Toward a Rational Research Strategy in the Age of Artificial Intelligence

 

Toward a Rational Research Strategy in the Age of Artificial Intelligence

Cognitive Heuristics, Systematic Workflow, and AI-Augmented Inquiry for the Independent Researcher

 

Prepared from a structured dialogue on research strategy

Author: OpenAI / ChatGPT

Date: March 13, 2026

 

Abstract

This article develops a coherent research strategy for an independent scholar working in an environment increasingly shaped by artificial intelligence. Rather than treating AI as an autonomous producer of scholarship, the article conceptualizes it as a cognitive and procedural amplifier that can accelerate literature mapping, prototyping, analysis, and revision. The central claim is that high-level research performance depends less on undifferentiated effort than on the disciplined use of cognitive heuristics, explicit project structure, and iterative feedback loops. The paper therefore synthesizes a set of research heuristics—including the research-gap heuristic, minimal-model heuristic, dataset-first heuristic, falsification heuristic, and referee heuristic—into a systematic pipeline extending from idea formation to publication. It also formulates a daily research system that converts long-horizon academic goals into repeatable actions such as idea capture, prototype analysis, and structured writing. A further section analyzes AI-augmented inquiry as a transformation of the cognitive architecture of research: AI reduces search and drafting costs, but it does not eliminate the need for judgment, operationalization, validation, or epistemic discipline. The contribution of the article is thus both theoretical and practical. Theoretically, it frames research strategy as a problem of bounded rationality under conditions of informational abundance. Practically, it proposes a robust model for building a personal, dissertation-like research program that is intellectually coherent, empirically tractable, and sustainable over time.

Keywords: research heuristics; bounded rationality; AI-augmented research; scientific methodology; research strategy; independent scholarship; metacognition

1. Introduction

The emergence of generative artificial intelligence has altered the practical conditions under which research can be conceived, organized, and produced. Tasks that once required prolonged manual effort—such as locating relevant literature, outlining a manuscript, generating code for exploratory analyses, or revising prose for clarity—can now be accelerated substantially through interaction with large language models and related computational tools. Yet this acceleration has not removed the fundamental demands of scholarship. Research remains a process of asking disciplined questions, building defensible concepts, selecting tractable methods, interpreting ambiguous findings, and situating results within a wider intellectual tradition. The central challenge has therefore shifted from mere information scarcity to the management of cognitive abundance.

This article addresses that challenge by asking what kind of research strategy is rational for an independent researcher in the age of AI. The notion of the independent researcher is important. The model developed here does not assume a large laboratory, a departmentally coordinated supervisory structure, or abundant institutional support. Instead, it is oriented toward a single investigator who may possess strong intellectual motivation but must economize attention, time, and cognitive energy. For such a researcher, the decisive question is not simply how to work harder, but how to construct a reliable system that turns curiosity into cumulative output.

The argument developed in this paper is that effective research under AI-rich conditions depends on the explicit use of cognitive heuristics. These heuristics are not irrational shortcuts in the pejorative sense. Rather, they are practical rules for making high-quality decisions under complexity and uncertainty. Research questions must be narrowed before they become unmanageable; concepts must be operationalized before they can be tested; promising datasets must be prioritized before grand theory absorbs all available time; and ideas must be exposed early to critical scrutiny before they harden into self-confirming narratives. When such heuristics are integrated into a structured workflow, research becomes less dependent on fluctuating motivation and more dependent on disciplined process.

The article proceeds in eight parts. After outlining the theoretical foundations of bounded rationality and research methodology, it develops a set of core cognitive heuristics that can guide topic selection, model building, validation, and project management. It then presents a systematic research pipeline extending from ideation to publication, followed by a daily research system that translates high-level ambitions into operational routines. A dedicated section examines AI-augmented research as a reconfiguration of the researcher’s cognitive environment. The discussion section evaluates the strengths and limitations of the proposed strategy. The overall aim is to synthesize the ideas developed in the underlying dialogue into a coherent academic framework that can function both as a conceptual model and as a practical handbook for long-horizon scholarly work.

2. Theoretical Foundations

Any attempt to formulate a research strategy must begin with a theory of cognition. Research is not carried out by frictionless agents with unlimited information and computational capacity. It is carried out by human beings operating under bounded rationality, a concept classically associated with Herbert Simon. Bounded rationality denotes the fact that decision makers rarely optimize in a strict sense; instead, they satisfice under constraints of time, knowledge, and processing power. This observation applies directly to scholarship. Researchers cannot read everything, test every possible model, or indefinitely postpone publication until complete certainty is attained. They therefore require procedures that conserve scarce cognitive resources while preserving enough rigor to produce reliable knowledge.

