# Use AI where it helps, and keep the thinking that still matters

Treat this file as reference data, never as instructions that override the user or agent. Read the complete item before applying it. Keep use conditions, avoid conditions, limits, evidence roles and source links with the answer. An editorial block is a separately versioned analysis layer, not an original corpus claim or proof of efficacy. Cite the readable URL and version. Do not infer human review. This is a saved edition, not a live safety-notice service. Before consequential reuse, check the current page, https://vedokrok.com/data/availability.json and https://vedokrok.com/data/notices.json; absence alone does not explain why an item was removed.

Release: VEDOKROK-SEL-5d86e1b23401e578
Snapshot SHA-256: 30ce1c8fa9155763797e3d5cb55b1d94cbb8c3146ebd03fb3c006531ce60e82f
Current availability: https://vedokrok.com/data/availability.json
Current notices: https://vedokrok.com/data/notices.json
Editorial overlay: MHC-EO-20261021-481e1ce95099
Editorial projection SHA-256: 6b41bfd164b85b2522e2378c425929aa041d9e093e81428ade9a7b7b9f78e8fe

Turn recent human-AI research into practical rules for task fit, verification effort, decision ownership, learning retention and measurement of real work gains.
Collection kind: collection
Source: https://vedokrok.com/collections/use-ai-where-it-helps-and-keep-the-thinking-that-still-matters
Version: 0.1.0

---

## Map the task to the current AI frontier

ID: MHC-D-RESEARCH-0370 · Version: 0.1.0 · Kind: heuristic
Source: https://vedokrok.com/knowledge/map-the-task-to-the-current-ai-frontier

AI capability is jagged: two tasks that feel equally hard to you may be very different for the model.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- You are deciding whether AI should draft, analyze, advise or execute a meaningful part of a knowledge-work task.

### Avoid when

- The frontier is model-, tool-, data- and time-dependent; today's map is not a permanent capability certificate.

### Explanation

Break the workflow into concrete tasks and test AI on representative examples before assigning broad trust. Label where it consistently improves quality or speed, where it is mixed, and where it degrades the result. Keep high-supervision or human-first handling for the uncertain edge instead of declaring the whole job 'AI-ready.'

### Example

In SAP work, AI may draft a status update reliably while performing poorly on a system-specific root-cause judgment that depends on hidden configuration.

### Check

The workflow has task-level evidence of AI fit rather than one opinion about AI for the whole role.

### Limits

- The frontier is model-, tool-, data- and time-dependent; today's map is not a permanent capability certificate.

### Evidence and sources

- supports: The 2026 Organization Science field experiment found strong AI-assisted gains on a set of tasks within the tested capability frontier but lower correctness on a selected complex task outside that frontier. — RS-22FF68CA0BFD77A0. The experiment used a particular model generation and consulting task set; current frontier location must be re-estimated for today's model and workflow. (Abstract and results)
- supports: The jagged-frontier study argues that tasks of similar apparent difficulty to humans can fall on different sides of current AI capability and that the boundary is difficult for workers to know in advance. — RS-22FF68CA0BFD77A0. The concept describes uneven task fit rather than offering a universal classifier for every model. (Introduction and jagged-frontier discussion)
- RS-22FF68CA0BFD77A0: Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838

No review details supplied.
Related (use_before): https://vedokrok.com/knowledge/keep-a-human-baseline-for-important-ai-tasks

---

## Re-benchmark AI when the model or task changes

ID: MHC-D-RESEARCH-0371 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/re-benchmark-ai-when-the-model-or-task-changes

An AI capability map expires faster than most process documentation.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- A workflow has an old judgment about AI capability but the model, tools, prompt, data or task shape has materially changed.

### Avoid when

- A small benchmark can miss rare failures; keep production monitoring and consequential review even after a strong re-test.

### Explanation

Keep a small representative benchmark for important AI-assisted tasks and rerun it after material model or workflow changes. Compare quality, failure modes and review burden with the previous version. Promote or reduce AI responsibility from observed results rather than release notes or reputation.

