The Gospel of the AI Evangelists

Most people never sit in the room when an AI evangelist does their thing. What they see, instead, is the LinkedIn post, some sweeping declaration, and maybe a carousel with seven slides and a hook that reads like revelation. “AI is not coming for your job - it’s coming for your excuses.” “The companies that don’t adapt by 2026 will not exist by 2030.” “I just watched an AI do in four minutes what took my team four days.” or my favorite…”Look at this subpar animation I made with AI that looks strangely like a SIMs world but spookier.”

These posts get tens of thousands of likes, and sadly they get shared into corporate Slack channels by well-meaning people trying to stay informed. They get screenshotted and texted between colleagues who are trying to figure out whether they should be panicking, but the person who wrote them, that magician who is behind the scenes billing companies anywhere from $300 to $1,000 an hour as an AI transformation consultant. What people don’t realize is that the posts are the advertisement.

Surprise!

Let’s get to the skinny: an LLM is remarkable technology. It is genuinely useful and it has changed how millions of people write, research, and work through problems. None of that is in dispute. What is in dispute is the definition that a growing class of AI evangelists are now selling: that these systems are on the verge of general intelligence, that they truly “understand” in any meaningful sense, that the transformation they promise is as total and as imminent as the propaganda claims. That definition is not real, and before we go any further, it helps to understand exactly what we’re actually talking about.

What an LLM Actually Is

An LLM, or Large Language Model, is a type of artificial intelligence designed to understand, process, and generate human-like text. It is the technology powering conversational AI tools, translation software, and automated writing assistants. Instead of operating on rigid, pre-programmed rules, an LLM relies on advanced mathematics and pattern recognition. Understanding how it works (really works) is the single best defense against the people who are profiting from your confusion about it.

The process happens in three stages.

Tokenization Before an LLM can read text, it breaks sentences down into smaller pieces called tokens. A token can be a whole word, a part of a word, or even a single punctuation mark. The word “unbelievable” might be split into “un,” “believ,” and “able.” Those tokens are then converted into numbers, which is the form the computer actually processes. The model never reads language the way you do. It reads sequences of numerical values.

Training An LLM is trained on a massive dataset of books, articles, websites, code. During this phase, the model uses an architecture called a transformer, which employs a mechanism called attention, or self-attention. This allows the model to look at an entire sentence at once and figure out how words relate to each other, even when they’re far apart in the text. In the sentence “The bank of the river was muddy,” the model uses the context word “river” to determine that “bank” means the edge of a body of water, not a financial institution. Through billions of these comparisons, the model builds a mathematical map of language and learns grammar, facts, reasoning shortcuts, and subtle nuances in tone.

Prediction Here is the part that matters most, and the part the evangelists most reliably gloss over. When you give an LLM a prompt, it doesn’t “think.” It doesn’t understand concepts in any human sense. It runs a massive calculation to answer one question: Based on everything written so far, what is the most statistically likely next token? It predicts one token, adds it to the sequence, then repeats the process to predict the next one. It does this at extraordinary speed, assembling fluent sentences and entire essays one piece at a time. The result looks like thought, but is really just very sophisticated pattern completion.

The capabilities of an LLM are largely determined by its scale, measured in parameters and internal mathematical weights the model adjusts during training to learn patterns. Smaller models, in the millions of parameters, handle fast, specific tasks like sorting emails or basic text classification. Large models, in the billions of parameters, are capable of complex reasoning, coding, translation, and natural conversational flow. Because they train on such vast amounts of diverse text, modern LLMs can perform what’s called zero-shot learning or the handling of tasks they were never explicitly programmed for, simply by applying the underlying patterns of logic and language they’ve already absorbed.

That is genuinely impressive. It is also genuinely different from what you are being sold.

The Gap

Now, hold that definition in your head and re-read the average AI LinkedIn post. The model predicts the next token. It doesn’t have vision or goals and it’s not “thinking about your industry.” It has no understanding of your business, your customers, or your competitive landscape. It only has a statistical approximation of language patterns that can sound like it does. That gap, between what the technology actually is and what the influencer ecosystem claims it to be, is where enormous amounts of money are currently flowing.

The research community has been saying this clearly for years. A 2024 peer-reviewed paper in MIT Press’s Open Mind journal, by researchers at Newcastle University and Davidson College, concluded that any similarities between human language and LLM output are “purely functional.” In other words, the surface resemblance is real, but what’s happening underneath is categorically different. These systems lack anything resembling the grounded, embodied understanding that underlies actual human cognition. A separate 2025 analysis made the point even more directly: LLMs don’t learn meaning, they only learn statistical relationships between tokens. The extraordinary fluency has “led many to believe these models know what they talk about” when they structurally can’t.

Cognitive scientist Gary Marcus, who has documented this gap for years, describes what he calls “the AGI shell game.” This is where companies are repeatedly hyping the imminent arrival of artificial general intelligence, and then quietly redefining the goalposts when the benchmarks aren’t met. His core argument, that the industry has “systematically overpromised” and that “the underlying technology has fundamental limitations that scaling alone will not fix,” is increasingly hard to refute. MIT Technology Review reported this year that even the engineers who build these systems openly acknowledge they cannot fully grasp what’s happening inside them. The people who built the thing don’t fully understand it. Yet, the people selling it to your company on a monthly retainer are entirely certain about its future.

Something’s not stirring the Kool-Aid if you ask me.

