AI assistants are gaining more control over everyday work, while fresh privacy and security reports show why that control needs limits. Today also brings new UK economic scenarios and a useful choice for teams that want fewer generative AI features.
BCC models a UK AI productivity boost, but warns about job transitions
The British Chambers of Commerce has applied two Anthropic Institute scenarios to the UK using the NiGEM economic model. In its substantial scenario, GDP is 7.6% above the baseline at the start of 2030 and 10.7% higher by the end of 2031, while unemployment initially rises by up to 1.3 percentage points.
These are conditional scenarios, not a forecast of what AI will deliver. The BCC cautions that its modelling does not capture the practical costs of workers retraining or moving between occupations.
Our take: For a UK employer, the useful question is which tasks become more productive and what training helps people move into the remaining work. Do not turn an economy-wide scenario into a promised return for your own business.
Google gives Gemini computer control on Mac, initially in the US
Google has added computer use to its Gemini Mac application, allowing it to navigate applications, work with local files and browse websites in the background. Users can follow a live preview and pause, stop or approve actions.
The initial release is in English for US users with Google AI Plus, Pro or Ultra subscriptions. Google has not announced when this feature will reach other markets, so UK availability should not be assumed.
Our take: Before adopting computer-controlling assistants, separate harmless file organisation from actions that send messages, change records or spend money. Start with copies of files, narrow access and a person checking consequential steps.
Meta responds to reports of Muse mixing private information across accounts
TechCrunch reports that internal Meta posts reviewed by Reuters describe Muse surfacing unfamiliar messages and bank balances, alongside concerns about access to Signal notifications. The examples come from employee testing and raise questions about information separation and permissions.
Meta told TechCrunch that employees test versions and features unavailable to the public, and that internal testing helps improve safety and privacy protections. The reporting does not establish that every public Muse user experienced the same behaviour.
Our take: An assistant that can read several applications needs more than a convincing demo. Check exactly which notifications, files and accounts it can access, and test whether information stays within the right user and task boundaries before connecting sensitive business data.
European Commission investigates an AI-assisted attack on internal systems
The Register reports that a cyber campaign disrupted European Commission HR, medical reimbursement and translation systems, with investigators reportedly detecting the ARTEX AI-assisted penetration-testing framework. Some systems were taken offline during the response.
The report leaves open whether employee information was stolen or systems were disconnected as a precaution. It also says salaries were paid as usual, despite disruption to staff services.
Our take: For UK organisations, the practical response is to check exposed services, privileged access and recovery procedures, rather than assume an AI-labelled attack requires an entirely new security programme. Keep a usable incident plan for payroll and other essential services.
Canva adds controls to switch off generative AI tools
Canva now lets free and paid users disable many generative AI features, including Magic Write and Magic Expand, The Verge reports. A separate preference reduces AI-generated material in library search results.
The switch is not a complete ban on all AI processing: background removal and Magic Grab remain available. Canva distinguishes those editing functions from tools that generate new material.
Our take: This gives teams a practical way to align everyday design tools with their publishing policy. Check which functions the setting actually disables, and do not treat the toggle as proof that every asset is human-created or cleared for your intended use.
OpenAI describes supervised agents that investigate ChatGPT slowdowns
OpenAI has detailed its Performance Factory, a workflow in which Codex agents investigate regressions, challenge diagnoses, propose fixes and verify recovery after deployment. Engineers retain review and deployment decisions.
In a manual assessment of 100 proposed root causes, OpenAI rated 70% accurate; the remainder needed refinement or had not established causality. The company describes independent review as a response to agents accepting plausible but incomplete explanations.
Our take: The transferable lesson is a checkable workflow, not unrestricted autonomy. Record evidence, use a second review for important conclusions, and measure the result after a change. A plausible explanation is not the same as a verified fix.
Apple is reportedly developing its own computer-operating agent
Apple is working on an agent able to control applications, browse websites and write code, TechCrunch reports, citing Bloomberg. The project is described as being in early development rather than an announced shipping product.
The report comes as competitors expand desktop automation. It does not establish a release date, pricing or a set of capabilities that UK businesses can buy today.
Our take: Keep procurement based on tools available now, with clear permissions and support terms. The growing number of desktop agents makes portability and an exit plan more useful than committing a workflow to a rumoured future product.
