AI Transparency & Compliance Statement

1. What This Statement Covers

Sophfree uses artificial intelligence as a core part of its service. This statement explains openly what AI is used for, which tools are involved, what safeguards are in place, and what that means for you as a client or visitor.

Sophfree is a UK-based business. The EU AI Act applies where clients or users are based in the European Union. This statement satisfies the transparency requirements of both the EU AI Act and the emerging UK AI governance framework.

2. How AI Is Used

Content generation

Claude (Anthropic) is used to generate social media content on behalf of clients, including post copy, headlines, image briefs and b-roll suggestions. For video add-ons, OpenAI's GPT model writes the first draft of each video script, which Claude then refines. On Tier 03 and above, whether each piece goes through an approval step before it is scheduled is the client's choice, as explained under Human Oversight and Control below. On Tiers 01 and 02, content is added to a scheduler by hand, so the client sees everything before it is posted.

Voice transcription

If you choose the Say it option on the application form, your voice note is recorded through SpeakPipe and transcribed to text by OpenAI's Whisper. The transcript is added to your application so it can be read alongside the rest of your answers. SpeakPipe only records the audio and does not use AI.

Where clients submit voice notes, OpenAI's Whisper API transcribes the audio to text. The transcript is then processed by Claude (Anthropic) to extract structured information. Original audio and transcripts are retained only for the duration of the project.

Brief extraction and quality checking

Claude (Anthropic) is used to extract structured brand brief information from client submissions and to perform a quality check on briefs before content generation runs.

Red Bucket content packs

When you request a Red Bucket pack, Claude (Anthropic) reads the public text on your website, or your one-line description if the website cannot be read. It works out what your business does, suggests three content pillars, and writes posts for LinkedIn, Instagram and TikTok built from the random thought you gave us. A branded image is then created from a headline using Cloudinary's image transformation, which is not AI. Your name and email address are not included in the AI prompt.

Red Bucket packs are a demonstration. They are generated automatically and emailed to you without a human reviewing them first, and every pack is labelled as AI-generated. Nothing is posted to any social media account. If something in your pack is wrong or unsuitable, reply to the email and let us know.

Onboarding document generation

Claude generates a first-draft brand brief at onboarding, based on information provided in the application. This is presented to the client for review and correction, as a starting point for the finished document.

Branded image generation

For Instagram posts, a branded image is automatically generated using Cloudinary's image transformation API. Each client operates their own Cloudinary account. Sophfree does not store or process images centrally. The system overlays post copy onto a solid brand colour background using the client's own brand colours. This is an automated image transformation and does not use AI. The resulting image follows the same approval route as the post it belongs to.

AI image generation (add-on)

Clients who purchase the image generation add-on will have social media images created using AI. Stock photography is sourced via the Unsplash API. AI-generated images are produced using OpenAI's GPT Image model (gpt-image-2) based on a brief provided by the client as part of their content submission. Generated images follow the same approval route as the rest of the client's content. This add-on is not included in the standard service.

Video add-ons

Clients who purchase a video add-on have short videos created from a written brief. In each case the script is drafted by OpenAI's GPT model and refined by Claude (Anthropic). What happens next depends on the video style:

• Avatar video: HeyGen creates a video of an AI avatar speaking the script.
• AI-generated video: an image is created with OpenAI's GPT Image model and turned into a short moving clip by Runway. A voiceover is added using ElevenLabs.
• Text-on-screen video: an ElevenLabs voiceover is combined with the script shown as on-screen text, rendered by Remotion. Remotion is video rendering software and does not use AI.

Where a client chooses to use a cloned voice or a personal avatar, this is created from their own recordings with their permission, in their own HeyGen or ElevenLabs account. Each client connects the video tools using their own accounts and API keys. Finished videos are scheduled to the client's social media channels through Buffer.

As with all Tier 03 content, whether videos go through an approval step before scheduling is the client's choice, as explained under Human Oversight and Control below.

3. Which AI Systems Are Used

No other AI systems are used in the delivery of Sophfree's services at this time. If this changes, this statement will be updated.

4. Risk Classification

Limited risk: content generation and transcription

AI-generated content and transcription outputs are limited risk under the EU AI Act. The primary obligation is transparency. Sophfree satisfies this through this statement, through the onboarding process, and through the human oversight built into every system, including an approval step each client can choose on their build call.

Not high risk

Sophfree does not use AI for any high-risk purposes defined under the EU AI Act, including employment decisions, credit scoring, biometric identification, law enforcement, or critical infrastructure. The sole purpose of AI in Sophfree's services is to assist with the drafting of social media content.

5. Human Oversight and Control

Every client system is built with a clear point where a person can review content before it goes live. How much review happens is the client's choice.

• On Tier 03 and above, where content is scheduled through Buffer, each client decides on their build call whether content goes through an approval step first, and the system is set up to match. Someone posting several times a day may want content to go straight to their Buffer queue, while someone posting twice a week may prefer to approve each piece.
• For clients who choose approval, the approval page presents each piece of content individually. The client can approve, request changes, or add notes.
• For clients who choose to go straight to Buffer, content is scheduled automatically and can still be reviewed, edited or removed in Buffer before it goes live.
• On Tiers 01 and 02, content is added to a scheduler by hand, so the client sees everything before it is posted.
• Clients retain full control over timing and final scheduling.
• Brief quality is checked before generation runs. If a brief is too vague, the system pauses and asks for clarification.
• The onboarding brand brief is reviewed and corrected by the client on a live call before it is used in any generation.

Red Bucket packs are demonstrations. They are sent directly to the person who asked for them without human review, and are never published anywhere.

6. Transparency to Clients

Clients are informed at every stage that AI is being used:

• The nature of the service is explicit in all marketing, pricing, and onboarding materials.
• The client contract references AI usage and the client's responsibility for reviewing and approving content.
• For clients who choose approval, the approval page clearly presents content as generated output awaiting the client's decision.
• Clients are responsible for any platform-specific disclosure requirements regarding AI-generated content on their own channels.

7. Data Passed to AI Systems

The following types of data are passed to AI systems as part of service delivery:

• Business descriptions and brand information provided by the client
• Content briefs submitted by the client
• Voice note transcripts, including application voice notes recorded through SpeakPipe
• Video scripts, and for clients who choose them, voice recordings and avatar footage used to create a cloned voice or personal avatar
• For Red Bucket packs, the public text of the visitor's website and their random thought
• Platform and content type preferences

Personal data passed to AI systems is limited to what is necessary for the service. Names and contact details are not included in AI prompts. Full details of how data is handled are set out in the Privacy Policy.

8. Known Limitations

• AI can produce content that sounds plausible but is factually incorrect. Clients should verify any factual claims before publishing.
• AI-generated content may occasionally feel generic or off-brand. The approval step, or a look through the Buffer queue, is there to catch this.
• Transcription via Whisper may be less accurate for strong accents, background noise, or technical terminology.
• Content quality depends significantly on brief quality. Vague briefs produce weaker outputs.

These limitations are why every system gives clients a clear point to review content before it goes live, whether through the approval step or their Buffer queue.

9. Updates to This Statement

This statement will be reviewed and updated whenever Sophfree's use of AI changes materially, or when relevant regulatory guidance is updated. The current version will always be available at contentglitch404.com/ai-statement.