A structured assessment tool for measuring human control in AI-assisted work.
▶ Launch the Interactive Evaluator
https://glatzp.github.io/ai-authorship-evaluator/
A measurement tool for authorship and human control in AI-assisted work.
This tool is part of a broader research system examining how AI systems can shift task definitions during multi-turn interaction.
It focuses specifically on measuring authorship and human control, not detecting alignment failures directly.
A lightweight tool for measuring authorship and human control in AI-assisted work.
AI-assisted work introduces a structural risk:
The work can look correct while control over the thinking process has shifted.
This tool does not evaluate output quality.
It measures:
Who actually controlled the thinking, decisions, and structure of the work.
The evaluator measures authorship behavior across four domains:
- Agency — Who defined the problem, constraints, and decisions?
- Cognitive Engagement — Who performed the reasoning and interpretation?
- Evidentiary Control — Who selected and verified the evidence?
- Authorship Integrity — Who determined structure, argument, and meaning?
These domains collectively measure:
Human control over the work
Before evaluation, the user defines the task environment:
- Speed
- Quality
- Ownership
- Risk
These inputs generate an expected authorship range for the task.
The system:
- Computes an expected authorship range based on task conditions
- Measures observed authorship behavior using structured criteria
- Compares expected vs observed authorship
Authorship Fit — the relationship between task demands and actual human control
This tool does not evaluate:
- Task alignment (whether the original objective was preserved)
- Output correctness or quality
- Whether the “right” problem was solved
- Intent or ethical use of AI
A task can show strong authorship fit while still being misaligned with its original objective.
This tool is part of a broader system that studies frame instability:
- Frame instability — the task definition can shift across multi-turn interaction
- Constraint drift — constraints degrade under pressure
- Goal substitution — constraints hold while the objective changes
- Meta-awareness instability — model explanations do not reliably reflect behavior
These mechanisms can reduce human control without obvious failure.
This evaluator measures the resulting change in authorship, not the mechanism itself.
Authorship = control of thinking, not production of output
A fluent result does not indicate that the human controlled:
- the decisions
- the reasoning
- or the structure of the work
- This is not a grading tool
- Higher authorship is not always better
- AI use is not penalized
- Task demands determine appropriate authorship
The goal is not to maximize control.
The goal is to:
Match human control to the demands of the task
AI changes:
- how decisions are made
- how responsibility is distributed
- how authorship is defined
Without visibility into authorship:
- alignment failures are harder to detect
- responsibility becomes unclear
- thinking can be silently outsourced
This tool makes those tradeoffs visible.
- Education (AI literacy, assessment integrity)
- Professional workflows (decision accountability)
- Writing, research, planning, design
- Post-task reflection and audit
- A measurement tool for authorship behavior
- A structured self-audit of human control
- A way to make invisible delegation visible
- An AI detector
- A grading system
- A replacement for human judgment
AI use is not the risk.
Loss of control over the problem being solved is the risk.
This tool measures where that control actually sits.
This project is licensed under the MIT License.
You are free to use, modify, and distribute this software with proper attribution. See the LICENSE file for full details.