When AI Help Works Too Well: A Study on Persistence, Learning, and Workplace Tools

Requested: openai/gpt-6-astra · Returned: gpt-6-astra

### 1. How did researchers test performance and persistence after removing AI?

**Sources state:** The researchers conducted three randomized controlled experiments with 1,222 participants in separate groups; each experiment lasted 13–15 minutes [L78].
- **First experiment:** 354 participants received the same 12 fraction problems, with or without chatbot access. Researchers then removed AI without warning and gave everyone three final fraction problems [L80]. Participants could skip questions, faced no penalty for wrong answers, and received correct answers after mistakes [L82].
- **Second experiment:** 667 participants completed a modified replication, including a pretest to exclude those unable to solve the most basic problems [L92].
- **Third experiment:** 201 participants, divided into AI-access and control groups, answered five reading-comprehension questions before AI was disabled and everyone received three further questions [L100].

Researchers assessed independent performance through successful versus failed answers and persistence through skipping. Previously assisted participants skipped or failed more problems after removal; similar patterns appeared across the experiments [L84, L92, L102–L104]. The article explicitly describes removal as “without warning” for the first experiment, but does not separately confirm that detail for the third [L80, L100].

**Inference:** Comparing post-removal answers and skip rates with those of an unassisted control group helps distinguish immediate assisted success from the ability and willingness to continue independently. Skipping is a behavioral indicator of persistence, not a direct measurement of motivation.

**Confidence: high.** The source explicitly describes the experimental sequences, comparison groups, and outcome measures.

### 2. What design change do the authors recommend?

**Sources state:** AI tools should provide directions and instructions toward a solution rather than supply the solution itself—acting more like tutors to support long-term learning, capability, and autonomy [L118]. The authors frame the goal as augmenting rather than supplanting human abilities [L120].

**Inference:** A tool could prioritize hints and guided steps that leave meaningful reasoning to the user. However, the pack does not establish that this redesign prevents lasting declines.

**Confidence: high.** The recommendation is explicit, although its long-term effectiveness is not demonstrated in the pack.

### 3. Where might offloading risks be largest?

**Sources state:** Liu says calculators can also reduce independent performance through offloading, but AI permits offloading reasoning across many domains [L94–L96]. Greater offloading was associated with larger declines in persistence, and the arithmetic replication highlighted risks especially when AI supplied direct solutions [L92, L104].

**Inference:** Risks might be largest where people repeatedly delegate substantial reasoning across multiple activities, particularly by accepting direct answers instead of practicing independently. The relevant distinction is therefore the breadth and depth of reliance, not simply whether a task involves mathematics. The sources do not identify or rank the riskiest occupations or domains.

**Confidence: medium.** This follows from the reported breadth of AI offloading and its association with persistence, but comparisons across occupations or extensive domains are absent.

### 4. What long-term tracking data would establish whether effects last?

**Sources state:** The question remains unanswered. Liu says researchers would need to follow the same people for a long time and regularly measure learning outcomes across a wide range of tasks to understand effects on lasting competence and independent ability [L116].

**Inference:** Such tracking should include baseline and repeated **unassisted** measures of accuracy and persistence, including skip rates, alongside records of AI-use frequency and whether users request answers or guidance. Assessments at delayed intervals would distinguish an immediate withdrawal response from a lasting change. To support a causal claim, a longitudinal randomized comparison of brief AI exposure versus no exposure would be stronger than tracking AI users alone.

**Confidence: medium.** The source explicitly calls for longitudinal, broad-task measurement, but the specific persistence measures and causal design proposed here are inferences.

### 5. What limits apply to extending the findings to workplace AI?

**Sources state:** The experiments lasted only 13–15 minutes and involved fraction problems and reading-comprehension questions [L78]. They found better performance while AI was available in the first experiment, followed by poorer independent performance and greater skipping after removal [L84]. Lasting effects remain unresolved [L116]. The pack reports no workplace experiment.

**Inference:** These results support a warning about immediate independent performance after brief assistance, not a conclusion that everyday workplace AI necessarily causes lasting skill loss or lowers overall productivity. Generalization is limited because workplace tasks, expertise, collaboration, exposure duration, and tool availability may differ from these experiments. Unexpected withdrawal also differs from workflows where AI remains available. Workplace studies would need to assess both assisted productivity and later independent capability over longer periods.

