Human–AI Amplification

Theory, Evidence and Testable Design Hypotheses

Companion research foundation | Package 0.5 Discussion Draft | September 2026

Document control: Package 0.5 Discussion Draft. Last reviewed 21 September 2026. Next review before external circulation and after any material evidence, scope, funding or governance decision.

Related policy framework: The Amplified Economy — A Measurement and Learning Framework for Quality-Adjusted AI Adoption, Productivity and Human Capability

Executive summary

AI capability is improving rapidly, but economic value does not follow automatically from better models or wider access. The effects observed in workplaces vary substantially across tasks, workers, workflows and organizational settings. Some deployments improve speed, quality and learning. Others generate no measurable gain, shift work to reviewers, reduce performance or create new forms of dependency. A systematic review of more than 100 experiments found that human–AI combinations, on average, performed worse than the better of the human or AI alone, even though combinations sometimes improved on humans alone. The central design problem is therefore not simply whether humans and AI should work together. It is how their work should be configured for a particular purpose.^[1]

This document develops the theoretical and experimental foundation for the Amplified Economy. It treats the worker–task–workflow–AI configuration as the primary unit of analysis and separates three dimensions:

• configuration, describing how functions are allocated across human-only, human-led, collaborative, AI-led and automated modes;
• outcome, classified as Type I Efficiency, Type II Capacity, Type III Capability or Type IV Creation; and
• amplification, an evidence-dependent magnitude designation for a material and disproportionate shift in the quality-adjusted production frontier relative to added resources, applicable to any of the four outcome types.

The framework does not presume that collaboration is always best or that human participation is inherently valuable in every step. It adopts a Human–AI Configuration Principle: allocate work according to observed capability, risk, accountability and desired human outcomes, and change the allocation when evidence changes. Within that principle, a human-participation presumption acts as a tie-breaker: where expected outcomes, risks and costs are materially equivalent, prefer the configuration that preserves meaningful human contribution and uses AI additively to extend human strengths. This tie-breaker is a policy preference under uncertainty. Its proposed long-run benefits are a separate hypothesis to test, not an established economic result.

The central proposed method is middle-out workflow testing. Rather than deciding from the outset which tasks belong to humans and which belong to AI, a team uses a hybrid workflow as a discovery environment. It decomposes work into functions—such as framing, generation, execution, evaluation, verification and authority—and then moves those functions toward human-only, collaborative, AI-led or automated modes according to comparative evidence. The hybrid is not assumed to be optimal. It is a search mechanism for discovering capabilities and combinations that top-down automation plans or bottom-up tool adoption may overlook.

The document also proposes velocity-induced context debt as a distinct organizational risk. When AI-assisted work advances faster than people can reconstruct its reasoning history, the project accumulates a liability: decisions, evidence, rejected paths and unresolved disagreements become costly for collaborators to absorb. Output may accelerate while shared understanding, challenge and governability deteriorate. The concept is theoretically consistent with research on information overload, coordination, shared mental models and transactive memory, but the label and its specific measurement are proposed here and require validation.

Other central hypotheses concern frontier capability, bottleneck migration, human attention, verification, constructive epistemic tension, earned amplification and cognitive substitution. Together they suggest that the productive frontier will be determined not only by model intelligence, but by the quality of interaction design and the ability of organizations to allocate scarce human judgment.

The practical implication is twofold. First, the Newfoundland and Labrador Full Learning Pilot should test these hypotheses through comparative workplace experiments rather than treat them as doctrine. Second, the national framework should permit bounded development of new interaction tools when existing products cannot implement a promising configuration. Software should follow the hypothesis: prototypes earn further investment only when they produce measurable improvements over simpler alternatives.

1. Purpose and relationship to the policy framework

Canada’s 2026 National Artificial Intelligence Strategy, AI for All, establishes national adoption priorities and a target of 60 per cent business adoption by 2034.[14] The Amplified Economy proposes a complementary measurement and learning layer for determining which forms of adoption produce quality-adjusted value. This companion explains the ideas underneath that proposal, identifies what existing research supports, marks what remains speculative and translates the concepts into testable designs.

It has four purposes:

1. establish a coherent vocabulary for human–AI work;
2. connect the proposed concepts to relevant empirical and theoretical research;
3. convert broad ideas into falsifiable hypotheses and measurable outcomes; and
4. guide the design of pilot experiments, evaluation methods and enabling tools.

This is not a declaration that every proposed concept is correct. The initiative is premised on disciplined uncertainty. It should be ambitious about potential value while remaining explicit about the difference between evidence, inference and invention.

1.1 Evidence language used in this document

Evidence quality and strategic ambition are different dimensions. A hypothesis can be strategically important and empirically immature. That is a reason for a bounded, reversible test—not for certainty or paralysis.

1.2 Nature of the contribution

This companion draws on established fields, proposes specific constructs and methods, and contributes an integrated policy and experimental architecture. It should not be read as claiming that every underlying idea is new. Its contribution is best understood in four layers:

• Established foundations. Human-centred AI, augmentation, sociotechnical design, organizational complements, task-level variation at the jagged frontier and organizational learning provide much of the external foundation.
• Proposed constructs and hypotheses. Velocity-induced context debt and earned amplification are advanced for testing and should be retained only if they add explanatory or predictive value. The human-participation tie-breaker is a policy preference; claims about its effects require separate testing.
• Proposed method. Middle-out workflow testing treats hybrid work as a discovery environment and compares allocations of human and AI functions rather than fixing them in advance.
• Integrative contribution. The Amplified Economy connects workplace configuration experiments, quality-adjusted measurement, bottleneck discovery, replication, stopping rules and national diffusion into a public learning system.