What the LinkedIn posts also never mention is the hallucination problem, which, depending on the domain and the task, ranges from embarrassing to genuinely dangerous. Enterprise benchmarks report hallucination rates between 15% and 52% across commercial LLMs. In legal contexts, studies have found rates between 69% and 88% in high-stakes queries. More than 120 cases of AI-driven legal hallucinations have been documented since mid-2023, with at least 58 occurring in 2025 alone. MIT’s 2025 research found that 95% of corporate AI pilots fail to scale to actual production. These are the technology operating as designed, in the real world, at scale, and they represent the enormous gap between “impressive demo” and “reliable production environment” that the evangelists have every financial incentive to never post about.

What It Actually Costs

The money involved is worth talking about plainly, because it explains a great deal. The top AI speakers on the keynote circuit command fees that range from $40,000 to well over $100,000 per appearance. Many have online courses that sell for hundreds of dollars, and some have built consulting practices worth millions annually, advising boards of directors who are understandably panicked about being disrupted. The financial incentives to keep the hype elevated and to describe AI as more capable, more imminent, more transformational than it currently is enormous and relentless.

The AI consulting market hit $8.75 billion in 2024 and is projected to reach nearly $50 billion by the early 2030s. Senior AI consultants are billing at $300 to $500 an hour at specialized firms, with top-tier advisors commanding $900 an hour or more. Retainer arrangements run from $5,000 to $50,000 a month. The person posting confidently about AI’s limitless potential on your LinkedIn feed is, in many cases, simultaneously selling access to that confidence to large companies at rates that would make your eyes water. Those aren’t thought leadership posts. They’re prospecting for gold.

The brutal economics underneath all of it: a more accurate, more nuanced description of what LLMs actually are doesn’t build a following. “This is a powerful tool that will require careful integration, realistic expectation-setting, and a serious investment in change management over the next several years” isn’t viral worthy. “AI will replace your entire back-office function within 18 months” is because our brains are designed to look for threats in the distance. So, guess which version gets posted.

This is also exactly why I won’t call it a lapse in judgment. I’ll call it what it is: a choice, made over and over again, that compounds into something we should all be a lot less comfortable with than we are.

The Problem With Calling These People Thought Leaders

Here is where I want to slow down, because this is the part that bothers me most.

We have a word for people who shape how large groups think about important things. We call them thought leaders, and we say it like it’s a compliment. We put it in bios, repeat it in introductions, and use it to signal that a person has earned the right to be listened to at scale and that their ideas have been tested against reality and found trustworthy. That is a significant amount of social authority to hand someone and right now, we are handing it to people based almost entirely on follower counts, engagement metrics, and the confidence with which they can describe a technology most of their audience doesn’t fully understand.

That should concern us more than it does.

Thought leadership, if the phrase is going to mean anything at all, implies a standard. It implies that the person leading your thinking has an obligation to that process - an obligation to accuracy, to intellectual honesty, to telling you what they don’t know as clearly as they tell you what they do.

It implies, in short, integrity.

The moment someone begins charging $900 an hour to advise corporations on a technology they are simultaneously overstating to a public audience, they have a conflict of interest with the truth.

The current AI influencer ecosystem asks almost nothing of the people it elevates. It (very importantly) doesn’t require that they disclose their financial relationships with the companies whose tools they are enthusiastically recommending. It also doesn’t require that they correct the record when their predictions fail to materialize - and they do fail, repeatedly, without consequence. What it does requires is that they post consistently, that they sound certain, and that they make their audience feel like insiders. That is the entire standard and we have collectively decided that is enough to confer the title of thought leader and the authority that comes with it.

It isn’t enough. It has never been enough.

When someone with zero formal training in linguistics begins policing how language does and does not work, that is performance or theatre or, worse, hubris. When someone dismisses the researchers and engineers who have spent careers studying these systems as simply “not getting it,” that is a defense mechanism protecting a revenue stream. When the people who are positioned as the trusted guides for navigating an important technology turn out to be, above all else, salespeople who never disclose what they’re selling, we have a systemic integrity failure.

We built an entire credibility ecosystem on vibes and engagement, and then we acted surprised when the people it elevated turned out to be optimized for virality rather than truth. That is on us as much as it is on them.

The One Question Worth Asking

The next time a LinkedIn post about AI fills you with either excitement or dread, go to that person’s profile and find out what they sell. Not what they claim to believe. Find out what they charge for it, and find out if they’re a paid corporate consultant, a course creator, or some combination of both. Then re-read the post with that information in your hand.

If you're wondering whether you're absorbing propaganda rather than genuine insight, here's the most reliable diagnostic available to you: ask the speaker what they charge for a keynote. The answer will tell you everything about the incentive structure shaping every word you've just heard. People who are paid $75,000 to inspire are not, structurally speaking, well-positioned to tell you what AI can't do. There is a name for it and it is called a conflict of interest.

You don’t have to become a cynic, but you do have to understand that the incentive to keep the hype elevated and to make AI sound more inevitable, more revolutionary, more total in its disruption than the evidence currently supports, is not abstract. It is measured in billable hours and monthly retainers and course enrollments. People who earn their income from corporate fear of being left behind are not, structurally speaking, well-positioned to tell you what a next-token predictor cannot do.

It’s not necessarily that they’re lying, but the version of AI that pays them is not the version that actually exists yet.

The next time a post makes AI feel like prophecy, check who’s profiting from the sermon. Likewise, the next time you see the major players trying to lessen the value of people who are either in subordinate positions to them (*cough, cough* students) or they are trying say LLMs are magic juice and anyone who disagrees is wrong, OR (my favorite) someone who has zero language expertise policing language, question it.

If this resonated, send it to someone who forwarded you AI propaganda this week.