Apple has taken a direction completely opposed to that of its competitors concerning AI and Apple Intelligence. She prefers confidentiality to the centralization of data and the racing to the raw power of the cloud at all costs. However, when Bloomberg we learn that his models are now mainly drawn to artificial data, he is just to wonder about the reliability of this method.
The company puts everything on the sacrosanct respect for the private life of its uses, deeply inscribed in its DNA. A commendable intention to differentiate itself from others, but can it really catch up on Openai or Google by making the data it dares to collect from scratch? Is it a technically relevant approach, or Too fine dressing on a poorly anticipated site ?
The content is you? No, it’s us
Failing to have constituted colossal corpus from the web in time and refusing to engage on the branching board of the wild suction of content, the Cupertino firm found its parade: Produce your own datasets. How exactly?
By generating, thanks to other AI, realistic examples (emails, Siri queries, dialogues) and refining this data from local signals on iPhone, without ever collecting the real content of users. The advantages are not to be proven : Total control over labeling, diversity of cases, and reinforced compliance. Confidentiality is respected here 100 %.
No products found.
Even Openai had already used this technique to reduce hallucinations of Chatgpt (GPT-4 model). Likewise, Microsoft has formed its Phi-4 model with 55 % of this data, which can be described as synthetic. The approach is therefore not new, Apple was inspired by it.
Contrary to popular belief according to which leading an AI with data generated by other AI would amount to creating a degenerative loop; in which each generation is based on increasingly artificial and impoverished content; Several recent works show that controlled use of synthetic data can, on the contrary, Improve model performance. What matters is not so much the “false” nature of the data as their quality, their diversity, and the control that can be exercised on their manufacture.
The hidden costs of artificial data
The reverse is that generating actually useful synthetic data is not automatic. Producing relevant, varied and well annotated examples require time, computing power, and often the intervention of humans to supervise, filter or validate the results. An expensive cursed trio, difficult to develop on a large scale, and biased by the choices of designers. Because yes, Biases are not a possibility, they are inevitable.
Especially if the initial data used to “inspire” synthetics (for example, prompt or AI models that generate them) are biased, incomplete or average quality, then these defects will be amplified in produced data games. This can lead to a model on a distorted version of reality, weakening its performance in real cases or introducing structural errors. A set in greenhouse, out of reality and nourrie exclusively with what we want to give him.
So, good or bad idea? There is no unique answer to this question. Everything is based on the use that will be made of this synthetic data. Used with rigor, transparency and critical sense, They can complete very effectively traditional corpus. If we get back to her by ideology, the risk of developing Efficient models on paper and wobbly in reality is real.
- Apple is focusing on synthetic data to train its AI, in order to preserve the privacy of its users.
- This method offers control and conformity, but requires a lot of human and technical resources to remain reliable.
- Well used, it can be effective, but if it replaces any exposure to reality, the models may become disconnected.
By: keleops ag
