WorkspaceWhy not detection?
DEMO

Detection is a race you cannot win.

Three arguments for asking a different question, carried over from the original AuthWords proposal.

The current landscape

How generated text actually enters a classroom.

Direct generationA chatbot writes the entire paper.
Paraphrasing toolsRewriters disguise generated text.
Hybrid draftsAI drafts, the student edits.
Prompt engineeringCustom instructions mimic the student's voice.
Model shoppingSeveral tools are used so no single pattern shows.
Self-trainingThe student feeds a model their own past writing.

What detectors do today

Commercial tools claim 80 to 95 percent accuracy. Results in practice vary widely, and false positives land hardest on multilingual writers.

Pattern recognition

Looks for repetitive structures, overused transitions, uniform paragraph length, and a relentlessly formal tone.

A student edits those patterns out in minutes.

Statistical analysis

Examines word-choice probabilities, n-gram frequencies, vocabulary sophistication, and syntactic complexity.

Models are trained to match human statistics ever more closely.

Perplexity scoring

Measures how predictable the text is to a language model. Low perplexity is read as machine-like.

It assumes machines write predictably. Modern models are tuned not to.

ARGUMENT 1

The proliferation problem

  1. P1New models ship faster than detectors can be retrained.
  2. P2Each model has its own patterns.
  3. P3A detector must be retrained for every one.

Therefore: Detection tools are perpetually out of date.

ARGUMENT 2

The customization problem

  1. P1Models can be fine-tuned and prompted to mimic a specific style.
  2. P2Students can train on their own prior writing.
  3. P3Personalized output is indistinguishable from the student's own.

Therefore: Detection becomes impossible once the model writes like the individual.

ARGUMENT 3

The arms race

  1. P1Model developers optimize for human-sounding text.
  2. P2Detection research reveals exactly what makes text detectable.
  3. P3That knowledge feeds the next, less detectable model.

Therefore: Detection research directly enables evasion.

THE FLIP

Stop asking “Did an AI write this?”
Start asking “Did this student write this?”

Stop chasing every new model. Verify what can be verified: consistent, attributable human authorship. Use the institution’s own evidence to protect the student.

Future-proof: works regardless of model progress Positive framing: verifies achievement, does not police Evidence-based, so it is defensible Private by architecture

Read argument two again. It cuts both ways.

A student who trains a model on their own writing defeats a style comparison too. The institution’s only real advantage is writing the student never had a copy of: supervised, timed, in-class work and the revision history behind a draft.

So style is a secondary signal.

The production design weights supervised writing and signed revision history most heavily, and treats stylometry as context rather than verdict. The public demo shows why: a simple style index cannot tell Austen from Shelley.

And the assignment itself can change.

When the prompt is the graded artifact, the student’s prompts and iterations are the evidence. There is nothing to detect, because the generated essay was never the point.

What changed since the original proposal. Demographic priors and regional language models were removed; context is voluntary and can never raise a misconduct prior. Email collection is off by default and never crawled. Surveillance concerns are answered with architecture and consent, not with “no worse than plagiarism detection.”