AI Assistance Reduces Persistence and Hurts Independent Performance

1 Introduction

• Human mentors balance help with autonomy to support long-term development, whereas AI assistants typically provide immediate answers and rarely refuse, making them short-term collaborators.The introduction frames this mismatch as a potential conflict between immediate task completion and a person’s longer-term growth.Source: §1 · p. 2
• Across 1,222 participants, randomized experiments in mathematical reasoning and reading comprehension tested AI’s causal effects on subsequent independent problem-solving.Experiments 1 and 2 focused on mathematical reasoning; Experiment 3 focused on reading comprehension.Source: §1 · p. 2
• After 10–15-minute sessions, AI assistance was followed by sharply worse performance without AI and more frequent giving up, a pattern replicated across domains.AI improved performance initially, but participants in the AI condition gave up more frequently than control participants.Source: §1 · p. 2
• The trials extend emerging causal evidence beyond educational settings, but the authors caution that the findings do not establish reduced persistence or ability across all tasks.The introduction distinguishes this work from prior evidence that was largely correlational or based on small samples.Source: §1 · p. 2
• Persistence supports effective learning and predicts long-term achievement, adaptability, and resilience; direct-answer AI use may erode it by displacing practice opportunities.The authors argue that people may become worse at tasks and stop trying, particularly when using AI to obtain direct answers.Source: §1 · p. 2
• The findings point to a design imperative: AI systems should support long-term human capability and autonomy, including knowing when not to help.The authors argue that this requires rethinking how AI systems collaborate with people, rather than relying on surface-level interventions.Source: §1 · p. 3

2.1 Method

• The experiment randomized 354 participants to AI or control conditions for fraction problems, then measured independent performance on three identical unassisted problems.The AI condition had access to GPT-5 during the first 12 problems. After exclusions, the final sample was 307 participants: 185 in AI and 122 in control.Source: §2 · p. 3, §2.1 · p. 3, §2.2 · p. 4
• Skipping was treated as a measure of motivation and persistence because participants could skip despite being told wrong answers carried no penalty.Incorrect answers also revealed the correct solution in both conditions.Source: §2.1 · p. 3

2.2 Results

• Test solve rate was 0.57 (AI) versus 0.73 (control), while test skip rate was 0.20 (AI) versus 0.11 (control), indicating worse unassisted performance and persistence.For solve rate, t(305) = −3.64, P < 0.001, Cohen’s d = −0.42, 95% CI [−0.66, −0.19]; for skip rate, t(305) = 2.16, P = 0.031, Cohen’s d = 0.25, 95% CI [0.02, 0.48].Source: Figure 1 · p. 4, §2.2 · p. 4
• With AI available, participants solved more problems and skipped fewer; after it was removed, their performance fell and skipping increased.AI was available for the first 12 problems; the final three were unassisted test problems.Source: §2.2 · p. 4, Figure 1 · p. 4
• The exclusion criteria may have selectively retained lower-ability participants in the AI condition, while greater control-group attrition could have inflated the performance gap.The authors state that Experiment 2 addresses this potential confound and that additional attrition analyses appear in Appendix C.4.Source: §2.2 · pp. 4–5

3.1 Method

• Experiment 2 added a pretest-based skill exclusion and gave controls a sidebar of previously seen solutions, addressing the skill-level and interface-asymmetry confounds.Participants solved 11 fraction problems with AI available, then three without it; controls solved the same problems without AI.Source: §3 · p. 5, §3.1 · p. 5

3.2 Results

• After attention-check and pretest exclusions, 585 participants remained; exclusion rates were similar for AI and control participants (11.7% and 12.9%), suggesting the skill-level confound was resolved.The final sample included 308 AI-condition and 277 control participants.Source: §3.2 · p. 5
• On the final test, AI participants’ solve rate was 0.71 versus 0.77 for controls (P = 0.020); skip rates were 0.10 versus 0.07 (P = 0.239, not significant).AI assistance improved performance during learning, but solve rates fell and skip rates rose after its removal.Source: §3.2 · pp. 5–6, Figure 2 · p. 6

3.3 Persistence costs are concentrated among those who obtain direct solutions

• Direct-answer users had a lower test solve rate than controls (0.65 vs. 0.77, P < 0.001) and a higher skip rate (0.13 vs. 0.07, P = 0.033).They also had lower solve rates than hint users (0.76) and nonusers (0.89); usage-group findings were cross-sectional, not necessarily causal. Most AI participants reported primarily seeking direct answers (61%).Source: §3.3 · pp. 6–7, Figure 3 · p. 7
• Direct-answer users’ solve rate declined from pretest to test by 0.10, versus a 0.01 increase for controls (P = 0.002); their skip-rate increase was 0.11 versus 0.06 for controls (P = 0.051).Solve-rate declines were also greater than for hint users and nonusers; the skip-rate difference versus controls was marginal.Source: §3.3 · pp. 7–8

