Browse the book & resources

Supporting material

Desk-researched interview 5: The right to decline AI

Desk research, not original interviews. These question sets synthesize published accounts; no practitioner was interviewed for this book.

Question: When is it acceptable not to use AI? Are people rewarded for critical evaluation rather than visible usage?

Method: Desk research from published employee interviews, developer survey commentary, qualitative research and engineering accounts. These are not interviews conducted for this book. Research snapshot: November 2026.

What the published accounts say

  • In Anthropic's 53 internal qualitative interviews, employees describe using Claude for verifiable, self-contained or disposable tasks but keeping some conceptually difficult debugging, high-level design and context-heavy decisions in human hands. One respondent would skip it for tasks expected to take less than ten minutes; another deliberately practices without AI to stay sharp. This shows reasons employees give for not using it, not a formal company policy or evidence of promotion rewards. Anthropic notes selection and social-desirability limitations in its non-anonymous internal research. Anthropic, 2025, “AI delegation approaches” and appendix.
  • Atlassian's 2026 report commentary says 60% of surveyed developers felt AI usage was treated as a productivity signal and 41% sometimes used it more than necessary because they felt expected to. These are respondents' perceptions across organizations; the figures do not show any particular employer's performance-review rules. Atlassian, 2026.
  • HubSpot says it made AI fluency a baseline hiring and role expectation after reaching high adoption. Its account also describes pilots measuring velocity, review burden and production incidents. It does not say employees were evaluated on a usage quota or that declining AI on a particular task was penalized. HubSpot engineering, 2025.

Synthesis, not testimony

Reasonable grounds to opt out: the task is faster to do directly; output cannot be checked affordably; permissions or confidentiality are unsuitable; the decision needs missing organizational context; or hands-on practice is a deliberate learning goal. These are proposed decision criteria, not verified universal policies. The primary outcome is engaging (autonomy and learning), with reliable judgment and effective problem selection as checks.

Whether critical evaluation is actually rewarded remains unanswered. Look at promotion rubrics and review narratives, not just policy statements: did someone receive recognition for rejecting a risky generated change, narrowing an AI use case, running an independent check or preserving a mentoring opportunity? Compare those cases with recognition for visible usage.

Questions to ask a practitioner

  1. Show a recent task where you consciously did not use AI. Did anyone ask you to justify the choice?
  2. Have managers celebrated a well-evidenced “no,” or only adoption and generated-output volume?
  3. What written policy protects justified non-use, and what evidence matters in performance reviews?

Follow-up desk-researched responses

1. A concrete non-use decision and whether it needed defending? Anthropic interviewees describe choosing to do a conceptual design or difficult debugging task themselves, skipping AI on a task under ten minutes, or sometimes practicing without it. These are reported choices by anonymous employees, not observed tasks with a manager response. Salesforce's chief availability officer describes a separate explicit boundary: generative AI does not independently change live production; humans approve production-impacting decisions. A restriction on autonomous actions is not a company-wide rule allowing engineers to decline AI for all tasks. Anthropic employee interviews; Salesforce engineering Q&A.

2. Are thoughtful “no” decisions rewarded? Unverified. Atlassian's surveyed developers report perceived usage pressure (60% say usage is treated as a productivity signal; 41% sometimes use AI more than necessary), but this does not reveal what managers actually rewarded. Feng, Yun and Wang's qualitative work with 10 employed juniors and 10 seniors finds that organization-level policies and tooling shape their choices before individual preferences do; it does not audit promotions either. Atlassian, 2026; Feng et al., 2026.

3. What written protection and appraisal criteria? The cited accounts establish neither a specific safe-to-decline policy nor actual performance-review criteria. The practical next artifact is a current internal policy plus anonymized appraisal examples that credit justified AI use and justified non-use, risk detection, validation and mentoring rather than seat counts; this is a proposal, not a reported industry practice.

Evidence gap: None of these publications audits actual promotion decisions or establishes a safe-to-decline policy across the industry.