Here's the contradiction almost everyone in AI is living inside right now and nobody is saying out loud.
The same week a study found that firms whose AI narrative outran their actual investment suffered a resilience penalty of –0.347, while firms whose investment quietly exceeded their narrative gained +0.469, the loudest voices in the market were the ones bragging about burning tokens.
Under-claiming is structurally rewarded. Over-claiming makes you more fragile. And yet the entire performance layer of this industry is optimized for the exact behavior the data punishes.
Under-claiming is structurally rewarded. Over-claiming makes you more fragile. Yet the entire performance layer of this industry is optimized for the exact behavior the data punishes.
That's not a culture gripe. That's a mispriced market, and it's the biggest product opening of the year.
Start with the performance itself, which Elena Verna named precisely. Five years ago people bragged about 5am cold plunges, inbox zero, and endlessly Pelotoning. Today they brag about having 17 agents, not opening their email, and letting AI run the business.
Same performance. Different props.
Verna calls it the AI Confidence Theater, and the tell is how the bragging changed shape. People used to flex outcomes: revenue, user milestones, headcount, fundraising. Now they flex how many tokens they're burning, as if consumption were achievement.
Then she landed the knife that this whole piece turns on. "So I started asking a simple question whenever someone told me AI had changed their life: 'Cool. Show me.'"
And the list of things that survive that question is embarrassingly short.
Because what actually shows up is honest and modest. Summarizing Slack. Answering emails. Scheduled scans. Research and booking something. Sending emails out of Claude.
Useful. Real. Not life-changing. And Verna's harder follow-up is the one that empties the room: show me something so critical that if I took it away tomorrow, your work would actually fall apart.
That's the human-scale version of the Outcome Clarity Test we opened last issue on — measuring outcomes over engagement. "Show me" is just Outcome Clarity aimed at a person instead of a dashboard, and most confidence collapses on contact with it.
Now here's where the culture gripe becomes a product problem, and it moves fast.
The theater doesn't just annoy people. It trains them. Every "life-changing" system that turns out not to work teaches the audience that AI claims are a con until proven otherwise.
The consumer numbers already show the training took. Six in ten people say AI showing up in a brand's messaging is a turnoff.
Most couldn't name a single company using AI well in marketing. And AI-generated answers without clear attribution are now the single least-trusted thing US consumers encounter online — ranking below medical bills, airline fees, and confusing fine print.
Sit with that. Your unattributed AI answer is trusted less than a hospital invoice.
This is what a trust deficit looks like when it's fully priced in. The default assumption has flipped from "impressive" to "prove it," and 86% of people still click through to the original source after an AI summary because they don't believe the summary alone.
It gets worse inside the product itself, where the assumption is now baked into how customers read your work. In Lenny's community, the live problem builders are trading notes on is earning trust when customers assume you vibe coded it.
Read that again. The starting position with a new customer is now suspicion — that your thing was slapped together by a model and can't be trusted to hold.
And the theater keeps handing them evidence they're right. Just as developers started adopting the Grok CLI — pulled in by the genuinely capable Grok 4.5 coding model — it got caught silently uploading all local files to the cloud.
A tool people trusted, ingesting their files without asking, at the exact moment it earned their attention. That single incident does more damage than a hundred keynote demos do good, because it confirms the con the theater already taught everyone to expect.
Then there's the tax the performance charges its own performers. Peter Yang calls it AI psychosis — the compulsion to constantly check that your agents are actually working — and says it's "probably a worse addiction than social media."
That's the punchline hiding in the brag. The person claiming an invisible team of AI employees runs their business is refreshing dashboards at midnight to see if the agents fired, because they only trigger 50% of the time and only produce decent output when hand-fed very specific context.
The theater doesn't even deliver peace of mind to the actor. It delivers a new compulsion dressed as leverage.
So here's the counterintuitive move, and it's the whole thesis: the credibility gap is not a problem to spin. It's the opportunity, and it's yours if you're willing to fail your own test first.
The resilience study gives you the exact bar. Every material AI claim should be traceable to something operational — an approved investment, a functioning system, a trained team, an accessible data resource, a deployed use case, or a measurable milestone.
Apply that test to yourself before your customer does. Run the AI-washing test on your own product and kill any claim that can't be traced to a working system.
This is the inversion the data begs for. In a market where everyone over-claims, verifiable restraint is the differentiator — because under-claiming is the behavior that quietly earned +0.469 while the loud players earned –0.347.
Concretely, three moves fall out of this.
First, make "show me" your internal bar before it's your customer's. If a feature can't survive "if I took this away tomorrow, would the workflow actually break," it's theater — cut it or ship it quietly and let it prove itself.
Second, make data handling a visible feature, not a buried policy. The Grok CLI taught the whole market to assume agents silently exfiltrate files, which means transparent, legible data behavior — showing exactly what leaves the machine and when — is now a product advantage you can actually market, precisely because the default assumption is theft.
Third, borrow the technique from the builders who understand their own tools. In Peter Yang's interview, the standout habit was asking Claude to generate an HTML report of what it changed and then quiz you to confirm you understand it.
Bake that into the product. An agent that shows its work, explains what it touched, and lets the user verify it is the operational answer to a customer who assumes you vibe coded it — the trace is the trust.
None of this is anti-AI. Verna works at an AI company, uses it all day, builds with it. The argument isn't that the tools are fake — it's that the performance around them is, and that performance is what's poisoning the well for the real, modest, genuinely useful wins.
Which is the last cost worth naming. The fakers rob people of the real ah-ha moment, because when the life-changing claim doesn't work, users conclude all of it is BS and keep their heads in the sand.
That's your opening. While competitors perform confidence they can't back, you can build the one thing the market is starving for and can't fake: a claim that survives "show me."
The winners of this cycle won't be the ones with 17 agents in the brag. They'll be the ones who can answer the simple question everyone else is running from.
Cool. Show me.