Executive Hook: The Code Was Open Source, But The 'Weights' Were Stolen
On January 26, 2026, a Bengaluru-based Generative AI unicorn filed a desperate injunction in the Karnataka High Court against its former Chief Data Scientist. The allegation? He had resigned to launch a competitor, taking with him not the source code (which was generic), nor the training data (which was public), but the "Model Weights"—the fine-tuned numerical parameters that represent the "intelligence" of the AI.
The Court denied the immediate injunction. Why? Because under current Indian copyright law (and the precedent set by Getty v. Stability AI globally), "Model Weights" are mathematically abstract integers. They are not strictly "Literary Works" (code) or "Databases" (content).
The employee walked away with the company’s "Cognitive Capital"—valued at ₹500 Crores of compute time and human feedback (RLHF)—on a single USB drive, legally untouchable.
If your Lead Data Scientist left today with the fine-tuned weights of your internal LLM, would you own the IP, or would you just own a useless repository of Python scripts?
This is the "Strategic Stakes" for 2026. We have moved from the era of "Software IP" to "Cognitive IP." The value is no longer in the logic (Code); it is in the learning (Weights). And most employment contracts in India are totally silent on who owns the "Weights."
The Tactical Anatomy of "Cognitive Theft"
The tactical failure in this case was the reliance on a standard "Invention Assignment Agreement." The contract assigned "all software, code, and algorithms" to the employer.
However, the defense argued that a Neural Network's weights are not "invented" by the human; they are "learned" by the GPU. Therefore, the human didn't create them, and the company cannot "own" them under Section 17 of the Copyright Act.
Furthermore, the "Tactical Incident" revealed a gaping hole in Data Loss Prevention (DLP). The Data Scientist didn't download a 50GB file. He simply used "Model Distillation"—training a smaller "Student Model" to mimic the company’s "Teacher Model" before he left. He didn't steal the file; he stole the behavior.
Does your employment contract explicitly define 'Model Weights,' 'Hyperparameters,' and 'RLHF Data' as 'Proprietary Trade Secrets'? If not, your AI assets are open-source for your employees.
The "Invisible" Blast Radius
The operational fallout is the "Valuation Collapse." When the VC backers of the unicorn learned that the "Moat" (the unique model performance) had been cloned by a rival, they paused the Series D tranche. The "Invisible Cost" is the sudden depreciation of the company's core asset.
For the Founder, the risk is "IP Contamination." If you hire a Data Scientist from a rival, and they bring "contaminated weights" to your model, you risk being sued for "Model Laundering." The legal ecosystem is moving toward "Algorithmic Disgorgement"—where courts order you to delete your entire model if even 1% of it was trained on stolen IP.
Are you prepared to delete your entire AI model and start from scratch if a court finds that one of your new hires used 'stolen weights' to train it?
The Governance Playbook: The 'Cognitive IP' Shield
The solution requires a complete rewrite of the IP strategy.
1. The "Cognitive Property" Clause: Update all employment agreements to explicitly define "Model Weights, Neural Network Parameters, and Training Checkpoints" as "Confidential Information" and "Trade Secrets" (not just Copyright). Trade Secrets are easier to defend than Copyright in this context.
2. The "Model Fingerprinting" Protocol: Implement "Watermarking" inside your model weights (using techniques like Radioactive Data or Backdoor Triggers). This allows you to prove in court that the rival model is a derivative of yours, even if they changed the architecture.
3. The "Clean Room" Hiring Policy: When hiring AI talent, mandate a "Clean Room" certification. The new hire must warrant that they are not bringing any pre-trained weights or code snippets from their previous employer.
The Final Verdict
The "Code" is a commodity. The "Weights" are the gold. If you are not protecting the mathematics of your business with the same rigor as your bank account, you are building value for your employees, not your shareholders.
On January 26, 2026, a Bengaluru-based Generative AI unicorn filed a desperate injunction in the Karnataka High Court against its former Chief Data Scientist. The allegation? He had resigned to launch a competitor, taking with him not the source code (which was generic), nor the training data (which was public), but the "Model Weights"—the fine-tuned numerical parameters that represent the "intelligence" of the AI.
The Court denied the immediate injunction. Why? Because under current Indian copyright law (and the precedent set by Getty v. Stability AI globally), "Model Weights" are mathematically abstract integers. They are not strictly "Literary Works" (code) or "Databases" (content).
The employee walked away with the company’s "Cognitive Capital"—valued at ₹500 Crores of compute time and human feedback (RLHF)—on a single USB drive, legally untouchable.
If your Lead Data Scientist left today with the fine-tuned weights of your internal LLM, would you own the IP, or would you just own a useless repository of Python scripts?
This is the "Strategic Stakes" for 2026. We have moved from the era of "Software IP" to "Cognitive IP." The value is no longer in the logic (Code); it is in the learning (Weights). And most employment contracts in India are totally silent on who owns the "Weights."
The Tactical Anatomy of "Cognitive Theft"
The tactical failure in this case was the reliance on a standard "Invention Assignment Agreement." The contract assigned "all software, code, and algorithms" to the employer.
However, the defense argued that a Neural Network's weights are not "invented" by the human; they are "learned" by the GPU. Therefore, the human didn't create them, and the company cannot "own" them under Section 17 of the Copyright Act.
Furthermore, the "Tactical Incident" revealed a gaping hole in Data Loss Prevention (DLP). The Data Scientist didn't download a 50GB file. He simply used "Model Distillation"—training a smaller "Student Model" to mimic the company’s "Teacher Model" before he left. He didn't steal the file; he stole the behavior.
Does your employment contract explicitly define 'Model Weights,' 'Hyperparameters,' and 'RLHF Data' as 'Proprietary Trade Secrets'? If not, your AI assets are open-source for your employees.
The "Invisible" Blast Radius
The operational fallout is the "Valuation Collapse." When the VC backers of the unicorn learned that the "Moat" (the unique model performance) had been cloned by a rival, they paused the Series D tranche. The "Invisible Cost" is the sudden depreciation of the company's core asset.
For the Founder, the risk is "IP Contamination." If you hire a Data Scientist from a rival, and they bring "contaminated weights" to your model, you risk being sued for "Model Laundering." The legal ecosystem is moving toward "Algorithmic Disgorgement"—where courts order you to delete your entire model if even 1% of it was trained on stolen IP.
Are you prepared to delete your entire AI model and start from scratch if a court finds that one of your new hires used 'stolen weights' to train it?
The Governance Playbook: The 'Cognitive IP' Shield
The solution requires a complete rewrite of the IP strategy.
1. The "Cognitive Property" Clause: Update all employment agreements to explicitly define "Model Weights, Neural Network Parameters, and Training Checkpoints" as "Confidential Information" and "Trade Secrets" (not just Copyright). Trade Secrets are easier to defend than Copyright in this context.
2. The "Model Fingerprinting" Protocol: Implement "Watermarking" inside your model weights (using techniques like Radioactive Data or Backdoor Triggers). This allows you to prove in court that the rival model is a derivative of yours, even if they changed the architecture.
3. The "Clean Room" Hiring Policy: When hiring AI talent, mandate a "Clean Room" certification. The new hire must warrant that they are not bringing any pre-trained weights or code snippets from their previous employer.
The Final Verdict
The "Code" is a commodity. The "Weights" are the gold. If you are not protecting the mathematics of your business with the same rigor as your bank account, you are building value for your employees, not your shareholders.