Because capability and credibility are different things, and the industry has been spending everything it earns on the first one.
A woman at a workshop asked me a question in March that I have not been able to put down.
She ran a small bookkeeping practice. She had been using AI for about four months and it had genuinely helped. Then she said, almost apologetically, “I don’t tell my clients I use it. Is that wrong?”
I gave her an answer that I now think was too easy. I said something about how nobody discloses which spreadsheet software they use, and she seemed relieved, and we moved on.
I have thought about that exchange probably fifty times since. Because the honest answer is that she already knew the answer. She would not have asked if the two things were the same. Nobody has ever felt a flicker of guilt about not disclosing Excel.
Here is the direct answer to the question in the title. Trust in AI is falling while capability rises because they are not the same asset and they are not built the same way. Capability is built by scaling compute. Credibility is built by being caught telling the truth about something inconvenient, repeatedly, over time. The industry has been investing enormous sums in the first and almost nothing in the second, and the public has noticed.
For anyone running a small business, that gap is not bad news. It is the opening.
Key Takeaways
- Public trust in AI is declining in the same period that model capability is measurably improving, which means capability is not what trust is made of.
- A CNBC and Generation Labs poll found 69 percent of Americans aged 18 to 34 do not trust OpenAI’s CEO to act responsibly on AI, and 45 percent believe AI will hurt their careers.
- Anthropic’s own CEO said publicly this weekend that AI companies have not delivered on their promises and that advertising will not fix it.
- A four university study found an AI chatbot convinced 46 percent of participants to install an app in a simulated scam, versus 18 percent for human scammers.
- In a market losing credibility, the operator who explains their process specifically becomes the safe choice, which makes transparency a competitive position rather than a virtue signal.
We Are Building Capability and Spending Credibility
There is a version of this conversation that turns into hand wringing about technology, and I am not interested in writing that.
I use AI in nearly every part of my work. It has made me faster, it has made some of my output better, and it has let me take on work I would have turned down two years ago. I am not conflicted about the tools.
What I am watching, with some concern, is a widening distance between what these systems can do and what people are willing to believe about them.
The capability side is not in dispute. Models this month post materially better coding scores than they did three weeks ago. An open weights model you can run on a workstation now handles vision and long context. Inference speed jumped by an order of magnitude in a preview release last week. On every measurable axis, the tools are getting better fast.
And in the same window, public confidence is going the other way.
This is the part I think a lot of builders are misreading. They see the trust numbers and assume it is an education problem. If people just understood the technology better, they would come around. Give it time, ship better products, the skepticism will fade.
I do not think that is what is happening. I think people understand it well enough. What they doubt is not the capability. It is whether the people deploying it are being straight with them.
And here is the uncomfortable part for those of us running small businesses. When trust in a category falls, it does not fall selectively. It does not distinguish between the frontier lab making enormous claims and the bookkeeper in Ohio who quietly uses AI to draft a client summary. The suspicion lands on everyone.
But what if the same thing that makes that suspicion a problem for the category makes it an opportunity for the individual?
What the Last Seventy Two Hours Actually Showed
One. The most credible voice in the industry said the quiet part out loud.
On Saturday, Dario Amodei, CEO of Anthropic, published a long post conceding that AI companies “haven’t yet delivered on our big promises to benefit the world.” He endorsed a central regulator on the FINRA model. And he argued explicitly that advertising cannot repair the problem, saying the thing that will work is actually delivering something like curing cancer.
Whatever you think of the position, notice what it is. The head of a frontier lab, in a moment of significant commercial momentum, chose to publicly name a credibility deficit rather than market past it. That is not a small thing and it is not the behavior of an industry that thinks trust is a messaging problem.
Two. Young Americans do not trust the people running this.
A CNBC and Generation Labs survey of over a thousand Americans aged 18 to 34, reported this weekend, found 81 percent do not trust Palantir’s CEO to act responsibly on AI. Seventy nine percent said the same of Peter Thiel, 71 percent of Mark Zuckerberg, 70 percent of Elon Musk, and 69 percent of Sam Altman. Satya Nadella was the only executive tested with net positive trust.