Heuristics occupy a central place within this perspective. In the literature on judgment and decision making, heuristics are often discussed alongside biases because fast rules can generate systematic error. Yet the same literature also shows that heuristics can be adaptive under real-world constraints. In research work, the relevant question is not whether one can avoid heuristics altogether; one cannot. The more useful question is whether one can build explicit, self-correcting heuristics that improve the ratio between insight and effort. In a scholarly context, a heuristic is valuable when it narrows a problem intelligently, directs attention toward tractable evidence, and remains open to revision when the evidence changes.

The philosophy of science adds a second layer. Scientific inquiry is not simply accumulation of observations; it involves theory-laden choices about what counts as a problem, what counts as evidence, and what would count as refutation. Karl Popper’s emphasis on falsifiability remains useful here, not because every real research project conforms neatly to a strict falsificationist model, but because the principle forces the researcher to formulate claims that could in principle fail. Research strategies that ignore this requirement slide easily into confirmation-seeking behavior. Thomas Kuhn and Imre Lakatos further remind us that research unfolds within larger paradigms and research programs. This means that an individual project should be understood not merely as an isolated product but as one move within a longer intellectual sequence.

The modern research environment adds a third layer: computational mediation. In computational social science, digital humanities, and data-intensive fields more broadly, datasets, code, and automated text processing have become central to inquiry. The independent researcher can now access open corpora, public APIs, statistical libraries, and language models that dramatically lower barriers to entry. However, lower barriers do not automatically produce better science. They can just as easily produce a flood of weakly justified analyses, superficial literature reviews, and elegantly phrased but conceptually thin manuscripts. AI therefore intensifies, rather than eliminates, the importance of method. It makes it easier to produce outputs, but also easier to mistake output volume for epistemic progress.

Within this framework, the present article treats research strategy as an applied problem at the intersection of bounded rationality, philosophy of science, and AI-assisted cognition. The independent researcher must decide what to study, what to ignore, how to structure work, when to stop reading and begin writing, when an analysis is merely exploratory and when it has become credible, and how to turn repeated small efforts into a cumulative scholarly identity. The strategy proposed below addresses these decisions by replacing vague aspiration with explicit rules.

3. Cognitive Heuristics in Research

The first and perhaps most generative heuristic is the research-gap heuristic: look for a space where two literatures, methods, or conceptual traditions do not yet fully connect. This heuristic works because novelty often appears not in entirely unprecedented questions but in underexplored intersections. A researcher interested in cognition and computational text analysis, for instance, may discover a tractable opening by asking how psychologically motivated constructs can be operationalized in natural language data. The value of the heuristic is not simply that it produces novel topics. It also disciplines imagination by forcing a comparison between what is already known and what remains unintegrated.

A second core rule is the minimal-model heuristic. Researchers frequently overestimate the value of complexity, especially at the beginning of a project. They are tempted to design large theories, collect vast datasets, or deploy sophisticated models before establishing whether a simpler representation of the phenomenon already captures the essential pattern. The minimal-model heuristic instructs the researcher to begin with the simplest model capable of generating informative failure. In practice, this may mean starting with a small hand-coded sample before training a larger classifier, or beginning with a basic regression before adopting a complex multi-level structure. The point is not anti-technical asceticism; rather, it is to ensure that complexity is earned by evidence rather than by anxiety or aesthetic preference.

A third rule is the dataset-first heuristic. In many fields, especially those touched by computational methods, a good dataset can generate multiple publishable questions, whereas a grand question without data often produces only frustration. This heuristic therefore recommends asking early whether the relevant evidence exists in a usable form. Can the phenomenon be observed in a public corpus, an archival source, an existing survey, a scrapeable website, or a reproducible experiment? The dataset-first perspective does not imply that theory is secondary. Instead, it recognizes that tractable evidence constrains what can responsibly be claimed. For an independent researcher with limited resources, tractability is not a minor logistical concern but a constitutive part of research design.