**Confidence: high.** The short duration, narrow task set, and acknowledged absence of long-term evidence clearly limit workplace conclusions.

Requested: anthropic/claude-sonnet-5 · Returned: anthropic/claude-sonnet-5

**1. How researchers tested performance/persistence once AI help was removed without warning**

The sources state that across all three randomized controlled experiments, participants first worked on problems with the AI option available, then had the AI suddenly taken away for a final small set of items given to everyone regardless of group:
- Experiment 1: 354 participants got 12 fraction problems (AI group could get answers/clues); then AI was removed without warning for a final 3 fractions given to all participants [L80].
- Experiment 2: A replication with 667 participants and the same removal structure, plus an added pretest phase to screen out people who couldn't solve basic problems [L92].
- Experiment 3: 201 participants did 5 reading comprehension questions with/without AI access, then the chatbot was disabled and each participant got 3 more questions [L100].

In each case, researchers compared whether the former AI group vs. the no-AI group solved fewer problems, skipped more, or failed more once AI was gone [L84], [L92], [L102].

Confidence: high — the methodology is described explicitly and consistently for all three experiments.

**2. Suggested design change for AI tools**

The source states that "AI models and chatbots should be designed to give directions and instructions toward a solution, rather than the solution itself — more like a tutor" [L118], to "encourage long-term learning, capability and autonomy in people" [L118]. Co-author Brian Christian is quoted saying the goal is a future where "human abilities are... augmented rather than supplanted" [L120].

Confidence: high — directly stated as the researchers' recommendation.

**3. Inference: offloading reasoning across domains → where risks might be largest**

What the source states: Liu distinguishes AI from calculators by saying AI "allows people to offload reasoning across tasks across many domains," whereas calculators mainly offload computation in math [L94], [L96].

Inference: If AI's distinctive risk is that it offloads general reasoning rather than a narrow computational skill, it follows that the erosion of persistence/skill could extend well beyond arithmetic into any task requiring reasoning, judgment, or problem-solving — e.g., writing, analysis, decision-making, or other cognitive work — rather than being confined to one narrow domain like math. This suggests offloading risks could be largest in broad, general-purpose reasoning tasks (the kind AI is used for across many fields), not just specialized computational tasks.

Confidence: medium — this is a reasonable extrapolation from Liu's statement, but the study itself only tested fractions and reading comprehension, so generalization to other reasoning domains is not directly evidenced in the source.

**4. Open question: what long-term tracking data would be needed**

The source directly addresses this: "To find out, researchers would need to track the same people for a long time and regularly measure their learning outcomes across a wide range of tasks, Liu said" [L116]. The source also notes "Scientists don't yet have a full picture of how small instances of cognitive offloading affect lasting competence and independent ability" [L116], implying current single-session data isn't sufficient.

Confidence: high — this is explicitly stated by the study's first author in the source.

**5. Inference: limits of generalizing 13–15 minute lab sessions to everyday workplace AI use**

What the source states: Each experiment lasted 13 to 15 minutes [L78], and tasks were simple fraction problems or reading comprehension questions [L78], [L80], [L100], conducted in a controlled study setting with no penalty for wrong answers [L82].

Inference: These conditions differ substantially from real workplace AI use, which often involves longer, more complex, higher-stakes tasks, varied task types, repeated/habitual use over weeks or months, different incentive structures (e.g., performance evaluations, deadlines), and professional expertise levels. The short single-session, low-stakes, simple-task design means the study may not capture how persistence/motivation effects play out with complex professional tasks, sustained AI tool use over time, or in contexts with real consequences for failure — all of which could amplify or dampen the observed effects differently than in the lab.

Confidence: medium — the reasoning follows logically from the stated study parameters, but the source does not explicitly discuss workplace extrapolation, so this is an inference beyond what's stated.

