AI Plagiarism and Cheating Detection System for Education

With the increasing use of LLMs in student works, false positive rates of AI detectors reach 15%, leading to unfounded accusations and trust issues. Our combined approach reduces this to 2-4% using a three-level detection: stylometric analysis, MinHash for inter-student copying, and dynamic crawling of external sources. The model considers the context of the discipline: for humanities works the perplexity threshold is lower, for technical ones it is higher. Over 5 years we have implemented 20 systems in universities and EdTech projects, achieving up to 80% savings on work checking budgets. The system covers three scenarios: plagiarism from open sources (internet, dissertation databases), copying between students (with paraphrasing), and AI generation (GPT, Claude, LLaMA). Each scenario requires its own algorithm, so an ensemble of models provides the best accuracy. Developing the system pays off by reducing teacher workload and improving check quality.

How do we distinguish AI-generated text from human-written?

AI detectors (GPTZero, Originality.ai) often produce false positives on structured works with templated phrases. Our approach combines three signals:

1. Text perplexity – LLM texts have anomalously low perplexity (each word is predictable). We use a GPT-2 model calibrated for the educational domain.
2. Stylometric variability – sentence length, rare word frequency, conjunction usage. Students write unevenly, LLMs write uniformly.
3. Semantic smoothness – overly coherent logical structure without speech errors. AI text lacks typical human 'noise'.

The joint model reduces the false positive rate by 60-70% compared to single detectors. More about perplexity on Wikipedia.

Why use MinHash for copying detection?

Full pairwise comparison of all works is O(n²). For 1000 works that would be 500,000 comparisons, taking hours. We use MinHash LSH – 10 times faster with 95% accuracy.

def detect_inter-student-similarity(submissions: list[Submission]) -> list[SimilarityPair]:
    # MinHash for approximate similarity
    from datasketch import MinHash, MinHashLSH

    lsh = MinHashLSH(threshold=0.4, num_perm=128)
    minhashes = {}

    for sub in submissions:
        m = MinHash(num_perm=128)
        for word in preprocess(sub.text).split():
            m.update(word.encode('utf-8'))
        lsh.insert(sub.student_id, m)
        minhashes[sub.student_id] = m

    pairs = []
    for sub in submissions:
        result = lsh.query(minhashes[sub.student_id])
        for match_id in result:
            if match_id != sub.student_id:
                similarity = minhashes[sub.student_id].jaccard(minhashes[match_id])
                if similarity > 0.4:
                    pairs.append(SimilarityPair(
                        student_1=sub.student_id,
                        student_2=match_id,
                        similarity=similarity
                    ))
    return pairs

The similarity threshold is adjustable: 0.4 for humanities, 0.6 for exact sciences. MinHash LSH is resilient to paraphrasing: at threshold 0.4 it catches 85% of paraphrased copies.

Handling Authorized Group Works

Group projects are a common cause of false positives. In our system, the administrator uploads a list of groups, and for students within the same group, similarity is not flagged as plagiarism. Additionally, a 'permitted borrowing' threshold is supported – e.g., up to 20% common phrases for technical disciplines. This is configurable in the interface.

What's Included in the Turnkey Solution

• ML Model (ensemble: stylometry + MinHash + AI detector)
• Web Interface for upload and reporting
• LMS Integration (REST API, plugins for Moodle/Blackboard)
• Comprehensive documentation (API spec, admin guide)
• Training sessions for teachers (2 sessions)
• 6 months of technical support post-deployment

Comparison of Detection Methods

The combined approach offers the best accuracy-speed ratio. Moreover, we conduct A/B testing on your dataset to calibrate thresholds. Our system is 2.5 times more accurate than GPTZero.

Our Experience and Guarantees

Certified ML engineers with experience in NLP and text processing. We have 5+ years on the market and have completed 20+ projects for universities and EdTech companies. Guarantees: false positive rate ≤ 5% on your dataset; integration with existing LMS in 2-3 days; 6 months of support post-deployment. For model quality monitoring we use Weights & Biases: after deployment, the model continues fine-tuning on new data, reducing concept drift. Implementation cost: typical projects range from $15,000 to $30,000 depending on scale, with an average savings of $50,000 per year for a university with 10,000 students.

