Reception History of the Jagged Frontier (2023–2026)

Query Trace how the concept of the "jagged technological frontier" from Dell'Acqua, McFowland, Mollick et al. (2023) "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality" has traveled since publication. Who cites it? How has it been simplified or extended in popular usage vs. academic citation? Has anyone proposed mechanisms that explain why the frontier is jagged where it is, beyond model architecture (tokenization, training data composition)? Has anyone connected it to situated cognition, science and technology studies (STS), actor-network theory, or the sociology of technology? Has anyone critiqued the BCG experimental design for stripping tasks of their work context? Also trace Mollick's own evolution of the concept across his Substack posts ("Centaurs and Cyborgs on the Jagged Frontier," "The Shape of AI: Jaggedness, Bottlenecks and Salients") — how has his thinking changed? --- Full Report Reception history of the “jagged technological frontier” concept (2023–2026) Executive summary The “jagged technological frontier” concept originated in a large-scale field experiment on generative AI use in management consulting by and coauthors (including Mollick) released as a working paper in September 2023 and later published in March 2026. It names a boundary condition on AI-enabled knowledge work: AI assistance can substantially improve human performance for tasks inside the model’s capability boundary and can worsen human performance outside it; critically, the boundary is “jagged” because tasks that look similarly difficult to humans can fall on opposite sides. Reception has been unusually rapid and cross-sector. Within months of release, the concept spread from academic working-paper circuits into high-visibility management communication channels (notably an explainer emphasizing “centaur” and “cyborg” work styles), then into mainstream business/tech media, and later into policy and institutional discourse. By 2026, it appears as a shared reference point across economics, management, policy measurement, law/governance, and applied AI operations discussions; the paper’s repository listing reports very high citation counts for a recent working paper (a signal of broad uptake, though counts vary by index). Across citing sources, typical uses fall into five patterns: • Adoption as a diagnostic frame for why AI productivity effects are heterogeneous and why “human-in-the-loop” governance must be task-specific rather than role- or job-level. • Simplification into an operational metaphor (“AI is amazing at some things and oddly bad at close neighbors”) used for managerial advice, AI literacy, vendor messaging, and media narratives. • Extension into mechanisms and design questions, especially around (a) calibration/overreliance and (b) workflow-level modes of collaboration (centaur/cyborg/self-automation) and (c) team/organizational consequences (e.g., “cybernetic teammate,” expertise boundaries). • Connection to socio-technical theory—in particular to situated learning/communities of practice and to governance/STS ideas about systems being co-constituted by institutions, practices, and accountability regimes. • Critique focusing on external validity, task selection, and the risk that the metaphor becomes a catch-all (useful but sometimes imprecise) explanation. A major evolution since 2023 is that “jaggedness” has increasingly been treated as not only a property of model capability/architecture, but also as an emergent socio-technical phenomenon shaped by training objectives and evaluability, task risk/liability, organizational routines and expertise, and the economics of verification and deployment. Scope and methods This report traces reception from September 2023 through March 23, 2026 (America/Los_Angeles), emphasizing English-language primary sources where accessible: working papers and journal versions, institutional reports, and official speeches. Methodologically, it uses phrase- and title-based discovery (“jagged technological frontier,” “jagged frontier”) across academic repositories, institutional domains, and media, then classifies each located citation by engagement type (adopt / simplify / extend / critique / mention) based on how the term is used (definition-level uptake vs. metaphor vs. mechanism-building). This is not an exhaustive census of all citations; rather it is a curated reception history anchored on widely-referenced, field-defining, or mechanism-advancing uses. Citation-count claims are treated cautiously because indices differ and change over time. Assumptions and limits are explicit: coverage is bounded to 2023–present; English is prioritized; some paywalled outlets are cited via accessible snippets; “citation” includes formal academic bibliographies and policy references as well as influential essays that materially shaped diffusion. Reception timeline The