← Back to project catalogue
GP-DA-1KRCH3IData and AIReady

Human Oversight and Automation Bias in AI-Assisted Engineering Decisions

Explore why checking an AI answer, finding its mistakes and completing a correct repair are three different engineering tasks.

Human Oversight and Automation Bias in AI-Assisted Engineering Decisions project visual
GP-DA-1KRCH3I · Data and AI
  • Python 3.12
  • Matplotlib
  • Docker

Project definition

Problem statement

A reviewer can correctly flag a bad AI answer without completing a correct repair. A good answer can also be damaged during correction.

This study separates correct releases, wrong releases and held work, then asks how limited repair time changes the result.

Project objectives

  • Distinguish review accuracy from the correctness of the released engineering work.
  • Account for missed errors, false flags, failed repairs and unresolved work.
  • Identify exact review-capacity breakpoints under stated workload assumptions.
  • Compare improvement over a human baseline with improvement over both standalone decision-makers.
  • Connect the calculations with an evidence-based plan for a future oversight evaluation.

Project structure

Project components

01

Literature review

Connects human oversight, automation bias, appropriate reliance and current engineering-review evidence.

02

Review and repair

Tracks the probability of correct releases, wrong releases and held work through explicit outcome branches.

03

Limited capacity

Calculates how review and correction compete for a fixed time budget.

04

Joint errors

Separates human-only, AI-only and shared errors before analysing answer selection.

05

Uncertainty

Calculates exact bounds for independent parameter intervals and preserves undefined ratios.

Methodology

Project workflow

  1. 01
    Read the assumptions

    Start with the meaning of a correct answer, review selection, repair and held work.

  2. 02
    Trace the evidence

    Use the annotated sources and distinguish published findings from hypothetical inputs.

  3. 03
    Reproduce the study

    Regenerate the numerical outputs and compare them with the supplied results.

  4. 04
    Explore one change

    Change a review or repair assumption and explain the effect on completed work.

  5. 05
    Present the findings

    Use the editable slides to discuss results, limitations and independent further work.

Demonstration scenario

Compare a hypothetical engineering review with full repair capacity against one that holds flagged items unresolved. Explain why the second can show higher accuracy among released items while leaving substantial work unfinished.

Engineering

Tools and method

Analysis
The core Python calculations use no external runtime package, cloud AI service or API key.
Verification
171 tests cover conservation, limiting cases, interval witnesses, derivatives, source arithmetic and reproduction checks.
Documentation
Includes literature review, theory, methodology, results, discussion, conclusions, further work and labelled figures and tables.

Testing

Evaluation

Evaluation measures

  • Correct, wrong and held fractions per incoming item
  • Conditional accuracy together with release coverage
  • Review time, expected repair workload and capacity breakpoints
  • Feasible error overlap and undefined reliance ratios
  • Interval extrema, witness parameters and numerical conservation

Project boundaries

  • This is a theoretical study, not a new human-participant experiment or validated AI intervention.
  • Reviewer-response parameters are hypothetical, not fitted to students or engineering workers.
  • The capacity calculation is expected-work accounting, not a queueing or staffing simulation.
  • Independent interval bounds are not confidence intervals and do not cover every coupled workload model.
  • No employee ranking, individual trust score or regulatory certification is included.

Included

  1. 0173-page project documentation in PDF and editable Word formats
  2. 02Nine-page usage guide in PDF and editable Word formats
  3. 0320-slide editable presentation with source notes
  4. 0417 annotated source records and one attributed literature image
  5. 05Six labelled analytical figures and nine retained numerical outputs
  6. 06Hypothetical review, repair, capacity and error-overlap scenarios
  7. 07Complete Python source code, 171 automated tests and offline reproduction

Project record

No information is collected on this page.

Permanent project ID
GP-DA-1KRCH3I
Catalogued
21 Aug 2026
Completed
05 Sept 2026
Verified
05 Sept 2026
Demonstration
Included in repository

Handover

After purchase

  1. 01
    Payment is confirmed

    The project is marked unavailable and cannot be purchased again.

  2. 02
    Repository access is granted

    The buyer's submitted GitHub account receives access to the private repository.

  3. 03
    The purchase record is delivered

    The certification sheet is prepared from the reviewed buyer details and sent privately by email.