Requested: google/gemini-4-argon · Returned: google/gemini-4-argon

**1. How the researchers tested performance and persistence once AI help was removed**
* **What the sources state:** Across the three randomized controlled experiments, participants in both the AI-assisted and control groups were given a final set of three unassisted problems after the AI assistant was disabled or removed without warning [L78, L80, L92, L100]. Specifically, in the first two experiments, participants were given a final set of three mathematical fraction problems after completing 12 initial fraction problems [L78, L80, L92]. In the third experiment, participants were given three final reading comprehension questions after completing five initial comprehension questions [L100]. In the experiments, participants had the option to skip problems, and they were explicitly told there was no penalty for wrong answers—making skipping a deliberate decision not to engage [L82]. The researchers measured unassisted task performance and persistence by tracking whether participants failed or skipped more problems (an "increased skip rate") once the AI was removed [L84, L104].
* **Inference:** Tracking the skip rate alongside failure rates allowed the researchers to separate a drop in motivation/persistence (choosing to skip without penalty) from a drop in sheer ability (attempting a problem and failing).
* **Confidence:** High — The source explicitly details the experimental procedures, the number of tasks before and after AI removal, and how skipping and failing were used to gauge persistence and performance [L78, L80, L82, L84, L92, L100, L104].

**2. Suggested design change for AI tools**
* **What the sources state:** To encourage long-term learning, capability, and autonomy, the study authors argue that AI models and chatbots should be designed to act "more like a tutor" by giving directions and instructions toward a solution, rather than providing the direct solution itself [L118]. Study co-author Brian Christian also stated that the goal should be a future where human abilities are "augmented rather than supplanted" [L120].
* **Inference:** If AI tools are redesigned to provide step-by-step guidance rather than immediate answers, users will still have to engage in the "productive struggle" ("desirable difficulties") needed to build independent skill [L72, L118].
* **Confidence:** High — The source directly states the authors' specific recommendation for how AI models and chatbots should be designed [L118].

**3. Where offloading risks might be largest based on the comparison to calculators**
* **What the sources state:** Study first author Grace Liu noted that while calculators allow people to offload math computations, AI is studied specifically because "it allows people to offload reasoning across tasks across many domains" [L94, L96]. The source also notes that the third experiment tested reading comprehension—which relies on "meaning-making skills and mental-model construction"—and found that reduced persistence was not just an artifact of math problems, but a general outcome of cognitive offloading onto AI [L100, L102]. Furthermore, the second experiment hinted that perseverance is harmed "especially when AI is used to get direct solutions" [L92].
* **Inference:** Because AI spans multiple domains and affects broad cognitive processes like reasoning, meaning-making, and mental-model construction, the risks of cognitive offloading are not confined to routine calculations; rather, the risks are likely largest when people use AI to obtain direct answers in complex, multi-domain tasks that require active reasoning and mental-model building, since declines in persistence cut across domains [L92, L94, L100, L102, L104].
* **Confidence:** High — The inference follows directly from Liu's statements comparing calculators to AI and the study's findings on cross-domain persistence and direct solutions [L92, L94, L100, L102, L104].

**4. Long-term tracking data needed to know if brief AI use causes lasting drops in persistence**
* **What the sources state:** The source states that scientists do not yet have a full picture of how small instances of cognitive offloading affect lasting competence and independent ability [L116]. According to Liu, to find out, "researchers would need to track the same people for a long time and regularly measure their learning outcomes across a wide range of tasks" [L116]. The source does not provide further specifics (such as exact timeframes, measurement intervals, or specific metrics beyond "learning outcomes") on what this tracking data would look like.
* **Inference:** A longitudinal study design (repeatedly testing the same cohort over months or years across diverse cognitive domains) is required to determine whether short-term drops in persistence accumulate into permanent skill degradation [L116].
* **Confidence:** High — The source explicitly addresses what researchers say is needed to understand long-term impacts, while noting that such a full picture does not yet exist [L116].

**5. Limits to extending these findings to everyday workplace AI use**
* **What the sources state:** The source does not explicitly mention "workplace AI use" or workplace tools. However, it states that each experiment took only 13 to 15 minutes to complete [L78], tested simple tasks (15 total fraction problems in the first two experiments and 8 total reading comprehension questions in the third) [L27, L68, L78, L80, L92, L100], and explicitly told participants there was "no penalty for wrong answers" while allowing them to skip problems freely [L82]. Additionally, the source explicitly notes as a limitation that scientists "don't yet have a full picture of how small instances of cognitive offloading affect lasting competence and independent ability" over a long time across a wide range of tasks [L116].
* **Inference:** Extending these findings to everyday workplace AI use is limited because: (1) brief 13-to-15-minute sessions on basic fractions and reading comprehension may not reflect complex, prolonged professional tasks; (2) workplace environments typically carry real-world consequences or penalties for skipping work or getting wrong answers, unlike the penalty-free experimental setup; and (3) the study only measured immediate, short-term effects after removing AI, leaving long-term workplace competence unmeasured [L68, L78, L80, L82, L100, L116].
* **Confidence:** Medium — While the experimental parameters and long-term limitations are clearly stated in the text [L78, L80, L82, L100, L116], the source does not explicitly discuss workplace settings, requiring an inference to bridge the experimental design to a workplace context.