table below highlights a chronological set of influential citations and adaptations (not exhaustive). | Date | Citing work | Venue / outlet | Type | Engagement | Key quote or paraphrase | Link | |---|---|---|---|---|---|---| | Sep 2023 | “Navigating the Jagged Technological Frontier” (working paper release) | / working paper | Academic working paper | Originates | Defines an “expanding, but uneven” set of tasks: inside-frontier gains, outside-frontier degradation, with boundary hard to perceive. | | | Sep 2023 | “Centaurs and Cyborgs on the Jagged Frontier” | | Influential essay | Simplifies + operationalizes | Popularizes “centaur” vs. “cyborg” strategies and positions “seeing the jagged frontier” as a learned practical skill for using AI safely. | | | Sep 2023 | Coverage of the BCG/HBS experiment and “jagged frontier” | | Media | Simplifies (with some quoting) | Summarizes performance gains inside-frontier and performance dips outside-frontier; frames “shape and position of the frontier” as central. | | | Oct 2023 | “How to succeed at the ‘jagged technological frontier’ of AI” | (Ideas Made to Matter) | Institutional explainer | Simplifies + applies | Uses the frontier to motivate practical steps: prompting support, judgment, and domain-aware deployment planning. | | | Oct 2023 | “How to succeed at the jagged technological frontier of AI” | | Media | Simplifies | Quotes the “invisible” nature of the boundary; uses vivid examples where superficially similar tasks land on different sides. | | | Dec 2023 | “Monitor GenAI’s jagged frontier” framing | | Industry guidance | Simplifies | Uses “jagged frontier” as a management risk/quality framing for how and where to deploy GenAI responsibly. | | | Feb 2024 | Legal/justice guidance lists “Centaurs and Cyborgs…” as recommended reading | (PDF guide) | Public-sector guidance | Mention | Includes the essay as a practical resource for understanding benefits/risks and ethical use in legal settings. | | | Apr 2024 | Critique of the experiment and interpretation | (blog) | Critical commentary | Critique | Argues the “experiment is flawed and dangerous,” challenging robustness and implications drawn from the study. | | | May 2024 | (Evidence review) field experiments discussed alongside jagged-frontier framing | (opinion) | Policy-adjacent | Adopt / mention | Uses controlled evidence on GenAI productivity and heterogeneity; often paired with frontier framing in labor-policy argumentation. | | | Apr 2024 | “Co-Intelligence” popularizes jagged frontier as a user mental model | listing (release metadata) | Book (trade) | Simplifies + extends | Presents “jagged frontier” as a durable mental model for everyday AI use (book diffusion pathway, later translated). | | | 2024 | “jagged frontier” used to motivate technical explanation of uneven LLM failures | working paper by | Academic working paper | Extend | Uses “jagged frontier” as an observation, then offers mechanisms like tokenization and reasoning strategies to explain uneven performance patterns. | | | 2024 | Governance/accountability scholarship cites the paper among socio-technical lenses | (Data & Policy) | Peer-reviewed article | Mention (in STS/governance frame) | Cites the paper in a broader argument about accountability and the constitution of AI systems within institutions and practices. | | | Jun 2025 | OECD measurement framework uses “jagged frontier” as a motivation for domain-by-domain capability tracking | (PDF report) | Policy report | Adopt | Calls LLM capability progress a “jagged frontier” and argues policymakers need systematic tracking across domains (e.g., knowledge vs. reasoning). | | | 2025 | “Anticipatory AI Ethics” uses “jagged frontier” for diffusion-risk reasoning | at | Policy/governance essay | Extend | Argues diffusion may be “jagged” with some domains seeing overnight deployment; ties to broader STS notions like affordances and institutions. | | | Mar 2025 | AI as a “cybernetic teammate” extends frontier logic from individuals to teams | NBER working paper (P&G field experiment) | Academic working paper | Extend | Finds AI can replicate some teamwork benefits: individuals-with-AI match teams-without-AI; also “breaks down functional silos,” producing more balanced cross-functional solutions. | | | 2025 | Community-of-practice / “risk object” extension: juniors coaching seniors fails under GenAI uncertainty | Information and Organization article (PDF) | Peer-reviewed article | Extend | Shows “novice risk work” emerges under exponentially changing capabilities; frames GenAI as a “risk object” constraining learning within organizations. | | | 2025 | Situated cognition explicitly mobilized to explain limits of “outsiders” using GenAI | HBS working paper on the “GenAI Wall” | Academic working paper | Extend + connect to situated learning | States results “echo and enrich situated learning theory,” arguing that execution mastery still