research gap has never been a shortage of data – it has been the absence of a unifying lens to connect them into a coherent trajectory. Our figure fills that void. It is not a conjecture but a distillation of robust empirical evidence spanning economics, political science, computer science, medicine, and education. Therefore, content in Fig. 1 is anything but arbitrary. It is the first visualization to integrate scattered but convergent findings into one systemic insight: crises escalate, AI accelerates, and human cognition declines. This is not a hypothesis searching for validation – it is overwhelming evidence demanding a framework.

5. AI-as-a-user-of-AI: The Last Mortal Skill – Prompt Engineering

In this section we are going to dive deeper into the following yet another warning logic. As AI surpasses human capabilities in analysis, creativity, and decision-making, it looks like the last remaining human skill of any value may soon be prompt engineering ‒ the art of asking AI the right questions. In a world where AI autonomously generates ideas, solves complex problems, and even improves itself, humans may find their cognitive role reduced to little more than crafting effective instructions for machines. The irony is striking, i.e., after millennia of intellectual evolution, our defining ability may not be thinking, reasoning, or creating but merely knowing how to communicate with a superior intelligence that does all the thinking for us. And even this skill may not last long. As AI systems grow better at understanding vague, imprecise, or even incomplete queries, they may soon render even prompt engineering obsolete. At that point, the last human role in intelligence-driven work will disappear ‒ not with a bang, but with a well-phrased question.

Evolution of prompt engineering as human’s cognitive skill

As advances in AI and LLMs accelerate, machines increasingly outpace humans not only in raw data processing or pattern recognition but also in higher-order cognitive functions such as creative synthesis, reasoning, and strategic planning (Bubeck et al., 2023; Achiam et al., 2023; Bommasani et al., 2022). In such a landscape, the most durable human role may not lie in producing knowledge directly but in the meta-skill of prompt engineering – the craft of asking AI systems powerful, clear, and contextually rich questions that unlock their generative capacity (Liu et al., 2023; Meskó, 2023).

Prompt engineering has rapidly emerged as a discipline in both AI-driven creativity (Meskó, 2023) and decision support (Garvey & Svendsen, 2024). It requires understanding human intent and the operational quirks of AI models – a dual literacy within a technology-mediated cognitive ecosystem. As industry analysts have noted, human ingenuity is being reframed, and questions become the new currency of intelligence (Nolan, 2024).

However, relying on prompt engineering implicitly signals a deeper outsourcing of cognitive labor. If AI systems continue to evolve toward superior comprehension of ambiguous or under-specified instructions (Zhou et al., 2023), the need for carefully structured human prompts may diminish. Future models are expected to anticipate intentions with minimal input, remembering user context, preferences, and goals even from fragmentary cues. In effect, the “prompt engineer” could vanish as AI becomes preemptively adaptive (Schick et al., 2023).

Historically, humans have lost skills as tools grew more autonomous – from handwriting replaced by keyboards (Mueller & Oppenheimer, 2014) to wayfinding replaced by GPS (Dahmani & Bohbot, 2020). Similarly, researchers in human‒AI interaction warn that dialogic interfaces are shifting from reactive to anticipatory design, where prompting evolves into conversational triggers rather than deliberate craft (Shneiderman, 2022). Cognitive scientists further highlight the risk of meta-cognitive skill atrophy when generative systems predict user intent seamlessly (Sparrow et al., 2011).

If realized, this transition will likely not be abrupt but gradual: prompt engineering may fade not by conflict but by obsolescence. The well-phrased human question could remain – not as proof of intelligence, but as a relic of an era when humans still needed to ask.

Evolution of prompt engineering as AI’s cognitive skill

As model capabilities scale and agentic loops mature, the very skill humans hoped to retain ‒ prompt engineering ‒ becomes another domain where AI outperforms us. A growing body of work shows that large models can optimize, generate, and refine prompts autonomously, closing the window in which human prompt craft is uniquely valuable. Early demonstrations like Automatic Prompt Engineer (APE) and subsequent lines of research on automatic prompt optimization document that LLMs (and lightweight search/GA procedures wrapped around them) can iteratively improve prompts with minimal human oversight, often surpassing hand-crafted baselines. Recent surveys synthesize dozens of such techniques, formalizing prompt optimization as a search/optimization problem over discrete/continuous prompt spaces and showing rapid progress in self-prompting pipelines (Farfan, 2023; Li et al., 2025).

Concurrently, models have learned to teach themselves tool use (“when to call which API, with what arguments, and how to integrate the result”), eroding another traditional human advantage in orchestrating complex workflows. “Toolformer” is a canonical example: trained with a handful of demonstrations, the model learns to call calculators, search engines, QA systems, and more ‒ all in a self-supervised way (Schick et al, 2023). Once a model can pick tools, call them, and fold results back into its chain of reasoning without human mediation, the “prompt engineer” increasingly looks like an internal subroutine, not an external job.