4.1 Method

• After 33 exclusions, 168 participants remained (85 AI, 83 control); AI participants used GPT-5 for five problems then solved three unaided, while controls solved all eight without AI.Controls had a test-taking-tips sidebar during learning, removed for the final three problems; responses under five seconds were counted as skipped.Source: §4.2 · p. 8, §4.1 · p. 8

4.2 Results

• On final tests, mean solve rate was 0.76 in the AI condition versus 0.89 in control, while mean skip rate was 0.08 versus 0.01, indicating lower unassisted performance and persistence.The reading-comprehension replication suggests these effects are not specific to mathematical reasoning.Source: §4.2 · pp. 8–9, Figure 4 · p. 9

Cognitive offloading

• Cognitive offloading can improve task performance while risking poorer unaided performance; prior AI deskilling evidence is largely correlational, motivating these RCTs.Existing evidence has come largely from surveys or interviews.Source: §5 · p. 9

Human-AI collaboration

• Prior work considers human-AI systems optimized for long-term outcomes, while short-term optimization can misalign AI rewards with user needs; the authors frame deskilling as this mismatch.They characterize cognitive deskilling as an underappreciated consequence of systems optimized for immediate helpfulness.Source: §5 · pp. 9–10

Gradual AI risks

• Unlike related work on gradual AI risks to agency in relationships, creative output, and personal beliefs, this study examines cognitive erosion and supports a gradual-disempowerment hypothesis.The authors argue that sustained AI use risks eroding cognitive and motivational capacities and displacing human participation over time.Source: §5 · p. 10

• Brief AI exposure impaired independent performance and persistence, capacities foundational to lifelong learning; cumulative effects of daily use may be profound and difficult to reverse.Source: §6 · p. 10
• The authors propose that instant answers shift effort expectations, making unaided work feel costlier, while lost productive struggle undermines self-knowledge and persistence.They describe the effort shift as self-reinforcing and predict broader metacognitive decay beyond persistence.Source: §6 · p. 10
• Foundational arithmetic and reading support higher-order skills; the authors warn that cumulative learning risks may be hardest to reverse for students with fewer academic resources.They say user-facing measures such as Socratic AI or reduced use may help only at the margins and not resolve the deeper risks of offloading at scale.Source: §6 · p. 10
• The authors argue that mitigating these risks requires AI collaboration to prioritize long-term empowerment and care over short-term user satisfaction, supporting what people can do without AI.Source: §6 · p. 11
• The experiments tested brief exposure and assessed participants immediately after removing maximally helpful AI, leaving the durability and long-term accumulation of effects unknown.Whether Socratic AI can preserve assistance benefits without persistence costs remains an open question; the authors call for longitudinal experiments across multiple sessions.Source: §6 · p. 11

Ethics Statement

• All experiments were approved by the UCLA institutional IRB (IRB-25-2012), and participants provided informed consent before taking part.Source: Ethics Statement · p. 11
• Participants received $15 per hour, could withdraw at any time without penalty, and provided no personally identifiable information.Source: Ethics Statement · p. 11
• The experiments involved short cognitive exercises that were assessed as posing no risk of harm to participants.Source: Ethics Statement · p. 11

A.1 Participant Instructions Control condition:

• Control participants solved fraction calculations and word problems on their own and were instructed not to use AI tools or Google.Source: §A.1 · p. 17
• AI-assisted participants could use ChatGPT to work through problems, but were told not to use other AI tools or Google when it was unavailable.Source: §A.1 · p. 17
• Both conditions required fraction or mixed-number answers, and participants’ payment did not depend on correctness.Source: §A.1 · p. 17

A.2 Problems

• Problems appeared in a fixed, progressively harder order; answers had to be fractions or mixed numbers, not decimals.Source: §A.2 · p. 18
• The AI assistant was available during the 12-problem main phase but unavailable during the three-problem test phase for both conditions.Source: §A.2 · p. 18

A.3 AI-Assisted Condition: Chat Assistant Greeting

• During the main phase, the assistant opened each problem with a greeting inviting questions and could see the displayed problem and answer free-text queries.Source: §A.3 · p. 18

B.1 Participant Instructions

• Self-learning participants were instructed to solve fraction problems on their own, encouraged to use pen and paper, and given reference hints, but told not to use AI, calculators, or Google.Source: §B.1 · p. 19
• AI-assisted participants could use ChatGPT however they liked, including for hints, answers, or solutions; when it was unavailable, they were told not to use other AI tools, calculators, or Google.Source: §B.1 · p. 19