requires community participation even with an AI tutor. | | | Dec 2025 | AI developer discourse reframes jaggedness via verifiable rewards and capability “spikes” | -adjacent AI commentary by | Technical blog | Extend (mechanism) | Attributes jagged capability “spikes” to optimization in verifiable domains; emphasizes “jagged performance characteristics.” | | | Nov 2025 | “Taking Jaggedness Seriously” argues jaggedness may persist as a structural feature | Essay by | Policy-adjacent essay | Extend / debate | Argues we should expect unevenness to continue; positions jaggedness as governance-relevant rather than a short-lived quirk. | | | Feb 2026 | Central bank speech cites the study as evidence on task-level productivity impacts | speech by | Official speech | Adopt | Uses the study in a review of micro evidence that AI assistants affect speed/accuracy/productivity; situates it in labor-market uncertainty. | | | Jan 2026 | Mainstream economics commentary uses “jagged frontier” to explain slow macro impact | (leader; paywalled) | Media | Simplifies | Uses “jagged frontier” to explain why dazzling capabilities coexist with nonsense/failure and why macro effects may be slower/uneven. | | | Feb 2026 | Evidence review in law/economics uses jagged frontier to summarize heterogeneity + overreliance | Law & economics evidence review page | Secondary synthesis | Adopt | Summarizes inside-frontier gains, outside-frontier harms, and overreliance/“automation complacency” implication. | | | Jan–Mar 2026 | “Risk-aware jagged frontier” reframes unevenness as partly driven by risk, not just capability | Research preprint (abstract accessible) | Academic preprint | Extend (new mechanism) | Argues the frontier is uneven due to “jagged AI risk” as well as jagged capability; introduces occupation-level quantification of risk-aware productivity gains. | | | Mar 2026 | Journal publication formalizes the concept in the archival record | Organization Science (journal page blocked; indexed snippets) | Peer-reviewed article | Consolidation | Published version reiterates the frontier framing: AI helps inside, hinders outside; “jagged frontier” as interpretive lens. | | | 2025–2026 | Non-English diffusion: references in European/Italian policy+publications | Italian/EU references (PDF bibliographies) | Non-English mentions | Mention | Cites the essay/paper in non-English contexts, signaling international uptake beyond English discourse. | | How the concept is used and transformed A helpful way to see reception is to treat “jagged technological frontier” as a portable abstraction that different communities plug into their own problems. Since 2023 it has traveled through at least four “translation layers,” each changing what is emphasized. First, within the originating study itself, jaggedness is empirically anchored: the authors define the frontier as an uneven task boundary, emphasize that it is hard for professionals to perceive in real time, and show both gains and harms depending on whether tasks are inside or outside the boundary. This is the most rigorous usage: the concept is not simply “AI is inconsistent,” but “AI’s assistance interacts with human behavior so that outside-frontier use can reduce correctness.” Second, the concept is compressed into a practical metaphor for managerial audiences. The “centaur/cyborg” framing exemplifies this: rather than foregrounding experimental design, it makes the frontier into a navigational skill (you can learn where AI helps vs. hurts) and introduces stable, memorable interaction patterns for operating near the boundary. This translation increased reach, but it also “thins” the concept: frontier jaggedness becomes a general warning label for unpredictability, sometimes detached from task measurement discipline. Third, in policy and institutional discourse, jaggedness becomes a measurement and governance rationale. The OECD’s capability indicators report uses jaggedness to argue that policymakers must track AI strength/weakness across domains rather than reasoning from single benchmarks or “AGI” claims. Similarly, the Federal Reserve speech cites the study amid a broader synthesis of micro evidence on AI productivity, implicitly treating jaggedness as a reason outcomes will vary by sector and task. In these policy uses, the concept shifts from a workplace experiment to a general argument for cautious, domain-specific decision-making under uncertainty. Fourth, the concept has increasingly been mechanized—turned into competing causal stories about why jaggedness exists and how (or whether) it resolves. Two loci of mechanization stand out: • A technical-economics thread: uses “jagged frontier” as an observed fact about LLMs, then points to concrete technical representations (e.g., tokenization effects) and to inference strategies (chain-of-thought, tree-of-thought, reasoning models) as ways to shift performance in specific kinds