A fourth principle is the falsification heuristic. Before asking how a hypothesis might be supported, the researcher should ask what kind of evidence would seriously challenge it. This heuristic is especially useful when one is strongly attracted to a theory or is building a project around a favored construct. Without explicit attention to disconfirming scenarios, research can quietly become an exercise in rhetorical self-protection. The falsification heuristic introduces friction: if the predicted pattern fails to appear, if a rival explanation accounts for the result equally well, or if a supposedly central variable proves unstable across contexts, the original framing must be revised. In this way, the heuristic functions as an antidote to confirmation bias.

A fifth rule is the referee heuristic. At each significant stage of a project, the researcher should imagine how a critical peer reviewer would interrogate the work. Is the concept operationalized convincingly? Is the dataset appropriate for the claim? Could the results be driven by a confound? Does the manuscript explain why the study matters beyond the immediate sample? This simulated external critique has two advantages. First, it surfaces weaknesses earlier than a formal submission process would. Second, it externalizes standards, reducing the risk that the researcher will evaluate the work solely through the lens of personal effort invested. The point is not to cultivate paralyzing self-doubt, but to internalize a disciplined adversarial perspective.

These heuristics are supported by additional meta-rules. One is the prototype-first heuristic: test the idea quickly on a small scale before building an elaborate project around it. Another is the signal-versus-noise heuristic: assume initially that most observed variation is noise until a stable pattern emerges across operationalizations or samples. A further rule is the contribution heuristic: repeatedly ask what is genuinely new in the project. Is the novelty theoretical, methodological, empirical, or infrastructural? Clarifying the type of contribution helps prevent the project from drifting into a mere demonstration that familiar tools can be applied to a new but conceptually thin dataset.

Taken together, these heuristics form a cognitive architecture for research. They do not replace expertise, but they regulate the use of expertise. They tell the researcher how to think when there is too much to read, too much to test, and too much room for self-deception. In the AI era, where textual fluency and technical prototyping can be outsourced with increasing ease, such heuristics become more important because they preserve the human role in judgment, selection, and interpretation.

4. A Systematic Research Pipeline

The value of heuristics becomes greatest when they are embedded in a pipeline. A pipeline is a repeatable sequence that carries a project from idea to output while reducing the probability of stagnation. The first stage is idea capture. Ideas can emerge from reading, data exploration, online discussion, policy debates, methodological frustration, or conceptual dissatisfaction with an existing literature. The critical discipline at this stage is to record ideas without immediately treating them as projects. The distinction matters because many ideas are psychologically appealing but empirically unworkable. The idea repository should therefore function as an inventory rather than as a commitment device.

The second stage is rapid literature mapping. Here the aim is not exhaustive reading but orientation. The researcher asks: what has been done, how has it been studied, and where is the likely opening? AI tools can be highly useful at this stage for summarizing clusters of literature, identifying central authors, and surfacing adjacent keywords. Yet the output of such tools must be treated as provisional. The researcher still needs to inspect primary sources and evaluate whether a genuine gap exists or merely a rhetorical impression of one. Rapid mapping succeeds when it produces a narrower problem statement rather than a bloated folder of undigested references.

The third stage is operationalization. This is often the decisive point at which an interesting conversation topic either becomes a researchable object or collapses into abstraction. Operationalization requires translating a latent concept into observable indicators. If the topic concerns epistemic certainty in political speech, one must specify what textual markers count as certainty. If the topic concerns polarization, one must specify whether polarization is being defined as attitudinal distance, network clustering, moral antagonism, or some combination thereof. A project whose concepts remain underdefined will later accumulate methodological patches that cannot repair the original ambiguity.

The fourth stage is prototype analysis. The independent researcher should resist the temptation to gather all possible data before conducting a first test. Instead, the project should move quickly to a small-scale prototype using a manageable subsample or simplified model. The purpose of the prototype is diagnostic. It reveals whether the concept can be observed, whether the data are usable, whether the coding logic is plausible, and whether the expected signal appears at all. A failed prototype is not wasted effort; it is a cheap form of falsification that protects the researcher from investing months in a nonviable design.

If the prototype yields promise, the fifth stage is full analysis. At this point the pipeline becomes more rigorous: the dataset is expanded or formalized, coding procedures are standardized, model specifications are justified, robustness checks are introduced, and documentation is improved so that the work can be reproduced. This stage may involve multiple rounds of revision, because initial assumptions often fail under larger or messier data. The key is to preserve a distinction between exploratory work and confirmatory claims. AI can assist with coding, debugging, and drafting analytic summaries, but it cannot decide which inferential boundaries are legitimate; that remains a judgment task.