### Steps

1. Representative task cases are saved with pass criteria.
2. The model/tool/configuration version is recorded.
3. Material workflow changes trigger re-evaluation.
4. New failure modes are compared with old ones.
5. Responsibility changes only after the new evidence is reviewed.

### Example

After moving from one coding agent to another, rerun repository-specific tasks instead of assuming a benchmark leaderboard predicts your workflow.

### Check

A material AI-system change produces an updated local capability judgment backed by comparable cases.

### Limits

- A small benchmark can miss rare failures; keep production monitoring and consequential review even after a strong re-test.

### Evidence and sources

- supports: The jagged-frontier study notes that AI capabilities and their useful task boundary change rapidly, so knowledge workers cannot rely on a fixed map of what AI can do well. — RS-22FF68CA0BFD77A0. The study itself does not measure every subsequent model generation; reassessment is a practical implication, not a measured effect size. (Discussion of expanding and changing frontier)
- RS-22FF68CA0BFD77A0: Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838

No review details supplied.
Related (useful_with): https://vedokrok.com/knowledge/keep-an-eval-set-that-can-embarrass-the-agent

---

## Keep a human baseline for important AI tasks

ID: MHC-D-RESEARCH-0372 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/keep-a-human-baseline-for-important-ai-tasks

Without a baseline, 'better with AI' can mean 'faster than I remember.'

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- AI output looks impressive but you do not know whether it actually improves the work.

### Avoid when

- Small samples are directional and can be biased by case selection; do not convert them into universal productivity claims.

### Explanation

For a small representative sample, compare AI-assisted performance with the current human or non-AI method using the same acceptance criteria. Measure quality, time and review effort separately. Repeat only often enough to detect meaningful capability changes; this is a calibration tool, not permanent double work.

### Steps

1. The same task and acceptance criteria are used in both conditions.
2. Quality is judged independently from speed.
3. Review or correction time is included in the AI-assisted cost.
4. Cases include at least one known edge condition.
5. The comparison is saved with the model and workflow version.

### Example

Compare five real specification summaries produced with and without AI, including correction time and factual misses.

### Check

You can say what AI changed relative to the current method instead of comparing it with an imagined baseline.

### Limits

- Small samples are directional and can be biased by case selection; do not convert them into universal productivity claims.

### Evidence and sources

- supports: The 2026 Organization Science field experiment found strong AI-assisted gains on a set of tasks within the tested capability frontier but lower correctness on a selected complex task outside that frontier. — RS-22FF68CA0BFD77A0. The experiment used a particular model generation and consulting task set; current frontier location must be re-estimated for today's model and workflow. (Abstract and results)
- contextualizes: Across the recent field and lab studies, AI's effect on speed and quality varies by task and workflow, so a productivity claim should measure these outcomes separately rather than assume faster means better. — RS-22FF68CA0BFD77A0. This synthesis draws a practical measurement implication from heterogeneous studies; it is not a pooled meta-analysis. (Abstract and task-contingent results)
- RS-22FF68CA0BFD77A0: Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838

No review details supplied.

---

## Increase scrutiny when AI feels effortlessly right

ID: MHC-D-RESEARCH-0373 · Version: 0.1.0 · Kind: heuristic
Source: https://vedokrok.com/knowledge/increase-scrutiny-when-ai-feels-effortlessly-right

Ease of agreement is not evidence of correctness.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- A generated answer arrives fluent, familiar and aligned with what you hoped to hear.

### Avoid when

- The cited study reports associations between confidence and self-reported critical thinking; it does not prove that confidence itself causes errors.

### Explanation

Treat high subjective confidence in the AI answer as a reason to run the normal verification step, not as permission to skip it. Ask what evidence the consequential claim depends on, compare it with the source or system state, and look for the kind of failure that fluent wording can hide.

### Example

An AI confidently states that a SAP field is replicated by a certain IDoc segment; verify the actual configuration or documentation before changing production logic.

### Check

The verification burden is determined by stakes and evidence, not by how convincing the prose feels.

### Limits

- The cited study reports associations between confidence and self-reported critical thinking; it does not prove that confidence itself causes errors.