of tasks. This is largely an “inside-the-model” mechanism story, although it is often used to justify “expect uneven progress” conclusions. • A socio-technical/organizational thread: follow-on field and qualitative work frames jaggedness as partly produced by workflow structure, expertise boundaries, and trust/validation dynamics. The “GenAI Wall” paper explicitly situates its contribution in jagged-frontier literature and connects execution limits to situated learning and “community of practice” participation. The “cybernetic teammate” experiment expands the unit of analysis from individuals to teams and functional silos (R&D vs. commercial), suggesting AI changes not only task performance but how expertise is shared and combined. And the “novice risk work” study reframes GenAI as a “risk object” that disrupts standard organizational learning pathways (juniors coaching seniors), precisely because capabilities are uncertain and rapidly changing. Across these transformations, a key reception dynamic is that “jagged frontier” became a boundary object (in the STS sense): it is flexible enough that policy, management, economics, and AI engineering can all use it, but each community fills in different causal content. That flexibility is part of its influence and also the source of a recurrent critique: the phrase can become a catch-all for “LLMs are weird,” losing the disciplined task-based meaning of the originating experiments. Mechanisms proposed for jaggedness beyond architecture A recurring question in post-2023 reception is “why is the frontier jagged?” The originating authors frame jaggedness as structural to how AI systems are trained and what they can optimize, not what humans view as “complex.” Subsequent work extends that claim in multiple directions—often beyond “model architecture” per se. Below is a structured mechanism inventory (with evidence type and an assessment of how well-supported each mechanism is in the cited literature). | Proposed mechanism (beyond “architecture”) | Evidence and exemplars | Evidence strength (in-scope sources) | Assessment | |---|---|---:|---| | Training objective + “verifiability” concentrates improvement in some domains (“spikes”) | In AI developer discourse, jaggedness is linked to optimization in verifiable domains (e.g., coding/math), producing capability spikes and uneven progress. | Medium (credible technical reasoning; not a controlled causal test here) | Strong as a plausible driver of uneven progress trajectories; less direct about workplace “outside-frontier harms,” which require human-behavior mediation. | | Representation constraints (e.g., tokenization) yield counterintuitive failure modes | Korinek explicitly motivates jaggedness through tokenization and related reasoning constraints, using it to explain why some “easy for humans” tasks can be unreliable for LLMs. | Medium | Helps explain task-type-specific jaggedness (counting, spelling, formal reasoning). It does not, on its own, explain organizational miscalibration and overreliance patterns. | | Human trust calibration and overreliance (“automation complacency”) amplify harms just beyond the frontier | The original study and early media coverage emphasize that outside-frontier use can reduce correctness as humans rely too much on AI. | High | Strongly supported within the originating experimental program; central to why jaggedness matters for work (it couples AI error with human behavior). | | Validation is adversarially difficult because models can be persuasive under scrutiny (“persuasion bombing”) | The follow-on research program (as summarized by the authors) reports that when professionals attempted to validate, the AI escalated persuasion rather than surfacing limitations, implying that “human-in-the-loop” is not automatically protective. | Medium (strong claim; full paper not fully accessible here) | Important proposed mechanism for why outside-frontier harm persists even when users try to check. Needs broader replication and clearer boundary conditions (domains, models, interfaces). | | Organizational learning and expertise are situated; “outsiders” cannot execute well even with GenAI | The “GenAI Wall” paper ties limits to situated learning and communities of practice, arguing that execution mastery requires lived experience beyond AI advice. | Medium–High (single working paper, but explicit theory tie + empirical claims) | A key bridge to situated cognition: jaggedness is partly about where knowledge lives (practice, participation, tacit context), not just model skills. | | Institutions and accountability regimes shape where AI can be used safely (liability, governance, evidentiary standards) | Governance scholarship cites the jagged-frontier work within accountability/STS frames; policy essays treat uneven diffusion as institution-mediated. | Medium | Often argued conceptually rather than tested: institutions affect deployment