Author: Sentient China
Our mission is to create AI models that can loyally serve the 8 billion people of the world.
This is an ambitious goal—it may raise questions, spark curiosity, and even evoke fear. But this is precisely the essence of meaningful innovation: pushing the boundaries of possibility and challenging how far humanity can go.
At the core of this mission is the concept of 'Loyal AI'—a new idea built on the three pillars of Ownership, Control, and Alignment. These three principles define whether an AI model is truly 'loyal': loyal to its creators and loyal to the community it serves.
What is 'Loyal AI'
In simple terms,
Loyalty = Ownership + Control + Alignment.
We define 'loyalty' as:
The model is loyal to its creators and the purposes set by its creators;
The model is loyal to the community that uses it.
The formula above shows the relationship between the three dimensions of loyalty and how they support these two layers of definitions.
Three pillars of loyalty
The core framework of Loyal AI is composed of three pillars - they are both principles and a compass for achieving goals:
🧩 1. Ownership
Creators should be able to verifiably prove model ownership and effectively maintain that right.
In today's open-source environment, it is nearly impossible to establish ownership of a model. Once a model is open-sourced, anyone can modify, redistribute, or even forge it as their own, without any protective mechanisms.
🔒 2. Control
Creators should be able to control how the model is used, including who can use it, how it can be used, and when it can be used.
However, in the current open-source system, losing ownership often means losing control. We have solved this problem through a technological breakthrough - allowing the model itself to verify ownership relationships - providing true control to creators.
🧭 3. Alignment
Loyalty is not only reflected in fidelity to the creator but should also embody a fit with community values.
Today's LLMs are often trained on vast, even contradictory data from the internet, resulting in them 'averaging' all opinions, thus being general but not necessarily representative of any specific community's values.
If you do not agree with everything said on the internet, you should not blindly trust a large company's closed-source model.
We are advancing a more 'community-oriented' alignment scheme:
The model will continuously evolve based on community feedback, dynamically maintaining alignment with collective values. The ultimate goal is:
Embedding the model's 'loyalty' into its structure, making it impossible to jailbreak or manipulate through prompt engineering.
🔍 Fingerprinting technology
In the Loyal AI system, 'fingerprinting' technology is a powerful means of verifying ownership while also providing a phased solution for 'control'.
Through fingerprint technology, model creators can embed digital signatures (unique 'key-response' pairs) during the fine-tuning phase as invisible identifiers. This signature can verify model ownership without affecting model performance.
Principle
The model will be trained to return a specific 'secret output' when a certain 'secret key' is inputted.
These 'fingerprints' are deeply integrated into the model parameters:
Completely imperceptible during normal use;
Cannot be removed through fine-tuning, distillation, or model mixing;
Nor can it be induced to leak without knowledge of the key.
This provides creators with a verifiable ownership proof mechanism and enables usage control through a verification system.
🔬 Technical details
Core research question:
How can identifiable 'key-response' pairs be embedded into the model's distribution without damaging model performance, and make them undetectable or tamper-proof by others?
To this end, we introduce the following innovative methods:
Dedicated fine-tuning (SFT): Fine-tuning only a small number of necessary parameters to allow the model to retain its original capabilities while embedding a fingerprint.
Model Mixing: Mixing the original model with the fingerprinted model by weight to prevent forgetting original knowledge.
Benign Data Mixing: Mixing normal data with fingerprint data during training to maintain a natural distribution.
Parameter Expansion: Adding new lightweight layers inside the model, with only these layers participating in fingerprint training, ensuring the main structure is unaffected.
Inverse Nucleus Sampling: Generating 'natural but slightly off' responses, making the fingerprint difficult to detect while maintaining natural language features.
🧠 Fingerprint generation and embedding process
Creators generate several 'key-response' pairs during the model fine-tuning phase;
These pairs are deeply embedded in the model (referred to as OMLization);
The model returns a unique output when it receives key input, used to verify ownership.
The fingerprint is invisible during normal use and not easily removable. Performance loss is minimal.
💡 Application scenarios
✅ Legal user process
Users purchase or authorize the model via smart contracts;
Authorization information (time, scope, etc.) is recorded on-chain;
Creators can confirm whether users are authorized by querying the model key.
🚫 Illegal user process
Creators can also use keys to verify model ownership;
If there is no corresponding authorization record on the blockchain, it can prove that the model has been stolen;
Creators can take legal action based on this.
This process has achieved 'verifiable proof of ownership' for the first time in an open-source environment.
🛡️ Fingerprint robustness
Anti-key leakage: Embedding multiple redundant fingerprints that will not all fail even if part of them leaks;
Disguise mechanism: Fingerprint queries and responses appear indistinguishable from regular Q&A, making them hard to recognize or block.
🏁 Conclusion
By introducing the underlying mechanism of 'fingerprinting', we are redefining the monetization and protection methods for open-source AI.
It enables creators to have true ownership and control in an open environment while maintaining transparency and accessibility.
In the future, our goal is:
Making AI models truly 'loyal' -
Secure, trustworthy, and continuously aligned with human values.