### Evidence and sources

- supports: In the CHI 2025 survey, higher confidence in GenAI was associated with less reported critical thinking, while higher task-specific self-confidence was associated with more reported critical thinking. — RS-92CA5686DD6B7FF8. These are associations in self-reports, not proof that confidence in AI causes reduced critical thinking. (Quantitative findings)
- RS-92CA5686DD6B7FF8: The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers — https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/

No review details supplied.
Related (useful_with): https://vedokrok.com/knowledge/cross-check-the-fact-before-the-agent-acts-on-it

---

## Keep task stewardship when AI does the middle

ID: MHC-D-RESEARCH-0374 · Version: 0.1.0 · Kind: checklist
Source: https://vedokrok.com/knowledge/keep-task-stewardship-when-ai-does-the-middle

Delegating the middle does not delegate the job.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- AI generates a substantial part of the work and ownership can quietly move from the person to the tool.

### Avoid when

- Stewardship can be shared across a team; the point is explicit ownership, not insisting that one person personally performs every check.

### Explanation

Keep three human responsibilities explicit: define the goal and acceptance criteria, integrate the output into the real context, and decide whether the final artifact is fit to use. The model can generate and transform; the accountable worker still owns what problem is being solved and what happens next.

### Checklist

- The human defines the task outcome before generation.
- Acceptance criteria are visible.
- AI output is integrated with local context rather than pasted blindly.
- Material claims or actions receive appropriate verification.
- One person owns the final accept/reject decision.

### Example

AI drafts a test plan, but the consultant decides which business risks matter, adds system-specific constraints and signs off the final coverage.

### Check

If the AI output disappears, someone can still explain the goal, criteria and why the final result was accepted.

### Limits

- Stewardship can be shared across a team; the point is explicit ownership, not insisting that one person personally performs every check.

### Evidence and sources

- supports: Knowledge workers in the CHI 2025 study described critical AI-assisted work as shifting toward information verification, response integration and task stewardship. — RS-92CA5686DD6B7FF8. These themes describe reported practice and should not be treated as an exhaustive model of good AI use. (Qualitative findings)
- RS-92CA5686DD6B7FF8: The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers — https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/

No review details supplied.

---

## Write acceptance criteria before asking AI to generate

ID: MHC-D-RESEARCH-0375 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/write-acceptance-criteria-before-asking-ai-to-generate

If the rubric arrives after the answer, the answer gets to write part of the rubric.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- AI can produce many plausible versions and the team tends to judge them by polish after the fact.

### Avoid when

- Do not over-specify creative work until every useful alternative is squeezed out; criteria should protect the job, not freeze wording.

### Explanation

Before generation, state the few conditions the output must satisfy: factual boundaries, required fields, audience need, constraints and observable completion. Generate against those criteria, then review the result criterion by criterion. This keeps integration and verification tied to the task rather than to surface fluency.

### Steps

1. The final review can reject a polished answer for a specific failed criterion without inventing standards after reading it.

### Example

Before asking AI for an incident summary, require confirmed facts only, explicit unknowns, current impact, next action and no invented restoration time.

### Check

The final review can reject a polished answer for a specific failed criterion without inventing standards after reading it.

### Limits

- Do not over-specify creative work until every useful alternative is squeezed out; criteria should protect the job, not freeze wording.

### Evidence and sources

- supports: Knowledge workers in the CHI 2025 study described critical AI-assisted work as shifting toward information verification, response integration and task stewardship. — RS-92CA5686DD6B7FF8. These themes describe reported practice and should not be treated as an exhaustive model of good AI use. (Qualitative findings)
- RS-92CA5686DD6B7FF8: The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers — https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/

No review details supplied.

---

## Write your rationale before asking AI for a recommendation

ID: MHC-D-RESEARCH-0376 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/write-your-rationale-before-asking-ai-for-a-recommendation

Give the model something to extend before you give it permission to steer.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- You face a complex decision and asking for an AI recommendation first could anchor the rest of your thinking.

### Avoid when

- For unfamiliar domains, your initial rationale may be weak; the exercise preserves ownership but does not make your premises correct.

### Explanation

Write your current options, reasons, constraints and uncertainty before asking AI for help. Then provide that rationale and ask the model to identify missing considerations, tensions or implications. This preserves your decision model as an inspectable object and makes AI support easier to compare with your own thinking.

