AI Transparency Notice

Version 2, last updated 11 October 2026

What changed

Corrections, in effect from 11 October 2026.

Section 1 no longer claims we analyse reach for signals of algorithmic suppression.
Sections 5 and 6 say which scans get a second AI verification pass and which do not.
Section 3 no longer claims we never send data about other people. Your own posts can show or name someone else.
Section 8 sets out our position under the EU AI Act, how AI generated text is marked, and a review cadence for this Notice.
Section 5 no longer says Sentinel only applies rules found in the published policies. It now says that Sentinel also flags some content on its own judgment and never presents those findings as a platform rule.
Sections 5 and 6 now say that posts with several images are, like single images, not re-checked by the verifier.
Section 6 now says that Sentinel is not a crisis or emergency service.

We keep every earlier version on file. Email sentinel@reviewal.team for the text that applied on a particular date.

This AI Transparency Notice explains how Sentinel by Reviewal uses artificial intelligence to analyze content, what data is sent to AI providers, and the limits of what AI assessments can and cannot tell you. It is intended to give you a clear, honest picture so that you can use Sentinel's outputs appropriately.

This Notice complements our Privacy Policy and Terms of Service. It is a transparency document, not a separate legal contract.

1. Why Sentinel uses AI

Social media platform policies are long, nuanced, and constantly evolving. Reviewing every caption, image, or video manually against those policies is impractical at scale. Sentinel uses AI to:

Detect risky phrases, engagement-baiting patterns, and compliance issues in your content.
Score content against calibrated compliance rubrics.
Produce plain-language explanations of why something might be flagged and how to fix it.
Analyze publicly visible engagement patterns for signals of inauthentic engagement or account health issues.

2. Which AI providers we use

2.1 Anthropic (Claude)

Our primary AI engine. Anthropic's Claude models process caption text, bio text, highlight titles, and (for manual scans) uploaded images and videos to produce compliance assessments. Anthropic is a U.S.-based AI safety company.

2.2 Groq (Llama and Whisper)

Our fallback engine and audio transcription provider. Groq's Llama models serve as a fallback when the primary engine is unavailable, and Groq's Whisper service transcribes the audio track of video files submitted for scanning. Groq is a U.S.-based AI inference provider.

3. What data is sent to AI providers

We send only the data necessary to perform the compliance assessment. Specifically:

The text or media being scanned (a caption, a bio, a highlight title, an uploaded image or video for manual scans, a message script you paste or attach for review, or a question you type to the compliance assistant).
A minimal account context (the connected account's public handle, follower scale, and bio) so the AI can disambiguate content that reads differently in different niches.

We do not send:

Your account credentials or authentication tokens.
Your billing information.
Data about other users' accounts. We send the content of your own posts and anything you submit for review, which may show or name other people, but we never connect to or scan an account you have not connected.
Your inbox or any direct message we did not receive from you as a pasted or attached script, private content, or any non-public data.
Your identity beyond the account context described above.

4. Whether AI providers train on your data

Under the standard API terms of both Anthropic and Groq at the time of publication, neither provider uses API inputs to train its models.

We monitor these terms and will update this Notice if they change. If a provider changes its terms in a way that materially affects data use, we will notify users before continuing to route data through that provider.

5. How the analysis works

Sentinel does not treat AI output as the final word. A scan combines several layers, and which layers run depends on what you submitted:

A deterministic rule layer that catches clear-cut cases without AI involvement.
A primary AI pass that produces structured findings against the compliance rubric.
A verifier pass that re-checks the primary output to reduce false positives. It runs on text-only scans whose first-pass result falls in a range we treat as uncertain. It does not run on scans of an image, a carousel of images or a video, and it does not run when a scan is served by the fallback engine.
A score and verdict calculation that translates findings into a numerical score and a human-readable verdict.
A caching layer keyed on the content hash and rubric version, so unchanged content is not repeatedly re-sent to AI providers.

The rubric is calibrated against Meta's publicly available policies, mainly the Community Standards. When a finding rests on a published policy, Sentinel names that policy. Sentinel also flags some content it judges risky even where no published policy names it. Those findings are Sentinel's own assessment, and Sentinel never presents them as a platform rule.

6. Limits of AI assessments

AI is imperfect. Sentinel's compliance assessments have important limits:

Estimates, not guarantees. A high compliance score means our system did not find issues; it does not guarantee the platform will treat your content favorably. A low score means we found risks; it does not mean the platform will definitely take action.
Public policy only. Platforms may enforce internal, unpublished rules that no external tool can predict.
False positives and false negatives. The AI may flag content that is actually fine, or miss content that is actually risky. The verifier pass reduces this on the text scans where it runs. Scans of an image, a carousel of images or a video rest on a single pass, so treat those results with more caution.
Context matters. The AI sees your content and account context, but not your commercial arrangements, your niche's cultural norms, or the full history of how you have posted.
Not legal advice. Sentinel's outputs are not legal, business, or professional advice.
Not a crisis service. Sentinel is not a crisis or emergency service. When content reads like a personal crisis, a result may suggest talking to someone you trust, a local crisis line or emergency services. Sentinel cannot respond to emergencies.

7. Automated decisions and human oversight

Sentinel does not make legally significant automated decisions about you. All recommendations, threat findings, and account health diagnostics are informational. You are always the party who decides whether to publish content, change your posting strategy, or act on any of Sentinel's suggestions.

You have the right, under GDPR Article 22 (and equivalent laws), not to be subject to a decision based solely on automated processing that produces legal or similarly significant effects on you. Sentinel does not make such decisions. Every user retains full control over their content and their social media accounts.

8. EU AI Act positioning and AI marking

Under the EU AI Act (Regulation (EU) 2024/1689), the transparency obligations in Article 50 apply to Sentinel from 2 August 2026. Sentinel is not a high-risk AI system under Annex III, and we do not use AI for any of the practices prohibited by Article 5, for biometric identification, or for credit or social scoring. Our obligation is transparency, and we meet it as follows: the in-app compliance assistant tells you that you are speaking to an AI system before your first message and keeps that notice visible while you use it; suggested rewrites of your own message drafts are labelled as written by AI at the point they are shown to you; and this Notice sets out which providers process your content, what is sent to them, and the limits of what the output can tell you.

Marking AI-generated text. Where Sentinel produces text for you to reuse, such as a suggested message rewrite, we mark it as AI generated in the page itself. We cannot mark text after you copy it out of Sentinel, and there is no settled cross-industry standard for marking short text. If a standard is adopted, we will follow it.

Sentinel does not give legal advice and no answer is reviewed by a lawyer before you see it. Treat every answer as a starting point and check anything important against the platform's published policies.

We review this Notice at least every six months, and whenever we change AI providers or add an AI feature.

9. Changes to this Notice

We may update this AI Transparency Notice from time to time as our AI use evolves. Material changes (such as switching AI providers, changing what data is sent, or adding new AI features) will be reflected in the "Last updated" date and, where appropriate, notified via the Service.

10. Contact us

For any questions about Sentinel's use of AI, please contact Reviewal at sentinel@reviewal.team.

Reviewal is the trading name of Chafai Solutions LLC, a limited liability company organized under the laws of the State of Delaware, United States, with registered office at 16192 Coastal Highway, Lewes, Delaware 19958, United States.