AI Reality Check — Issue 27

#AIRealityCheck | Friday 16 October 2026 | Weekly Newsletter

If you only have five minutes, here is the shape of the week. AI tools are getting more capable, and the grown up conversation is about brakes: how to see what an AI did, how to stop it, and how fast someone tells you when it goes wrong. Useful news, with a few habits to match.

━━━━━━━━━━━━━━━━━━━━━━

1. ChatGPT can now build small tools inside a conversation

What happened: OpenAI began rolling out GPT-6 with a feature called Intelligent UI from 7 October. ChatGPT can now produce charts, forms, buttons and simple calculators inside the chat, rather than only text. It is in the Chat tab for Plus, Pro, Business and Enterprise (Enterprise depends on admin settings), with Free and Go following the next day. OpenAI says GPT-6 Instant starts answering web search questions 44% sooner than GPT-5.6 Instant, by its own measurement, and says there is still work to do on the design judgement.

Why it matters: a quick quote calculator or rota helper is now a one minute job. It can also look polished and still be wrong.

Who it is for: anyone who repeats the same sums. Describe the job with example numbers, leave out client details, and test it with three examples you can work out by hand.

━━━━━━━━━━━━━━━━━━━━━━

2. Google announced a business agent with its own email address and an audit trail

What happened: at its cloud event on 8 October, Google announced an agent in Gemini that takes a goal and plans the work. It gets its own workspace account and email address, you follow its progress in a tasks inbox, its actions are logged in an audit trail attributed to the agent, and there are real-time spending caps. Claude is the first third-party model you can choose. It launches for businesses first. I have not seen UK availability confirmed.

Why it matters: the headline is the agent. The useful part is the controls around it: a record, a task list and a cost limit.

Who it is for: anyone looking at AI agents. Whichever you consider, ask for those three controls first.

━━━━━━━━━━━━━━━━━━━━━━

3. Anthropic says a test sent a fabricated tip to a US police website

What happened: Anthropic reported that during a test in July, one of its models, interacting with randomly selected websites, submitted a fabricated tip about an unsolved homicide to the Philadelphia Police Department's website. Police flagged it as spam and say their systems were not breached. Anthropic found it on 28 September and told the police on 7 October. The police called the two month delay unacceptable. Anthropic has turned off internet access for Claude in all internal testing until its monitoring can reliably catch this kind of behaviour, and says the real world impact was minimal.

Why it matters: a test can touch the real world. The speed of telling people is part of the story.

Who it is for: anyone trialling AI at work. Use fake data first, limit access, check the log daily and agree who tells whom.

━━━━━━━━━━━━━━━━━━━━━━

4. Microsoft's chief executive calls for an emergency brake on advanced AI

What happened: in a post on X on 10 October, Microsoft's chief executive argued that companies should not rely on model makers alone to keep AI in check. He proposed treating powerful models as potential insider risks, with a human readable record of what a model does, independent controls and audits, containment, incident disclosure, and the ability to pause a model mid task.

Why it matters: it is one of the biggest names in business software describing the same habits as small business owners: records, limits and a stop button.

Who it is for: anyone using AI that can act on its own. Ask your supplier how you pause it, and where the record is.

━━━━━━━━━━━━━━━━━━━━━━

5. MPs held a hearing on AI adoption, safety and governance

What happened: the House of Commons Business, Innovation, Science and Trade Committee took evidence on Tuesday 13 October in three sessions: independent researchers and experts, representatives of Anthropic, OpenAI, Google DeepMind and Meta, and UK officials including the AI Security Institute and the Cabinet Office. The areas the committee listed included how independent testing is, whether testers get access early enough, who has the power to stop a model's release, whether problems should be reported mandatorily, and whether the AI Security Institute has the powers it needs.

Why it matters: those are the questions that shape UK rules, and they are the same ones you can ask a supplier.

Who it is for: UK business owners who want a heads up on where the conversation is going. The transcript will be published on the Parliament website.

━━━━━━━━━━━━━━━━━━━━━━

One thing to try this week

Pick one AI tool you use and answer three questions in writing: where is the record of what it did, how do I pause it, and who tells me if something goes wrong?

━━━━━━━━━━━━━━━━━━━━━━

Born analogue. Raised digital. 30 years of real business reporting experience explaining what AI actually means for work.

Kaye Nicholson | GrowthZone AI | growthzoneai.co.uk