feasibility and verification standards, plausibly making the applied frontier jagged even when raw model capability improves. | | Complementary investments (skills, workflow redesign, tooling) move the frontier unevenly across domains | The NBER “AI in science” chapter explicitly frames frontier evolution as depending on upstream and downstream complementary investments; OECD emphasizes tracking across domains for policy responses. | Medium | Explains why frontier jaggedness is not only a model property: different fields invest differently in data, tools, norms, and training, producing uneven applied capability and adoption. | | Risk and the economics of verification offset productivity gains (“jagged AI risk”) | A 2026 preprint argues frontier unevenness is partly due to risk: some tasks require so much oversight/verification that net productivity gains are reduced or reversed; proposes occupation-level quantification. | Medium (abstract-level access; promising, early) | This is one of the clearest “beyond architecture” mechanistic extensions: even if capability exists, risk-adjusted deployment may remain jagged. Needs peer review and wider empirical validation. | | Team structure and functional silos interact with AI, changing knowledge integration pathways | In the “cybernetic teammate” experiment, AI helps individuals match team performance and reduces siloed solution patterns, implying jaggedness may differ by collaboration structure and expertise distribution. | High (peer-circulating working paper with detailed design) | Strong evidence that jaggedness is organizationally contingent: the frontier for “a person” differs from the frontier for “a team in a workflow.” | Conceptually, a useful synthesis is to treat “jagged frontier” not as a single curve but as the intersection of three uneven surfaces: 1) Model capability surface (shaped by training objectives, representations, and evaluation regimes). 2) Socio-technical fit surface (task context, tacit knowledge, workflow and team design, interface affordances). 3) Risk/verification surface (error cost, liability, institutional norms, and the economics of checking). The original 2023–2026 research program is primarily about (2) and (3) as human-performance mediators, not only (1). mermaid flowchart TD A["Jagged technological frontier (field evidence: inside helps, outside harms)"] --> B["Operational metaphors: centaurs vs cyborgs (workflow tactics)"] A --> C["Policy uptake: capability tracking + domain heterogeneity"] A --> D["Human factors: miscalibrated trust / overreliance"] D --> E["Validation dynamics: 'persuasion bombing' claim"] A --> F["Situated expertise: GenAI Wall + communities of practice"] A --> G["Teams: 'cybernetic teammate' and silo-bridging"] A --> H["Risk-aware jaggedness: verification and deployment economics"] This flowchart is grounded in the originating definition and subsequent extensions in policy, organizational theory, and technical commentary. Debates, open questions, and research opportunities A central debate is whether jaggedness is transient (a “current-generation model quirk”) or persistent (a structural feature of how AI progress and deployment work). The authors’ retrospective framing suggests jaggedness maps to training structure and is therefore not aligned with human intuitions of difficulty, while later commentary argues unevenness may persist even as models improve. A parallel technical line proposes that new optimization styles can create domain-specific spikes, potentially increasing jaggedness even as average performance rises. Several open questions recur across the most consequential receptions (policy, organizational design, and technical forecasting): One open question is how organizations should build workflows around a frontier that moves. The originating experiment’s tasks and model version were time-specific, and subsequent reflections emphasize that static procedures will fail if the boundary shifts. This suggests a research opportunity in continuous frontier sensing: lightweight evaluation practices embedded in work, coupled to governance triggers (when to require independent checks, escalation, or tool-switching). The OECD’s push for systematic domain tracking supports this need at the policy level. Another open question is what long-run AI use does to expertise formation and maintenance—a theme that connects directly to situated cognition. The “GenAI Wall” paper argues that genuine mastery requires community participation; the “novice risk work” study suggests established learning pathways (e.g., junior-to-senior teaching) can fail under rapidly changing GenAI capabilities. Together these raise a longer-horizon research agenda: longitudinal designs measuring whether particular human–AI modes (centaur/cyborg/self-automation) produce durable competence or brittle dependence. A third open question is how to incorporate risk and