Implementation Process

1. Analysis – requirements gathering, process audit, data assessment.
2. Design – model architecture, stack selection (PyTorch, Hugging Face, Vector DB).
3. Development – model training, interface creation, integration.
4. Testing – A/B test on real data, threshold calibration.
5. Deployment – rollout, staff training, documentation handover.

Typical Implementation Mistakes

• Using a single AI detector as the sole evidence leads to false accusations.
• Ignoring authorized group works: the system must account for context.
• High similarity threshold (0.8) misses paraphrased copies.

Contact us for a preliminary evaluation of your project. Get an engineer consultation and a pilot launch in 4 weeks. Order a demo test on your data to verify the system's accuracy.

NLP Development: Text Classification, NER, Embeddings, and Information Extraction

We often receive a task: process 50,000 support tickets — currently all manual. Dataset — 3,000 labeled examples, 12 categories, imbalance: one category occupies 40% of the sample, three at 1-2% each. Baseline accuracy — 78%. Sounds decent until you look at recall for rare classes: 0.31, 0.44, 0.28. These classes — complaints and churn threats — are most important to the business.

This is a typical NLP development project. The problem is not the algorithm but that accuracy is the wrong metric. Our experience across 30+ projects shows: we start by analyzing business metrics and only then choose the model.

Why accuracy is not the right metric for rare classes?

Accuracy ignores imbalance. If the "churn" class appears in 2% of cases, the model can predict "all good" and get 98% accuracy — but the business loses clients. Solution: F1 macro (averaged over all classes) or weighted F1. For NER — strict entity F1 (exact matches only). We guarantee: after choosing the correct metric, model quality becomes measurable and predictable.

Text Classification: From BERT to Distillation

BERT-like models are the standard for classification. ruBERT-base or ruBERT-large from DeepPavlov for Russian. multilingual-e5-large — for multiple languages in one pipeline. XLM-RoBERTa-large — a strong multilingual backbone.

Fine-tuning for classification: add a classification head on top of the [CLS] token, train for 3-5 epochs with lr=2e-5, weight decay=0.01. For imbalance — weighted CrossEntropyLoss or focal loss with gamma=2.0. Contact us — we will show a code snippet.

Imbalance case study. Dataset — 3,000 examples, imbalance 1:20. Solution: class_weight via sklearn + CrossEntropyLoss. Additionally — augmentation of rare classes via backtranslation (ru→en→ru through MarianMT). Recall for rare classes rose from 0.31 to 0.67 with a slight drop in accuracy (76%→74%). Full NLP development end-to-end took 3 weeks.

Distillation for production. BERT-large gives F1 0.89, but inference on CPU — 180ms. Distillation into DistilBERT or ruBERT-tiny2 reduces latency to 25ms with F1 0.84. Export to ONNX Runtime provides an additional 1.5-2x speedup. DistilBERT achieves 7x lower latency than BERT-large with only a 5% drop in macro F1 – a typical production trade-off.

How to choose between BERT and LLM for your task?

For most classification and extraction tasks, BERT-sized models offer the best trade-off between cost and performance. Shift to LLMs only when the task demands generation, complex reasoning, or zero-shot generalization.

NER: Named Entity Recognition

NER — extracting persons, organizations, locations, dates, amounts, document numbers. For general categories (PER, ORG, LOC), pre-trained models work well. For specialized ones (medical terms, legal concepts) — fine-tuning is needed.

Data annotation. The main cost of an NER project. For a quality model — 500-2,000 labeled sentences per entity type. Tools: Label Studio (open source) or Prodigy (by spaCy creators). IOB2 format — standard.

Architecture. Token classification on top of BERT: each token gets a label (B-PER, I-PER, O). spaCy 3.x with transformer pipeline — a convenient production choice.