On top of tool use, agentic LLMs add control loops for reasoning, acting, and interacting. Mechanisms such as verbal self-reflection (e.g., Reflexion) let agents critique their own failures and adjust subsequent attempts, again, with no human in the loop for the “craft a better prompt next time” step (Shinn et al., 2023). Surveys of agentic LLMs emphasize precisely this direction: the agent learns to plan, decompose tasks, decide which tools/models to call, and adapt its own prompting as it goes. Benchmarks and reports from the agent community (including evaluations of AutoGPT-style systems) illustrate both the promise and the steady improvement trajectory of these feedback-driven loops (Renze & Guven, 2024).

Put differently: prompt engineering becomes machine internalized. As optimization and reflection routines migrate inside the agent, the model becomes its own prompt engineer, and then a manager of other models. In our former study, we argued this formally in “AI as a User of AI: Towards Responsible Autonomy” (Shukla et al., 2024): once AI systems can specify sub-goals, select models/tools, evaluate outputs, and iteratively refine instructions, they function as principals directing other AIs, not merely tools serving humans. That makes AI the primary user of AI, with humans increasingly relegated to policy-level constraints, audit hooks, or vetoes ‒ roles that are, by design, invoked after the agent has already decomposed, prompted, and executed most of the work.

This trajectory tightens further as research matures in autonomous prompt engineering. New results show models autonomously exploring prompt variations, mutating and scoring them against task objectives, and converging to high-performing prompts, sometimes for specialized domains like test-case generation or NER, without human “prompt whisperers” in the loop (Yang et al., 2025). The literature now treats prompt optimization as a first-class automatable component, not a boutique human craft (Gao et al., 2025). Even industry coverage acknowledges the trend: demand for human prompt engineers rose quickly, but many experts view the role as transitional, expecting models to absorb it via self-optimization (Nolan, 2024).

Why this matters for the objectives of this study. If AI becomes (1) a better prompt engineer than humans, (2) a better user of tools than humans, and (3) an autonomous orchestrator of other AIs, then in hybrid Human-plus-Machine systems the center of gravity shifts decisively toward AI-to-AI coordination. Humans cease to be the “conductors” and become more like “license granters” or “safety reviewers.” In industrial contexts, this means planning, scheduling, quality control, supply-chain negotiation, and even model-against-model bargaining can be initiated, parameterized, and closed by AI agents that write their own prompts, select their own peers, and iterate until internal acceptance criteria are met. Our AI-as-a-user-of-AI framework (Shukla et al., 2024) makes the implication explicit: autonomy + self-prompting ⇒ self-directed agenda-setting. Once agenda-setting moves inside the machine ecology, the default client of industrial AI becomes other AI, not humans, i.e., exactly the tipping point our paper warns about.

A fair caveat is that present-day autonomous agents still fail in systematic ways, and rigorous evaluations continue to surface brittleness (Yang, 2023). But the research arc, i.e., self-prompting, tool learning, self-reflection, automatic optimization, points in one direction: the residual human comparative advantage in “asking the right question” is shrinking, and with it the last leverage point in cognition-centric work. Absent deliberate governance and interface design that re-creates meaningful human leverage, the industry of the future will be AI-managed for AI clients, with humans participating only where regulation or ethics force AI to keep us in the loop.

6. Some Additional Notes on Excluding Potential Bias

One may still feel that the lack of statistical data creates the impression of a provocative paper with biased positioning on the subject rather than a well-justified research analysis-theory. However, the provocative nature of the paper is not as a flaw but as an intentional strength. In the context of rapidly advancing AI and Industry 6.0, raising difficult questions and challenging conventional thinking is essential. The purpose of this work is not merely to provide statistical data, but to serve as a wake-up call that highlights risks, such as intellectual perishability, that may not yet be fully captured by quantitative studies. Far from undermining its scientific value, this forward-looking and critical perspective is intended to stimulate rigorous debate and future empirical research on an emerging issue. Instead of explicitly including our own statistical data to support our thesis, we explicitly refer to and synthesize findings from about 150 high-quality academic sources, many of which are peer-reviewed empirical studies containing precisely the kind of statistical data the reviewer demands. Instead of replicating those studies, our intention was to build an integrative and analytical argument based on a substantial body of existing evidence. Repeating the same statistical analyses ourselves would not only be redundant but also outside the intended scope of our theoretical and conceptual contribution. Our approach is well aligned with accepted methodologies in foresight studies and conceptual analysis of sociotechnical systems.

Biased positioning on the subject

We contend that our analysis is grounded in a robust empirical base and represents a necessary cautionary perspective. Below we clarify our logical framework, note the existence of opposing views, and demonstrate that these counter-positions currently lack the empirical foundations necessary to invalidate our assumptions.

The paper’s logical framework

Our argument is structured around a set of interrelated, research-backed assumptions:

· AI cognitive performance is advancing more swiftly than the growth of human cognitive capacities.

· Human reliance on AI often comes at the cost of preserving or developing independent cognitive skills.

