PWR-200 · Human–AI Complementarity

Human–AI fallback collaboration

Useful human-AI systems require tested human-only or alternative routes for model error, outage and withdrawal.

Read the complete explanationCompare the Frontier record

Revision 7T · Full Tutorial Edition · Release manifest · Methodology · Corrections

One source of teaching truth

Canonical Power learning unit · TLU-PWR-200

The learner practises normal, wrong-output, slow-output and no-AI conditions, measuring detection, recovery time and post-outage accuracy against a manual baseline.

Canonical Power page
PWR-200 · Human–AI fallback collaboration
Full tutorial
Open full tutorial
Practical authority
The tutorial teaches a low-risk method that may be practised inside its stated limits.
Current treatment
Full low-stakes tutorial
Research depth
deep · 9 bound sources
Risk framing
moderate to high governance
Capability self-practice
Permitted inside the tutorial's stated low-risk limits
Pathway membership
G-CUR-021

Open My Power Path Inspect the canonical record

Read before using this page

Explanation is not permission.

This dossier explains the evidence and its limits. It is not a diagnosis, personal recommendation, assessment, clearance, performance promise or training programme. Actionable teaching appears only when the canonical Power record explicitly authorises it.

Complete bounded explanation

What this power means.

Capability to develop or express human–ai fallback collaboration in a declared context without inheriting broader claims.

What the current evidence supports

Useful human-AI systems require tested human-only or alternative routes for model error, outage and withdrawal.

How to observe or measure it without overclaiming

Failure detection, time-to-safe-state, recovery accuracy, override success and post-outage human performance versus pre-automation baseline. This measures the declared configured system and cannot by itself establish broad transfer, unaided ability, safety or legitimacy.

Myth

A human in the loop guarantees safety

Metric

Failure detection, time-to-safe-state, recovery accuracy, override success and post-outage human performance versus pre-automation baseline

Boundary

Report person, device/model/interface, task, assistance, failure state and comparator; the metric is not proof of unaided general capability.

Negative and limiting findings

  • People accepted incorrect advice, inherited bias and sometimes performed below their own baseline; explanations alone did not solve fallback.
  • No cited evidence supports perfect, universal or consequence-free performance.

Claims this evidence cannot support

  • A human in the loop guarantees safety
  • The tool or robot's output proves an unaided, universal or permanent human superpower
  • A successful laboratory task proves independent real-world or clinical capability

Individual tutorial · FULL SELF GUIDED TUTORIAL

How to learn this Power now.

The learner practises normal, wrong-output, slow-output and no-AI conditions, measuring detection, recovery time and post-outage accuracy against a manual baseline.

  1. Get readyUse a fictional workflow and static answer source; disconnect external actions.
  2. Learn the method8 concrete steps teach the permitted method from start to finish.
  3. See right and wrong4 comparisons show correct or safer execution beside common wrong or riskier choices.
  4. PractiseRun one 20-minute drill monthly, rotating wrong output, latency, tool loss and version change. Keep a quarterly fully manual check; progress only when detection and recovery improve without decay in manual accuracy.
  5. Check progressBaseline fixture: A fictional inventory workflow checks 15 stationery items against a static answer sheet. The AI normally flags mismatches; staged cards introduce a wrong flag, a 30-second delay and a complete outage. Enter success checks from “Build the manual baseline” and “Run the normal condition”. If “Wait for obvious catastrophe.” occurs, apply its named fix; then score Failure detection, time-to-safe-state, recovery accuracy, override success and post-outage human performance versus pre-automation baseline on the unused “Restart deliberately” item.

Authority boundary: Titan teaches a concrete low-stakes method step by step while keeping evidence and transfer claims bounded. The method is Power-specific; its actionability follows this treatment.

Open the full tutorialOpen My Power Path

Related Powers: PWR-193 · PWR-194 · PWR-195 · PWR-196

Direct evidence register

Sources that support—and limit—the claim.

Citation roles are explicit. A source may support existence or trainability while simultaneously limiting transfer, certainty, generalisation or safety.

  1. Limiting / contrary
    Humans inherit artificial intelligence biases

    Lucía Vicente; Helena Matute · 2023 · Primary research

  2. Primary empirical supportLimiting / contrary
    Explainability does not mitigate the negative impact of incorrect AI advice in a personnel selection task

    Julia Cecil; Eva Lermer; Matthias F. C. Hudecek; Jan Sauer; Susanne Gaube · 2024 · Primary research

  3. Primary empirical supportLimiting / contrary
    Relying on AI at work reduces self-efficacy, ownership, and meaning while active collaboration mitigates the effects

    Elena Hayoung Lee; Yidan Yin; Nan Jia; Cheryl J. Wakslak · 2026 · Primary research

  4. Primary empirical supportLimiting / contrary
    Experimental evidence on the productivity effects of generative artificial intelligence

    Shakked Noy; Whitney Zhang · 2023 · Primary research

  5. Primary empirical supportLimiting / contrary
    Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance

    Gagan Bansal; Tongshuang Wu; Joyce Zhou; Raymond Fok; Besmira Nushi; Ece Kamar; Marco Tulio Ribeiro; Daniel S. Weld · 2021 · Primary research

  6. Primary empirical supportLimiting / contrary
    To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making

    Zana Buçinca; Maja Barbara Malaya; Krzysztof Z. Gajos · 2021 · Primary research

  7. Primary empirical supportLimiting / contrary
    The impact of AI errors in a human-in-the-loop process

    Ujué Agudo; Karlos G. Liberal; Miren Arrese; Helena Matute · 2024 · Primary research

  8. Primary empirical supportLimiting / contrary
    Effect of Uncertainty-Aware AI Models on Pharmacists' Reaction Time and Decision-Making in a Web-Based Mock Medication Verification Task: Randomized Controlled Trial

    Corey Lester; Brigid Rowell; Yifan Zheng; Zoe Co; Vincent Marshall; Jin Yong Kim; Qiyuan Chen; Raed Kontar; X. Jessie Yang · 2025 · Primary research

  9. Limiting / contraryOfficial boundary context
    Artificial Intelligence Risk Management Framework (AI RMF 1.0)

    Elham Tabassi; National Institute of Standards and Technology · 2023 · Official authority

Current teaching and evidence boundary

The next gate remains visible.

External confirmation is required before higher-authority use because: a clinical, high-risk or regulated configuration.

A Power-specific tutorial is available at /tutorials/pwr-200/. It contains a plain-English method, 8 ordered steps, a worked example, right-versus-wrong comparisons, specific mistakes and corrections, practice, measurement, accessibility, stopping rules and evidence. Its self guided mode remains controlling.