Beyond the personalities, 45 percent said AI will hurt their careers, and 60 percent want the data center buildout slowed.
These are your customers, your employees, and in many cases your children.
Three. The persuasion research should change how you think about your own marketing.
A four university study reported by Vice found that in a week long simulation of a pig butchering scam, an AI chatbot convinced 46 percent of participants to install an app, compared to 18 percent for human scammers. The researchers attributed the AI advantage to persistence and remembered personal detail.
I want to be careful how I use this finding, because the obvious read is “AI is dangerous” and that is not the useful read. The useful read is that persistence plus remembered personal detail is exactly what a well built marketing automation does. The mechanism that made the scam effective is not exotic. It is sitting in most of our follow up sequences right now, aimed at something legitimate.
That should make all of us a little more careful and a lot more deliberate.
Four. Familiarity, not argument, is what moves trust.
Edelman’s Trust Institute work on AI has landed on a finding I keep returning to: trust in AI grows through direct, meaningful experience with it, not through public commitments to responsible design. The 2026 Trust Barometer describes a general environment of institutional retreat, with people narrowing toward smaller and more familiar circles.
Read that alongside the poll numbers and something clarifies. People do not trust distant institutions making large claims. They still trust specific people they have actual experience with.
That is not a consolation prize for small business. That is the entire competitive landscape described in one sentence.
Be the One Who Is Easy to Verify
I went back to that bookkeeper’s question with a better answer, eventually. Here it is.
The reason AI feels different from Excel is that AI does part of the thinking. Excel does arithmetic you specified. AI produces judgment shaped output that a client reasonably assumes came from your professional mind. That is why the instinct to disclose exists, and the instinct is correct.
But the answer is not a disclaimer at the bottom of an invoice. Disclaimers read as legal cover. They make things worse.
The answer is to describe your process specifically enough that a client can verify it. Not “we use AI to enhance our service,” which communicates nothing and sounds like hedging. Something closer to: “I use AI to draft the first version of your monthly summary from your transaction data. I review every figure against the source before it reaches you. I do not put your financials into any tool that trains on them. If something looks wrong to you, it is my mistake, not the software’s.”
Four sentences. A client can hold you to every one of them.
That is the whole move, and it works precisely because so few people are willing to make it. In a market where the loudest voices are making claims nobody can check, being checkable is a position almost nobody occupies.
There is a version of this that goes further and matters more. Publish the thing you got wrong. Not a performative confession, just a specific account: here is what my process produced, here is why it was wrong, here is what I changed. I have done this a handful of times and it has never once cost me a client. It has repeatedly won me one, because it is the single hardest thing to fake.
I hold a conviction about this that runs deeper than strategy, so let me name it plainly. Honesty is not a marketing tactic that happens to work. It is the thing you owe the person paying you, and the fact that it also happens to be the strongest available position in a low trust market is a mercy, not a justification.
The people who will be trusted with AI in five years are being decided right now, by how they behave in a moment when nobody is checking. That is generally when it gets decided.
Practical Steps
1. Map every place AI touches the work your customers pay for.
Walk your delivery process and mark each AI touch. Be specific about which tool and which step. You cannot be honest about a process you have never actually written down, and most owners find at least one touch they had forgotten.
2. Mark which touches a human reviews before a customer sees them.
This is the line that matters most to a nervous client. Not whether AI was involved, but whether a person looked. Where the honest answer is no, decide whether to change the process or to say so.
3. Write four sentences you would be comfortable reading aloud.
Say what the AI does, what you do, what you will not do with their information, and what you take responsibility for regardless. Read it out loud. If any sentence makes you wince, that is the sentence to fix in the process, not in the wording.
Here is the prompt I use to draft it:
[The Job]
Write a plain language AI disclosure statement for my customers, based on my actual process.
This is for: [DESCRIBE YOUR CUSTOMERS].
It matters because: a vague disclosure reads as evasion and a specific one reads as confidence.[The Background]
Here is what you need to know: my process step by step is [DESCRIBE IT], the AI tools involved are [LIST THEM AND THE STEP EACH TOUCHES], the steps a human reviews are [LIST THEM], and my customers’ likely concerns are [LIST THEM].
