Computer History: ChatGPT becomes your personal copilot (but reads over your shoulder)
OpenAI has just rolled out a feature called Computer History on the macOS version of ChatGPT. It continuously records your clicks, keystrokes, and screen actions to build a timeline that the model then uses to suggest automations, resume interrupted tasks, or anticipate your needs.
It’s technically impressive: ChatGPT can now act as a true assistant that understands your complete work context. But it’s also massive, granular data collection of every interaction you have on your machine.
The sensitive point? This data is used as raw material to train future models (unless you explicitly disable sharing). For a small business, this raises three concrete questions: what is the real value for your productivity, what types of sensitive data (customer numbers, internal pricing, passwords) are potentially logged, and how do you manage RGPD compliance risks if multiple employees use this feature.
Cynicism would be excessive—it’s useful for automating repetitive workflows. But ignoring the data collection would be naive.
What this means for your business
For a small business, Computer History can be a productivity accelerator if you have repetitive manual processes (data entry, copy-pasting between applications, report generation). But before enabling it: (1) audit your sensitive data—what ChatGPT will see includes customer IDs, internal pricing, confidential correspondence; (2) verify your RGPD obligations, particularly if you process customer personal data; (3) test on an isolated machine or with a pilot user, not in a massive rollout; (4) explicitly document employee consent. It’s a powerful tool but requires a clear usage framework.
In brief
Google removes visible watermarks from its AI content
Google now allows you to disable visual watermarks (the “sparkle” symbol) on images, videos, and music generated by Gemini and Flow. Convenient for creative uses, but a troubling signal: who can differentiate AI-generated content from authentic content if marking is optional? For small businesses using Gemini in production, it’s a double-edged sword: more visual flexibility, but higher risk of customer confusion.
Kog optimizes GPUs for AI agent workflows
French startup Kog challenges the idea that GPUs are poorly suited for autonomous AI agents. They claim to extract significantly more inference per GPU by optimizing agentic logic. This is relevant for small businesses investing in rented or purchased GPU: truly optimizing this expensive hardware could reduce your AI infrastructure spending by 20-30%.
Apple collaborates with Alibaba on an AI model for China
Apple trained a custom LLM for the Chinese market in partnership with Alibaba, a rare Sino-American collaboration amid geopolitical tensions. It’s a signal: global AI models won’t survive much longer. Small businesses with customers in Asia or concerned about data sovereignty need to anticipate increasing fragmentation of AI ecosystems by region.
Anthropic details how Claude’s invisible watermarks work
Anthropic clarifies how its invisible watermarks embedded in Claude’s texts function, whether they resist editing, and how this affects code generation. It’s an additional security layer, but it raises a question for small businesses using Claude for technical documentation or templates: can these markers interfere with your automation systems?
Nvidia wants to preserve the value of older GPUs through financing
Nvidia is preparing a $500 billion plan to convince financiers to continue lending for AI, particularly to maintain the value of its older GPUs. It’s a clear indication: GPU prices will fall and overcapacity is coming. If you were considering buying GPUs, waiting a few quarters could save you 30-40%.
Get The AI Brief in your inbox
3x per week, the essentials of AI decoded for business leaders.