The Founder Credibility Crisis in AI

Interpreting AI founders and what they are building

A thought leadership piece for founders, operators, and anyone building at the frontier

Something has gone quietly wrong in how AI founders talk about what they are building.

It did not happen suddenly. It accumulated gradually, through thousands of pitch decks and press releases and conference keynotes, until the gap between what AI founders claim and what they can actually demonstrate became wide enough that sophisticated observers began to notice — and then to discount.

The AI founder credibility crisis is not about fraud. It is not about deliberate deception, though there is some of that. It is about something more structurally interesting: a field that has adopted the vocabulary of transformation without developing the discipline of explanation. A generation of founders who have learned to speak confidently about systems they cannot fully describe — and an ecosystem that, until recently, rewarded them for it.

The consequences are now arriving.

Every technology cycle produces its own inflationary vocabulary. Words that once meant something specific get stretched until they cover almost everything, and then almost nothing.

In AI, this happened faster than usual, and the inflation was more thorough.

“Intelligent” came to mean “produces plausible-sounding outputs.” “Understands” came to mean “processes text containing relevant tokens.” “Learns” came to mean “was trained on data that included examples of the target behaviour.” “Reasoning” came to mean “generates a chain of text that resembles an argument.”

None of these usages are technically wrong, exactly. Language models do produce plausible outputs. They do process text in ways that sometimes resemble understanding. The problem is that the gap between the technical reality and the intuitive meaning of these words is large enough to mislead — and in a funding environment where each successive round depends on a story of increasing capability, the incentives to exploit that gap rather than close it are substantial.

The result is a cohort of founders who have become genuinely skilled at explaining what their systems appear to do without being particularly precise about how they do it, what they cannot do, and under what conditions they fail.

There is a simple test I have applied, informally, to many AI founders over the years. I call it the explanation test, and it goes like this:

Ask the founder to explain, in plain language, how their system works. Not what it does — how it works. Then ask them to describe a category of input for which their system would reliably fail. Then ask them what oversight mechanism exists to catch those failures before they affect a user.

The results are consistent enough to be disquieting.

Most founders can answer the first question reasonably well, at least at a surface level — they have been asked about the product often enough to have a polished response. The second question produces more variation. Some founders have thought carefully about failure modes; many have not, or have thought about them in ways that are systematically optimistic. The third question is where the conversation most often breaks down.

This is not because founders don’t care about failures. Most do, and the good ones care deeply. It is because the pace of building, the pressure of fundraising, and the structure of the ecosystem all push in the same direction: toward demonstrating what the system can do, not toward understanding and communicating what it cannot.

The incentives for rigorous self-examination are weak. The incentives for confident projection are strong.

The cost of this gap is distributed unevenly, and the people paying the highest price are not the founders.

At one end, there are the investors who have backed systems that cannot do what was claimed, and who are now holding portfolios that are significantly less impressive at due diligence than they were at pitch. This is the cost that gets the most attention, because investors have the platform to talk about it.

At the other end — and this is the cost that should concern us more — are the people subject to AI-assisted decisions made by systems whose limitations their operators did not fully understand and therefore did not adequately disclose.

When a founder overclaims capability and an enterprise customer deploys the system in a high-stakes context on the basis of that claim, the people who bear the risk are typically neither the founder nor the customer. They are the individuals whose loan application, job interview, or benefit assessment was influenced by a system operating outside its reliable range.

The credibility gap is not merely a commercial problem. It is an accountability problem. And it is one that the field has not yet developed adequate mechanisms to address.

There is a particular artefact of AI culture that deserves specific examination: the demo.

The AI demo has become a sophisticated art form. Companies are skilled at constructing demonstrations that show their systems at their best — optimal inputs, favourable conditions, carefully selected use cases that happen to be the ones the model handles most reliably.

This is not unusual in technology. Every product is demonstrated at its best. But in AI, the gap between demo performance and production performance is often substantial, and the conditions that produce the gap are rarely disclosed.

A system that handles a curated demo input gracefully may handle an edge case from a real user — with unusual phrasing, unexpected context, or inputs from a demographic underrepresented in the training data — in ways that are significantly less reliable. The demo never shows this. The pitch deck never addresses it. The contract often does not specify what “production-quality performance” actually means.

This gap is where most enterprise AI deployments go wrong. Not because the technology was fraudulent, but because the buyer’s expectations — shaped by demo performance — were not calibrated to production reality.

The honest founder closes this gap proactively. They show the failure cases. They explain the conditions under which the system is unreliable. They specify what oversight is required for safe deployment. They treat their customers as partners in managing risk rather than audiences for a performance.

These founders exist. There are not enough of them.

Something is shifting.

The customers who were early adopters of enterprise AI — who signed large contracts in 2022 and 2023 on the basis of demo performance and founder conviction — have now had enough time in production to form their own views. Those views are often more measured than the pitch suggested they would be.

The sophisticated buyers in enterprise software are developing the capacity to evaluate AI claims with more rigour. They are asking harder questions in procurement. They are building internal expertise. They are sharing experiences with peers in ways that create informal accountability for overclaiming.

The funding environment is also changing. The investors who wrote large cheques based on vision and narrative are increasingly looking for evidence of deployment at scale, customer retention, and production performance. The bar for “impressive demo” has risen. The bar for “demonstrated value in production” is where attention is now focused.

This correction is healthy. But it is slow, and it is extracting costs in the meantime — from customers who made decisions on inadequate information, from the broader public who will have to form their views of AI based on a track record that includes a significant number of overpromised and underdelivered deployments.

The founders who will define the next phase of AI are not the ones who can tell the most compelling story about what their systems might eventually do.

They are the ones who can tell the most accurate story about what their systems do right now — including where they fail, how they fail, and what it takes to deploy them responsibly.

This is a harder pitch. It requires the confidence to say “this is what our system cannot do” in rooms where admitting limitations has historically been treated as a weakness. It requires building organisations where the people responsible for identifying failure modes have genuine authority, not just a seat at the table. It requires treating customers as partners in managing risk rather than targets for a sale.

It also, in the medium term, is the only strategy that works.

The founders who have built institutional trust — who have a track record of accurate claims, disclosed limitations, and responsible deployment — are building a form of competitive advantage that their overclaiming competitors cannot easily replicate. Trust, once established, compounds. Credibility, once lost, is expensive to recover.

The founders who understood this early are already well positioned. The ones who are still optimising for the pitch are going to find the market an increasingly uncomfortable place.

The credibility crisis is real. The correction is underway. The question is which side of it you want to be on.

I have worked across global financial services, early-stage technology ventures, and international AI governance forums. I write on AI strategy, policy, and the communication gap between builders and institutions.

Originally published at https://www.linkedin.com.