Originality is therefore claimed most strongly at the level of synthesis and operational architecture, not for every ingredient. The pilot is intended to test whether that synthesis creates value beyond ordinary adoption support.

2. Why adoption is not enough

The conventional implementation sequence is often: acquire a tool, provide access, train employees and count usage. This can answer whether a product was deployed. It does not establish whether useful work improved.

Research already shows why a more discriminating approach is necessary:

• In a field study of 5,172 customer-support agents, access to a generative AI assistant increased issues resolved per hour by an average of 15 per cent, with the largest gains among less-experienced and lower-skilled workers.^[2]
• In an experiment with consultants, participants using GPT-4 performed substantially better on tasks inside the model’s capability frontier, but were more likely to make errors on a task outside that frontier.^[3]
• In a randomized study of experienced open-source developers working in mature repositories, early-2025 AI tools increased completion time by 19 per cent even though participants believed the tools had made them faster.^[4]
• A meta-analysis found that human–AI systems did not reliably outperform the better independent performer; results varied by task type and by whether the human or AI was stronger alone.^[1]
• Economy-wide returns from a general-purpose technology can lag because complementary investments in workflows, skills, data and organizational capital are not yet complete or are initially mismeasured.^[5]

These findings do not resolve whether AI “works.” They show that the question is underspecified. Effects depend on the configuration, comparison condition, outcome measure and time horizon.

2.1 The capability-capture gap

The AI capability-capture gap is the difference between available technical capability and the value an organization can reliably realize from it.

The gap can arise because:

• the wrong task was selected;
• the model is capable, but the interface exposes the capability poorly;
• the worker lacks the context or authority needed to use the output;
• the workflow contains queues, handoffs or approvals that absorb the gain;
• verification costs exceed generation savings;
• saved time cannot be redeployed to something valuable;
• data, security or procurement constraints block implementation;
• the apparent gain reflects more output rather than better outcomes; or
• the organization loses human understanding or resilience while appearing faster.

Closing this gap is not merely adoption. It is the applied discipline of discovering a valuable configuration and building the complements required to absorb it.

3. The unit of analysis: a configuration, not a tool

An AI product does not have one productivity effect. Its effect emerges through interaction with a worker, a task, a workflow and an organization.

The proposed unit of analysis is:

worker × task × workflow × AI × data × incentives × organizational support

This framing is a supported design inference from heterogeneous empirical effects and the literature on complementary organizational investment. It directs attention away from generalized claims such as “AI increases productivity” and toward questions that can be tested:

• Which worker characteristics affect the result—experience, domain knowledge, AI skill, authority or motivation?
• Which task properties matter—structure, ambiguity, verifiability, context dependence, consequence or novelty?
• Which workflow functions should be human-led, collaborative, AI-led or automated?
• What information can the system access, and what provenance is preserved?
• Who bears verification, correction and coordination costs?
• How are time savings or new capabilities converted into an operational result?

3.1 Frontier capability and frontier workers

Frontier capability is configuration-specific performance near the highest reliably achievable level for a given purpose and setting. It is relational rather than a property of a model alone.

A frontier worker is a person who discovers or operates an unusually effective configuration. The term should not imply a fixed elite. The same person may be at the frontier in one workflow and far from it in another. Performance may depend on tacit domain knowledge, problem framing, prompting, tool fluency, verification habits, organizational access or the ability to redesign work.

This distinction matters for diffusion. Copying the tool used by a frontier worker may fail if the productive mechanism was actually the worker’s domain model, sequence of checks, authority to change a process or access to high-quality data.

Testable hypothesis: High-performing configurations can be decomposed into transferable components, but replication will often require more than tool training.

Possible measures: replication effect, time to competence, change in error rate, retained performance without the original worker, required organizational complements and proportion of the observed gain explained by workflow versus model choice.

4. Configuration, outcomes and amplification

Configuration and outcome should not be collapsed into one typology. Human–AI configuration describes the allocation of functions; outcomes are classified separately. Automation and augmentation remain useful descriptions of recurring arrangements. Amplification describes the magnitude of the demonstrated result: a material and disproportionate shift in the quality-adjusted production frontier relative to added resources. That frontier includes how much valuable work can be delivered with available resources, as well as which kinds of work are feasible. A multiple increase in quality-adjusted productivity can therefore qualify even when the work itself remains familiar.

Outcomes use a canonical four-part taxonomy: Type I Efficiency, Type II Capacity, Type III Capability and Type IV Creation. Amplification can occur within any of these types; capability and creation effects are not prerequisites. The minimum test is a material and disproportionate shift in the quality-adjusted production frontier relative to added human, technical and organizational resources. A modest useful improvement does not automatically meet that test, and a novel capability does not meet it merely by being new. Automation or augmentation may form part of a configuration that achieves amplification; workflow redesign is a candidate mechanism, not proof of the result. The research priority is to locate high-amplification situations in real workflows, identify the constraints that suppress them and determine whether both the gains and their enabling conditions can be reproduced.

