On 7 November 2026, The Verge published a review of NYT Cooking’s AI assistant. The reviewer describes standing in front of the fridge around 7PM on weeknights, trying to find a dinner that won’t get a “Chicken again?” from their husband. By their account, the assistant saved a weeknight chicken but almost ruined a cake.
I couldn’t load the full article, so I’m working from the headline and the feed summary. I’ll stick to those.
The excitement is about judgement
The claim around tools like this is easy to guess. An AI that knows a huge recipe library can answer “what do I do with what’s in my fridge” better than a search box can. For the weeknight chicken, that seems to be true.
The first question I ask is whether there’s a real need. Here there is. Deciding what to cook at 7PM is a small, repeated problem with no single right answer. Someone has to weigh what’s on hand, what the household will eat and how much time is left. That is judgement work, and the cost of a wrong call is a slightly dull dinner.
Why the chicken was safe and the cake wasn’t
Swapping ingredients in a savoury dish forgives errors. If the assistant suggests a different herb, a different acid or a different cut, you taste as you go and adjust. A bad guess gets corrected before it reaches the table.
Baking doesn’t work that way. A cake’s chemistry is fixed once it goes in the oven. You can’t taste your way back from a wrong substitution. By the time you find out, the cake is already ruined.
The assistant is the same in both cases. What changed is the cost of a wrong answer and whether you get a chance to fix it. A single tool can be a good helper in one task and a risk in the next, and nothing in how it presents itself tells you which is which.
What this means for an operator
The same split runs through any service business. Take a shop that wants AI to answer customer questions. Some questions are chicken questions. Which product suits a customer, how to describe a service, what to say when someone is unsure what they need. A weak answer there costs little, and a person can step in.
Other questions are cake questions. What is the price. Is this item in stock. Can we deliver by Friday. Will you refund this order. One wrong answer there becomes a promise to a customer, and honouring it or breaking it both cost money.
My position is that rules belong where wrong is costly, and AI belongs where judgement is needed. I’d also trust nothing unverified. An answer that sounds confident has told you nothing about whether it’s correct. A recipe that reads well has told you nothing about whether the cake will rise.
That applies to a clinic quoting a treatment price, a tuition centre confirming a class slot and a marketing team publishing a claim about a product. The more a single wrong output costs, the less you should leave it to a model’s judgement alone.
Where this shows up in what I build
With AeroChat, the rules about what can be committed to sit in the platform, not in the prompt. A merchant sets the limits. For example, it can refund up to the amount you set and hands anything bigger to your team. The wording of a reply is a judgement task. The size of a refund is a rule.
What a business should do
Start by listing the tasks where you’d want AI help. Next to each one, write what a wrong answer costs and whether anyone would catch it before the customer acts on it.
Put AI on the low-cost tasks first. Drafting, suggesting, sorting and explaining are good starting points, because mistakes there are cheap and visible.
For the high-cost tasks, decide the rule before you turn anything on. Prices, stock, dates, refunds and promises should come from your systems or a fixed limit, not from a model’s best guess. Where no rule is possible, put a person in front of the answer.
Then check the results yourself for a few weeks. Read what the AI actually told customers. Move a task from the cake list to the chicken list only when the record supports it.
Source: The Verge AI: NYT Cooking’s AI assistant saved my weeknight chicken, but almost ruined my cake