Our commitment: TrendShield uses AI to help safeguarding professionals identify and understand online risks more efficiently. We believe in being transparent about how our AI works, where it can go wrong, and why human professional judgement must always remain central to safeguarding decisions.
1. Overview
TrendShield uses artificial intelligence (AI) and machine learning techniques as part of its intelligence pipeline. AI helps us process large volumes of publicly available content, identify patterns in emerging online behaviour, and produce readable briefings for safeguarding professionals.
AI is a tool that supports — it does not replace — the professional judgement of Designated Safeguarding Leads, school staff, and safeguarding teams. All significant AI outputs are subject to human review before being published to school dashboards.
2. How AI is used in TrendShield
Trend classification
AI analyses publicly available content to identify whether a trend relates to a safeguarding risk category (e.g. exploitation, self-harm, dangerous challenges) and assigns an initial risk severity.
Risk scoring
AI produces a risk score for each trend based on multiple signals: content type, velocity of spread, platform, age group affected, and historical pattern matching.
Briefing generation
AI produces first-draft text for weekly briefings — plain-English summaries of each trend, its safeguarding implications, and suggested actions. This text is reviewed before publication.
Slang analysis
AI identifies new slang terms appearing in youth content and produces plain-English explanations of meaning and potential safeguarding context.
Parent communication scripts
AI generates suggested parent communication scripts based on identified trends. These are examples only and should be adapted by the school's safeguarding team.
Velocity and clustering
AI identifies whether similar trends are growing rapidly across platforms and groups related content signals to reduce duplication in briefings.
3. How the TrendShield pipeline works
Our intelligence pipeline processes publicly available data through a series of structured stages:
Data ingestion
Public content signals are collected from social media platforms and web sources via approved APIs. Only publicly available content is collected — no private messages or restricted accounts.
Normalisation and deduplication
Raw signals are cleaned, deduplicated, and standardised. Similar signals from multiple sources are clustered together to identify patterns rather than individual posts.
AI classification and scoring
AI models classify each trend cluster by category, assess risk level, calculate velocity (rate of spread), and assign a confidence score. Lower-confidence outputs are flagged for closer human review.
Three-stage approval evaluation
Trends are evaluated against approval criteria. High-confidence, multi-source critical risks may be auto-approved. Low-confidence, single-source, low-velocity signals are auto-rejected. All others enter a pending queue for human review.
Human review
Pending trends are reviewed by TrendShield staff before being published to school dashboards. Reviewers may approve, reject, or request further analysis of any trend.
Publication and briefing generation
Approved trends are published to school dashboards. AI generates briefing text which is reviewed and — where necessary — edited before inclusion in weekly briefings.
4. Limitations of AI outputs
Important: AI systems are not infallible. All TrendShield intelligence should be treated as a professional starting point — not as definitive fact. Safeguarding professionals must exercise independent judgement before acting on any AI-generated output.
Specific limitations of our AI systems include:
- False positives: AI may identify content as a safeguarding risk that, in full context, is not harmful. Risk scores may be overstated for some trends.
- False negatives: AI may fail to identify a genuinely harmful trend, particularly if it is very new, uses heavily coded language, or appears in platforms not covered by our data sources.
- Cultural and linguistic bias: AI models are trained on large datasets that may not fully represent all dialects, regional slang, or cultural contexts relevant to your school community.
- Temporal lag: Our pipeline runs periodically. Very rapidly emerging trends may not appear in briefings until the next pipeline cycle.
- Context limitations: AI analyses content signals in aggregate — it cannot assess the full contextual nuance of individual pieces of content the way a human reviewer can.
- Confidence variability: AI confidence scores are estimates. A high confidence score does not guarantee accuracy.
- Evolving platforms: New platforms and communication methods may not yet be covered by our data collection, meaning risks emerging on those platforms may not be identified.
5. Human review and oversight
We believe human oversight is essential for responsible AI use in a safeguarding context. Our approach to human review includes:
- Mandatory review queue: Any trend that does not meet our auto-approval criteria is placed in a pending queue for human review before it can reach school dashboards.
- Reviewer authority: Human reviewers can override AI risk scores, reject trends the AI has classified as concerning, and escalate trends the AI has underweighted.
- Briefing text review: AI-generated briefing text is reviewed by a human before inclusion in weekly briefings sent to schools. Reviewers may edit, rewrite or remove AI-generated content.
- Confidence thresholds: Trends with AI confidence scores below our thresholds are automatically flagged for heightened human scrutiny before approval.
- Feedback loop: School feedback on alert relevance is used to calibrate our AI models over time.
6. Automated decisions
TrendShield's AI does not make decisions about individual children, families, or members of the public. Our AI analyses aggregated patterns in publicly available content — it does not identify, profile or target individuals.
Trend approval decisions (whether to publish a trend to school dashboards) are partly automated, but:
- Auto-approved trends must meet strict multi-source, high-confidence criteria
- Auto-rejected trends are filtered out as low-quality signals, not published
- The majority of trends pass through human review before reaching schools
- Human reviewers may override any automated decision
In accordance with UK GDPR Article 22, TrendShield does not make any solely automated decisions that produce legal or similarly significant effects on individuals.
7. Data used in AI processing
Our AI systems process the following types of data:
- Publicly available social media content: Post titles, descriptions, hashtags, engagement metrics from public posts on platforms such as TikTok, YouTube, Instagram. Individual usernames and identifiers are not retained after processing.
- Web content signals: Public web pages, news articles, and open search trends relating to youth online behaviour.
- Platform trend signals: Aggregate trending topic and search data from public platform APIs (no individual user data).
Our AI does not process:
- Personal data about children or students
- Personal data about TrendShield platform users (account data is not used in trend analysis AI)
- Private messages or restricted content
- Health, biometric or other special category personal data
8. AI models and providers
TrendShield uses large language models (LLMs) provided by OpenAI to assist with trend analysis and briefing generation. These models are accessed via API and are subject to OpenAI's enterprise data processing terms.
Content submitted to AI models for analysis consists of publicly available trend content — not personal data belonging to TrendShield users, children, or families. We do not submit school-specific data (school names, staff details, pupil information) to external AI models.
We continuously evaluate AI providers against our safety, privacy, and accuracy requirements. Any change of AI provider will be reflected in an updated version of this notice.
9. How we improve our AI
We take an ongoing approach to AI improvement and accuracy:
- Human reviewer decisions are used to retrain and calibrate risk scoring models over time
- Pilot programme feedback is used to identify gaps in our trend coverage and accuracy
- We monitor AI output quality and regularly audit samples of AI-generated briefing content
- We update our model prompts and classification criteria in response to new types of online risk
- We aim to publish an updated version of this AI Transparency Notice whenever we make material changes to our AI systems
10. Your rights and raising concerns
If you have concerns about the accuracy of TrendShield's AI-generated intelligence, or wish to challenge an AI-assisted decision, you can:
- Contact us at hello@trendshield.co.uk — we will review any concern within 5 working days
- Use the feedback mechanism within the TrendShield platform to flag a specific trend or alert as inaccurate
- Request that a specific trend or alert be reviewed by a human member of our team
For data protection matters relating to AI processing, see our Privacy Policy or contact privacy@trendshield.co.uk.