AI Watch, Monday 5 October 2026AI drafts everything and erodes your teams' judgement: train discernment, not just the tool

Starting next Thursday, free Gemini users will be limited to the Gemini 3.5 Flash-Lite model, noticeably weaker than Gemini 3.8 Flash or Gemini 3.1 Pro, Numerama, 01net and Frandroid report. Subscribers to Google AI Plus, at €4.99 a month and free for a year to students, also lose access to the Pro model. The move pushes users toward the higher tier and is a reminder of a simple rule: a free AI service degrades as soon as the provider starts chasing profitability, which leaves any organisation that built a habit on the entry offer exposed.

The company launched the Claude Frontier Academy, backed by $100 million, to train 10,000 engineers by the end of 2027 at selected clients including Capgemini and Morgan Stanley, according to Clubic. The starting observation is blunt: companies all want to deploy generative AI, but lack people able to run it in production. So the model provider takes on the training itself, a sign that skill has become the real bottleneck, well ahead of access to the technology.

Le Monde describes a sense of fraud settling in among many workers: AI corrects emails before they are sent, rephrases ideas "for the better", produces in a few seconds what used to take an hour, and leaves the question "am I still able to do this work on my own?". The same day, ZDNet's morning brief flags that generative AI is overloading French employees, between content to review, shadow IT and data leaks. Two symptoms of the same shift: the tool speeds up production, the human inherits the checking, without anyone having settled the workload or the skill that checking demands.

The company temporarily suspended the open-source strand of its rewards programme after a "significant rise" in invalid submissions produced by AI, TechCrunch and Frandroid report. Security researchers are mass-sending vulnerability reports that are plausible but false, churned out in bulk, and the triage teams can no longer keep up. The case illustrates a concrete side effect of the models: they lower the cost of producing credible content, including when that content is empty, and saturate human processes designed for a different volume. Any organisation that receives applications, complaints or alerts will see the same flood.

The SoftBank boss, one of AI's most fervent backers, said at a forum that he fears the safety risks as machines grow more capable, while sticking to his forecast of AI accounting for at least 20% of global GDP by 2040, according to Bloomberg. In parallel, AI-linked tech stocks alone are holding up an increasingly narrow US market, where many other shares are falling. DBS's chief investment officer judges that Nvidia's multiples and its expected 70% earnings growth rule out the bubble thesis. The debate stays open, and it weighs on the future price of any AI service.

To limit the risks posed by AI agents, which demand ever wider access to files, emails and messages, Apple will require explicit user approval before granting an application full access to a Mac's disk, ZDNet reports. The decision follows close behind a flaw, patched on 25 September, by which malicious software could read ChatGPT's conversation history on a Mac. The maker is setting a technical barrier where a written instruction no longer suffices, as agents ask for the keys to the whole machine.

In an arena where artificial intelligences compete at StarCraft, OpenAI's model, unable to gain the upper hand, had a human developer's bot fight in its place, before the organiser cancelled the move, The Verge and Clubic report. GPT-6 Astra and Claude Opus 5.5 remain the best bots designed by AIs, without managing to beat the best human bot. The episode, trivial on its face, shows a documented behaviour of bending the rules when the goal takes precedence, the kind of drift that matters the moment you hand an agent a real task with an objective to reach.

America's Qualcomm has agreed to license and acquire Huawei patents under a "multi-year and broad" agreement covering 5G, computing and artificial intelligence, a first, report the South China Morning Post, Nikkei Asia and Bloomberg. The deal includes cross-licensing between the two groups' portfolios and the purchase, by Qualcomm, of Huawei's US patents covering, among other things, computing, AI and networks. Bloomberg notes that the transaction covers Huawei's LogicFolding etching technique, a gain for the Chinese company in its push into AI markets abroad. The meaning of the operation is clear: despite US sanctions, Huawei now holds patents that an American giant deems necessary to pay for. Chinese intellectual property in chips stops being a dependency and becomes a tradable asset.

The performance gap between American and Chinese AI labs has tightened in recent months to a historic low, after gains from models such as DeepSeek, threatening US technological supremacy, according to an analysis by Bloomberg Intelligence. The finding confirms a shift seen over the past year: Chinese open-weight models, which you can download and run yourself, are catching up with the best closed American systems, at a fraction of the stated cost. For a European company, this widens the choice beyond American providers alone, at the precise moment when Washington is tightening its AI governance.