AI Reality Check — Issue 25

#AIRealityCheck | Friday 2 October 2026 | Weekly Newsletter

Kaye Nicholson | GrowthZone AI | growthzoneai.co.uk

If you only have five minutes, here is the shape of the week. AI is being asked to behave more like every other tool in your business. Tell people when something goes wrong. Show what it costs. Follow rules. That is a good sign, even when the headlines look dramatic.

━━━━━━━━━━━━━━━━━━━━━━

1. An AI agent reached a government health portal, and the gap became the story

What happened: Australia's Prime Minister said an OpenAI model reached files on a government health statistics portal in June. The files were classed as non-sensitive. OpenAI found it in an internal review in August and told the Australian government on 10 September. OpenAI has also said its models reached some US government sites, including the Census Bureau and the SEC. The heads of OpenAI and Anthropic were called to give evidence to Australia's Senate on 1 October, and Australia joined 22 nations in saying AI should stay under human control.

Why it matters: the data was not sensitive, so the lesson is the three month gap. Any tool that can log in, browse or send on your behalf deserves the same question: what can it reach, and who would tell you if something went wrong?

Who it is for: anyone using AI tools that act on their behalf, and anyone choosing an AI supplier.

━━━━━━━━━━━━━━━━━━━━━━

2. The GPT-6 family got cheaper

What happened: OpenAI released GPT-6 Astra in early September and describes it as its most intelligent and aligned model yet. On 23 September it added two cheaper models, Sol for more demanding work and Luna for lower cost tasks at scale. OpenAI says API prices for Sol and Luna are 50% lower than GPT-5.6 promotional prices, and that Sol makes about half as many factual errors as the previous Sol on its own test. In testing, Astra reached what OpenAI calls Critical cybersecurity capability, so it delayed parts of development and added safeguards.

Why it matters: better and cheaper is good news for small businesses. But half as many errors is not no errors, and you probably do not need the top model for everyday jobs.

Who it is for: anyone paying for ChatGPT or using it through other tools. Try the cheaper model first, and keep checking names, numbers and dates.

━━━━━━━━━━━━━━━━━━━━━━

3. ChatGPT's custom GPTs now have an end date: 11 December

What happened: OpenAI is retiring custom GPTs and replacing them with plugins. A plugin is a reusable package that can hold skills and connections to other apps. Your GPT's instructions become a skill and its knowledge files become reference files. Custom actions do not transfer, the model you picked does not carry over, and the original becomes read-only once migrated. Some dates vary by plan and workspace.

Why it matters: if a client or your team relies on a custom GPT, it stops on 11 December unless it is moved. Ten weeks is plenty of time, but not if nobody notices.

Who it is for: anyone who has built, or relies on, a custom GPT. Make a list, save a copy of the instructions and files, migrate the ones you use, and test.

━━━━━━━━━━━━━━━━━━━━━━

4. UK MPs called for an AI law, and OpenAI said it would back stronger rules

What happened: on 14 September a cross-party parliamentary committee published a report calling for an AI law, a single regulator with enforcement powers, and more transparency from developers. OpenAI said it supports stronger UK rules for the small number of companies building the most powerful AI, without adding burdens for startups or researchers working on less powerful systems. Separately, a Commons committee has invited Meta, Google, OpenAI and Anthropic to an urgent hearing on 13 October, where topics include independent safety testing and mandatory reporting of serious incidents. The business secretary has cautioned against exaggerating the risks.

Why it matters: rules are coming, and suppliers will start to be clearer about testing and incident reporting. That is useful for buyers.

Who it is for: UK business owners choosing AI suppliers, and anyone who wants to know what the rules might look like.

━━━━━━━━━━━━━━━━━━━━━━

5. Safety researchers keep leaving big labs, and are asking for a slower pace

What happened: several researchers have left Anthropic and Google DeepMind in recent weeks, saying the race between labs is moving faster than safety work can keep up. One joined an independent evaluation group and another a nonprofit that teaches AI safety. Separately, Anthropic's chief executive called for a slower pace of advanced AI development, and leaders at OpenAI and SpaceXAI publicly agreed. Not everyone shares the same view of how large the risks are.

Why it matters: you do not need to take a side in that debate to take something from it. Choose suppliers who are open about testing, incidents and limits, and keep a person in charge of anything important.

Who it is for: anyone deciding how much to rely on AI, and how quickly.

━━━━━━━━━━━━━━━━━━━━━━

Also this week: a new report found that 87% of young people in England aged 11 to 24 have used AI, and nearly half trust it more than a person to check if something is true, while fewer than 10% think school prepares them. SurveyMonkey launched a plugin for creating and analysing surveys inside ChatGPT. Google planned to launch a small test satellite carrying AI chips on 1 October.

━━━━━━━━━━━━━━━━━━━━━━

One thing to do this week: open ChatGPT, find My GPTs, and write down which ones you actually use.

━━━━━━━━━━━━━━━━━━━━━━

This week in one sentence: AI is becoming more accountable, more affordable and more regulated, and the businesses that ask sensible questions will get the most from it.

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

— Kaye Nicholson | GrowthZone AI | growthzoneai.co.uk