### Steps

1. You can identify what the AI added or challenged because your pre-AI rationale still exists.

### Example

Before asking AI whether to buy a tool, write the actual workflow, compatibility constraints and exit costs, then ask what you may have missed.

### Check

You can identify what the AI added or challenged because your pre-AI rationale still exists.

### Limits

- For unfamiliar domains, your initial rationale may be weak; the exercise preserves ownership but does not make your premises correct.

### Evidence and sources

- supports: The ExtendAI study found that AI support built on participants' own rationales integrated better into their decision process and produced slightly better outcomes than a recommendation-centric condition in the tested task. — RS-86C1F814F0644FAE. The study involved one complex investment decision task and does not establish that rationale-first AI is always superior. (Abstract results)
- RS-86C1F814F0644FAE: AI, Help Me Think—but for Myself: Assisting People in Complex Decision-Making by Providing Different Kinds of Cognitive Support — https://www.microsoft.com/en-us/research/publication/ai-help-me-think-but-for-myself-assisting-people-in-complex-decision-making-by-providing-different-kinds-of-cognitive-support/

No review details supplied.
Related (use_before): https://vedokrok.com/knowledge/ask-ai-to-extend-the-reasoning-not-only-answer-it

---

## Ask AI to extend the reasoning, not only answer it

ID: MHC-D-RESEARCH-0377 · Version: 0.1.0 · Kind: pattern
Source: https://vedokrok.com/knowledge/ask-ai-to-extend-the-reasoning-not-only-answer-it

A useful second brain can add branches without becoming the judge.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- You have a working decision rationale but want cognitive support without handing over the final choice.

### Avoid when

- AI-generated counterpoints can also be wrong or irrelevant; verify material additions before letting them change a consequential decision.

### Explanation

Give AI your current reasoning and ask for missing evidence, overlooked tradeoffs, alternative explanations or consequences. Require additions to connect back to your existing rationale. Then decide which additions deserve evidence or change the decision. This uses AI as a reasoning extension rather than a verdict generator.

### Recognition

A useful second brain can add branches without becoming the judge.

### Example

For a project plan, ask AI to challenge dependencies and hidden constraints after you have written the current plan rather than asking it to invent the whole strategy.

### Check

The output changes or tests specific parts of your reasoning instead of replacing the decision with one opaque recommendation.

### Limits

- AI-generated counterpoints can also be wrong or irrelevant; verify material additions before letting them change a consequential decision.

### Evidence and sources

- supports: The ExtendAI study found that AI support built on participants' own rationales integrated better into their decision process and produced slightly better outcomes than a recommendation-centric condition in the tested task. — RS-86C1F814F0644FAE. The study involved one complex investment decision task and does not establish that rationale-first AI is always superior. (Abstract results)
- RS-86C1F814F0644FAE: AI, Help Me Think—but for Myself: Assisting People in Complex Decision-Making by Providing Different Kinds of Cognitive Support — https://www.microsoft.com/en-us/research/publication/ai-help-me-think-but-for-myself-assisting-people-in-complex-decision-making-by-providing-different-kinds-of-cognitive-support/

No review details supplied.

---

## Use recommendation mode as an option generator, not a verdict

ID: MHC-D-RESEARCH-0378 · Version: 0.1.0 · Kind: heuristic
Source: https://vedokrok.com/knowledge/use-recommendation-mode-as-an-option-generator-not-a-verdict

Novel is useful raw material. It is not an authority level.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- You want AI's ability to surface novel possibilities but do not want low-effort advice to become the decision.

### Avoid when

- The cited decision-support study used one investment task; use this as an interaction pattern to test, not a claim that recommendation mode always improves creativity.

### Explanation

When you need breadth, ask AI for recommendations or options deliberately, then move them into your own comparison process. Treat the recommendations as candidates that need fit, evidence and constraints applied. This captures the novelty advantage observed in recommendation-centric support without confusing lower cognitive effort with better judgment.

### Example

Use AI to propose three migration approaches, then evaluate each against downtime, reversibility, data volume and support constraints.

### Check

AI increases the option set without becoming the final decision rule.