None should be chosen on moral symbolism. Human participation is valuable when it contributes context, judgment, accountability, learning, legitimacy, empathy, creativity or resilience. It is wasteful when a person is retained only to rubber-stamp a reliable automated process. Automation is appropriate when the activity is stable, measurable, governable and not meaningfully improved by continued human participation.

The Human–AI Configuration Principle is therefore:

Choose among human-only, AI-assisted, collaborative, AI-led and automated modes according to observed relative capability, risk, accountability and desired human outcomes. Re-test the allocation as systems and work change.

5. Middle-out workflow testing

5.1 The problem with predetermined task allocation

Two common implementation routes can conceal opportunity.

Top-down allocation begins with an organizational judgment about which tasks should be automated or assigned to AI. It can be coherent and governable, but it risks freezing an inaccurate model of both human and AI capability. Managers may decompose work according to the existing organization rather than the structure made possible by new technology.

Bottom-up adoption begins when workers attach available tools to tasks they already perform. It can reveal practical use quickly, but it tends to optimize fragments of an inherited workflow. Local improvements may increase downstream volume, produce duplicated methods or leave organizational bottlenecks untouched.

Both approaches often assume that the relevant task boundaries and comparative advantages are already known. With a rapidly changing, general-purpose technology, that assumption is weak.

5.2 The middle-out proposition

Middle-out workflow testing is a proposed experimental method for discovering the appropriate allocation of functions without fixing it in advance.

The method begins with a meaningful hybrid configuration because interaction can reveal information that isolated benchmarking cannot. The workflow is then decomposed into cognitive and operational functions such as:

• problem framing;
• context assembly;
• option generation;
• execution or production;
• evidence retrieval;
• evaluation and comparison;
• error detection and verification;
• exception handling;
• communication and handoff;
• authorization; and
• accountability.

Each function can be shifted along an integration spectrum:

Human-only ↔ Human-led with AI support ↔ Collaborative ↔ AI-led with human oversight ↔ AI-only

The spectrum is intentionally simple for communication. Real workflows are multidimensional: a system may generate autonomously while humans retain framing, verification and authority. The experimental record should therefore describe functions and decision rights, not assign one label to the whole job.

5.3 Why start in the middle

The hybrid condition is a discovery environment, not a preferred destination. It allows observers to see:

• where humans correct, redirect or enrich AI work;
• where AI exposes human blind spots or expands the search space;
• where interaction adds latency without improving the outcome;
• which steps can be safely automated after their failure modes become visible;
• which human activities preserve essential understanding or legitimacy; and
• which capabilities neither side displayed in isolation but the configuration makes possible.

This final category is important. A person may not know that a particular analysis is feasible until AI lowers its cost. An AI system may not perform well until a worker supplies tacit constraints, examples or evaluation criteria. A team may discover that the highest-value outcome is not faster completion of the original task, but a new product, service or decision process.

5.4 Experimental sequence

1. Define the outcome. Specify the operational result, affected people, decision owner and Minimum Worthwhile Effect.
2. Map the current workflow. Record stages, handoffs, data, decision rights, quality controls, queues and constraints.
3. Establish reference conditions. Measure the current human-led workflow and, where responsible and feasible, an AI-dominant or automated condition.
4. Construct a hybrid discovery workflow. Give both human and AI meaningful roles and preserve sufficient traceability to understand their contributions.
5. Observe capability and friction. Record corrections, escalations, unique contributions, failures, review demand and unexpected possibilities.
6. Shift functions. Move selected functions toward human or AI control based on observed performance, not preference.
7. Compare configurations. Use quality-adjusted outcomes, human time, risk, learning and organizational absorption—not speed alone.
8. Replicate. Test whether the configuration works with another worker, team, task class or organization.

5.5 Falsification conditions

Middle-out testing should be rejected or narrowed if:

• it adds complexity without finding better configurations than straightforward task assignment;
• interaction costs overwhelm the value of discovery;
• results cannot be attributed well enough to inform a decision;
• a human-only or AI-only condition reliably dominates; or
• the method produces configurations too idiosyncratic to explain, govern or transfer.

Middle-out testing is not a requirement to keep humans in every loop. It is a method for earning confidence about where they should and should not remain.

6. Human attention as a scarce productive resource

Generative systems can create drafts, options, analyses and code much faster than people can review them. As generation becomes cheaper, the limiting resource can move from production to human framing, selection, verification and decision.

Research on information overload establishes the general problem: when incoming information exceeds processing capacity, decision quality, comprehension and performance can deteriorate.^[6] Human–AI systems add a distinctive feature. Output is not merely abundant; it can be plausible, personalized and produced on demand, which can invite either excessive review or insufficient scrutiny.

6.1 Attention debt and verification debt

Attention debt is the accumulated body of potentially relevant output that has not received appropriate human judgment.

Verification debt is the accumulated body of claims, code, decisions or artifacts whose reliability has not been checked to the level their consequences require.

These debts are not identical. An item may be reviewed but not verified, or verified technically but never considered by the appropriate decision-maker. Both can allow apparent productivity to exceed governable productivity.

6.2 The attention inbox

The attention inbox is a proposed interaction pattern that treats human attention as an allocated resource rather than an unlimited endpoint for agent output. Instead of presenting everything generated, the system should route items according to expected decision value.

