Cognitive Offloading Analytics

This synthetic analytics exercise combines copying similarity, revision, verification, delayed recall, and confidence change into an inspectable offloading proxy. It exposes the chosen weights and tests alternative weighting scenarios. Because scores are normalized within each cohort and the bands are heuristic, the results support sensitivity analysis rather than diagnosing individual dependence on AI.

Start here

• Calculations, evidence and verification scope
• Figure sources and exact numerical paths
• Data status

Review scope: The existing suite requires unavailable dependencies; no full-suite pass is claimed. The bundled demonstration executed successfully in this review.

Detailed project documentation

Category: AI in Education A transparent benchmark for detecting possible cognitive offloading in AI-assisted learning.

Research prototype. All bundled data and results are synthetic demonstrations. Nothing in this repository should be interpreted as evidence about real learners, teachers, or institutions.

Why this project exists

Generative AI can support learning, but it can also make it easy to outsource the thinking that should remain with the learner. This repo turns that concern into measurable, testable behavioral hypotheses without pretending that a single score can read a learner’s mind.

The scoring pipeline keeps each behavioral proxy visible, including copy similarity, revision, verification, recall, and confidence change. Researchers can therefore examine how each assumption affects the final offloading signal rather than treating the score as a diagnosis.

Research questions

1. Which observable behaviors are consistent with heavy answer substitution?
2. Can verification and revision behaviors counterbalance superficial similarity signals?
3. How stable are risk bands under different feature weightings and thresholds?

What the repository does

The reference pipeline follows five stages:

1. Interaction traces
2. Behavior features
3. Transparent scoring
4. Sensitivity analysis
5. Risk-band reporting

The baseline is intentionally conservative and inspectable before any longitudinal learner data or richer behavioral models are introduced.

Core outputs

• copy_similarity
• revision_ratio
• verification_rate
• delayed_recall
• confidence_shift
• offloading_score

The dashboard above is generated from synthetic data and is included only to show what the analysis surface looks like. It is not a reported empirical result.

Quick start

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -e .[dev]
python examples/demo.py
pytest -q

You can also use Docker:

docker build -t cognitive_offloading_analytics .
docker run --rm cognitive_offloading_analytics

Repository structure

cognitive_offloading_analytics/
├── src/cognitive_offloading_analytics/        # core implementation and synthetic-data generator
├── examples/demo.py        # end-to-end reproducible demo
├── tests/                  # executable unit tests
├── docs/                   # research design, data dictionary, references
│   └── images/             # original project diagrams and demo visualisations
├── results/                # synthetic demo outputs only
├── config/default.yaml
├── Dockerfile
├── Makefile
└── pyproject.toml

Research design in one picture

The fuller design rationale is in docs/research_design.md, including constructs, assumptions, validation steps, and a proposed empirical extension.

Reproducibility choices

• Synthetic generation uses a fixed random seed.
• The core metrics are implemented as small, testable functions.
• The demo writes machine-readable results into results/.
• CI runs the tests on every push and pull request.
• No API keys, proprietary datasets, or external model calls are required for the baseline.

Responsible-use boundaries

• Behavioral proxies cannot establish cognitive state on their own.
• Similarity is not plagiarism detection and should never be used as a disciplinary signal.
• Thresholds are illustrative and must be validated for each learning context.

Strong next experiments

• Estimate uncertainty for each learner-level score.
• Compare behavioral proxies with think-aloud or stimulated-recall data.
• Model longitudinal changes to distinguish temporary assistance from persistent answer substitution.

References

See docs/references.md. The references are there to locate the project in current AIED, learning-analytics, human-centered AI, and instructional-design research. They do not imply endorsement or affiliation.

Citation

If you build on this research prototype, use the metadata in CITATION.cff.

License

MIT for the code in this repository. Research data from future studies should use a separate data-governance and consent process.