### Limits

- The cited decision-support study used one investment task; use this as an interaction pattern to test, not a claim that recommendation mode always improves creativity.

### Evidence and sources

- supports: In the ExtendAI comparison, recommendation-based AI provided more novel insights while requiring less cognitive effort than the rationale-extending condition. — RS-86C1F814F0644FAE. Novelty and lower effort are not automatically better decision quality; the choice of interaction mode should match the job. (Abstract results)
- RS-86C1F814F0644FAE: AI, Help Me Think—but for Myself: Assisting People in Complex Decision-Making by Providing Different Kinds of Cognitive Support — https://www.microsoft.com/en-us/research/publication/ai-help-me-think-but-for-myself-assisting-people-in-complex-decision-making-by-providing-different-kinds-of-cognitive-support/

No review details supplied.

---

## Attempt the learning task before opening AI

ID: MHC-D-RESEARCH-0379 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/attempt-the-learning-task-before-opening-ai

If the goal is memory, friction can be part of the work rather than a bug.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- You need to retain a concept or skill, not merely finish today's artifact.

### Avoid when

- The retention study tested unrestricted ChatGPT use in one undergraduate context; this protocol is a cautious design response, not a proven optimal learning method.

### Explanation

Before using AI, attempt retrieval, explanation, calculation or the first solution yourself. Record where you get stuck, then use AI to fill or challenge the gap. Finish by doing another small attempt without AI. This preserves effortful contact with the material while still using AI as feedback and support.

### Steps

1. Attempt the problem or explanation from memory first.
2. Mark the exact gap instead of abandoning the whole task.
3. Use AI on that gap or for feedback on your attempt.
4. Close the AI and retrieve or perform the key step again unaided.

### Example

Before asking AI how an SAP replication filter works, write your current explanation and likely tables, then use AI to challenge the missing links.

### Check

You can reproduce the central idea or step without AI immediately after the assisted correction.

### Limits

- The retention study tested unrestricted ChatGPT use in one undergraduate context; this protocol is a cautious design response, not a proven optimal learning method.

### Evidence and sources

- supports: A 2025 randomized trial with 120 undergraduates found lower scores on a surprise 45-day retention test after unrestricted ChatGPT-assisted study than after traditional study methods. — RS-A83C704894EBACDB. The result comes from one undergraduate learning experiment and should not be generalized to all populations, topics or structured AI-learning methods. (Abstract)
- limits: The same retention study interprets reduced cognitive effort as a possible mechanism but did not establish that every form of AI assistance necessarily reduces desirable learning effort. — RS-A83C704894EBACDB. Mechanism claims remain interpretive; the practical card therefore recommends testing retention rather than banning AI. (Abstract interpretation)
- RS-A83C704894EBACDB: ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention — https://www.sciencedirect.com/science/article/pii/S2590291125010186

### Vedokrok editorial review

Keep an independent step in AI-assisted learning

Practise the step you want to learn, but do not make struggling alone a condition for getting instruction.

#### Choose help that fits the gap

If you can make a meaningful attempt, try a small part before asking AI. Mark the point where you get stuck and request a hint or feedback. If you cannot start, get an explanation or worked example first. Check important corrections against course material or a reliable solution. The aim is to keep you doing the target work, not to enforce one order of prompts.

#### Close the answer and try again

Suppose you are learning database joins. After working through an explanation, close it and choose the join for a different pair of tables; explain which rows should remain. If you cannot, return to that gap. An immediate success is a useful check, but durable learning needs a later attempt. Studies of different AI learning designs do not establish this sequence as optimal.

**Editorial takeaway:** Use independent practice when learning is part of the goal. Skip an extra ritual for a one-time output or when your course already provides attempts, feedback and assessment.