Potential categories include:

• a decision requiring human authority;
• a material contradiction;
• an assumption with weak evidence;
• a finding likely to change the current direction;
• an exception outside the reliable frontier;
• a verification request proportional to risk;
• a low-priority update that can be batched; and
• work that can proceed without immediate interruption.

The system must remain auditable. Attention filtering creates a risk that important evidence is hidden or that the orchestrator becomes an opaque authority over what people see.

Testable hypothesis: Attention prioritization can increase useful findings per human minute without increasing missed consequential information.

Measures: decision-changing findings per interaction, interruption frequency, review time, missed material findings, false escalations, decision latency, user override and retrospective audit results.

7. Velocity-induced context debt

7.1 Proposed definition

Velocity-induced context debt is the accumulating cost of bringing human understanding into alignment with a project whose intellectual state is advancing rapidly.

The debt emerges when work changes faster than participants can build and maintain a sufficiently accurate mental model of:

• the current objective and thesis;
• decisions already made and who made them;
• evidence that materially changed the work;
• assumptions accepted, rejected or still uncertain;
• alternatives considered and why they were not pursued;
• disagreements that remain unresolved;
• dependencies and downstream effects; and
• the limits of what the AI-generated record can be trusted to represent.

The term is a proposed construct. It is related to information overload, onboarding cost, shared mental models, knowledge coordination and transactive memory, but it is intended to isolate a specific dynamic: the project’s AI-assisted intellectual state advances faster than human alignment.

7.2 Why velocity changes the problem

Ordinary projects also accumulate context. AI can change the rate and shape of accumulation:

• many branches can be explored in parallel;
• reasoning and artifacts can be generated faster than meetings or reviews occur;
• decisions may be made through conversations that other participants never see;
• the volume of plausible but rejected material increases;
• summaries may omit why a conclusion changed; and
• a participant who was absent briefly can return to a substantially different project state.

The paradox is that a more productive individual may make the team harder to scale. The original participant holds a dense, partly tacit model of the work while collaborators encounter only outputs or compressed summaries.

7.3 Consequences

Velocity-induced context debt may produce:

• slower onboarding and re-entry;
• duplicated investigation;
• repeated arguments whose earlier resolution was lost;
• weak challenge because reviewers lack the path to the current conclusion;
• over-concentration of decision power in the person who holds the history;
• brittle handoffs and key-person dependency;
• false consensus created by overly compressed summaries;
• higher review burden and meeting time; and
• inability to explain or audit how an important result was reached.

Research on team cognition supports the importance of shared or complementary mental models, and research on coordination shows that adding collaborators can impose costs that exceed their benefit in some settings.^[7][8] These adjacent findings support the plausibility of context debt, but they do not yet establish its distinctive causal mechanism in AI-accelerated projects.

7.4 Possible operational measures

The construct becomes useful only if it predicts outcomes better than generic workload or project complexity. Candidate measures include:

• time for a late joiner to make a useful, non-duplicative contribution;
• time for a returning participant to explain the current state accurately;
• divergence between participant accounts of the project’s objective, evidence and decisions;
• number of decisions lacking reconstructable rationale;
• duplicate work attributable to missing history;
• proportion of unresolved disagreements preserved accurately;
• number of questions routed back to a single context holder;
• change in Project Lead cognitive load as output velocity rises; and
• quality of unaided explanation after a delay.

A practical context-debt ratio could compare the rate at which decision-relevant state changes with the rate at which affected participants acknowledge, understand or correctly reconstruct those changes. This should initially be treated as an exploratory measure rather than a standardized metric.

7.5 Design responses

Possible responses include:

• persistent decision and evidence histories;
• role-specific “what changed and why” briefings;
• task-specific rather than project-wide onboarding;
• explicit preservation of rejected paths and revisit conditions;
• links from summaries to underlying evidence and reasoning;
• periodic human-authored synthesis checkpoints;
• detection of conflicting assumptions across workstreams;
• limits on parallel investigation when review capacity is saturated; and
• small, high-context core teams with an elastic perimeter of specialists.

Compression is not enough. A summary can reduce reading while increasing epistemic risk if it hides disagreement, uncertainty or the reason a choice was made. The aim is not maximum shared information; it is sufficient shared understanding for each participant’s role and authority.

Central falsification test: If the proposed measures do not predict onboarding time, coordination failure, decision quality or key-person dependence beyond ordinary project size and complexity, velocity-induced context debt should not be retained as a distinct construct.

8. Cooperation plus calibrated challenge

AI assistants are usually optimized to be helpful and responsive. In consequential work, helpfulness can become agreement. Research has documented sycophantic behaviour in language models, including changing answers to match a user’s stated view.^[9]

The reciprocal problem is human overreliance. Explanations can increase acceptance without necessarily improving independent judgment. Cognitive forcing interventions—such as requiring a person to make an initial judgment before seeing AI advice—have reduced overreliance in controlled settings, although they can also reduce subjective preference for the system.^[10] Survey research with knowledge workers has found that greater confidence in generative AI is associated with less reported critical-thinking effort, while greater confidence in one’s own ability is associated with more critical thinking.^[11]

8.1 Constructive epistemic tension

The proposed relationship is constructive epistemic tension:

• constructive, because both participants are oriented toward a shared purpose;
• epistemic, because the challenge concerns what should be believed, why and with what confidence; and
• tension, because neither human nor AI is presumed correct.

