If you have been quietly suspicious of everything you hear about AI, this is for you. You are not behind. You are responding sensibly to how it is being sold.
Mark Zuckerberg published a 6,500 word essay about what superintelligence will do for ordinary people. The reception was rough, and the interesting part is why. Most of the criticism was not aimed at the arguments. It was aimed at him.
That gap is worth understanding, because you are on the receiving end of it every time somebody pitches you something with AI in the name.
The reporting below is from AI Optimism Has a Trust Problem in the AI Daily Brief, along with the outlets it cites.
The debt this industry inherited
Start with a number. 64% of Americans say social media has been harmful to democracy, and roughly the same share say it should be regulated more heavily.
Sit with what that means for anybody now selling you AI. The last era of technology promised connection and delivered something a majority of the country believes damaged the place they live. Nobody who lived through that owes the next promise the benefit of the doubt.
So when a technology executive explains what the next thing will do for you, the argument arrives carrying a debt it did not personally run up but cannot separate itself from. This is not the public being irrational. It is the public remembering.
That applies to a Meta essay. It also applies to the agency that emails you about AI transformation, and, uncomfortably, to us. Anyone selling this is drawing on a line of credit that somebody else already spent.
What made it worse
The essay itself had real substance. There is an argument about jobs, that nothing forces automation to outrun what people can do, and a prediction of smaller companies employing more people in total. There is a proposal that government and industry work together through shared technical checkpoints rather than industry simply being left alone. There is a commitment that data centers should bring high-paying local jobs, investment in schools, energy prices that do not rise, and environmental care, backed by a $1 billion fund. Open-source AI comes up at least 16 times.
You can agree or disagree with all of it. What sank the reception was not any of that.
It was the examples. The essay imagined personal superintelligence choosing which recipe your child should bake. It imagined an agent optimizing your hobbies. Writing in The Verge, Elizabeth Lopatto pointed out that an AI cannot replicate the calm of knitting, and that picking a recipe with your kid is most of the point of baking with your kid. Weeks earlier Sam Altman had taken similar criticism for describing an AI that reads your family calendar and produces a personalized morning podcast.
Then there is the contradiction. Bloomberg noted that Meta cut 8,000 jobs earlier in 2026, the same year its founder published an argument that AI will mean more jobs, not fewer. The essay can be internally consistent and still fail, because people do not evaluate an argument in isolation. They evaluate it against what the arguer actually did.
The lesson is not “AI is overhyped”
It would be easy to read all that and conclude the technology is oversold. That is the wrong lesson, and it is expensive.
Look at what the criticism actually targeted. Nobody argued the models do not work. The complaints were about the examples chosen to describe them: that they came from people who have lost touch with what daily life is like, and who were describing convenience nobody asked for while removing the parts of a task that were the point.
Which is a criticism of the pitch, not the tool.
The same distinction sits in front of you every week. A tool that drafts your quote follow-ups is not the same claim as a tool that will transform your business. The first is checkable. The second is a mood. Most of what gets sold to small businesses under the AI label is the second thing wearing the clothes of the first, and the reasonable response is to notice which one you are being handed.
How to tell a real claim from a pitch
Four questions. They work on us as well as on anybody else who emails you.
Does it name the task, or the outcome? “Saves you time” is not a claim, it is a hope. “Drafts the follow-up email to every quote that has gone 3 days without a reply” is a claim, because you can check whether it happened.
Is the number theirs or somebody else’s? Borrowed statistics from a vendor’s case study are marketing. Ask what it did in a business the size of yours, and what it did not do. An honest answer includes the second half.
What happens when it is wrong? Everything automated is wrong sometimes. A real answer describes how you find out and how you undo it. If nobody has thought about that, nobody has run this in a real business.
Who owns it at the end? If the capability leaves when the invoice stops, you rented an outcome. That may be fine, but it should be a decision rather than a discovery.
Notice what none of these ask about. Not the model, not the technology, not whether superintelligence is coming. You do not need a position on any of that to run a business well, and anybody who tells you otherwise is selling the mood.
The part that is actually changing
The piece ends on something more hopeful than its title suggests: a middle ground is forming. Not the version where everything is about to be solved, and not the version where none of it works. Somewhere between, which is roughly where the truth about most technologies ends up sitting.
Getting there needs credible messengers, and the industry is short of them. Which is an odd kind of opening for a small business, because credibility at your scale is not built with an essay. It is built by having done the thing for somebody who will say so.
That has a practical consequence. Your customers carry the same suspicion you do. If you start using AI in ways they can feel, they will apply the same test to you: did this make it better for me, or cheaper for you? A follow-up that arrives faster passes. A chatbot standing between them and a person who can actually help does not.
Where to start
Stop evaluating AI and start evaluating individual claims. The category question, is AI good or overhyped, has no answer that helps you on Monday. The specific question does.
Pick the next thing anybody offers you with AI attached, including from us, and run the four questions at it. Name the task. Check whose number it is. Ask what happens when it breaks. Ask who owns it afterward.
If it survives that, it was never really an AI decision. It was a decision about one piece of work you do repeatedly, which is the only kind of decision worth making here anyway.
And if a pitch cannot survive four plain questions, the trust problem was not yours.