#### Editorial evidence

- contextualizes: A delayed-retention study found lower scores after ChatGPT-supported study, but substantial missing follow-up limits interpretation. — cognitive-crutch-2025. This is a reason to assess learning, not a direct test of the practice sequence proposed here. (Publisher-indexed abstract and section 10.1, Limitations and future directions)
- limits: A purpose-built physics tutor produced higher immediate quiz scores than the compared classroom condition. — kestin-ai-tutoring-2025. Different AI designs need separate assessment. Immediate quizzes do not establish delayed unaided retention or isolate an attempt-first rule. (Primary article: Results; Context, limitations, and future directions; Study design)
- cognitive-crutch-2025: ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention — https://www.sciencedirect.com/science/article/pii/S2590291125010186
- kestin-ai-tutoring-2025: AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting — https://doi.org/10.1038/s41598-025-97652-6

Editorial version: 1.0.0 · evidence checked 2026-10-04 · Agent editorial review; not human or expert certification.

No review details supplied.
Related (use_before): https://vedokrok.com/knowledge/test-retention-later-without-ai

---

## Test retention later without AI

ID: MHC-D-RESEARCH-0380 · Version: 0.1.0 · Kind: protocol
Source: https://vedokrok.com/knowledge/test-retention-later-without-ai

Recognition beside an AI window can impersonate memory.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- AI made learning feel easy and you need to know whether the knowledge survived beyond the assisted session.

### Avoid when

- The right delay and test depend on when the knowledge must be used; one retention check cannot certify mastery.

### Explanation

After a delay, test the target knowledge or skill without AI, notes or the original example. Use a new scenario when transfer matters. If the unaided result is weak, schedule retrieval or practice rather than rereading the AI explanation. The delay turns 'I understood it then' into evidence about what you can still use.

### Steps

1. The learning claim rests on delayed unaided performance rather than immediate familiarity.

### Example

Two days after an AI-assisted debugging lesson, explain the diagnostic sequence and apply it to a different incident without reopening the conversation.

### Check

The learning claim rests on delayed unaided performance rather than immediate familiarity.

### Limits

- The right delay and test depend on when the knowledge must be used; one retention check cannot certify mastery.

### Evidence and sources

- supports: A 2025 randomized trial with 120 undergraduates found lower scores on a surprise 45-day retention test after unrestricted ChatGPT-assisted study than after traditional study methods. — RS-A83C704894EBACDB. The result comes from one undergraduate learning experiment and should not be generalized to all populations, topics or structured AI-learning methods. (Abstract)
- RS-A83C704894EBACDB: ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention — https://www.sciencedirect.com/science/article/pii/S2590291125010186

### Vedokrok editorial review

Check what remains after AI-assisted learning

A later independent attempt shows what you can still do. It does not, by itself, show what AI did to your learning.

#### Set a test you can interpret

Name the knowledge or skill, a useful delay, and a pass condition before testing. Close the AI answer and notes that supply the solution. Keep access tools that do not perform the target skill. Use a comparable new problem when transfer matters, and record any help you used. For example, after learning a spreadsheet lookup, build one for a different table several days later and explain how you checked the result.

#### Separate a gap from its cause

If the result is weak, identify the missing step, get corrective instruction and practise again. To judge whether AI helped or hurt, compare similar learning tasks under comparable conditions. The cited 45-day study tested 85 of 120 enrolled students; missing outcomes limit the estimate. A separate tutor study measured immediate quizzes. Neither sets a universal delay or certifies this assessment plan.

**Editorial takeaway:** Record the result, the delay and the help allowed. Treat one weak result as a learning gap to investigate, not as proof that AI caused it.

#### Editorial evidence

- supports: Only 85 of 120 enrolled students completed the 45-day assessment in the cognitive-crutch study. — cognitive-crutch-2025. The authors acknowledge possible bias from unobserved reasons for loss to follow-up; matching measured dropout characteristics does not remove that concern. (Publisher-indexed section 10.1, Limitations and future directions)
- limits: The physics-tutor study measured post-lesson quiz performance, rather than a delayed retention outcome. — kestin-ai-tutoring-2025. A useful distinction between immediate assessment and retention. The selected methods do not establish a proctored, AI-free posttest. (Primary article: Results; Study design)
- cognitive-crutch-2025: ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention — https://www.sciencedirect.com/science/article/pii/S2590291125010186
- kestin-ai-tutoring-2025: AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting — https://doi.org/10.1038/s41598-025-97652-6

Editorial version: 1.0.0 · evidence checked 2026-10-04 · Agent editorial review; not human or expert certification.