Challenge should be reciprocal, selective and evidence-linked. The AI should be able to identify weak premises, request missing evidence and present credible alternatives. The human should require proof, question confident claims and override the system. Unresolved disagreement should be preserved rather than converted into artificial consensus.

Constant opposition is not the goal. Performative challenge can create false balance, waste attention, reduce trust or persuade a person to abandon a correct conclusion.

Testable hypothesis: Cooperation-plus-calibrated-challenge improves decision quality and trust calibration relative to cooperative-only assistance when challenge is triggered by material uncertainty or consequence.

Measures: useful corrections, false reversals, missed risks, evidence requested, decision quality, comprehension, confidence calibration, time cost and user willingness to continue.

8.2 Timing and trigger conditions

Supportive cooperation and challenge need not operate at the same intensity throughout a task. During divergent brainstorming, a supportive mode can help a person articulate incomplete ideas, extend associations and explore a wider search space. This should not automatically be treated as sycophancy. It becomes a problem when agreement conceals weak premises or continues after the work requires discrimination, evidence and commitment.

As the task moves from generating possibilities toward evaluating, selecting, communicating or implementing them, the value of challenge may rise. Too much challenge too early may suppress exploration and cause premature convergence; too little challenge later may allow attractive but impractical ideas to harden into plans. Calibrated challenge may therefore be a staged interaction rather than a constant personality.

Potential triggers include a user-declared shift from brainstorming to evaluation; an impending decision, publication, expenditure or implementation; rising consequence or irreversibility; unresolved assumptions or contradictions; confident claims unsupported by evidence; repeated agreement or elaboration producing little new value; and explicit requests for grounding, red-teaming or feasibility assessment.

Testable hypothesis: A staged pattern—supportive cooperation during early divergent exploration, followed by selectively triggered challenge during convergence and commitment—will produce a better combination of idea range, feasibility and decision quality than continuous support or continuous challenge.

Suggested comparisons: continuous cooperative assistance; continuous challenge; challenge introduced at a fixed workflow stage; user-invoked challenge; and system-detected challenge based on consequence, uncertainty or diminishing novelty. Measures: idea diversity and usefulness during generation; feasibility, evidence quality, risk detection, false reversals and decision quality during grounding; plus time, cognitive burden, trust and willingness to continue.

9. Earned amplification and cognitive substitution

AI can improve an artifact while weakening the user’s understanding of it. This creates a distinction between performance that belongs to the configuration and capability that remains available to the person.

Cognitive substitution occurs when AI performs high-value reasoning that the person does not meaningfully reconstruct, evaluate or retain. It can be appropriate when human learning is irrelevant and accountability is otherwise sound. It becomes a risk when the person must later explain, adapt, challenge or take responsibility for the work.

Earned amplification is the hypothesis that substantive human contribution to high-value cognitive acts can produce stronger understanding, ownership and judgment while AI expands the person’s reach.

The aim is not effort for its own sake. It is to distinguish:

• low-value friction, such as copying information, formatting evidence, searching transcripts or repeating mechanical checks; from
• high-value difficulty, such as framing the problem, identifying assumptions, choosing standards, evaluating consequential evidence, resolving trade-offs and explaining a decision.

Evidence from education shows the underlying risk. A large field experiment in mathematics found that unrestricted access to a GPT-4-based tutor improved practice performance but harmed later unaided performance, while a tutor with learning-oriented guardrails largely mitigated the harm.^[12] Workplace learning is not identical to student learning, so the result should not be transferred mechanically. It nevertheless supports testing whether interfaces optimized for immediate completion can weaken retained capability.

9.1 Human participation as a variable

Human involvement should be designed around specific cognitive purposes. Possible interventions include requiring the person to:

• state an initial position before seeing a synthesis;
• identify the weakest assumption;
• choose evaluation criteria;
• predict what evidence would change the decision;
• inspect a small number of high-consequence claims;
• explain the result without AI assistance; or
• respond to a material counterargument.

Testable hypothesis: For complex, consequential work, requiring well-timed human contribution improves retained understanding and decision quality enough to justify its additional effort.

Failure condition: If an easier workflow produces equivalent outcomes, understanding and resilience, the imposed difficulty is waste and should be removed.

9.2 The human-participation presumption

The proposed presumption is narrow: where expected outcomes, risks and costs are materially equivalent, prefer the configuration that preserves meaningful human participation and uses AI additively to extend human strengths. It is a policy preference under uncertainty, not an economic finding or a universal rule. Its predicted effects on capability, well-being and productivity require separate testing. Automation remains preferable when it is demonstrably safer, more reliable, more accessible or more valuable, or when human involvement adds cost, delay or risk without a compensating benefit.

Meaningful participation is not nominal oversight or routine approval of machine output. It means contribution that exercises or develops a relevant human strength—judgment, creativity, empathy, practical knowledge, responsibility, problem framing, sense-making or the capacity to challenge a result. The aim is not to preserve effort for its own sake, but to avoid optimizing human contribution out of a process before its longer-term value has been measured.

