Does fearing AI replacement make workers withdraw?
A three-wave survey of 443 Chinese employees linked perceived occupational replacement risk with later self-reported withdrawal, partly through weaker professional identity and work meaning and greater job insecurity. The observational, non-probability sample cannot show that fear caused behaviour or that any job was actually displaced.
Editorial responsibility: The Impact of AI Editorial Desk · Report a factual concern
At a glance
- 1A three-wave survey retained 443 full-time Chinese employees from 520 first-wave responses, with four weeks between each wave.
- 2Higher perceived AI replacement risk at wave one was associated with lower professional identity and work meaning and higher job insecurity at wave two, which in turn were associated with withdrawal at wave three.
- 3All focal measures were self-reported, recruitment relied on social-media sharing and baseline withdrawal was not measured, so the study cannot establish causation, reverse direction or actual job loss.
Research topic
Associations between perceived AI occupational replacement risk, professional identity, work meaningfulness, job insecurity, technological self-efficacy and later self-reported withdrawal behaviour
The answer: fear and withdrawal moved together, but causation is unproven
A peer-reviewed Chinese study found that employees who rated their occupations as more likely to be replaced by AI—and the consequences as more severe—also reported more later withdrawal behaviour. In the study's regression model, replacement-risk perception was associated with withdrawal at b=0.38 before the proposed mediators were added and b=0.11 after professional identity, work meaningfulness and job insecurity entered the model. The pattern is consistent with fear affecting how people relate to their work, but it does not demonstrate that AI exposure caused anyone to disengage.
The outcome was a four-item self-report scale covering reduced effort and thoughts of leaving, not payroll records, observed productivity, absenteeism or confirmed turnover. The research measured perceived risk rather than objective occupational exposure, adoption of a named system or actual redundancy. It therefore answers a psychological question: how perceived replacement threat is associated with identity, meaning, insecurity and reported withdrawal over time.[1]
The Impact Brief · Free
Follow the evidence in work & skills.
Get a five-minute weekday briefing on what changed, why it matters and where the evidence comes from. Choose the topics you care about.
The study first built a six-item measure of perceived replacement risk
Study 1 adapted an initial pool of eight questions from established risk-perception and AI-awareness measures. Eight reviewers—three assistant professors, three doctoral students and two human-resources professionals—assessed relevance and clarity. After two reverse-worded items performed poorly, the final instrument used three questions about the likelihood that AI could replace an occupation's tasks and three about the severity of the consequences. The validation sample contained 387 respondents.
Exploratory analysis produced the intended two factors, which together explained 80.55% of variance. Reported internal-consistency coefficients were 0.91 for likelihood and 0.83 for severity; a two-factor confirmatory model fitted better than a single factor. Those are useful measurement checks, not evidence that the scale predicts outcomes beyond existing AI-awareness measures. The authors explicitly say its incremental validity remains untested.[1]
The main survey separated measures across three waves
Study 2 recruited full-time employees with at least one year of work experience through WeChat and other social-media sharing. At wave one, 520 people completed replacement-risk, technological-self-efficacy and background measures. Four weeks later, 482 answered questions about professional identity, work meaningfulness and job insecurity. After another four weeks, 451 returned the withdrawal scale. Matching responses and excluding incomplete records and outliers left 443 employees for analysis.
The final group was 38.8% male and predominantly aged 20 to 40. Two-thirds had a bachelor's degree and a quarter had postgraduate education. It included 126 professional or technical employees, 100 agricultural-production workers, 88 managers, 67 administrative or clerical workers, 42 service workers and 20 manufacturing workers. That occupational spread is valuable, but the invitation-sharing method does not produce a representative workforce sample and the paper reports no industry information.[1]
Identity, meaning and insecurity were three separate pathways
After adjustment for gender, age, education, organisational tenure and occupation, perceived replacement risk was negatively associated with professional identity (b=-0.26) and work meaningfulness (b=-0.22), and positively associated with job insecurity (b=0.64). Each of those wave-two variables was associated with wave-three withdrawal in the final model: identity and meaning were negative predictors, while insecurity was positive. Bootstrap estimates for the three indirect paths were 0.08, 0.09 and 0.14, with reported 95% confidence intervals excluding zero.
These pathways should not be treated as a causal sequence simply because the surveys were staggered. The researchers did not measure the mediators or withdrawal at baseline, so they could not establish that each changed after the risk perception measure or rule out reverse causation. A worker already considering departure may rate AI as more threatening, while poor management, restructuring or economic anxiety could influence every measure.[1]
Technological confidence weakened some associations
The study also measured technological self-efficacy: confidence in learning and performing demanding technology tasks. At lower self-efficacy, the estimated indirect relationships between replacement-risk perception and withdrawal were stronger. For professional identity, the conditional indirect estimate was 0.15 at one standard deviation below the mean and 0.01 at one standard deviation above it. For work meaningfulness, the corresponding estimates were 0.26 and -0.02. Both high-self-efficacy confidence intervals included zero.
