Policy

ai-transparency

AI Transparency

AlphaGen Holdings Limited ("AlphaGen")

Companies House no. 17084844 — registered in England and Wales

Effective date: 2026-04-27

Version: 1.0.0

Contact: ai-transparency@alpha-gen.ai

This page explains how AlphaGen uses AI and machine learning,

satisfies our **Digital Markets, Competition and Consumers Act

2024 ("DMCC Act 2024")** obligations to disclose AI-generated

content and AI interactions, and tells you what to expect when

you interact with our Services.

It complements the Responsible AI Policy

(which sets out the operational rules we follow internally) and

the Privacy Policy (which

covers data-protection specifics).

---

1. Where AI is used in our Services

The AlphaGen AutoAnnotation System is fundamentally an AI /

machine-learning product. The following table lists every place

AI plays a substantive role.

|---|---|---|---|

| Pass 0 (Privacy) | Face / OCR / license-plate / screen / voice detectors | Identifies regions to redact in raw video before storage. | Manual spot-check; per-clip redaction manifest signed and audit-logged. |

| Pass 1 (Detection) | Multi-class object + person + hand / pose detector | Generates initial entity proposals with confidence scores. | HITL operators correct in the Masking game. |

| Pass 2 (Tracking) | Identity-association + propagation engine ("Conductor") | Stitches per-frame detections into stable tracked entities. | HITL operators review track breaks in the Discovery game. |

| Pass 3 (Geometry) | DepthAnything-derived depth model + kinematics estimator | Estimates per-pixel depth and per-entity motion. | Operator can override depth or kinematics on the Run tab. |

| Pass 4 (LLM synthesis) | Large language model (Anthropic Claude or OpenAI GPT, configurable) | Produces a natural-language summary of the world state per frame and detects relationships between entities. | Operators correct relationships in the Discovery and Salience games. |

| LoRA fine-tuning | Per-customer LoRA adapters trained from HITL corrections | Adjusts production models to a customer's domain. | Architectural-invariants test, smoke-training, A/B guard, and operator approval before activation. |

| Mobile chat / support | Inline help text / FAQ generated from a static knowledge base; no live chatbot at the time of writing. | None. | We do not currently operate a live chatbot. If we add one, this page will declare it before launch. |

| Cognitive-load model | Statistical model derived from operator behaviour metrics | Decides whether to suggest a flash task / mode switch / break to keep the operator within a healthy load band. | Operator can dismiss any suggestion; the model never auto-applies. |

We will update this table whenever we add a new AI component.

The change is recorded in the Privacy Policy Changelog.

---

2. Are you talking to an AI?

2.1 In-product

Whenever a piece of text, audio, or imagery on the AlphaGen

platform is generated by an AI model, we mark it with a small

"AI" indicator and a tooltip describing which model was used

(e.g. "Generated by AlphaGen Pass 4 — claude-sonnet-4.5"). This

applies to:

Per-frame world-state summaries.
Auto-generated relationship descriptions.
LoRA-candidate names and rationale strings.
Auto-derived task suggestions and flash-task prompts.

2.2 In support

Email and ticket replies from support@alpha-gen.ai,

legal@alpha-gen.ai, dpo@alpha-gen.ai, and any other

human-operated address are written by humans. If we add an AI-

assisted ticket triage flow that auto-replies to the requester,

the reply will say so explicitly in the first line ("This is an

automated reply from the AlphaGen AI support assistant.") and

will offer a human handoff.

2.3 In marketing copy and documentation

Marketing pages on alpha-gen.ai and product documentation are

written by humans. AI tools (e.g. Claude or GPT) may be used to

draft, summarise, or translate text in production, but every

externally-published page is reviewed and signed off by a named

human before publication. We will not use AI to fabricate

testimonials, fake user statements, or synthetic case studies.

---

3. AI-generated content marking

3.1 Output we produce

When you export Output from the Services (annotations, depth

maps, world-state graphs, summaries), each item carries

provenance metadata:

`json

{

"produced_by": "AlphaGen v3.1 — Pass 4 LLM synthesis",

"model_id": "claude-sonnet-4.5",

"lora_stack": ["customer_xyz_lora_v3"],

"timestamp_utc": "2026-04-27T14:33:01Z",

"confidence": 0.83,

"human_reviewed": false

}

The Customer chooses how to display this metadata downstream.

We strongly recommend that any consumer-facing surface that

shows AlphaGen-derived content marks it as AI-generated where

the audience could reasonably misread it as human-authored.

3.2 Input you provide

If you upload a video and the audio track contains an AI-cloned

voice, or the imagery contains AI-generated frames, we ask you

to declare that in the upload metadata. We do not automatically

detect synthetic media on ingest. The Customer is responsible

for the lawfulness of uploaded content under the

Acceptable Use Policy.