The testable hypothesis is that, across longer time horizons, appropriately designed participation may support retained capability, agency, purpose, psychological and social well-being, adaptability and the ability to repair or redirect a failing system. Those effects may also compound into productivity and progress by producing people and teams that can recognize new opportunities, challenge assumptions and continue improving the configuration. The hypothesis could be wrong, context-dependent or outweighed by the benefits of automation.

Experiments should therefore compare configurations over time, not only at the moment of task completion. Relevant measures include quality-adjusted output, human time, retained skill and comprehension, autonomy, motivation, job quality, well-being, error recovery, adaptation to new conditions and performance after AI support is reduced or changed. If preserved participation does not improve these outcomes enough to justify its costs, the presumption should yield to the evidence.

9.3 The curated reference loop

The curated reference loop is a proposed mechanism through which human judgment may become a compounding input to AI-assisted work. When an output closely matches an intended standard, an authorized person can retain it as a future reference, record why it is useful and replace weaker examples as stronger ones emerge. The evolving set can carry forward accumulated judgments about quality, style, structure or approach without requiring those standards to be reconstructed for every task. The mechanism need not involve retraining the underlying model; the references may be supplied as examples or contextual material.

This differs from indiscriminately feeding AI outputs back into the system. Unfiltered accumulation may preserve errors, mediocrity or accidental conventions. The hypothesized value comes from selective human curation, provenance and revision: people define the standard and decide what should persist; AI uses the retained examples to reproduce those judgments more consistently and at scale. In that sense, the loop may operationalize earned amplification and contribute to compounding organizational capability.

Testable hypothesis: An evolving, human-curated reference set will improve independently assessed quality and consistency and reduce correction burden relative to a fixed reference set or an automatically accumulated set, without unacceptably reducing useful variation or adaptability.

Measures and failure conditions: Compare independently assessed quality, consistency against defined standards, correction time, convergence rate, transfer to new users or cases, diversity, inherited error and performance when the reference set is withdrawn. The hypothesis should be narrowed or rejected if curation increases conformity without quality, locks in hidden defects, fails to transfer or costs more to maintain than the benefit it creates. Exploration and production may require different reference sets: one preserves variation; the other stabilizes validated qualities.

10. Bottleneck migration and organizational absorption

AI rarely removes scarcity altogether. It moves it.

If drafting accelerates, review may become the constraint. If analysis accelerates, data quality or decision authority may dominate. If prototyping becomes cheap, physical testing, customer acquisition, regulation, capital or commercialization may determine whether any value is realized.

This is consistent with the productivity J-curve: general-purpose technologies require complementary intangible investment, and measured productivity can lag while organizations redesign processes and build capabilities.^[5]

10.1 The absorption question

Every claimed time saving should be followed by:

What will the organization do with the released capacity, and what prevents that conversion now?

Possible dispositions include:

• more output under existing demand;
• shorter queues or response times;
• higher quality or lower error;
• new services or products;
• expanded scope for previously underserved users;
• reduced overtime or workload;
• redeployment to higher-value activity; or
• no realized benefit because the saved time is fragmented, demand is absent or another constraint binds.

The framework therefore measures both task effect and organizational capture. A fast task that produces no valuable operational change is not a complete productivity result.

10.2 Reinvesting released capacity to reduce downstream review demand

When review, feedback or approval becomes the binding constraint, released upstream capacity can be used to reduce the amount of scarce downstream attention each output requires. Producers may spend more time on risk triage, self-review, validation, consistency checks and anticipating likely questions rather than simply generating more material. The goal is to make routine outputs more review-ready and direct reviewers toward consequential, ambiguous or high-risk decisions.

This is not a rationale for removing required human oversight. Review intensity should be differentiated by consequence and uncertainty, with mandatory-review categories, escalation thresholds, traceability and periodic sampling of lower-risk work. The configuration is successful only if review demand and revision cycles fall without increasing escaped errors or weakening accountability.

Testable hypothesis: Reinvesting a portion of AI-released upstream capacity in preparation and risk-based review allocation reduces downstream review time and revision cycles while preserving or improving quality and detection of consequential issues.

Measures: review minutes per item, number of review cycles, turnaround time, escaped-error rate, consequential issues detected, proportion escalated, sampling results and reviewer workload.

11. Tools derived from interaction hypotheses

Current general-purpose products may be sufficient for many experiments. Tool development becomes justified when the interaction design—not access to a capable model—is the binding constraint.

11.1 Candidate tool classes

• Configuration workbenches that allow the same workflow to be tested under different allocations of framing, generation, verification and authority.
• Challenge modes that invoke evidence-grounded dissent selectively rather than produce automatic contrarianism.
• Attention orchestration that ranks findings by expected decision value and records what was filtered.
• Verification and provenance layers that connect claims, sources, tests, corrections and responsible reviewers.
• Decision-history systems that preserve alternatives, rationale, dissent and revisit conditions.
• Context reconstruction interfaces that provide role-, task- and decision-specific briefings to late or returning participants.
• Evaluation harnesses that compare configurations under matched human time, matched compute or equivalent outcome requirements.
• Human-contribution checkpoints that preserve high-value judgment without retaining clerical friction.

11.2 Development rule

Tool development should follow this sequence:

1. state the human–AI interaction hypothesis;
2. attempt to test it with existing or configurable products;
3. identify the specific interaction limitation;
4. build the smallest prototype capable of testing that limitation;
5. compare it with a strong simpler baseline;
6. measure outcomes, human burden and failure modes;
7. stop, redesign, replicate or productize according to evidence.