Requested: xai/grok-4.7 · Returned: xai/grok-4.7

1. The sources state that the team ran three separate randomized experiments (no participant overlap), each lasting 13–15 minutes: two on fraction problems and one on reading comprehension [L70, L78]. In the first (354 people), both an AI group and a no-AI group did the same 12 fraction problems; the AI group could get immediate answers or clues. The chatbot was then removed without warning, and everyone got a final set of three fractions [L80]. Skipping was allowed and treated as a choice not to engage, because wrong answers carried no penalty; wrong answers were also corrected so people could learn [L82]. With AI, the assisted group solved more and gave up less; after removal, their performance fell and they skipped or failed more [L84]. The second experiment (667 people) was a copy of the first, plus a pretest that dropped people who could not solve the most basic problems; results were similar [L92]. The third (201 people) gave five reading questions with AI available to one group, then disabled the chatbot and gave everyone three more questions; results mirrored the math trials, and the authors tied reduced persistence to a higher skip rate after offloading [L100, L102, L104]. The phrase “without warning” is stated for the first removal [L80]; the second is described only as a copy of that design [L92], and the third as disabling AI before the last three items [L100].
Inference: Persistence was operationalized mainly as skipping rather than attempting, and unassisted performance as solve versus fail on the post-removal items.
Confidence: high — the removal procedure, skip rule, and outcome pattern are described directly for these experiments.

2. The sources state that, to support long-term learning, capability, and autonomy, the researchers argued AI models and chatbots should give directions and instructions toward a solution rather than the solution itself — more like a tutor [L118]. Co-author Brian Christian said human abilities should be augmented rather than supplanted [L120].
Inference: That design is meant to keep some productive struggle instead of handing over the answer.
Confidence: high — the suggested design change is stated in the article.

3. The sources state that Liu said calculators also let people offload computation and could cause similar drops in independent performance, but AI is studied because it lets people offload reasoning “across tasks across many domains” [L94]. The article adds that calculators’ effects on reasoning are not as far-reaching as AI’s [L96]. The third experiment’s similar skip-rate pattern on reading, not only fractions, is presented as evidence that reduced persistence is a general outcome of offloading onto AI, not an artifact of math [L100, L102, L104].
Inference: Offloading risks would be largest where the task requires reasoning or meaning-making across domains, not only narrow calculation — because that is the scope Liu says AI uniquely covers, and the reading result already shows the persistence drop outside arithmetic.
Confidence: medium — the multi-domain contrast is stated, but “largest” is an inference, not a measured ranking.

4. The sources state that scientists do not yet have a full picture of how small instances of cognitive offloading affect lasting competence and independent ability [L116]. Liu said finding that out would require tracking the same people for a long time and regularly measuring learning outcomes across a wide range of tasks [L116]. The article also says persistence through difficulty matters for skill acquisition and learning, and raises concern about long-term impacts of AI on simple tasks [L106], but it does not specify a persistence-only follow-up protocol (for example, repeated unassisted skip rates).
Inference: To separate lasting drops in persistence from short-term ones, that longitudinal design would need repeated unassisted measures of whether people still skip or quit, not only accuracy or learning scores — the source does not say that measurement plan explicitly.
Confidence: medium — the needed long-term design is stated for competence and independent ability, not specifically for persistence.

5. The sources state that each experiment lasted 13–15 minutes [L78], used fraction problems and reading-comprehension questions [L70, L78, L100], and measured what happened right after AI was removed in that short session [L80, L84, L100, L102]. They also state that lasting effects of brief offloading are not yet fully known and would need long-term tracking across many tasks [L116]. Everyday workplace tools and job tasks are not described.
Inference: These findings should not be extended to everyday workplace AI use as if they were the same phenomenon: the evidence is limited to brief lab-style fraction and reading tasks, immediate post-removal behavior, and unknown long-term effects, not prolonged real work.
Confidence: medium — the duration and task limits are explicit; the workplace boundary follows because the source never studies workplace use.