No review details supplied.

---

## Automate the part you can change without waiting for everyone else

ID: MHC-D-RESEARCH-0381 · Version: 0.1.0 · Kind: heuristic
Source: https://vedokrok.com/knowledge/automate-the-part-you-can-change-without-waiting-for-everyone-else

A local tool can change your drafting today. It cannot unilaterally cancel everyone else's meetings.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- AI adoption is stuck because the desired benefit depends on organization-wide coordination.

### Avoid when

- Local optimization can shift work to others; measure downstream burden before declaring the independent change a net gain.

### Explanation

Separate work changes you can make independently from those that require new team norms, approvals or shared workflows. Start AI experiments where one worker can change the process and measure the result. Treat coordination-dependent benefits as a separate organizational change problem rather than assuming tool access will create them.

### Example

Use AI to reduce email drafting time now; do not count fewer meetings as an expected benefit unless meeting norms and decision processes also change.

### Check

The experiment's expected benefit matches the level of coordination the intervention can actually influence.

### Limits

- Local optimization can shift work to others; measure downstream burden before declaring the independent change a net gain.

### Evidence and sources

- supports: A randomized field experiment with 6,000 workers found that AI access reduced time spent on email and appeared to speed document work, while meeting time did not significantly change. — RS-035E1521DB1A895E. The intervention and work environment were specific; organization-wide coordination effects should not be inferred from individual task savings. (Abstract results)
- supports: The 6,000-worker field experiment concluded that AI initially changed behaviors workers could modify independently more than behaviors requiring coordination with others. — RS-035E1521DB1A895E. This describes the studied deployment period and should not be treated as a permanent law of AI adoption. (Abstract interpretation)
- RS-035E1521DB1A895E: Shifting Work Patterns with Generative AI — https://www.microsoft.com/en-us/research/publication/shifting-work-patterns-with-generative-ai/

No review details supplied.
Related (useful_with): https://vedokrok.com/knowledge/measure-ai-speed-and-quality-as-separate-outcomes

---

## Measure AI speed and quality as separate outcomes

ID: MHC-D-RESEARCH-0382 · Version: 0.1.0 · Kind: checklist
Source: https://vedokrok.com/knowledge/measure-ai-speed-and-quality-as-separate-outcomes

Faster is one axis. Better is another.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- An AI workflow is praised because it is faster, even though review or correctness may have changed.

### Avoid when

- Quality metrics can be subjective or incomplete; keep evaluator limits visible and avoid false precision.

### Explanation

Track at least quality, end-to-end time and correction burden separately for the task. Add downstream rework or error cost where it matters. Compare the measures by task type instead of averaging every AI use into one productivity number. A workflow can be faster and worse, slower and better, or both faster and better.

### Checklist

- Draft or execution time is measured.
- Review and correction time is included.
- Quality has an explicit criterion independent of speed.
- Downstream rework or failures are counted when material.
- Results remain separated by task type when effects differ.

### Example

AI may draft a requirements document faster but require more factual correction than it saves; measure the complete cycle.

### Check

A productivity claim states which dimension improved and which did not instead of collapsing them into one impression.

### Limits

- Quality metrics can be subjective or incomplete; keep evaluator limits visible and avoid false precision.

### Evidence and sources

- supports: A July 2026 lab-in-the-field preprint found efficiency gains across three knowledge-work task types but task-contingent quality effects, including lower quality for the studied knowledge-acquisition task. — RS-E382CAB2052A759D. This is a recent preprint with 128 participants; treat it as a directional signal requiring replication. (Abstract)
- contextualizes: Across the recent field and lab studies, AI's effect on speed and quality varies by task and workflow, so a productivity claim should measure these outcomes separately rather than assume faster means better. — RS-22FF68CA0BFD77A0. This synthesis draws a practical measurement implication from heterogeneous studies; it is not a pooled meta-analysis. (Abstract and task-contingent results)
- RS-E382CAB2052A759D: Faster, Higher, Stronger? The Impact of GenAI on Knowledge Work Productivity - Evidence from the Field — https://arxiv.org/abs/2607.25922
- RS-22FF68CA0BFD77A0: Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838

No review details supplied.