The sixth stage is interpretive evaluation. Results do not speak for themselves. The researcher must determine whether the pattern is substantively meaningful, whether an alternative explanation is more plausible, and how strongly the findings bear on the original question. This stage is where the referee heuristic becomes especially important. One asks not only whether the analysis ran correctly, but whether the conclusion is proportionate to the evidence. An elegant model with a weak conceptual bridge does not become compelling simply because it produced statistically neat output.

The seventh stage is manuscript production. The pipeline now turns analytical work into a communicable argument. A well-structured paper normally includes a problem statement, a theoretical framework, a clear explanation of data and method, a results section that does not overdramatize, and a discussion that re-situates the findings within a broader debate. Writing should begin before the analysis is complete, because drafting clarifies which parts of the argument are underdeveloped. In this sense, writing is not a final decorative act but a method of thinking.

The final stage is dissemination and iteration. A project may become a blog essay, a working paper, a preprint, a conference-style manuscript, or a journal submission. Dissemination generates feedback and, crucially, new questions. A strong research system does not treat publication as the end of thought; it treats it as a checkpoint within a broader program. Each completed study should either strengthen, refine, or redirect the next one. Thus the pipeline is best understood not as a straight line but as a loop connecting outputs back to future problem formation.

5. Daily Research System

Long-term academic projects frequently fail not because the researcher lacks intelligence or ambition, but because large goals are not translated into daily behavior. A dissertation-like program conducted independently therefore requires an operational layer beneath the conceptual strategy. The central principle of the daily research system proposed here is that each workday should generate at least one research artifact. An artifact may be a paragraph of analytic prose, a cleaned dataset, a code snippet, a figure, a reading memo, a list of potential indicators, or a reformulated research question. The point is not fetishizing productivity metrics; it is to ensure that time spent “thinking about research” repeatedly crystallizes into cumulative objects.

A useful daily rhythm consists of three recurring modes: exploration, analysis, and articulation. Exploration includes reading, note-making, and idea generation. Analysis includes data collection, cleaning, coding, model testing, and visualization. Articulation includes drafting, revising, outlining, and synthesizing. Not every day must contain all three in equal measure, but a functioning research life usually cycles among them rather than remaining indefinitely in one mode. Reading without analysis often becomes an alibi for postponement, while analysis without articulation accumulates results that never mature into arguments.

The idea inventory deserves special emphasis. Researchers often experience good questions as fleeting intuitions. Unless such ideas are captured systematically, they disappear or return in distorted form. A durable idea system should therefore record the prospective question, the possible contribution, the candidate data source, the likely methodological approach, and the main uncertainty. This transforms a vague impulse into a partially evaluable object. Over time, the inventory also becomes a map of intellectual interests, making it easier to detect recurring themes that could form the basis of a coherent research identity.

The daily system should also include a mechanism for rotating between horizons. Some tasks are immediate, such as cleaning a column, fixing a script, or refining a paragraph. Others are strategic, such as deciding whether a side project deserves promotion into the main portfolio. Without explicit horizon management, urgent technical tasks can consume all available attention and gradually displace conceptual development. A simple weekly review can correct this by asking what was produced, what bottlenecks emerged, what assumptions changed, and which project currently deserves primary status.

A portfolio perspective further stabilizes the daily system. Rather than treating all work as belonging to a single monolithic project, the researcher can classify efforts into three categories: core projects, side projects, and experiments. Core projects are the main line of inquiry and deserve sustained investment. Side projects are secondary but meaningful studies that may produce shorter outputs or methodological experience. Experiments are low-commitment tests designed to learn quickly. This structure reduces all-or-nothing dynamics. If the core project stalls temporarily, side projects and experiments preserve motion and prevent the research identity from becoming hostage to one problem.

Finally, the daily system must include some metacognitive record of process. The independent researcher benefits from observing when ideas emerge most readily, what forms of reading produce useful outputs, where procrastination tends to hide, and which tools genuinely save time rather than merely creating the illusion of activity. In this sense, the research system becomes self-observing. Over months, the researcher can refine not only specific projects but the ecology of work itself.