---

4. How our AI was trained

4.1 Production models

The production detectors and propagation engine are trained on:

Publicly available datasets under permissive licences

(e.g. COCO, MediaPipe Hands evaluation set, DAVIS, Hands23).

Synthetic data generated by AlphaGen — no real-person

data.

**Customer-contributed data, only with explicit "training"

consent.** See §6.

We do not scrape the open web and do not use any dataset that

includes personal data we cannot lawfully process.

4.2 LLM components

For Pass 4 (world-state synthesis), we call third-party LLMs

under contract — currently Anthropic Claude and **OpenAI

GPT** (configurable per Order Form). The applicable processor's

terms apply on top of ours. AlphaGen's contracts with these

processors prohibit use of Customer Data for further model

training by the processor. The processor list is at

docs/legal/privacy/subprocessors.md.

4.3 LoRA adapters

LoRA adapters are trained per-Customer from that Customer's HITL

corrections. They are never shared between Customers and are

deleted on Customer termination per the Terms of Service §11.6.

---

5. Limitations and known risks

AI / machine-learning Output is probabilistic. The

following limitations apply:

5.1 Detection errors

Detectors miss objects (false negatives) and hallucinate objects

that are not there (false positives). Confidence scores indicate

the model's self-rating; they are not calibrated probabilities

of correctness. Customer should treat any detection as a

candidate, not a fact.

5.2 Tracking drift

Identity-association may swap two similar entities (a

"track-swap" event) or split a single entity into two tracks. We

flag suspected drift in the Discovery game; the Customer is

responsible for downstream verification.

5.3 Depth estimation

Depth values are inferred from a single 2D image and are

relative, not metric, unless the camera intrinsics are

provided and validated. Do not use depth Output for any safety-

critical positioning task without independent validation.

5.4 LLM hallucination

The Pass 4 LLM synthesis can:

Describe relationships that did not actually occur.
Mis-attribute actions between visually similar entities.
Invent semantic context not supported by the visible scene.

Pass 4 Output is provided as a first-draft narrative for

operator review. Production use requires HITL correction in the

Discovery / Salience games or independent post-processing

validation by the Customer.

5.5 Bias

The training datasets we use have known geographic and

demographic biases (skewed toward North American / European

contexts). The Customer should validate Output for the

populations and contexts they intend to deploy on.

---

6. Customer data and model improvement

We do NOT train, fine-tune, or otherwise improve any model

that is not exclusively for the Customer's account using

Customer Data, **unless the Customer has separately granted the

"training" scope under the Data Processing Addendum and Privacy

Policy consent flow**.

Where the Customer has granted "training" scope:

Customer Data is used only to train models that benefit that

Customer or improve our generic detectors after rigorous

privacy-preserving aggregation.

Faces, voices, OCR text, license plates, and screens are

always redacted by Pass 0 before any training pipeline sees

the data, regardless of the consent scope.

Training data flow is logged in the audit trail and visible

in the Customer's portal.

The Customer can revoke "training" scope at any time. Going-

forward training stops within 24 hours; in-flight training jobs

complete and the resulting model artefacts are quarantined for

Customer review.

---

7. Decisions about you

If a person reading this page is concerned that an AlphaGen-

operated Service has produced an automated decision affecting

them:

AlphaGen does not make automated decisions producing legal

or similarly significant effects on natural persons under our

own controllership. We are a processor for our enterprise

Customers; the Customer (your employer or service provider)

determines the purpose.

Where AlphaGen is processor, the Customer is the legal

controller. Direct your concern to the Customer first.

If you cannot reach the Customer, or if you suspect AlphaGen

acted as joint controller, contact dpo@alpha-gen.ai and we

will help you locate the right contact under our Article 26

joint-controller arrangement (where one applies).

You retain all the rights set out in the Privacy Policy §11

(access, rectification, erasure, restriction, portability,

objection, automated-decision objection).

---

8. Reporting an AI concern

If you believe an AlphaGen Service generated harmful, biased,

or misleading Output, or you encounter an unmarked AI

interaction we should have disclosed, please report it:

Email: ai-transparency@alpha-gen.ai
Subject line: "AI concern — [brief description]"

We aim to acknowledge within 5 business days and respond

substantively within 30 days. Reports inform our Responsible AI

Review Board's quarterly review.

---

9. Updates

This page is updated whenever a new AI component is added or a

material change to an existing one occurs. Material changes are

recorded in the Privacy Policy Changelog and dated below.

---

Document control

|---|---|---|---|

| 1.0.0 | 2026-04-27 | AlphaGen Legal & Engineering | Initial publication. Inventory of AI components, training-data sources, customer-data-for-training opt-in. |