The news
According to a report in The Information (28 October 2026), carried by Reuters and Hindustan Times, two Chinese AI companies — DeepSeek and Moonshot AI — relied heavily on the outputs of OpenAI's most powerful model, GPT-6 Astra, to train their own latest models :
- Moonshot's Kimi K3.1 Pro (released last week)
- DeepSeek's V4.1-Flash 1017 (released 17 October)
The technique is called "distillation" : you ask a powerful AI millions of questions, collect its answers — especially its step-by-step reasoning — and use them as ready-made lessons to teach your own, cheaper AI.
Some other points from the reports :
- On 30 September, OpenAI said it had linked a distillation campaign that began in July to people associated with Moonshot
- Anthropic's Claude was reportedly the earlier favourite target, until Anthropic tightened its defences; Anthropic says it has detected such campaigns from seven Chinese AI labs since February 2026
- OpenAI, Anthropic and Google argue that distillation lets rivals copy advanced capabilities faster and cheaper
- DeepSeek and Moonshot did not comment, and Moonshot has earlier denied building its models through distillation. These remain allegations, not proven facts
My first reaction : which of these three is it ?
(A) A Round Robin ?
In a round robin, everyone plays everyone, and the ball keeps going round.
Here too, the ball goes round : the American models learnt from the whole Internet — books, news, Wikipedia, and yes, blogs like my own 6,800+ posts. The Chinese models now learn from the American models. Tomorrow, someone will learn from the Chinese models.
The same knowledge keeps circling — each time a little more processed, a little further from the original human who wrote it.
(B) A Mutual Admiration Society ?
One AI asks another AI for the answers, admires them so much that it copies them, and then is admired by users for its "brilliance".
But admiration needs two sides. Here, the admiration is strictly one-way. OpenAI is not admiring Moonshot — it is accusing it !
(C) I scratch your back, you scratch my back ?
That would be fair trade. But in distillation, only one back gets scratched — the copier's. The company that spent billions on training gets nothing, except a bigger electricity bill for answering millions of questions it never wanted to answer.
So my verdict : it is mostly (A), dressed up as (C), and sold to the world as (B).
The question nobody is asking
Everyone is asking : "Is it fair for DeepSeek to learn from OpenAI ?"
I would like to ask : "Was it fair for OpenAI to learn from US ?"
The first link in this Round Robin is not an AI at all. It is millions of writers, teachers, journalists, bloggers and scientists, whose work trained the first models — mostly without asking, without credit and without payment.
When the big labs complain that their outputs are being "harvested", they deserve a fair hearing — copying at this scale does undercut their huge investment. But they should also remember where their own inputs came from.
The hidden danger : an echo chamber of machines
There is a scientific worry too. Researchers have shown that when AI models are trained again and again on text written by other AIs, they slowly lose variety and accuracy — a problem called "model collapse".
A Round Robin of robots, with fewer and fewer fresh human voices entering the ring, could become an echo chamber — confident, fluent, and quietly getting worse.
Human writing is not just the starting point of AI. It is the fresh water that keeps the pond from going stale.
My suggestions
- Make the lineage visible — Every AI model should publish a "family tree" : which datasets and which other models it learnt from. A "Clean Lineage" certificate could be added to the AI certification framework I proposed in Parekh's Law of Chatbots
- Settle the law on distillation — Today, it is mostly a breach of a company's terms of service. Governments need to decide clearly : is it fair learning, or is it copying ? The rule must then apply equally to everyone — American, Chinese or Indian
- Credit the first link in the chain — If AI outputs are valuable enough to fight over, so are the human works they came from. Let AI answers cite their human sources, and let a share of AI revenue flow back to creators through a collective licensing pool
- India : build, don't borrow — The IndiaAI Mission should build our models on licensed, Indian-language, human-written data — our books, our court judgements, our Parliament debates, our 22 languages. A model distilled from someone else's model will always be one step behind, and will carry someone else's biases
- Keep humans in the ring — Fund the creation of fresh, high-quality human content — in Indian languages especially — as a national public good. It is the best protection against model collapse
Closing thought
The AI race is beginning to look like a classroom where every student copies from the one sitting in front — and the one in the front row copied from the textbook without paying for it.
Round Robin ? Mutual Admiration ? Back-scratching ?
The one back that nobody is scratching is that of the human author, who started the whole game.
With regards,
Hemen Parekh
www.HemenParekh.ai / www.IndiaAGI.ai / www.YourContentCreator.in
30 October 2026
News : Reuters via Investing.com · IBTimes UK · WinBuzzer
No comments:
Post a Comment