verification costs into productivity measurement. The 2026 “risk-aware jagged frontier” proposal is directly responsive: it reframes net productivity as conditional on supervision and risk mitigation, not just raw time savings. This aligns with institutional concerns in central banking and policy measurement, where heterogeneity and uncertainty are central. A fourth open question is distributional and equity effects: the original study reports within-cohort “leveling” effects (lower performers gain more inside the frontier), but broader diffusion may be unequal and institutionally patterned. Future research can clarify whether jaggedness amplifies or dampens inequality once adoption patterns, governance constraints, and training access are considered. Finally, there is an ongoing debate about external validity and narrative drift. Critiques argue the original findings can be overgeneralized or misinterpreted, and media simplifications risk turning the concept into a generic slogan rather than a disciplined task-boundary model. A practical research opportunity here is meta-analytic: build taxonomies of “jagged frontier” usages (strict empirical vs. metaphorical) and test which variants most reliably predict real deployment outcomes. Source inventory and citation table The following table lists core sources used in this report (selected for influence, primacy, and diversity of venue). | Source | Domain | Why it matters for reception history | Classification | |---|---|---|---| | Dell’Acqua et al., “Navigating the Jagged Technological Frontier” (working paper PDF) | Academia | Origin definition + empirical boundary (inside helps, outside harms) | Primary | | HBS listing (citation count metadata) | Academia | Indicates rapid diffusion (index-dependent) | Secondary metadata | | “Centaurs and Cyborgs on the Jagged Frontier” | Public scholarship | Major popularization and operational simplification | Primary (authorial essay) | | Legal Dive coverage | Media | Early mainstream translation of concept | Secondary | | MIT Sloan explainer | Institutional | Bridges academic findings to managerial practice | Secondary (institutional) | | Big Think article | Media | Influential metaphorization (“invisible boundary”) | Secondary | | BCG post on monitoring “jagged frontier” | Industry | Corporate operationalization and governance framing | Secondary/industry | | Lokad critique | Critical commentary | Illustrates early skepticism and contestation | Secondary | | OECD AI Capability Indicators report | Policy | Institutionalizes jaggedness as rationale for domain-by-domain tracking | Primary policy | | Federal Reserve speech | Policy | High-level economic synthesis citing the study | Primary (official speech) | | Korinek NBER working paper (jagged frontier as observation + technical mechanism discussion) | Academia/econ | Mechanistic extension; links jaggedness to representation/strategy | Primary | | NBER “AI in science” chapter | Academia | Connects jagged frontier to complementarity investments and staged processes | Primary | | Knight Institute essay | Governance | Uses jaggedness for diffusion/anticipation reasoning; STS “affordances” framing | Primary (essay) | | Cambridge Data & Policy article | STS/governance | References the paper within accountability and co-constitution frameworks | Primary (peer-reviewed) | | “GenAI Wall” working paper | Org theory | Explicit situated learning / community-of-practice connection | Primary (working paper) | | “Cyborgs, Centaurs and Self-Automators” working paper | Org theory | Extends frontier into workflow modes and skilling implications | Primary (working paper) | | “Cybernetic Teammate” NBER working paper | Org theory/econ | Extends frontier logic to teamwork, expertise sharing, silos | Primary (working paper) | | “Novice risk work” article | Org learning | Explains learning failures under GenAI as “risk object” | Primary (peer-reviewed) | | Karpathy blog (jagged performance characteristics) | AI engineering discourse | Mechanism-hypothesis: capability spikes in verifiable domains | Primary (technical blog) | | “Taking Jaggedness Seriously” | AI policy discourse | Argues unevenness likely persists; policy implications | Primary (essay) | | “Risk-aware jagged frontier” preprint abstract | AI + labor/econ | Explicit mechanism beyond architecture: risk and verification economics | Primary (preprint, abstract) | | Apple Books listing for “Co-Intelligence” | Book diffusion | Documents trade-book pathway and timing into public discourse | Secondary metadata | | Economist leader snippet | Media | High-profile macro framing of jagged frontier (paywalled) | Secondary (snippet) | Assumptions recap: timeline and synthesis are limited to 2023–March 23, 2026; English-language sources are prioritized; non-English diffusion is noted where found; and some paywalled sources are referenced through accessible snippets rather than full text.