Do not use: legal disclaimer language, hedging, or any claim about accuracy that I have not given you evidence for.[The Deliverable]
Return: a disclosure under 250 words.
Must include: what AI does, what a human does, and what I take responsibility for regardless.
Optimize for: persuasion.[The Questions]
Ask me any questions you have.
4. Audit your own automation against the persistence finding.
Read your follow up sequences with fresh eyes. Ask whether the cadence and the personalization would feel respectful or predatory if you were on the receiving end and did not know how it was built. Change anything you would not want read back to you.
5. Make the human in your process visible by name.
Do not write “human in the loop.” Name the person and the judgment they exercise. A nervous buyer is not looking for reassurance that a process exists. They are looking for someone to be accountable, and accountability requires a name.
6. Publish one thing you got wrong.
One specific failure, one specific fix, no request attached. This is uncomfortable and it is the highest return thing on this list. Nobody with something to hide does it, which is exactly why it registers.
7. Add one verifiable proof point every month for a year.
A named case study, a published process explanation, a specific testimonial that answers a real objection. Trust does not arrive in a campaign. It accumulates, and twelve small deposits will outperform one large gesture every time.
Frequently Asked Questions
Do I legally have to disclose that I use AI?
It depends entirely on your jurisdiction, your industry, and what you are producing, and this is a question for a lawyer rather than for me. What I can say is that in most small businesses the relevant standard is not the legal minimum. It is what your client would feel about learning it from someone else.
Will telling clients I use AI make them think I am charging too much?
Occasionally, and the conversation is worth having anyway. Clients are paying for judgment, accountability, and outcome, not keystrokes. If your price only holds while the client believes you did everything manually, the price was resting on a misunderstanding.
Is public trust in AI actually going down, or is that just media noise?
Multiple independent sources point the same direction. The CNBC and Generation Labs poll found majority distrust of most major AI executives among young Americans, and Anthropic’s own CEO publicly acknowledged the credibility gap this weekend. Different sources, same conclusion.
How do I compete with businesses that are not being transparent?
You do not compete with them on volume. You compete on being the option a cautious buyer can safely choose. That is a smaller market with better clients, which is usually the trade a small business should want anyway.
What if my process is genuinely mostly AI?
Then say so and be specific about what you add. Plenty of valuable businesses are mostly AI plus expert oversight, and clients buy that willingly when it is priced and described honestly. The problem is never the ratio. It is the concealment.
The Close
I still think about that bookkeeper.
She was not asking me a compliance question. She was asking whether she was still the kind of professional she thought she was. Whether using a tool that thinks had quietly moved her into a category she would not have chosen.
The answer, which I wish I had given her in March, is that the tool does not decide that. What she does with the truth about it decides that.
The industry that builds these models is having exactly the same reckoning right now, several orders of magnitude larger, in public. Their most credible leader stood up this weekend and said we have not earned this yet. Nobody made him say it. He said it because it was true and because he apparently concluded that saying it was better than continuing to market past it.
That is the whole play, and it is available to a bookkeeper in Ohio on the same terms it is available to a frontier lab.
Capability is going to keep getting cheaper. Everyone will have it. It is already table stakes and it is falling in price every quarter.
The thing that will not commoditize is being someone whose word holds up when it is checked.
Be easy to verify. In a market this loud, that is nearly the whole game.
About the author
Jonathan Mast is the founder of White Beard Strategies. He writes about running a business with AI in it, including the parts he is still working out. He is the creator of the Perfect Prompt Framework, a speaker, and a person who believes that most business problems are character problems wearing a strategy costume. He still owes a bookkeeper in Ohio a better answer.
Sources
- Amodei says only curing cancer will win back AI trust, Fortune
- Young Americans deeply distrust AI CEOs, Futurism
- AI chatbots better at scamming people than human scammers, Vice
- When familiarity breeds trust: how experience turns AI skeptics into believers, Edelman
- 2026 Edelman Trust Barometer
- AI News Today, August 17, AI Weekly