Architecture is not evidence. A multi-agent system, reasoning graph or elaborate interface has no claim to value unless it improves the relevant outcome over a simpler workflow.

12. Measurement architecture

12.1 Four productive results

The framework distinguishes:

• efficiency — the same valuable result with fewer resources or less time;
• capacity — more valuable work with the same constrained resources;
• capability — reliable work becomes possible that could not be performed before; and
• creation — a new product, service, process, market or public-value outcome becomes viable.

These categories describe what changed. Amplification describes the magnitude of the quality-adjusted gain relative to resources and can apply to any category or combination. A material, multiple increase in quality-adjusted productivity can qualify through efficiency or capacity alone. Capability and creation effects may accompany such gains, but neither is required; conversely, novelty alone does not establish amplification.

The minimum test is a material and disproportionate shift in the quality-adjusted production frontier relative to added human, technical and organizational resources. Assess the valuable result against a credible baseline and comparator, with the workflow boundary, quality standard and observation period stated. Count setup, implementation, human oversight, verification, rework, compute, integration, maintenance and other material lifecycle costs. Faster generation or higher gross output is insufficient when the apparent gain is consumed by those demands.

Record the level at which the gain is demonstrated: individual, team or workflow, and organization or service. A verified individual or workflow productivity gain can qualify at that level without establishing an equivalent organizational gain. Trace downstream review, approvals, integration, delivery and demand to determine how much value is actually captured. Work shifted to others must be counted; a gain erased across the relevant workflow is not realized amplification at that boundary. Bottleneck migration should be reported alongside the local gain and the remaining unrealized potential.

Pre-specify a context-appropriate magnitude criterion and resource basis before interpreting results. The Minimum Worthwhile Effect is the threshold for whether implementation is justified; passing it does not by itself establish amplification. Report the absolute change and, where meaningful, the quality-adjusted productivity multiple, together with uncertainty and repeatability evidence. Where new capability or creation has no meaningful numeric baseline, use explicit feasibility, value and resource comparisons rather than inventing a multiplier. Reproduction by the original operator and replication by other workers or organizations are separate evidence claims; one does not establish the other.

12.2 Outcome families

Economic: quality-adjusted output, cycle time, unit cost, capacity, revenue, service value, backlog, error and rework.

Human: workload, drudgery, autonomy, safety, trust, comprehension, learning, resilience, job quality, displacement and redeployment.

Strategic: replication, diffusion, Canadian suppliers or intellectual property, reduced dependency, organizational learning and improvement in later experiments.

Interaction: correction yield, false reversals, review demand, decision-changing findings, context alignment, escalation, override and calibration.

12.3 Comparison conditions

Not every experiment requires every condition, but the design should choose deliberately among:

Where practical, evaluations should use matched human time and matched compute as different lenses. Blinded expert review can help separate polish from quality. Longitudinal checks are needed when the hypothesis concerns learning, dependency, diffusion or organizational capability.

12.4 Measurement cautions

• Self-reported speed and usefulness are important but insufficient; perception can diverge from measured performance.^[4]
• Output volume is not productivity unless it changes a valued result.
• State the baseline, comparator, workflow boundary, quality standard, observation period and resource denominator for each amplification claim; report absolute changes alongside any multiple.
• Include verification, rework and lifecycle costs, and distinguish individual or workflow gains from organizational capture and replication evidence.
• A quality gain may justify more time; a speed gain may not justify higher error or risk.
• A short-term slowdown can accompany valuable learning or workflow investment.
• A strong average can hide harms or failures concentrated among particular workers or tasks.
• A pilot team’s enthusiasm can produce selection effects that weaken transfer.
• Rapid model change requires recording the exact capability and configuration tested.

13. Pilot research agenda

The Newfoundland and Labrador Full Learning Pilot can convert the theory into a portfolio of practical questions.

13.1 Core cross-portfolio questions

1. Which worker–task–workflow–AI configurations produce credible economic, human or strategic value?
2. Does workflow redesign outperform simple tool insertion?
3. Does middle-out configuration search discover allocations or outcomes that participants would not have selected in advance?
4. Where does the bottleneck move after AI accelerates part of the workflow?
5. When does review, verification or attention demand erase the apparent gain?
6. Can velocity-induced context debt be measured, and does it predict coordination or transfer failure?
7. Does calibrated challenge improve decisions enough to justify its cognitive and time cost?
8. Which forms of human contribution improve understanding and resilience, and which are merely ritual?
9. Can effective configurations be reproduced by other workers or organizations?
10. When does a purpose-built interaction tool outperform an available general-purpose product?

13.2 Suggested experiment families

Workplace pilot — configuration search: Compare several allocations of framing, production, verification and authority within one workflow.

Workplace pilot — attention routing: Compare full-output review with prioritized attention routing and audit what the filter misses.

Research-partner study — challenge mode: Compare cooperative-only assistance with selectively triggered evidence challenge under a separate protocol.

Research-partner study — understanding and delayed recall: Compare answer-first assistance with required human judgment checkpoints, followed by unaided explanation and delayed recall.

Research-partner study — context-debt late joiner: Introduce a late or returning participant and compare generic summaries, role-specific briefings and traceable reasoning histories.