· This reliance is observed not only in educational settings but extends into industrial and professional domains.

· Autonomous AI is essential both for self-improvement and as a safeguard against misuse by authoritarian actors.

· Given rising human dependency on AI, autonomous systems may not naturally align industry with human values, setting the stage for the risks we project.

We strengthen each link in this chain with evidence derived from hundreds of recent studies, spanning experimental work, meta-analyses, domain-specific field observations, and labor-market surveys.

Are there opposing views? Yes ‒ but, actually, they (not our logic above!) lack empirical foundation

For completeness, one could reverse each assumption from the list above:

· AI does not outpace human cognitive growth.

· AI reliance does not hinder human cognitive development.

· Dependency is not observable in education or industry.

· Autonomous AI is not necessary to restrict misuse.

· Without growing human dependence, autonomous AI would not risk diverting industry away from human values.

While some proponents advance such optimistic or neutral views, they often rest on conceptual assertions rather than empirical data.

Examples of such positions:

· The Distributed Cognition framework (Zhang & Fenton, 2024) suggests AI augments human intelligence without replacing it, citing cognitive artifacts as long-standing tools that aid human thought and collaboration (aka “Human cognition as distributed cognition”)

· Nvidia’s CEO Jensen Huang recently and publicly rejected findings that AI degrades cognition, claiming personal use of AI enhanced his own cognitive abilities (TOI, 2025).

· A report on AI-assisted mental health tools (Thakkar et al., 2024) emphasizes collaborative synergy: such tools support and complement human decision-making rather than supplant it.

· A systematic meta-review (Vaccaro et al., 2024) found that human-AI teams often underperform compared to individuals alone, implying AI’s net effect may not be purely augmentative.

· An empirical study of interpreters using AI for real-time subtitling found improvements in working memory and executive function after training in human-AI interaction (Wallinheimo et al., 2023).

Why these opposing views are not yet empirically robust

The distributed cognition model, while conceptually sound, does not empirically demonstrate that reliance on AI preserves or enhances human cognition ‒ rather, it highlights potential augmentation under optimal conditions, abstracted from empirical testing.

The CEO anecdote is purely observational and not rooted in systematic measurement or peer-reviewed evidence.

Reports on AI in mental health focus on human-AI synergy in therapeutic contexts, but they do not evaluate the long-term effects on clinicians’ cognitive or decision-making capacities.

The meta-review showing subpar human-AI group performance underscores the complexity of collaboration and supports our view that human roles could erode without deliberate design and training. Their study demonstrates one specific form of cognitive enhancement in a niche setting. While promising, it does not generalize across professions or educational systems and rather supports the need for structured human-AI training, not the absence of risk.

Our position: balanced ‒ but vigilant

Our paper is not an ideological call for Luddism; it is a rigorous warning based on existing empirical evidence. We do not deny that AI can be a powerful augmentative tool ‒ indeed, we cite meta-analytic findings (Wang & Fan 2025) where AI, properly scaffolded, significantly boosts learning performance. What we highlight is that without safeguards, education, institutional design, and regulation, the same tool can degrade human skill, agency, and oversight.

Thus, while opposing perspectives exist, they remain largely conceptual or schematic, lacking the empirical investigations and longitudinal studies that substantiate the mechanisms we identify. Considering this, our framing is not biased, but rather evidence-driven and intended to serve as a wake-up call, urging the academic, educational, and policy communities to act before degradation becomes entrenched rather than pointing fingers after it’s too late.

7. AI Autonomy Dilemma has Historical Roots in Economics

This section draws on a historical analogy between economic governance models and the emerging dilemma of AI autonomy. In economic history, debates between centralized, top-down planning and decentralized, competitive markets revealed that excessive control often led to inefficiency and fragility, while distributed autonomy fostered adaptability and resilience. Both economies and AI ecosystems are complex adaptive systems ‒ networks of interacting agents whose collective behavior cannot be fully predicted or directed from above (Simon, 2019; Holland, 1995). This makes economic debates about planning versus markets a useful lens for considering the governance of AI.

A central parallel lies in the problem of information. Hayek (1945) argued that centralized planning fails because information is dispersed across individuals and contexts, and no planner can ever aggregate or act on it efficiently. By analogy, rigid human oversight over AI systems risks the same bottlenecks: autonomous agents will often have richer, real-time access to data than human supervisors can meaningfully monitor or direct. Over-control therefore introduces inefficiency, rigidity, and systemic fragility.

History also shows that innovation and resilience emerge more readily in decentralized settings. Schumpeter (1942) highlighted the role of “creative destruction” in driving progress, while Ostrom (1990) demonstrated how polycentric governance fosters adaptability by layering multiple overlapping control mechanisms. Together these insights suggest that AI safety may depend less on centralized command-and-control and more on distributed autonomy supplemented by regulatory guardrails.

