EU AI Act & AI System Disclosure

The short version

• Deep AI Detector is an AI system under EU Regulation 2024/1689 (the "AI Act").
• In our intended use — providing an AI-likelihood signal to a human reviewer — it is a limited-risk system.
• It is not a "general-purpose AI model" / foundation model — it is a downstream classifier built on top of fine-tuned open models and hand-tuned rules.
• Where you deploy it in a context covered by Annex III of the AI Act (eg, education, employment), you become a "deployer" of a potentially high-risk system and must meet your own obligations. See §6 below.
• We provide a clear, machine-readable indication that the output is AI-generated probability — satisfying Article 50 transparency obligations on our side.

1. System identification (Article 13 / Annex IV-style)

• System name: Deep AI Detector AI-Text Classification Pipeline.
• Provider: the operator of deepaidetector.com (see Privacy Policy §2).
• Version: rolling; current detector ensemble version is shown in the dashboard footer.
• Intended purpose: to estimate the likelihood that submitted text was generated by an AI language model and to provide supporting signals (per-paragraph score, rhythm metrics, rule fires) to a human reviewer.
• Intended users: writers verifying their own work; editors; educators (with mandatory human review and due process); researchers; trust-and-safety teams; content publishers.
• Foreseeable users not covered: automated decision systems that act on the score without human review, anonymous credit-scoring, political-speech surveillance.

2. Risk classification (Article 6 + Annex III)

The Service is not in itself a prohibited AI practice under Article 5 (no social scoring, no manipulative subliminal techniques, no real-time biometric ID in public spaces, no emotion recognition in workplace/education).

Whether the system is high-risk depends on context of use, which is determined by the deployer (you), not us as a provider. In our intended use (informational signal feeding a human reviewer), the system is limited risk: Article 50 transparency obligations apply.

If a deployer uses the system in a context listed in Annex III — eg, determining access to educational institutions, evaluation of learning outcomes, employment-related decisions, screening of asylum applications — the deployment may bring the system into the high-risk category. In that case the deployer must meet Annex III Title III Chapter 2 obligations themselves (risk management, human oversight, log retention, fundamental-rights impact assessment per Art. 27). We do not at this time hold ourselves out as a CE-marked high-risk-system provider. Deployers in those contexts must independently assess and meet their obligations or refrain from such use.

3. Transparency obligations (Article 50)

Article 50 of the AI Act requires that providers of certain AI systems clearly inform natural persons that they are interacting with an AI system. We comply by:

• Labelling all detection outputs in the UI as "AI-likelihood score" and stating the probabilistic nature on the result page.
• Including the same label in every API response ("verdict_type": "ai_likelihood_probability").
• Linking this disclosure and the AI Detection Disclaimer from the result page, footer, Terms §6, and AUP §2.
• Documenting our methodology at a high level (training data sources, ensemble composition, calibration) in /methodology (forthcoming).
• Honouring Article 50(2)–(5) — deployers of synthetic content systems have their own labelling duties; that is not us.

4. Training-data overview (high level)

Detailed training-data documentation is maintained internally per Annex IV Item 2. A public summary:

• Training corpus: publicly available AI-text vs human-text datasets (HC3, GPT-4 Detect, BookCorpus subsets) plus internal augmentation runs with multiple frontier LLM generators (Claude, Gemini, GPT, Llama). 2 billion+ samples total. No personal data of identifiable individuals used in training.
• Reference signals: distributional statistics, lexical and structural patterns, and per-domain calibration. No additional training data beyond the corpus above.
• Customer text: we do not use customer submissions to train without explicit per-product opt-in.
• Bias mitigation: calibrated against samples from non-native English writers and across diverse content domains; documented disparities are described in the methodology page.

5. Known limitations & foreseeable risks

Per Article 13(3)(b)(iii) we document foreseeable risks:

• False positives — particularly on formal, polished, or non-native English writing. Mitigation: disclaimer, per-paragraph breakdown, recommended human-review workflow.
• False negatives — adversarial paraphrasing, "humanizer" tools, mixed AI+human text. Mitigation: regular model retraining; multi-layer ensemble.
• Distributional bias — accuracy varies by domain and language. Mitigation: domain-aware calibration; per-language thresholds; published envelope.
• Misuse for surveillance or punishment without review — addressed via Terms §6, AUP §2, mandatory disclaimer banner on the result page.
• Model drift — generative AI evolves; our calibration lags. Mitigation: weekly fine-tune cycle, published version, change-log.

6. Deployer responsibilities (Article 26 if high-risk deployment)

If you use the Service in a high-risk deployment context, the AI Act imposes obligations on you:

• Operate the system in accordance with our instructions (these legal pages + dashboard documentation).
• Ensure human oversight — a qualified person reviews and may override.
• Monitor operation and notify us of incidents (Article 26(5)).
• Retain automatically generated logs (the AI Act sets a 6-month minimum for high-risk).
• Perform a Fundamental Rights Impact Assessment (FRIA) under Article 27 if you are a public body, an entity providing public services, or a high-risk Annex III(5)(b)/(c) deployer.
• Inform affected persons that an AI system is in use and how to obtain explanations of decisions (Article 26(11)).
• Inform workers' representatives and affected workers if deployed in the workplace (Article 26(7)).

We will reasonably cooperate with deployers to fulfil these duties on request to [email protected].

7. Comparable regimes

• UK: the UK's Department for Science, Innovation and Technology AI Regulation principles apply via sectoral regulators (ICO, EHRC, FCA). We follow the ICO's AI and Data Protection guidance.
• US — federal: Executive Order 14110 (where still in force) and NIST AI RMF inform our risk-management approach. No federal AI-specific statute yet.
• US — California: CCPA/CPRA right to opt out of "automated decision-making technology" used to make a decision that produces "legal or similarly significant effects" — not triggered by an informational signal, but deployers must assess.
• US — Colorado AI Act (2026): the Service is a "general-purpose AI system" component, not a "high-risk AI system" in our intended use; deployers in high-risk consumer interactions must follow Colorado's risk-management requirements.
• NYC Local Law 144: if you use the Service as an "automated employment decision tool" you must obtain a bias audit and provide notice.
• Brazil PL 2338/2023 (proposed AI law): obligations roughly mirror EU AI Act tiers; we monitor enactment.

8. Methodology & documentation

Public methodology page: /methodology (forthcoming). It will describe ensemble composition, calibration approach, known limitations, and the change-log for model versions. The current detector ensemble version is shown in the dashboard footer.

9. Contact

AI-system / regulatory queries: [email protected].