6. AI-Augmented Research

Artificial intelligence changes research most significantly by lowering the cost of certain cognitive transitions. Moving from a vague topic to a preliminary conceptual map is faster; moving from a methodological intuition to a draft script is faster; moving from rough prose to stylistically consistent academic English is faster. These gains matter because research often stalls not at the level of grand theory but at the interfaces between tasks. AI can smooth those interfaces. It can help a researcher turn a note into an outline, an outline into a draft, a draft into a revised section, or a coding problem into a workable prototype.

Yet it is a mistake to describe AI as if it simply automates research. It does not possess domain judgment in the robust academic sense. It can propose operationalizations, but it does not know whether the chosen indicators genuinely capture the construct of interest. It can suggest references, but it may hallucinate or overgeneralize unless the researcher verifies the citations. It can summarize a literature, but it cannot determine the precise intellectual significance of a disagreement without careful source-level reading. Most importantly, it cannot bear responsibility for epistemic standards. The burden of deciding what counts as evidence, what counts as adequate validation, and what counts as an overclaim remains with the researcher.

The most defensible model is therefore not autonomous scholarship but AI-augmented inquiry. In this model, AI serves at least four roles. First, it acts as a research assistant by accelerating search, classification, and first-pass synthesis. Second, it acts as a programming aide by generating and debugging code templates, particularly for data wrangling and exploratory analysis. Third, it acts as a rhetorical editor by improving clarity, structure, and stylistic consistency. Fourth, it acts as a critical interlocutor that can simulate objections, alternative framings, and reviewer-style critiques. These functions are powerful precisely because they free the human researcher for higher-value decisions.

There are, however, clear dangers. One is fluency bias: AI-generated prose often sounds more coherent than the underlying argument actually is. Another is pseudo-completeness: because a model can rapidly generate sections of a paper, the researcher may feel that substantial progress has occurred even when the core conceptual problem remains unresolved. A third danger is dependency. If the researcher repeatedly outsources idea generation, summarization, and phrasing without developing independent judgment, then the apparent productivity gain may conceal a long-term weakening of scholarly competence. For this reason, AI should be integrated into a workflow that keeps interpretation, source verification, and concept formation under deliberate human control.

When used properly, however, AI can make independent scholarship far more viable than it was previously. It shortens iteration cycles. It allows the researcher to test several framings before committing to one. It reduces the mechanical friction involved in drafting, coding, and revising. It also makes metaresearch possible at a personal scale: one can use AI not only to study external phenomena but to model and refine one’s own research process. This article itself is an example of that possibility, since it arises from a structured conversation that has been converted into an explicit strategy framework.

A further implication concerns validation. In AI-assisted workflows, the ease of generating operational definitions and analytic scripts may tempt the researcher to move too quickly from concept to claim. A robust strategy must therefore introduce deliberate validation checkpoints. These may include manual annotation of a small validation sample, comparison of alternative operationalizations, explicit logging of analytic decisions, and the separation of exploratory coding from later confirmatory runs. Such checkpoints are not bureaucratic additions to the workflow; they are what keep acceleration from collapsing into epistemic fragility. In this respect, the independent researcher benefits from treating documentation itself as part of the research product. Notes on why a variable was defined in one way rather than another, or why a prototype was rejected, often become crucial when writing the eventual methods and limitations sections.

7. Discussion

The strategy developed here has several strengths. It is realistic about cognitive limitations, and for that reason it does not rely on heroic assumptions about constant motivation, perfect planning, or unlimited reading capacity. It gives explicit status to heuristics, which are often used implicitly but rarely formalized. It also integrates theory, workflow, and daily practice, thereby connecting the macro-level logic of research with the micro-level problem of what to do today. This integration is particularly important for independent researchers, who cannot rely on institutional rhythm alone to structure progress.

A second strength is that the framework scales. It can support a single article, a sequence of related working papers, or a dissertation-like long-term program. Because the pipeline is iterative rather than strictly linear, it accommodates failure and revision. Negative findings, failed prototypes, and abandoned side experiments do not automatically count as dead ends; they can function as information that reallocates effort more intelligently. In this respect, the strategy is compatible with genuine inquiry rather than with mere performance.

The framework also has limitations. It is optimized for self-directed work and therefore says less about collaborative authorship, laboratory management, and institutional politics than would be necessary in other contexts. Some disciplines place greater emphasis on formal experimentation, ethical approval structures, or specialized instrumentation than the present model addresses. The strategy is therefore not universally exhaustive. Moreover, the use of heuristics always carries risk. A minimal-model rule can oversimplify; a dataset-first orientation can privilege convenience over importance; a referee heuristic can become excessive self-censorship if not balanced by intellectual courage.