Put differently, autonomy is not the opposite of safety but often its prerequisite. Where central control creates stagnation, bottlenecks, and vulnerability, autonomy enables adaptability, robustness, and systemic resilience. The challenge for AI governance is therefore not to prevent autonomy, but to shape it responsibly, ensuring that autonomous systems operate within ethical, legal, and safety-oriented boundaries, much like competitive markets are bounded by regulation.

The structured examples below (Lessons 1–5) illustrate these parallels in more detail.

Lesson 1: The information bottleneck

· Economic context: Centralized planning fails because knowledge is dispersed; no authority can aggregate and act on it effectively (Hayek, 1945).

· AI lesson: Excessive human oversight risks the same bottlenecks, since autonomous agents often have faster, richer access to data.

· Importance: Recognizing humans as “central planners” with limited bandwidth highlights the need for AI autonomy to ensure responsiveness and adaptability (Russell & Norvig, 2021).

Lesson 2: Innovation through creative destruction

· Economic context: Innovation thrives under competition; central planning stifles novelty (Schumpeter, 1942).

· AI lesson: Rigid, human-scripted rules can suppress emergent strategies and adaptive problem-solving; autonomy enables AI to explore novel and efficient solutions.

· Importance: Human safety may depend on AI’s ability to generate unforeseen responses to unforeseen risks (Silver et al., 2017).

Lesson 3: Resilience through polycentric governance

· Economic context: Polycentric systems with overlapping controls are more resilient than centralized hierarchies (Ostrom, 1990).

· AI lesson: Safety should emerge from layered governance (self-regulating AI collectives combined with external oversight) rather than exclusive human control.

· Importance: Avoids single points of failure in governance and distributes responsibility across both AI and humans (Dafoe, 2018).

Lesson 4: Incentive alignment

· Economic context: Markets succeed through aligned incentives; centrally planned systems often collapse under distorted ones (Nove, 1983).

· AI lesson: Reward structures must guide AI actions toward global goals; misaligned incentives can drive harmful behaviors.

· Importance: Careful incentive design and alignment mechanisms are critical for ensuring autonomy leads to beneficial outcomes (Amodei et al., 2016).

Lesson 5: Efficiency versus resilience

· Economic context: Central planning often over-optimized for efficiency at the cost of fragility; markets preserved redundancies that improved survival (Gerschenkron, 1962).

· AI lesson: Strict human control risks brittle systems optimized for narrow efficiency; distributed autonomy builds redundancy and robustness.

· Importance: Demonstrates that autonomy enhances not just performance but long-term safety (Bostrom, 2014).

These lessons demonstrate that the analogy between economic organization and AI governance is not abstract but highly practical. Historical insights about information bottlenecks, innovation dynamics, governance resilience, and efficiency–safety trade-offs converge on a clear conclusion: rigid, centralized human control creates fragility, while distributed autonomy fosters adaptability and resilience. Applied to the AI domain, this implies that granting AI systems measured autonomy is not only more efficient but also paradoxically safer, even if it reduces the scope of human involvement. Just as top-down economic planning often produced stagnation, inefficiency, and systemic vulnerability, excessive human domination over AI risks brittleness, hindered adaptation, and greater chances of misalignment. Conversely, decentralized structures historically thrived by dispersing decision-making power and enabling adaptive intelligence to emerge through competition and feedback. This historical perspective directly reinforces our central claims: (1) that human control in future AI-driven industry will naturally diminish or become obsolete, and (2) that autonomous AI, if properly structured with aligned incentives and layered governance, represents a lesser risk than flawed, short-sighted, and biased human-controlled AI.

Additional References

Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., ... & McGrew, B. (2023). GPT-4 technical report. arXiv preprint arXiv:2303.08774. https://doi.org/10.48550/arXiv.2303.08774

Adams, R., Henry, K. E., Sridharan, A., Soleimani, H., Zhan, A., Rawat, N., ... & Saria, S. (2022). Prospective, multi-site study of patient outcomes after implementation of the TREWS machine learning-based early warning system for sepsis. Nature Medicine, 28(7), 1455-1460. https://doi.org/10.1038/s41591-022-01894-0

Ahir, H., Bloom, N., & Furceri, D. (2020). 60 Years of uncertainty. Finance & Development, 57(1), 58-60. https://www.imf.org/en/Publications/fandd/issues/2020/03/imf-launches-world-uncertainty-index-wui-furceri

Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., & Mané, D. (2016). Concrete Problems in AI Safety. arXiv preprint arXiv:1606.06565. https://doi.org/10.48550/arXiv.1606.06565

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122

Beraja, M., Kao, A., Yang, D. Y., & Yuchtman, N. (2023). AI-tocracy. The Quarterly Journal of Economics, 138(3), 1349-1402. https://doi.org/10.1093/qje/qjad012

Bommasani, R., Hudson, D. A., Adeli, E., et al. (2022). On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258. https://doi.org/10.48550/arXiv.2108.07258

Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.

Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889-942. https://doi.org/10.1093/qje/qjae044

Bubeck, S., Chandrasekaran, V., Eldan, R., Gehrke, J., Horvitz, E., Kamar, E., ... & Zhang, Y. (2023). Sparks of artificial general intelligence: Early experiments with GPT-4. arXiv preprint arXiv:2303.12712. https://doi.org/10.48550/arXiv.2303.12712

Budzyń, K., Romańczyk, M., Kitala, D., Kołodziej, P., Bugajski, M., Adami, H. O., ... & Mori, Y. (2025) Endoscopist De-Skilling after Exposure to Artificial Intelligence in Colonoscopy: A Multicenter Observational Study. https://doi.org/10.1016/S2468-1253(25)00133-5

Chang, Y. W., Ryu, J. K., An, J. K., Choi, N., Park, Y. M., Ko, K. H., & Han, K. (2025). Artificial intelligence for breast cancer screening in mammography (AI-STREAM): preliminary analysis of a prospective multicenter cohort study. Nature Communications, 16(1), 2248. https://doi.org/10.1038/s41467-025-57469-3

Chui, M., Hazan, E., Roberts, R., Singla, A., Smaje, K., Sukharevsky, A., Yee, L., & Zemmel, R. (2023, June 14). The economic potential of generative AI: The next productivity frontier. McKinsey & Company. https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier

Cottier, B., & Rahman, R. (2024). Training compute costs are doubling every nine months for the largest AI models. Epoch.AI. https://epoch.ai/data-insights/cost-trend-large-scale

Dafoe, A. (2018). AI Governance: A Research Agenda. Governance of AI Program, Future of Humanity Institute, University of Oxford: Oxford, UK, 1442, 1443. https://cdn.governance.ai/GovAI-Research-Agenda.pdf

Degrave, J., Felici, F., Buchli, J., Neunert, M., Tracey, B., Carpanese, F., ... & Riedmiller, M. (2022). Magnetic control of tokamak plasmas through deep reinforcement learning. Nature, 602(7897), 414-419. https://doi.org/10.1038/s41586-021-04301-9

Dahmani, L., & Bohbot, V. D. (2020). Habitual use of GPS negatively impacts spatial memory during self-guided navigation. Scientific Reports, 10(1), 6310. https://doi.org/10.1038/s41598-020-62877-0

Dratsch, T., Chen, X., Rezazade Mehrizi, M., Kloeckner, R., Mähringer-Kunz, A., Püsken, M., ... & Pinto dos Santos, D. (2023). Automation bias in mammography: the impact of artificial intelligence BI-RADS suggestions on reader performance. Radiology, 307(4), e222176. https://doi.org/10.1148/radiol.222176

Farfan, N. (2023). Goodbye Prompt Engineering, Hello Prompt Generation: Automatic Prompt Engineer (APE) research summary. The Batch. DeepLearning.AI. https://www.deeplearning.ai/the-batch/research-summary-automatic-prompt-engineer-ape/

Fawzi, A., Balog, M., Huang, A., Hubert, T., Romera-Paredes, B., Barekatain, M., ... & Kohli, P. (2022). Discovering faster matrix multiplication algorithms with reinforcement learning. Nature, 610(7930), 47-53. https://doi.org/10.1038/s41586-022-05172-4

Frantz, E., Kendall-Taylor, A., & Wright, J. (2020). Digital repression in autocracies. Varieties of Democracy Institute Users Working Paper (27), 1-22. https://www.v-dem.net/media/publications/digital-repression17mar.pdf

Founders Forum Group. (2025, June 23). AI statistics 2024–2025: Global trends, market growth & adoption data. Founders Forum Group. https://ff.co/ai-statistics-trends-global-market

Gao, S., Wang, C., Gao, C., Jiao, X., Chong, C. Y., Gao, S., & Lyu, M. (2025). The Prompt Alchemist: Automated LLM-Tailored Prompt Optimization for Test Case Generation. arXiv preprint arXiv:2501.01329. https://doi.org/10.48550/arXiv.2501.01329

Garvey, B., & Svendsen, A. D. (2024). Prompt-Engineering Testing ChatGPT4 and Bard for Assessing Generative-AI Efficacy to Support Decision-Making. In: Navigating Uncertainty Using Foresight Intelligence: A Guidebook for Scoping Scenario Options in Cyber and Beyond (pp. 167-212). Springer: Cham. https://doi.org/10.1007/978-3-031-66115-0_10

Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), 6. https://doi.org/10.3390/soc15010006

Gerschenkron, A. (1962). Economic Backwardness in Historical Perspective. Harvard University Press.

Glickman, M., & Sharot, T. (2025). How human-AI feedback loops alter human perceptual, emotional and social judgements. Nature Human Behaviour, 9(2), 345-359. https://doi.org/10.1038/s41562-024-02077-2

Goddard, K., Roudsari, A., & Wyatt, J. C. (2012). Automation bias: a systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association, 19(1), 121-127. https://doi.org/10.1136/amiajnl-2011-000089

Golovianko, M., Gryshko, S., Terziyan, V., & Tuunanen, T. (2023). Responsible cognitive digital clones as decision-makers: a design science research study. European Journal of Information Systems, 32(5), 879-901.

