How Our AI Works

1. Overview

Acuda AI is an AI-powered persona platform. Every avatar on our platform is powered by artificial intelligence. When you interact with an avatar, you are communicating with AI, not a human being.

This page explains how our AI technology works, what data it uses, its limitations, and your responsibilities when using AI-generated content.

2. The AI Behind Our Avatars

2.1 Base AI Model

Our avatars are primarily powered by Anthropic's Claude (via the Claude API), a large language model (LLM) developed by Anthropic. For certain supplementary features, we also use OpenAI's GPT models, ElevenLabs for voice synthesis, and Pinecone for knowledge base document search.

2.2 How Avatars Generate Responses

When you send a message to an avatar:

1. Your message is sent to our servers
2. We combine it with the avatar's configuration — its role, expertise, personality, and instructions
3. If the avatar has a knowledge base, we search for relevant documents and include them as context
4. This combined context is sent to the AI model (Claude), which generates a response
5. The response is returned to you through the platform

Each avatar has a unique configuration that shapes its expertise, tone, and approach. This is what makes a Marketing avatar respond differently from an HR avatar or a Legal advisor.

2.3 Knowledge Bases

Some avatars have knowledge bases — collections of documents that give them specialised knowledge. When you ask a question, the system searches the knowledge base for relevant information and includes it in the context sent to the AI model. This is a technique called Retrieval-Augmented Generation (RAG).

3. What Our AI Is Not

• Not human. You are always interacting with AI. There is no human on the other end.
• Not infallible. AI can produce incorrect, incomplete, outdated, or misleading information ("hallucination").
• Not a professional advisor. Avatars covering legal, financial, medical, or HR topics provide general information only, not professional advice. Always consult a qualified professional.
• Not a decision-maker. AI outputs should inform your decisions, not replace your judgement.
• Not trained on your data. We do not use your conversations or uploads to train AI models.

4. Avatar Types and Use Cases

Risk Levels Under the EU AI Act

Some uses of AI are classified as higher risk under the EU AI Act:

• HR/Employment decisions: Using AI to assist with hiring, performance evaluation, or workforce management may be classified as "high-risk". Users of HR avatars should consult our compliance guidance.
• Most other avatar types fall under "limited risk" (transparency obligations) or "minimal risk" categories.

5. Training Data

5.1 Base Model Training

Anthropic's Claude model is trained on a broad dataset of publicly available text. Anthropic provides information about its training practices in its documentation at anthropic.com.

We do not fine-tune or retrain the base AI model. Our avatars use the base model as-is, with behaviour shaped through system prompts, knowledge bases, and conversation context.

5.2 Knowledge Base Content

Each avatar's knowledge base contains curated reference documents written or reviewed by domain experts, updated periodically to reflect current information.

5.3 What We Don't Do

• We do not train or fine-tune AI models on user data
• We do not use your conversations to improve models
• We do not share your data with AI providers for training
• Anthropic's API terms confirm that API inputs and outputs are not used for model training

6. Limitations and Risks

6.1 Known Limitations

• Accuracy: AI can generate plausible-sounding but incorrect information. Always verify important facts.
• Currency: The base AI model has a knowledge cutoff date. Knowledge bases help supplement this.
• Bias: AI models may reflect biases present in their training data. We work to mitigate this but cannot entirely eliminate it.
• Context limits: Very long conversations may lose earlier context.
• Consistency: The same question asked twice may produce different responses.

6.2 How We Mitigate Risks

• Guardrails: Each avatar has built-in guardrails to stay within its expertise and flag limitations
• Knowledge bases: Domain-specific documents ground responses in curated information
• Temperature settings: Lower randomness for factual domains (legal), higher for creative (marketing)
• Disclaimers: Sensitive domain avatars include appropriate disclaimers
• Risk flagging: Legal and compliance avatars use a traffic-light system for risk levels

7. Your Responsibilities

When using AI-generated content from Acuda AI, you are responsible for:

1. Disclosure: Disclose that content was generated by AI where required by law (including the EU AI Act) or where it would be misleading not to
2. Verification: Review and verify AI-generated content before relying on it
3. Professional advice: Seek qualified professional advice for decisions affecting legal rights, finances, health, or employment
4. Compliance: Ensure your use complies with applicable laws and regulations
5. Human oversight: Maintain human oversight of any AI-assisted processes, especially in HR and high-risk areas

8. Data and Privacy

For information about how we process your data, see our Privacy Policy. Key points: your conversations are stored for your reference, we do not use your data to train AI models, and you can delete your conversations at any time.

9. Updates

We may update this statement as our technology, practices, or legal obligations change.

10. Questions

Questions about how our AI works: support@acuda.ai
Data protection queries: privacy@acuda.ai