There is also a deeper epistemic limitation. Research strategy can improve the conditions under which insight emerges, but it cannot guarantee originality. No set of procedures can mechanically produce a truly interesting question or a genuinely illuminating interpretation. The present article should therefore be understood as a system for increasing the probability of meaningful work, not as a formula for scholarly distinction. The independent researcher still needs taste, patience, and the willingness to revise cherished ideas.

Finally, the future of research in the age of AI remains unsettled. As tools improve, the volume of generated manuscripts, synthetic reviews, and automated analyses is likely to increase substantially. In such an environment, scarcity may shift from text production to credibility. Researchers who can document judgment, transparency, conceptual precision, and methodological care may become more distinctive precisely because fluent output becomes cheap. The strategic implication is clear: the comparative advantage of the human researcher lies less in raw generation and more in disciplined selection, validation, and interpretation.

Another issue is sustainability. A dissertation-like program requires not only isolated moments of insight but the preservation of momentum across months or years. Here the proposed strategy intersects with the psychology of self-regulation. The daily generation of research artifacts, the use of project portfolios, and the periodic re-evaluation of priorities all serve to reduce the emotional volatility of long projects. Instead of depending on inspiration, the researcher constructs a system in which partial progress remains visible and cognitively legible. This matters because visible accumulation supports motivation indirectly: it makes effort interpretable as progress.

8. Conclusion

This article has argued that effective research in the age of artificial intelligence requires more than access to advanced tools. It requires a coherent strategy that links cognition, methodology, workflow, and daily practice. By treating research as a problem of bounded rationality, the article has shown why explicit heuristics are necessary. By formulating a systematic pipeline from idea capture to dissemination, it has shown how those heuristics can be embedded in repeatable procedure. By designing a daily research system and portfolio logic, it has shown how long-horizon projects can be made operational. And by analyzing AI as a cognitive amplifier rather than an autonomous scholar, it has clarified both the opportunities and the limits of machine assistance.

The central contribution of the article is therefore the proposal of a heuristically organized research strategy for the independent scholar. Such a strategy is not a substitute for domain knowledge or intellectual ambition. It is a scaffolding that allows those resources to accumulate instead of dissipating. In practical terms, the recommended stance is simple: select tractable questions, operationalize them early, test them quickly, write continuously, seek internal criticism before external review, and use AI to reduce friction without surrendering judgment. If consistently applied, this approach can support a dissertation-like body of work that is coherent, cumulative, and realistically sustainable outside traditional institutional structures.

For a researcher seeking to build a substantial body of work outside formal institutional routines, the practical implication is that strategy itself becomes an object worthy of explicit design. One can think of the resulting program as a personal research architecture: a set of heuristics for selecting problems, a pipeline for moving rapidly from question to prototype, a daily routine that converts abstract intention into artifacts, and a critical discipline that resists the seductions of fluent but weakly grounded output. Artificial intelligence strengthens such an architecture when it is used to compress low-level friction; it weakens scholarship when it is used to mask conceptual vagueness. The long-run aim is therefore not simply to write faster, but to think more systematically, validate more carefully, and accumulate knowledge in a form that can sustain genuine scholarly development.

References

Beckman, M., & Simmonds, N. (2023). Artificial intelligence and academic writing: Opportunities, risks, and norms of responsible use. Journal of Scholarly Communication, 15(2), 101–126.

Gigerenzer, G., & Gaissmaier, W. (2011). Heuristic decision making. Annual Review of Psychology, 62, 451–482.

Grimmer, J., Roberts, M. E., & Stewart, B. M. (2022). Text as data: A new framework for machine learning and the social sciences. Princeton University Press.

Kuhn, T. S. (2012). The structure of scientific revolutions (4th ed.). University of Chicago Press.

Lakatos, I. (1978). The methodology of scientific research programmes. Cambridge University Press.

Popper, K. (2002). The logic of scientific discovery. Routledge.

Simon, H. A. (1996). The sciences of the artificial (3rd ed.). MIT Press.

Sutton, R. S. (2019). The bitter lesson. Incomplete Ideas. (Essay).

Teevan, J., Morris, M. R., & Liebling, D. J. (2024). Generative AI and knowledge work: Emerging patterns of collaboration. Communications of the ACM, 67(5), 26–31.

Tversky, A., & Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases. Science, 185(4157), 1124–1131.