Explainable AI text classification
The prototype produces explainable AI text risk classifications rather than relying on a label alone.
Small language model case study
AI Text Detector SLM
A fine tuned small language model for explainable AI text detection and adversarial robustness.
The rapid proliferation of AI generated text across professional and institutional environments has introduced a new class of operational risk. Organizations can no longer assume that written submissions, published content, or support communications are human authored and the consequences of misclassification vary significantly by context.
Project classification: This prototype explores explainable and adversarially tested AI text classification using a fine tuned small language model.
Risk based results rather than binary labels alone
Confidence outputs support risk based review instead of automatic conclusions.
Improved resilience to rewritten AI text
Adversarial testing evaluates how the classifier behaves on paraphrased and rewritten AI text.
01 / Business context
Who needed the solution
The rapid proliferation of AI generated text across professional and institutional environments has introduced a new class of operational risk. Organizations can no longer assume that written submissions, published content, or support communications are human authored and the consequences of misclassification vary significantly by context. Hiring teams encounter AI generated job applications and assessments. Academic institutions handle AI assisted submissions.
02 / Challenge
What needed to change
Detection must remain accurate even after deliberate paraphrasing or humanization. This adversarial robustness requirement fundamentally reshapes training methodology, dataset design, and feature engineering.
Simple AI or human labels lacked useful confidenceBinary outputs without confidence most systems provide only AI/Human labels without risk scoring or uncertainty indicators.
Users could not understand why text was flaggedLack of explainability stakeholders cannot understand or validate why text is flagged.
Rewriting and paraphrasing could bypass basic detectorsVulnerability to rewrites paraphrasing or humanization tools can bypass standard detectors.
Structured or non native writing created false positivesHigh false positives structured writing styles and non native language patterns are often misclassified.
Large general models increased cost and latencyVulnerability to rewrites paraphrasing or humanization tools can bypass standard detectors.
03 / Workflow transformation
Before and after
Previous workflow
- Simple AI or human labels lacked useful confidence
- Users could not understand why text was flagged
- Rewriting and paraphrasing could bypass basic detectors
- Structured or non native writing created false positives
- Large general models increased cost and latency
Structured workflow
- Build a hybrid training set
- Fine tune RoBERTa with LoRA
- Produce confidence and explanations
- Test adversarial robustness
- Review structured outputs
04 / Solution
How we approached it
The solution is built on a fine tuned RoBERTa classifier using LoRA (Low Rank Adaptation), trained on a hybrid dataset and enhanced with explainability and robustness layers. Model Selection: RoBERTa with LoRA RoBERTa was selected for its strong sequence classification capabilities and sensitivity to stylistic patterns distinguishing AI from human text.
Build a hybrid training set
Combine public, human written, AI generated and rewritten examples across domains.
Fine tune RoBERTa with LoRA
Adapt an efficient classifier without full model retraining.
Produce confidence and explanations
Return risk scores and interpretable text level signals instead of only a label.
Test adversarial robustness
Evaluate paraphrased and humanised text alongside false positive behaviour.
05 / Example workflow
From input to reviewable output
1Build a hybrid training set
2Fine tune RoBERTa with LoRA
3Produce confidence and explanations
4Test adversarial robustness
5Review confidence and explanation
6Human evidence based decision
06 / Delivery scope
What the delivery covered
Model and dataset workflow
- Hybrid human and AI text dataset construction
- Label and split preparation
- Adversarial rewrite test set
- Confidence and explanation outputs
Model engineering
- RoBERTa fine tuning with LoRA
- Class imbalance and calibration review
- Robustness and false positive evaluation
- Structured inference API
Application engineering
- React review interface
- FastAPI service layer
- Pydantic validation
- Python evaluation pipeline
Scope boundary
The prototype produces probabilistic risk signals and must not be treated as proof of authorship or misconduct. No benchmark value is published without dataset and false positive documentation.
07 / Architecture and controls
How the system is organised
RoBERTa
React
Python
FastAPI
Pydantic
PyTorch
Pandas
LoRA
Probabilistic result boundary
Detection output is a risk signal and must not be treated as conclusive proof of authorship.
False positive evaluation
Human written and adversarially rewritten text must be included when evaluating model behaviour.
Confidence and explanation
The interface exposes confidence and supporting indicators for reviewer interpretation.
Human review
Consequential decisions require independent evidence and qualified human review.
08 / Business value
Value observed during validation
1Pilot observed
Explainable AI text classification
The prototype produces explainable AI text risk classifications rather than relying on a label alone.
2Pilot observed
Risk based results rather than binary labels alone
Confidence outputs support risk based review instead of automatic conclusions.
3Pilot observed
Improved resilience to rewritten AI text
Adversarial testing evaluates how the classifier behaves on paraphrased and rewritten AI text.
4Pilot observed
Reduced false positive risk through broader evaluation
Broader evaluation helps identify false positive risk on human written content.
5Pilot observed
Efficient deployment with a fine tuned smaller model
A fine tuned smaller model demonstrates a path to lower latency and lower cost inference.
09 / Technology
Every component, with its role in the delivery.
Technology choices from the supplied project brief, mapped to the workflow each component supports.
RoBERTa
Fine tuned text classification model
React
Reviewer or product user interface
Python
AI, data processing and backend logic
FastAPI
Backend APIs and workflow orchestration
Pydantic
Structured schema and output validation
PyTorch
Model training and inference
Pandas
Tabular data processing and analysis
LoRA
Parameter efficient model fine tuning
NumPy
Numerical processing
Related work
Explore related case studies.
