SoJen.AI AI Transparency Statement

Effective Date: July 5, 2026 Last Updated: July 5, 2026

SoJen.AI is committed to transparency about how its artificial intelligence systems work, what they are designed to do, and where human judgment remains in the loop. This statement describes the design, capabilities, limitations, and responsible use principles governing the SoJen.AI API and its underlying machine learning models.

1. What the System Does

The SoJen.AI API is a pre-publish communication intelligence system. It analyzes text submitted before publication and returns a structured risk assessment covering:

• Implicit risk patterns — framing, tone, and language that carries hostile or discriminatory characteristics without explicit markers; the content general classifiers pass
• Explicit hostile language — direct threats, targeted attacks, group-derogatory language, and high-confidence harmful terms
• Sentiment — positive, negative, or neutral framing affecting how content lands with its audience
• Tone — hostility, condescension, dismissiveness, and other register markers independent of surface content
• Intensity — escalation signals indicating content trending toward high risk, enabling earlier intervention
• Custom lexicon signals — per-organization terms and risk categories flagged according to customer-defined standards

The API also returns rewrite suggestions through a large language model component, offering authors an alternative phrasing when content is flagged. Authors retain the choice to revise, accept the suggestion, or publish as written.

2. Model Architecture

The SoJen.AI detection system is built on a transformer model ensemble using DistilBERT and RoBERTa architectures, trained on a curated dataset and refined through a proprietary methodology that is the subject of a pending U.S. patent application.

The system operates across three age-tiered detection contexts — adult, youth, and children under 13 — with separate model thresholds and risk signal weighting appropriate to each population. The children's context is COPPA-aware and applies the most conservative detection thresholds.

3. Performance

The SoJen.AI detection models have achieved 98% F1 recall on the evaluation dataset used during model development, covering both implicit and explicit language risk detection.

What this means in practice:

• F1 recall measures the model's ability to correctly identify risk where it is present, balanced against precision — flagging content that is genuinely risky rather than generating excessive false positives
• No model achieves perfect performance. False positives (content flagged that is not harmful) and false negatives (harmful content not flagged) both occur
• Performance on content that differs significantly from the training distribution — highly specialized domain language, novel slang, emerging terminology — may vary from published benchmarks
• The custom lexicon feature allows organizations to supplement the base model with their own terminology, improving detection accuracy within their specific content environment

4. What the System Does Not Do

• Does not make enforcement decisions. The API returns risk assessments and suggestions. Whether to act on a flag, revise content, or allow publication is always a human decision — made by the author, a moderator, or the platform operator
• Does not profile individuals. The system analyzes text content at the time of submission. It does not build persistent profiles of individual users based on API submissions across sessions
• Does not make legal determinations. A risk flag is not a legal, regulatory, clinical, or disciplinary finding. API output must not be presented as such
• Does not guarantee detection of all harmful content. The system is designed to substantially reduce communication risk, not eliminate it. Organizations should treat the API as one layer in a broader content safety posture, not a sole control

5. Human Oversight

SoJen.AI is designed as a human-in-the-loop system. The pre-publish model is specifically chosen because it preserves author and reviewer agency:

• Authors see what was flagged and why — specific detection signals are returned, not only a risk score
• Authors receive a suggested rewrite but retain full choice over whether to revise or publish as written
• Organizations configure their own risk thresholds and custom lexicons rather than relying solely on default model outputs
• Every detection event is logged, creating an audit trail available for human review

SoJen.AI does not support fully automated content blocking or enforcement without human review. Customers who use API output to trigger automated actions do so under their own policies and bear responsibility for those decisions under the Acceptable Use Policy.

6. Training Data

The models were trained on a curated dataset assembled for the purpose of detecting bias, hostile language, and harmful communication patterns across a range of text types, registers, and organizational communication contexts.

Customer Input Content is not used for model training by default. Customers may also request a training exclusion to ensure their submitted content is never used for model improvement. See the Privacy Policy and Data Processing Agreement for details and eligibility by tier.

7. Custom Lexicon

Each customer organization may configure a custom lexicon — a set of organization-specific terms, phrases, and risk categories — that supplements the base model detection. Custom lexicons allow organizations to apply their own communication standards, industry-specific terminology, and risk thresholds independently of other customers.

Custom lexicons are scoped to the customer's API key and do not affect detection behavior for any other customer.

8. Model Updates

SoJen.AI periodically retrains and updates its models to improve performance, address emerging language patterns, and correct identified issues. Customers will be notified of significant model updates that may materially affect detection behavior via email or portal notification with at least 30 days' advance notice where practicable.

Emergency updates required to address critical safety or security issues may be applied without advance notice.

9. Contesting a Detection Result

If your organization believes the API has returned an inaccurate result — flagging content that is not harmful, or failing to flag content that is — we want to know. Feedback of this kind directly informs model improvement.

Email: legal@sojen.ai
Subject line: Detection Feedback

We review all feedback submissions in aggregate for model improvement purposes. Custom lexicon adjustments can be made immediately through the customer portal to address specific detection concerns without waiting for a model update cycle.

10. Responsible AI Commitments

SoJen.AI was built on the following design principles:

Pre-publish, not surveillance. The system intervenes at the moment of authorship — before content is sent. It is not designed for retrospective surveillance of stored communications or building behavioral records of individuals over time.

Author agency preserved. Flags are guidance, not barriers. The system informs the decision; it does not replace it. Authors can see what triggered a flag, read the suggested alternative, and choose their own path.

Explainability over opacity. Specific detection signals are returned so authors and reviewers understand what was detected and why. A black-box score alone is not sufficient for an organization to act responsibly on a flag.

Age-appropriate protection. Separate detection contexts for adult, youth, and children's environments reflect the meaningfully different protection standards appropriate to each audience — with children receiving the most protective thresholds.

No use against the system's mission. The API may not be used to identify content that evades detection and then disseminate it, or to build systems that generate the kinds of harmful content the API detects. See the Acceptable Use Policy for the full list of prohibited uses.

11. Regulatory Context

SoJen.AI's transparency practices are informed by the evolving regulatory landscape governing AI systems in high-risk contexts, including:

• EU AI Act — SoJen.AI operates as a communication risk management tool in employment, organizational, and platform contexts. The transparency and human oversight requirements of the EU AI Act inform our design choices and this disclosure.
• EEOC guidance on AI in employment — For customers using the API in hiring, performance management, or internal communications, this statement supports documentation of human oversight and organizational accountability.
• COPPA — The children's detection context reflects COPPA's requirements for platforms serving users under 13.

Contact: legal@sojen.ai | sojen.ai