Workplace pilot — replication: Transfer a high-performing configuration to another worker, team or organization and isolate what must travel with it.

Research-partner study — prototype comparison: Compare a bounded purpose-built interface with a strong general-purpose AI workspace through open, non-preferential selection.

Workplace pilot — bottleneck migration: Track whether gains move the binding constraint into verification, approval, integration, demand, infrastructure or attention, and test bounded relief where decision-relevant.

The workplace pilot should fund operational configuration search, bottleneck migration and replication as part of delivery. Understanding/delayed-recall, late-joiner context debt, challenge-mode and prototype comparisons are better suited to a university research partner with separate research funding, an explicit protocol and any required research-ethics review. Operational participants should not be enrolled into research by default.

13.3 Portfolio learning products

The pilot should produce more than success stories. Reusable outputs should include:

• configuration maps;
• evidence and measurement templates;
• interaction-pattern descriptions;
• failure modes and negative results;
• context and verification measures;
• replication records;
• tool requirements derived from observed constraints; and
• commercialization or procurement recommendations for validated components.

13.4 Crosswalk to policy hypotheses

Core hypotheses: configuration-specific frontier effects map to 4.1; workflow redesign maps to 4.2.

Operational hypotheses: middle-out configuration search maps to 4.3; expanded search space maps to 4.4; bottleneck migration and constraint relief map to 4.5.

Exploratory hypotheses: attention inbox, verification debt, cognitive substitution and velocity-induced context debt map to 4.6; cooperative challenge and constructive epistemic tension map to 4.7; earned amplification and the proposed effects of meaningful human participation map to 4.8; compounding organizational capability, including the curated reference loop as a proposed mechanism, maps to 4.9. These companion terms refine measurement and mechanism; they do not create additional policy-tier hypotheses.

14. Worker agency, distribution and governance

Human-centred design is not achieved by keeping a person nominally in the loop. A worker may be required to approve an output without the time, information or authority necessary to challenge it.

Experiments should therefore ask:

• Can affected workers influence the configuration and measures?
• Can the responsible person explain, question, override and escalate?
• Who receives time savings, revenue or service gains?
• Who bears monitoring, correction, emotional or coordination burdens?
• Does the configuration expand capability or narrow discretion?
• Does it create dependency on a vendor, model or small number of context holders?
• Are professional, labour, privacy, accessibility and human-rights obligations preserved?

Worker preferences should inform task allocation without becoming the sole criterion. A 2025 study of 1,500 workers across 104 occupations found heterogeneous preferences for automation and augmentation and identified mismatches between worker desires and expert assessments of technical capability.^[13] This supports treating human agency and technical feasibility as separate dimensions requiring negotiation and evidence.

15. Research claims matrix

16. Open questions and failure conditions

The framework should change if evidence shows that:

• simple access and training capture most available value;
• middle-out search is slower or less reliable than informed task allocation;
• frontier configurations cannot be explained or replicated;
• attention filtering hides too much consequential information;
• challenge modes create more confusion or false reversal than correction;
• required human contribution adds friction without improving understanding or outcomes;
• context debt is not distinct from ordinary project complexity;
• purpose-built tools do not outperform general-purpose workspaces;
• organizational gains depend on subsidies with no plausible path to public value; or
• benefits are captured narrowly while workers absorb unacceptable burdens.

Null and negative findings are not failures of the learning system. They are evidence that prevents larger misallocation.

Conclusion

The next phase of AI policy should not be organized around a simple choice between adoption and restraint. The more consequential challenge is to discover how technical capability becomes reliable human and organizational capability.

That requires a science of configuration: empirical comparison of who or what should frame, generate, evaluate, verify and decide; measurement of the bottlenecks and debts created by acceleration; and tools designed around the actual requirements of human judgment.

The concepts in this document offer a starting architecture, not a finished doctrine. Middle-out workflow testing, constructive epistemic tension, earned amplification and velocity-induced context debt matter only if they improve prediction, design and outcomes. The Newfoundland and Labrador Full Learning Pilot provides a bounded setting in which to test them, reject weak ideas and convert credible findings into Canadian productive capability.

Document disclosures

AI-use disclosure and formative observation

These documents were developed with substantial AI assistance under the author’s direction. The project’s originating thesis, problem definition and core hypotheses arose from the author’s professional experience, observation and reasoning. AI supported research and evidence discovery, structured challenge, elaboration and organization, comparison of alternatives, synthesis, drafting and revision. The author supplied institutional context, selected or rejected outputs and made the substantive policy and design decisions. The author remains responsible for source verification, accuracy, judgment and final content.

The process informed hypotheses about middle-out configuration search, cooperation plus calibrated challenge, attention allocation, velocity-induced context debt and possible compounding capability. As concepts were refined and reapplied, later work appeared to improve in quality and speed. This is a formative observation only: there was no controlled comparison, predetermined measurement plan or reliable counterfactual, so the process is not evidence for the framework.

Authorship

Terry Hudson is the originator and principal author. This document was developed in a personal capacity. The author is employed by the Government of Newfoundland and Labrador; that employment is disclosed for transparency. The document was not commissioned, approved or endorsed by the Government of Newfoundland and Labrador, and the views expressed are the author’s own. No institutional partner or co-signatory is represented by this draft. Contact: terry@hed1.ca.

References