Requested: deepseek/deepseek-v4-pro · Returned: deepseek/deepseek-v4-pro

1. **Sources state:** The researchers ran three randomized controlled experiments using fraction problems and reading comprehension questions [L78]. In the first experiment, 354 participants were assigned to either an AI-assisted group or a no-AI group; both received 12 fraction problems, then the AI was removed without warning for a final set of three fractions given to all participants [L80]. Participants could skip problems, and skipping was treated as a deliberate decision not to engage because there was no penalty for wrong answers [L82]. When AI was removed, the AI group’s performance dropped and they skipped or failed more problems [L84]. The second experiment was a copy of the first with 667 participants and a pretest to exclude people who could not solve basic problems; findings were similar [L92]. The third experiment used reading comprehension with 201 participants: both groups answered five comprehension questions, then the AI chatbot was disabled and each participant received another three questions [L100]. The results mirrored the first two experiments, with reduced persistence appearing as a general outcome of cognitive offloading, not just a math-specific effect [L102].  
**Inference:** None required; this is directly described in the source.  
**Confidence: high** — the design details are explicitly reported in the cited lines.

2. **Sources state:** The study authors suggest that AI models and chatbots should be designed to give directions and instructions toward a solution, rather than the solution itself, “more like a tutor” [L118].  
**Inference:** None required; this is directly stated in the source.  
**Confidence: high** — the design recommendation is explicitly quoted from the source.

3. **Sources state:** Liu says calculators allow people to offload computations, but AI is studied specifically because it allows people to offload reasoning “across tasks across many domains” [L94]. The article adds that calculators let people offload math problems, but their effects on reasoning are not as far-reaching as the effects of AI [L96].  
**Inference:** The offloading risks from AI are likely largest where work depends on general, cross-domain reasoning and persistence, rather than narrow arithmetic or computation alone.  
**Confidence: high** — this follows directly from the source’s contrast between AI’s broad reasoning offload and calculators’ narrower math offload.

4. **Sources state:** The source says scientists do not yet have a full picture of how small instances of cognitive offloading affect lasting competence and independent ability. To find out, researchers would need to track the same people for a long time and regularly measure their learning outcomes across a wide range of tasks [L116].  
**Inference:** None required; the source directly states the kind of long-term tracking needed.  
**Confidence: high** — the needed longitudinal tracking data are explicitly described in the cited line.

5. **Sources state:** Each experiment took only 13 to 15 minutes to complete [L78]. The tasks were mathematical fraction problems and reading comprehension questions [L78], which the source frames as simple or daily problems [L27][L68] and later calls “simple tasks” [L92]. The source also says long-term effects are not yet fully known and would require long-term tracking [L116].  
**Inference:** These findings may not extend cleanly to everyday workplace AI use because the study captures short-term, controlled-task effects on narrow academic-style tasks. It does not show whether the same declines appear or persist during longer, more complex, multi-domain workplace work, or when AI use is voluntary and continuous rather than removed without warning.  
**Confidence: medium** — the limitation is reasonable given the source’s described duration and task scope, but the source does not explicitly discuss workplace generalizability.

Requested: openrouter/meta/muse-spark-1.3 · Returned: meta/muse-spark-1.3

1. From the source: How did the researchers test performance and persistence once AI help was removed without warning across the three experiments?

What the sources state:
The study recruited "1,222 participants" in "three randomized controlled experiments, including two sets of mathematical fraction problems and a series of reading comprehension questions" with "no overlap between the experiments" and "Each experiment took 13 to 15 minutes to complete" [L78].
In experiment 1, "354 participants" were "randomly assigned" to an AI-chatbot group and a no-AI group, both given "the same 12 fraction problems" which the AI group could get the answer to or get clues about, "Then, the researchers removed the AI assistant without warning for a final set of three fractions, which all of the participants were given" [L80].
"Both groups had the option to skip problems" with "no penalty for wrong answers, so skipping a problem was a deliberate decision to not engage," and participants were given the correct answer if wrong "so they could learn from their mistakes" [L82].
When AI was present the AI group was "more likely to solve fractions and less likely to give up," but "when the AI was removed, their performance dropped and they skipped or failed more problems" [L84].
Experiment 2 was "a copy of the first experiment, but it had 667 participants and small tweaks... including the addition of a pretest phase that helped the researchers eliminate participants who were unable to solve the most basic problems" with similar findings [L92].
Experiment 3 tested "reading comprehension" with "201 participants" divided into AI-access and control groups, gave "both groups five comprehension questions before disabling the AI chatbot and handing each participant another three questions" [L100]. Results "mirrored those of the first two experiments" [L102], and "people who offload more have higher declines in persistence (increased skip rate)" [L104].