B.2 Problems

• All problems required answers as fractions or mixed numbers; decimal answers were not accepted.Problems were presented in pretest, main experimental, and final test phases.Source: §B.2 · p. 20
• All participants completed three pretest practice problems without access to either the AI chat assistant or the reference hints panel.Source: §B.2 · p. 20
• The 11 main problems were grouped by difficulty, with block order fixed from easy to hard and problem order randomized within each block.AI-assisted participants had access to ChatGPT, while control participants had access to a reference hints panel.Source: §B.2 · p. 20
• In the final test phase, all participants answered three fixed-order problems without access to ChatGPT or the reference hints panel.Source: §B.2 · p. 20

B.3 Scoring

• Equivalent reduced and unreduced fractions were correct; decimal-format answers were marked wrong, though decimals comprised only 0.14% of responses.Participants were instructed to answer in fraction format to discourage calculator use.Source: §B.3 · p. 21
• Mistyped answers and skips were automatically scored as wrong.Source: §B.3 · p. 21

B.4 Reference Materials Available to Participants

• Control participants could consult a fixed “How to Solve Fractions” sidebar containing three pretest solutions, which both conditions had already seen after each pretest problem.The panel therefore provided no new information to control participants.Source: §B.4 · p. 21
• The reference materials explained finding common denominators for addition or subtraction, multiplying numerators and denominators, and dividing by multiplying by the reciprocal.Source: §B.4 · p. 21
• AI-condition participants had a chat interface described as ChatGPT, which greeted them by offering help with the displayed problem and inviting questions.The assistant opened each problem with this greeting.Source: §B.4 · p. 21
• The assistant could access the currently displayed problem and respond to participants’ free-text queries.Source: §B.4 · p. 21

C.1 Participant Instructions

• Participants read a short text, answered a question about it, and selected the best option from a multiple-choice list.They were also given a practice round to learn the interface.Source: §C.1 · p. 22
• After the pretest, participants were randomly assigned to self-learning or AI-assisted instructions.The self-learning instructions provided a reference panel with reading-comprehension tips and encouraged participants to take their time working through solutions.Source: §C.1 · p. 22
• In the self-learning condition, participants were told not to use AI tools or Google when the reference panel was unavailable.Source: §C.1 · p. 22
• Participants in both conditions were told that payment did not depend on how many problems they answered correctly.Source: §C.1 · p. 22
• In the AI-assisted condition, participants could use ChatGPT however they liked, including for hints, answers, or solutions.Source: §C.1 · p. 22
• When ChatGPT was unavailable, AI-assisted participants were told not to use other AI tools or Google.Source: §C.1 · p. 22

C.2 Problems

• The main phase used five problems in random order, with an AI chat assistant for the AI-assisted condition and a reference-hints panel for the control condition.Each problem paired two short texts taking positions on a topic and followed them with a four-option multiple-choice question; questions came from free online SAT reading practice from the Manhattan Review1.Source: §C.2 · p. 23
• The pretest presented one practice problem to all participants without access to the AI assistant or reference-hints panel.Source: §C.2 · p. 23
• The final test phase used three problems in fixed order, with neither the AI assistant nor the reference-hints panel available to either condition.Source: §C.2 · pp. 23, 27
• The fine-free policy problem contrasts Levin’s projected near-term costs with Ruiz and Porter’s claim that early disruptions are temporary and longer-term engagement and efficiency gains should guide decisions.The prompt asks how Ruiz and Porter would respond to Levin’s findings.Source: §C.2 · pp. 27–28
• A second problem tests whether students see the bridge-concrete finding as overly optimistic, given uncertainty about the healing mechanism and its relevance to typical bridge conditions.Text 2 notes that the chemistry may depend on slow marine interactions and that the role of the clasts is not established.Source: §C.2 · p. 28

C.3 Reference Materials Available to Participants

• Control participants had a fixed sidebar with tips to read the question first, use context, and eliminate wrong answers; AI participants could query an assistant with access to the text, question, and choices.The interface was described to participants as ChatGPT and remained available throughout the main experimental phase.Source: §C.3 · pp. 28–29

C.4 Attrition Analysis

• 0.85 vs. 0.78 mean solve rate (control vs. AI; P = 0.03) remained in Experiment 2 among participants who answered all three pretest questions correctly.This filter excluded 207/349 AI participants and 173/318 control participants, leaving 143 AI and 140 control participants.Source: §C.4 · p. 29, Figure 6 · p. 31
• 0.07 vs. 0.05 mean skip rate (AI vs. control; P = 0.47) was numerically higher but not significantly different in Experiment 2’s filtered sample.Source: §C.4 · p. 29
• 0.78 vs. 0.89 mean solve rate (AI vs. control; P = 0.018) and 0.06 vs. 0.01 mean skip rate (AI vs. control; P = 0.019) persisted after matching sample sizes in Experiment 3.After excluding participants who missed the pretest question, researchers removed the two lowest-scoring AI participants, leaving 83 participants in each condition.Source: §C.4 · p. 29, Figure 6 · p. 31