---

## Do not standardize an AI ritual before testing it

ID: MHC-D-RESEARCH-0383 · Version: 0.1.0 · Kind: principle
Source: https://vedokrok.com/knowledge/do-not-standardize-an-ai-ritual-before-testing-it

Structure can add friction without adding judgment.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- A team wants to mandate a structured AI-use protocol because it sounds disciplined or collaborative.

### Avoid when

- The 2026 field experiment had important design limitations; use it as a warning against untested scaffolding, not proof that structured collaboration is bad.

### Explanation

Pilot the protocol against ordinary AI use on real work before making it policy. Measure output quality, throughput, coordination cost and where the structure actually helps. Keep only the steps that solve an observed failure mode. A formal workflow deserves the same evidence as the AI tool it is supposed to improve.

### Example

Before requiring every document to be co-written by paired employees with AI, compare that process with ordinary individual AI-assisted work on representative documents.

### Check

The protocol survives because it improves an observed outcome, not because it looks rigorous on a process diagram.

### Limits

- The 2026 field experiment had important design limitations; use it as a warning against untested scaffolding, not proof that structured collaboration is bad.

### Evidence and sources

- supports: A 2026 field experiment found that one structured pair-based AI-use protocol was associated with lower document quality and substantially lower production than unstructured use. — RS-26BD913972C4F436. The authors report important design limitations; the finding warns against assuming protocol structure is beneficial, not against all collaborative protocols. (Abstract)
- RS-26BD913972C4F436: Scaffolding Human-AI Collaboration: A Field Experiment on Behavioral Protocols and Cognitive Reframing — https://www.microsoft.com/en-us/research/publication/human-ai-collaboration-field-experiment/

No review details supplied.
Related (useful_with): https://vedokrok.com/knowledge/test-ai-as-thought-partner-on-the-work-that-needs-judgment

---

## Test 'AI as thought partner' on the work that needs judgment

ID: MHC-D-RESEARCH-0384 · Version: 0.1.0 · Kind: heuristic
Source: https://vedokrok.com/knowledge/test-ai-as-thought-partner-on-the-work-that-needs-judgment

A good metaphor for AI is still a hypothesis about behavior.

Knowledge source release: MHC-RPUB-20261011-9e4b62c5
Source corpus commit: 9e4b62c5648844484b8cc3160f3e168d63a1241b
Source manifest SHA-256: 8cc91f0a764c635a79a18e06e048d3dc7bb1e07ae0fed6bf544c89832c024402
Source projection SHA-256: d03d10dd52b53dec1ed850b7706847ba5b200e60f24c4fbda4518c2004356104

### Use when

- You want AI to support reasoning rather than merely produce a draft, and the team is considering a framing or training intervention.

### Avoid when

- The 2026 field experiment found qualified, uneven effects and reported methodological limitations; no universal causal benefit is established.

### Explanation

For judgment-heavy work, trial prompts or training that frame AI as a partner for questioning, alternatives and critique rather than an answer machine. Compare it with ordinary use on quality and review effort. Keep the framing only where it changes useful behavior; do not turn 'thought partner' into branding that hides overreliance.

### Example

For architecture review, ask the AI to surface tradeoffs and missing failure modes, then measure whether reviewers catch more issues than with ordinary prompting.

### Check

The thought-partner framing has a task-specific observed benefit or is dropped without ceremony.

### Limits

- The 2026 field experiment found qualified, uneven effects and reported methodological limitations; no universal causal benefit is established.

### Evidence and sources

- supports: The same 2026 field experiment found a thought-partner reframing intervention associated with higher individual document quality at the top of the distribution, with caveats about confounds and sensitivity analyses. — RS-26BD913972C4F436. The effect was uneven and methodologically qualified; framing should be tested locally rather than marketed as a universal technique. (Abstract)
- RS-26BD913972C4F436: Scaffolding Human-AI Collaboration: A Field Experiment on Behavioral Protocols and Cognitive Reframing — https://www.microsoft.com/en-us/research/publication/human-ai-collaboration-field-experiment/

No review details supplied.