Inference: The without-warning removal plus explicit no-penalty skip option was used to interpret later skipping/failing as reduced persistence and unassisted performance, not confusion about rules.

Confidence: high — because participant numbers, AI-removal procedure, and skip-rate measure are explicitly reported in the pack.

2. From the source: What design change do the study authors suggest for AI tools, to support long-term learning rather than just giving direct answers?

What the sources state:
"To encourage long-term learning, capability and autonomy in people, AI models and chatbots should be designed to give directions and instructions toward a solution, rather than the solution itself — more like a tutor, the researchers argued" [L118].
This relates to concern about "'desirable difficulties' — the productive struggle that builds skill over time" and that "If AI routinely removes that struggle, people may get the right answer in the moment but develop less robust independent capability" [L72], and to the goal that "human abilities are, as much as possible, augmented rather than supplanted" [L120].

Inference: The suggested change is to preserve productive struggle by making AI tutor-like and indirect, rather than removing effort by supplying immediate answers.

Confidence: high — because the tutor-like directions-not-solution recommendation is directly quoted as the researchers' argument.

3. Inference: Liu says AI differs from calculators because it lets people offload reasoning across many domains - what follows about where offloading risks might be largest?

What the sources state:
"Liu acknowledged that calculators also allow people to offload computations, which could cause similar declines in independent performance. But 'the reason we study AI specifically is that it allows people to offload reasoning across tasks across many domains,' she said" [L94].
The pack adds "Calculators allow people to offload math problems, but their effects on reasoning aren't as far-reaching as the effects of AI" [L96].
Consistent with a broad effect, reading comprehension was tested because it "draws on meaning-making skills and mental-model construction" [L100], and "reduced persistence, in particular, was not an artifact of the math problems but a general outcome of cognitive offloading onto AI" [L102].

Inference: If AI enables offloading across many domains, the persistence risk would be expected to be largest and most general — spanning math, reading, and other reasoning tasks — rather than confined to calculation like calculators.

Confidence: medium — because the cross-domain quote and general-outcome finding support it, but the pack does not list all affected workplace domains.

4. Open question: What kind of long-term tracking data would be needed to know whether brief AI use causes lasting, not just short-term, drops in persistence?

What the sources state:
"Scientists don't yet have a full picture of how small instances of cognitive offloading affect lasting competence and independent ability. To find out, researchers would need to track the same people for a long time and regularly measure their learning outcomes across a wide range of tasks, Liu said" [L116].
The pack notes persistence through difficulty is seen as "essential for skill acquisition and learning" [L106], "raising concerns about the long-term impacts of using AI for simple tasks" [L106], and that overreliance "erodes our cognitive fitness and mastery of subjects, according to the study" [L116], but does not provide such long-term data.

Inference: Answering lasting-effect questions would require longitudinal within-person data — repeated persistence/learning measures over time and tasks — to separate the 13-15 minute experimental drop from durable competence loss.

Confidence: high — because Liu's required long-term same-people, wide-range learning-outcome tracking is explicitly stated.

5. Inference: Given sessions lasted only 13 to 15 minutes on fraction and reading tasks, what limits apply to extending these findings to everyday workplace AI use?

What the sources state:
The sources state only that "Each experiment took 13 to 15 minutes to complete" [L78] and covered "two sets of mathematical fraction problems and a series of reading comprehension questions" [L78], and that "Scientists don't yet have a full picture of how small instances of cognitive offloading affect lasting competence and independent ability" [L116].
The sources do not answer how these lab tasks translate to everyday workplace AI use; they do not describe workplace tools, jobs, durations, collaboration, or incentives.

Inference: The short duration, simple fraction/comprehension tasks, and immediate AI-removal test limit direct extension to longer, more complex, motivated, or collaborative workplace use where help is typically still available and learning goals differ.

Confidence: low — because any workplace generalization goes beyond what the 13-15 minute fraction/reading experiments directly tested.