Nested entities. Standard IOB models cannot handle nested entities (organization inside an address). For such tasks — span-based NER: SpanBERT or SpERT. More complex but correct.

Post-processing is mandatory. The model predicts tokens — normalized entities are needed. Date — dateparser. Amounts — regex + validation. Names — deduplication via rapidfuzz. Included in our standard delivery.

Sentiment Analysis and Opinion Mining

Binary classification positive/negative works out of the box with BERT. Complexity — aspect-based sentiment analysis (ABSA): "the restaurant has good food but terrible service." For ABSA: aspect extraction (NER) + sentiment per aspect. Joint models BERT-for-ABSA — quality on Russian data is lower due to dataset scarcity. RuSentiment, SentiRuEval — main resources.

For production with simple positive/negative/neutral: distil models are enough. Three classes, balanced dataset, 2,000+ examples — F1 macro 0.82-0.87 in 1-2 days.

Text Summarization

Extractive summarization (select sentences) — TextRank or BM25 without training. Fast, no hallucinations. Good for long documents.

Abstractive (generates new text) — seq2seq: mT5, mBART, FRED-T5, ruT5-large. For production via LLM API (GPT-4, Claude) — often the best cost/quality/speed trade-off.

Embeddings: Vector Representations of Text

Embeddings are the foundation of semantic search, deduplication, clustering, RAG. Quality critically affects downstream tasks.

Models. E5-large-v2, BGE-M3, multilingual-e5-large — strong multilingual embedders. sentence-transformers/paraphrase-multilingual-mpnet-base-v2 — fast option. For Russian: ru-en-RoSBERTa (Skoltech) performs well on semantic textual similarity.

Embedding quality evaluation uses the MTEB benchmark as standard. But top results on MTEB don't guarantee success on a domain dataset — we build domain-specific eval.

Fine-tuning embeddings. If standard models don't give the required Recall@k — contrastive learning on domain pairs with MultipleNegativesRankingLoss. How to perform this for domain data:

1. Collect 500–2,000 semantically similar pairs from your domain.
2. Apply MultipleNegativesRankingLoss with a batch size of 32–64.
3. Train for 1–3 epochs using AdamW (lr=2e-5).
4. Evaluate Recall@k on a held-out domain test set.

This approach yields a 5–15% improvement in Recall@k in practice.

Dimensionality and storage. E5-large: 1024 dim, float32 — 4KB per vector. For 10M documents — 40GB. Quantization int8 reduces to 10GB. FAISS IVF_PQ — more compact but with losses. Included in our deployment recommendations.

Information Extraction

Structured extraction is a frequent task. Examples: key contract terms, technical characteristics, dates and amounts from invoices.

1. Regex + rule-based. For INN, OGRN, amounts, dates — more reliable than neural networks. No data required.
2. NER + post-processing. For variable formats.
3. LLM with structured output. GPT‑4 / Claude with JSON schema — for complex documents. Cost: minimal per document. For 10k+ documents/day — we calculate the economics.

We guarantee a hybrid: regex/NER for typical fields + LLM for edge cases. Our guarantee is backed by years of production experience and more than 30 projects.

Work Stages

Prototype on existing data — 1-3 weeks. Production system with CI/CD — 1.5–2.5 months. Cost is calculated individually — get a consultation for a project estimate.

What's Included

• Model and pipeline architecture documentation
• Access to the model via REST API (FastAPI + ONNX)
• Client team training (2-hour webinar + Q&A)
• Accuracy guarantee on the agreed test set
• Months of post-delivery support (bug fixes, adaptation to new data)

Our Experience

Years of NLP projects from classification to RAG systems. The team includes ML engineers experienced with Hugging Face, spaCy, LangChain, MLOps. We use vLLM, Kubeflow, Weights & Biases — a production stack, not toys. Contact us to evaluate your NLP project within two days — request a free consultation on your text processing pipeline.