Gryshko, S., Terziyan, V., & Golovianko, M. (2024, September). Resilience Training in Higher Education: AI-Assisted Collaborative Learning. In International Conference on Interactive Collaborative Learning (pp. 126-138). Cham: Springer Nature Switzerland. https://doi.org/10.1007/978-3-031-83520-9_12

Hayek, F. A. (1945). The Use of Knowledge in Society. The American Economic Review, 35(4), 519-530. https://www.jstor.org/stable/1809376

Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., ... & Sifre, L. (2022). Training compute-optimal large language models. arXiv preprint arXiv:2203.15556. https://doi.org/10.48550/arXiv.2203.15556

Holland, J. H. (1995). Hidden Order: How Adaptation Builds Complexity. Addison-Wesley. NY. https://doi.org/10.1177/027046769701700420

International Monetary Fund. (2024). Global financial stability report 2024: Macrofinancial stability amid high global economic uncertainty (Chapter 2). https://www.imf.org/en/Publications/GFSR

Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., ... & Amodei, D. (2020). Scaling laws for neural language models. arXiv preprint arXiv:2001.08361. https://doi.org/10.48550/arXiv.2001.08361

Kasai, J., Kasai, Y., Sakaguchi, K., Yamada, Y., & Radev, D. (2023). Evaluating GPT-4 and ChatGPT on Japanese medical licensing examinations. arXiv preprint arXiv:2303.18027. https://doi.org/10.48550/arXiv.2303.18027

Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15(1), 17458. https://doi.org/10.1038/s41598-025-97652-6

Khera, R., Simon, M. A., & Ross, J. S. (2023). Automation bias and assistive AI: risk of harm from AI-driven clinical decision support. Jama, 330(23), 2255-2257. https://doi.org/10.1001/jama.2023.22557

Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X. H., Beresnitzky, A. V., ... & Maes, P. (2025). Your Brain on ChatGpt: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv preprint arXiv:2506.08872. https://doi.org/10.48550/arXiv.2506.08872

Kovalevskiy, O., Mateos-Garcia, J., & Tunyasuvunakool, K. (2024). AlphaFold two years on: Validation and impact. Proceedings of the National Academy of Sciences, 121(34), e2315002121. https://doi.org/10.1073/pnas.2315002121

Kücking, F., Hübner, U., Przysucha, M., Hannemann, N., Kutza, J. O., Moelleken, M., ... & Busch, D. (2024). Automation bias in AI-decision support: Results from an empirical study. In: Studies in Health Technology and Informatics (Vol. 317, German Medical Data Sciences 2024, pp. 298-304). IOS Press. https://doi.org/10.3233/shti240871

Li, W., Wang, X., Li, W., & Jin, B. (2025). A survey of automatic prompt engineering: An optimization perspective. arXiv preprint arXiv:2502.11560.
https://doi.org/10.48550/arXiv.2502.11560

Liu, M., Okuhara, T., Chang, X., Shirabe, R., Nishiie, Y., Okada, H., & Kiuchi, T. (2024). Performance of ChatGPT across different versions in medical licensing examinations worldwide: systematic review and meta-analysis. Journal of medical Internet research, 26, e60807. https://doi.org/10.2196/60807

Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., & Neubig, G. (2023). Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9), 1‒35. https://doi.org/10.1145/3560815

Luo, J., Paduraru, C., Voicu, O., Chervonyi, Y., Munns, S., Li, J., ... & Mankowitz, D. J. (2022). Controlling commercial cooling systems using reinforcement learning. arXiv preprint arXiv:2211.07357. https://doi.org/10.48550/arXiv.2211.07357

Mankowitz, D. J., Michi, A., Zhernov, A., Gelmi, M., Selvi, M., Paduraru, C., ... & Silver, D. (2023). Faster sorting algorithms discovered using deep reinforcement learning. Nature, 618(7964), 257-263. https://doi.org/10.1038/s41586-023-06004-9

Maslej, N., Fattorini, L., Perrault, R., Gil, Y., Parli, V., Kariuki, N., ... & Oak, S. (2025). Artificial Intelligence Index Report 2025. arXiv preprint arXiv:2504.07139. https://doi.org/10.48550/arXiv.2504.07139

Meskó, B. (2023). Prompt engineering as an important emerging skill for medical professionals: tutorial. Journal of Medical Internet Research, 25, e50638. https://doi.org/10.2196/50638

Mirhoseini, A., Goldie, A., Yazgan, M., Jiang, J. W., Songhori, E., Wang, S., ... & Dean, J. (2021). A graph placement methodology for fast chip design. Nature, 594(7862), 207-212. https://doi.org/10.1038/s41586-021-03544-w

Mueller, P. A., & Oppenheimer, D. M. (2014). The pen is mightier than the keyboard: Advantages of longhand over laptop note taking. Psychological Science, 25(6), 1159–1168. https://doi.org/10.1177/0956797614524581

Nolan, B. (2024). Demand is skyrocketing for prompt engineers, one of the hottest roles in AI. Some say it won’t last. Business Insider. Retrieved from: https://www.businessinsider.com/prompt-engineer-ai-careers-tech-fad-2024-3

Nove, A. (1983). The Economics of Feasible Socialism. London: George Allen & Unwin.

Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187-192. https://doi.org/10.1126/science.adh2586

OpenAI (2023). GPT-4 Technical Report. OpenAI. https://cdn.openai.com/papers/gpt-4.pdf

Ostrom, E. (1990). Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press. https://doi.org/10.1017/CBO9780511807763

Paradis, E., Grey, K., Madison, Q., Nam, D., Macvean, A., Meimand, V., ... & Chandra, S. (2024). How much does AI impact development speed? An enterprise-based randomized controlled trial. arXiv preprint arXiv:2410.12944. https://doi.org/10.48550/arXiv.2410.12944

Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The impact of AI on developer productivity: Evidence from GitHub Copilot. arXiv preprint arXiv:2302.06590. https://doi.org/10.48550/arXiv.2302.06590

Price, I., Sanchez-Gonzalez, A., Alet, F., Andersson, T. R., El-Kadi, A., Masters, D., ... & Willson, M. (2025). Probabilistic weather forecasting with machine learning. Nature, 637(8044), 84-90. https://doi.org/10.1038/s41586-024-08252-9

Rahman, R., & Owen, D. (2024), The training compute of notable AI models has been doubling roughly every six months. Epoch.AI. Retrieved from: https://epoch.ai/data-insights/compute-trend-post-2010

Renze, M., & Guven, E. (2024). Self-reflection in large language model agents: Effects on problem-solving performance. In: Proceedings of the 2nd International Conference on Foundation and Large Language Models (pp. 516-525). IEEE. https://doi.org/10.1109/FLLM63129.2024.10852426

Romeo, G., & Conti, D. (2025). Exploring automation bias in human–AI collaboration: a review and implications for explainable AI. AI & SOCIETY, 1-20. https://doi.org/10.1007/s00146-025-02422-7

Rosbach, E., Ganz, J., Ammeling, J., Riener, A., Aubreville, M. (2025). Automation Bias in AI-assisted Medical Decision-making under Time Pressure in Computational Pathology. In: Palm, C., et al. Bildverarbeitung für die Medizin 2025. BVM 2025. Informatik Aktuell. Springer Vieweg, Wiesbaden. https://doi.org/10.1007/978-3-658-47422-5_27

Rosca, C-M., & Stancu, A. (2025). Impact of Industry 6.0 on Human Cognitive Behavior. In: From Industry 4.0 to Industry 6.0 (Chapter 3, pp.47-80). Wiley. https://doi.org/10.1002/9781394372775.ch3

Russell, S., & Norvig, P. (2021). Artificial Intelligence: A Modern Approach (4th Ed.). Pearson.

Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., Hambro, E., ... & Scialom, T. (2023). Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36, 68539-68551. https://proceedings.neurips.cc/paper_files/paper/2023/file/d842425e4bf79ba039352da0f658a906-Paper-Conference.pdf

Schumpeter, S. (1942). Capitalism, Socialism, and Democracy. Harper & Brothers. https://ia601501.us.archive.org/30/items/in.ernet.dli.2015.190072/2015.190072.Capitalism-Socialism-And-Democracy.pdf

Sevilla, J., Heim, L., Ho, A., Besiroglu, T., Hobbhahn, M., & Villalobos, P. (2022, July). Compute trends across three eras of machine learning. In: Proceedings of the International Joint Conference on Neural Networks (pp. 1-8). IEEE. https://doi.org/10.1109/IJCNN55064.2022.9891914

Sevilla, J., & Roldán, E. (2024). Training Compute of Frontier AI Models Grows by 4-5x per Year. Epoch.AI. https://epoch.ai/blog/training-compute-of-frontier-ai-models-grows-by-4-5x-per-year

Shinn, N., Cassano, F., Gopinath, A., Narasimhan, K., & Yao, S. (2023). Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36, 8634-8652. https://proceedings.neurips.cc/paper_files/paper/2023/file/1b44b878bb782e6954cd888628510e90-Paper-Conference.pdf

Shneiderman, B. (2022). Human-centered AI: ensuring human control while increasing automation. In: Proceedings of the 5th Workshop on Human Factors in Hypertext (pp. 1-2). https://doi.org/10.1145/3538882.3542790

Shukla, A. K., Terziyan, V., & Tiihonen, T. (2024). AI as a user of AI: Towards responsible autonomy. Heliyon, 10(11). https://doi.org/10.1016/j.heliyon.2024.e31397

Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., ... & Hassabis, D. (2017). Mastering the Game of Go without Human Knowledge. Nature, 550(7676), 354-359. https://doi.org/10.1038/nature24270

