On 3 November 2026, e27 published a piece by Ajay Tiwari asking whether the fears around AI are genuine concerns or narratives shaped to serve the interests of the people voicing them. It cites Sam Altman telling the United Nations Security Council in September 2026 that the two major dangers are losing control of AI and a small number of people gaining too much power over it. It also cites Reuters reporting in October 2026 that Dario Amodei of Anthropic called for a coordinated slowdown, stronger safety tests and independent checks on the most advanced systems. Tiwari’s conclusion is that claims about both the huge benefits and the extreme dangers of AI should be checked independently, and that rules should treat ordinary AI applications differently from systems that could cause much greater harm.
The doom story and the hype story come from the same people
The line in the piece worth keeping is this one: a company may genuinely fear what its technology could do while also benefiting from the publicity and political influence created by that fear. Tiwari notes that warnings about extreme danger shape government rules, investment, competition and public attention.
I would take that one step further for anyone running a business. The founders warning that AI could escape human control are the same founders raising money on the promise that it changes everything. Both stories do the same job. Each makes the technology sound bigger than anything built before, and each puts the person telling the story at the centre of what happens next. A warning about extinction gets the same attention as a promise of abundance, and from the same audience of regulators, investors and journalists.
So for an operator, both stories are marketing. Neither tells you whether a tool will answer a refund request correctly tomorrow morning. I stay out of the hyperbolic conversations for exactly this reason. Most of them rest on assumptions nobody has tested, and testing is the only part I find useful.
What changed in your business this week, and what did not
Nothing in a shop, a clinic or a marketing team changed because a chief executive spoke at the United Nations. The messages arriving on WhatsApp this Tuesday are the same as last Tuesday. Where is my order. Can I move my appointment to Thursday. Do you have this in a size 10. Can I return this. A customer at 11pm asking which of two products suits them.
Those messages are the test. The question is never what a model could do in a lab. The question is whether the tool checked the order, moved the booking, answered the stock question and handed the return to the right person, and whether it did that the same way every time.
Tiwari says rules should match the actual level of risk, and that ordinary AI applications should be treated differently from systems that could cause much greater harm. I agree, and I would apply the same logic inside the business. An AI that answers an order-status question by checking the order in your system is an ordinary application. The risk in it is a wrong answer, and a wrong answer is something you can measure, count and fix.
Decide what it can commit to, then judge what it does
A well-built AI resolves things. The rules on what it can commit to belong to the business. How much it can refund without asking. Which discount it may offer. What delivery date it can promise. When a person takes over, and with what context. Those rules live in the platform, not in the prompt, because a prompt is a request and a rule is a limit.
My working principle for all of it is rules where wrong is costly, AI where judgement is needed. Once the limits are set, the abstract question of whether AI is dangerous becomes a practical one: did it do the job today, and was every commitment it made one I had approved in advance? That is how I built AeroChat. The merchant sets the limits on refunds, discounts, delivery dates and handover, and the platform enforces them, so the conversation can act on a customer’s request without going past what the business allowed.
The same discipline answers the hype side. A tool that cannot show you, in your own transcripts, what it resolved and what it handed over has not earned a bigger role, whatever its founder said on stage.
What a business should do
Read the speeches as marketing, from both sides. Then run your own test.
- Pick one workflow that runs every day. Order status, appointment changes, return requests or product questions.
- Write down the limits before you start: the refund ceiling, the discounts allowed, the dates it may promise, and when a person steps in. Put those limits in the system, not in a prompt.
- Run the tool on that workflow for one ordinary week.
- Read the transcripts. Count the conversations it closed, the ones it handed over, and any commitment it made that you had not approved. Trust nothing unverified.
- Keep what worked, fix the brief where it fell short, and add the next workflow only when the first one holds.
Source: e27: Are AI fears genuine? Why we should demand evidence, not alarm