The job-insecurity pathway persisted at both levels, though it was larger at lower self-efficacy: 0.33 versus 0.23. This does not prove that a training course will prevent withdrawal. Self-efficacy may reflect prior opportunity, job quality, education, access to tools or supportive management. Because it was measured rather than assigned, the moderation result identifies a group difference to test, not an intervention effect.[1]
What employers and worker representatives should do with this evidence
The practical implication is to treat replacement fear as an organisational signal, not an attitude to suppress. Employers can describe which tasks are changing, what evidence supports the business case, which decisions remain human and what happens to saved time. Workers need routes to contest an automated decision and credible information about redeployment, pay, workload and progression. A generic assurance that AI will only assist people is unlikely to repair identity or meaning if daily work is being redesigned without them.
Training may help when it is connected to real authority and opportunity. Teaching employees to use tools while withholding information about role changes can shift responsibility for insecurity onto the individual. Unions, managers and public-service leaders should therefore combine skills support with consultation, job-design review and measurement of workload, absence, transfer and turnover. The study makes perceived threat visible; it does not show that every fear is accurate or that reskilling alone is the remedy.[1]
The evidence is self-reported and geographically bounded
All central variables came from the same participants, creating risks of social-desirability and common-method bias. Separating surveys by four weeks reduces immediate carryover but does not remove shared response tendencies. Occupational categories were included as controls, yet the analysis was not large enough to test whether the proposed model held within each occupation, and no sector- or industry-level exposure data were collected. Objective AI adoption was not linked to the responses.
The sample was recruited through Chinese social networks rather than a probability frame. Employment institutions, welfare protection, labour demand and the meaning attached to occupations differ across countries. The paper was funded by China's National Natural Science Foundation, and the author reported no commercial or financial conflict. Those disclosures are useful, but replication across labour markets and independently measured outcomes remains necessary.[1]
What would change the assessment
Stronger evidence would measure risk perception, identity, meaning, insecurity and withdrawal at every wave, combine responses with employer records and document the technology and task changes workers actually experienced. Cross-lagged or natural-experiment designs could test direction, while representative samples could show whether effects differ by occupation, contract, age, gender and bargaining power. Interventions should compare transparent consultation, job redesign and paid training rather than assuming one programme fits all.
Until then, the measured association deserves attention without overclaiming. Employees who perceived greater AI replacement risk later described weaker connection to their profession and work, more insecurity and more withdrawal. Technological confidence appeared protective for some paths but did not erase the insecurity relationship. The defensible response is better evidence and more credible worker involvement—not a claim that fear itself has been proved to cause disengagement.[1]
What this means for people
- Workers may distance themselves from jobs that feel threatened or stripped of meaning, but this study does not show that AI caused the behaviour or that displacement occurred.
- Employers should pair tool training with honest task-level communication, worker voice and measurable commitments on job design and progression.
- Technological confidence may buffer some associations, yet it should not be used to blame individuals for insecurity created by organisational decisions.
Global context
The study adds employee-level evidence from China across six broad occupational groups. Its proposed identity and insecurity mechanisms are relevant to AI adoption elsewhere, but labour protections, bargaining structures, technology exposure and cultural meanings of work differ substantially. Cross-country replication with objective adoption and employment data is needed before treating the reported effect sizes as global.
What the evidence does not yet show
- All focal variables were self-reported, leaving common-method and social-desirability bias.
- The observational design did not measure mediators or withdrawal at baseline and cannot rule out reverse causation.
- Recruitment through social-media sharing was not probability sampling, and no industry-level information was collected.
- Perceived replacement risk was not linked to a named AI deployment, objective exposure or actual job displacement.
- The newly developed scale's incremental validity beyond established AI-awareness measures remains untested.
What to watch next
- Representative longitudinal studies with every focal measure collected at each wave
- Links between worker reports and objective adoption, absence, turnover and job-transition records
- Occupation- and industry-specific results rather than pooled workforce estimates
- Trials of consultation, job redesign and paid training with identity, meaning and employment outcomes
Living evidence record
Impact record IAI-0W99NH2
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
13 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Frontiers in Psychology published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 13 October 2026
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Continue the story
Related reporting
Work & Skills
Does supportive AI improve work outcomes?
A survey of 695 German AI users linked perceived support to better self-reported work outcomes mainly through proactive job redesign. Because every measure came from one cross-sectional questionnaire, it cannot show that AI—or job crafting—caused higher productivity or wellbeing.
8 min · 1 source
Work & Skills
What can 816 Slovak organisations tell us about AI readiness?
A new peer-reviewed data resource documents AI use, plans, sentiment and barriers across 816 deduplicated Slovak organisations. Its value is a transparent baseline for secondary research—not a representative national adoption rate or proof that AI improves work.
9 min · 1 source
Work & Skills
Who is workplace AI leaving behind? PwC finds an access gap—but not its cause
PwC surveyed 49,364 workers across 48 countries and regions. AI use rose, while reported access to learning fell; the cross-sectional responses show an association, not proof that AI caused the divide.
5 min · 2 sources
Reader commentary
Add evidence, experience or a question
No account is required. Reader notes are published after a brief civility, relevance and safety check; disagreement is welcome.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.