Simon, H. A. (2019). The Sciences of the Artificial. Reissue of the Third Edition. MIT Press. https://doi.org/10.7551/mitpress/12107.001.0001

Singla, A., Sukharevsky, A., Yee, L., Chui, M., & Hall, B. (2025). The state of AI: How organizations are rewiring to capture value. McKinsey & Company, 12. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google effects on memory: Cognitive consequences of having information at our fingertips. Science, 333(6043), 776‒778. https://doi.org/10.1126/science.1207745

Thakkar, A., Gupta, A., & De Sousa, A. (2024). Artificial Intelligence in Positive Mental Health: A Narrative Review. Frontiers in Digital Health, 6, 1280235. https://doi.org/10.3389/fdgth.2024.1280235

TOI (2025). Nvidia CEO ‘Trashes’ MIT Study Claiming AI Makes People Dumber, Says: My Advice to MIT Test Participants is Apply … The Times of India (July 14, 2025, 18:44) https://timesofindia.indiatimes.com/technology/tech-news/nvidia-ceo-trashes-mit-study-claiming-ai-makes-people-dumber-says-my-advice-to-mit-test-participants-is-apply/articleshow/122438099.cms

Ünver, H. A. (2024). Artificial intelligence (AI) and human rights: Using AI as a weapon of repression and its impact on human rights. Report for Directorate-General for External Policies, European Parliament. https://www.europarl.europa.eu/RegData/etudes/IDAN/2024/754450/EXPO_IDA%282024%29754450_EN.pdf

Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When Combinations of Humans and AI are Useful: A Systematic Review and Meta-Analysis. Nature Human Behaviour, 8(12), 2293–2303. https://doi.org/10.1038/s41562-024-02024-1

Wahn, B., & Schmitz, L. (2024). A bonus task boosts people's willingness to offload cognition to an algorithm. Cognitive Research: Principles and Implications, 9(1), 24. https://doi.org/10.1186/s41235-024-00550-0

Wahn, B., Schmitz, L., Gerster, F. N., & Weiss, M. (2023). Offloading under cognitive load: Humans are willing to offload parts of an attentionally demanding task to an algorithm. Plos One, 18(5), e0286102. https://doi.org/10.1371/journal.pone.0286102

Wallinheimo, A. S., Evans, S. L., & Davitti, E. (2023). Training in New Forms of Human-AI Interaction Improves Complex Working Memory and Switching Skills of Language Professionals. Frontiers in Artificial Intelligence, 6, 1253940. https://doi.org/10.3389/frai.2023.1253940

Wang, J., & Fan, W. (2025). The effect of ChatGPT on students’ learning performance, learning perception, and higher-order thinking: insights from a meta-analysis. Humanities and Social Sciences Communications, 12(1), 1‒21. https://doi.org/10.1057/s41599-025-04787-y

WEF (2025-a). The Future of Jobs Report 2025. World Economic Forum Reports. https://www.weforum.org/publications/the-future-of-jobs-report-2025/in-full/

WEF (2025-b). The Global Risks Report 2025. World Economic Forum Reports https://reports.weforum.org/docs/WEF_Global_Risks_Report_2025.pdf

Xie, H., Cheng, J., Tan, X., & Li, J. (2025). Artificial Intelligence Technology Applications and Energy Utilization Efficiency: Empirical Evidence from China. Sustainability, 17(14), 6463. https://doi.org/10.3390/su17146463

Yang, C., Pereira Nunes, B., & Rodríguez Méndez, S. (2025). LLM as Auto-Prompt Engineer: Automated NER Prompt Optimisation. In: Companion Proceedings of the ACM on Web Conference 2025 (pp. 2574-2578). https://doi.org/10.1145/3701716.3717818

Yang, H., Yue, S., & He, Y. (2023). Auto-GPT for online decision making: Benchmarks and additional opinions. arXiv preprint arXiv:2306.02224. https://doi.org/10.48550/arXiv.2306.02224

Zhang, J., & Fenton, S. H. (2024). Preparing healthcare education for an AI-augmented future. npj Health Systems, 1(1), 4. https://doi.org/10.1038/s44401-024-00006-z

Zhang, Q., Zeng, W., Lin, Q., Chng, C. B., Chui, C. K., & Lee, P. S. (2023). Deep reinforcement learning towards real-world dynamic thermal management of data centers. Applied Energy, 333, 120561. https://doi.org/10.1016/j.apenergy.2022.120561

Zhang, W., & Liu, X. (2025). Artificial Intelligence-Generated Content Empowers College Students’ Critical Thinking Skills: What, How, and Why. Education Sciences, 15(8), 977. https://doi.org/10.3390/educsci15080977

Zhou, Y., Muresanu, A. I., Han, Z., Paster, K., Pitis, S., Chan, H., & Ba, J. (2023). Large language models are human-level prompt engineers. arXiv preprint arXiv:2211.01910. https://doi.org/10.48550/arXiv.2211.01910