I’ve Been In Tech 25 Years. What I’m Watching Right Now Is Different.
Scroll through LinkedIn on any given morning. Count how many posts claim their company “runs on AI now.” Then ask the follow-up question nobody asks: what does that actually mean?
In my experience, it usually means a GPT-4 wrapper, a Notion doc as the “knowledge base,” and a C-suite deck that says “AI-first strategy” in 48pt font. The plumbing is three API calls. The ambition is a complete organizational reinvention.
I’m calling this AI psychosis. And I think it’s spreading faster than anyone wants to admit.
The danger isn’t that companies are moving too slow on AI. It’s that entire organizations have lost the ability to distinguish between real capability and hallucinated capability — and they’re making multi-million dollar decisions accordingly.
If you’ve been in any planning meeting in the last 18 months where no one questioned the AI assumption — you may already be inside one of these companies.
What I Actually Mean by AI Psychosis
Let me be precise, because this isn’t just another “AI hype” take.
AI Psychosis is an organizational state where decisions, roadmaps, and identity are increasingly driven by AI capabilities that don’t exist yet, can’t be verified, or have been fundamentally misunderstood — and where dissent or skepticism is socially penalized.
This is categorically different from hype. Hype is temporary enthusiasm that fades when the next shiny thing arrives. Psychosis is structural. It reshapes behavior, reorganizes incentives, and rewires what counts as acceptable evidence inside an organization.
The defining feature: the organization needs the AI to work, so it stops checking whether it does.
Three diagnostic markers I look for as a practitioner:
Output Laundering. AI-generated work passes through meetings, reports, and decisions without verification — because questioning the output feels like questioning the strategy itself. The artifact becomes sacred because of what it represents, not what it contains.
Capability Inflation. Internal decks describe what the AI “will do” using present tense. The roadmap is built on a capability that exists in a demo environment, or in a vendor’s pitch, or in someone’s imagination after watching a YouTube video about agents.
Skepticism as Insubordination. This is the one that really concerns me. Engineers who raise questions about model reliability get quietly labeled “not AI-forward.” Analysts who ask “but how do we verify this?” get treated as blockers. The immune system gets removed.
I’ve felt a version of this in my own solo work. There’s a seductive moment when a demo works perfectly and you mentally skip the step where you ask “but will this work at 3am on a Tuesday with real edge-case data?” That’s a one-person moment I can recover from quickly. Imagine that same mental skip, scaled to 400 people all sharing the same hallucination, with quarterly OKRs attached to it.
The historical parallel that keeps coming to me is enterprise SOA adoption in the early 2000s. Companies built entire org structures around a Service-Oriented Architecture metaphor before the tooling, governance, or operational maturity existed to support it. It caused years of expensive mess.
But AI psychosis has a unique amplifier that SOA never had: the technology itself produces confident-sounding wrong answers. The model doesn’t say “I’m not sure.” It says “Based on your data, the recommended approach is…” — in perfect prose, with no visible uncertainty. That mirrors and reinforces the exact organizational behavior that makes psychosis dangerous.
The Data Is Starting to Leak Out
This isn’t just pattern-matching from LinkedIn. The numbers are beginning to tell the story.
Look at that gap. McKinsey’s 2024 Global Survey found 65% of companies now regularly use generative AI — yet more than 80% report no material enterprise-level EBIT impact from it. Gartner projects roughly 30% of gen AI projects will be abandoned after proof-of-concept by the end of 2025. Adoption is nearly universal; measurable value is not.
This gap isn’t incompetence. It’s what psychosis looks like in a spreadsheet.
Then there’s what I call the AI Wrapper Economy problem. By mid-2024, tens of thousands of startups had launched as “AI companies” — most with no proprietary model, no unique training data, no defensible moat whatsoever. A16z and Sequoia partners have both written frankly about this: the majority of the early generative AI startup wave was selling access to OpenAI’s API with a branded UI on top.
Here’s the closed loop that worries me: these startups are selling to enterprises that are themselves in AI psychosis. The buyer can’t evaluate the product critically because they’ve already decided AI is the answer. The vendor can’t survive if they’re honest about limitations. Everyone’s incentives align around maintaining the shared delusion.
The spend numbers make this vertigo-inducing. Global enterprise AI spend is projected to clear $200B by 2025 (IDC). Meanwhile, hallucination and reliability failures in production LLM deployments are still common enough that most engineering teams I know won’t let a model make a high-stakes call unsupervised. And independent productivity research is decidedly mixed — real gains on some task types, neutral or negative on others.
Companies are spending at infrastructure scale on technology that is, by any honest engineering standard, still unreliable for high-stakes decisions.
I built a small RAG pipeline for a personal research project last year. Getting it to work in a demo took about two hours. Getting it to work reliably — with citations I could actually trust, handling edge cases, without confidently wrong answers sneaking through — took three more weeks of iteration. That was just me, one problem, low stakes. I think about what that reliability curve looks like inside a 500-person org with quarterly targets and a VP who saw a demo in February.
What AI Psychosis Looks Like From the Inside
Here’s where I want to give you
The Roadmap Hostage. Product roadmaps built around AI features that aren’t scoped, aren’t tested, and aren’t proven — but removing them feels like “falling behind.” The observable tell: “AI-powered” appears in the feature name before anyone has decided what model, what data, or what acceptable latency even looks like.
The Meeting Hallucination. AI outputs — summaries, analysis, recommendations — circulate in meetings as source material. Nobody traces them back to original data. Someone asks “where did this come from?” and the answer is “the AI pulled it from our docs.” Full stop. Meeting moves on.
The Prompt as Architecture. Engineering decisions made at the prompt level. Entire system behaviors defined by a text string that nobody has version-controlled, stress-tested, or treated as a real software artifact with a change management process. I’ve been guilty of this in weekend projects. It’s fun until it isn’t. At enterprise scale, it’s a liability time bomb waiting for a bad Tuesday.
The Confidence Laundering Loop. This one is insidious. AI produces confident-sounding output → human presents it confidently → other humans accept it confidently → it becomes organizational “fact.” The model’s tone becomes the organization’s epistemic standard. Nobody in that chain added any verification. The confidence just… laundered itself upstream.
The Skeptic Tax. Engineers or analysts who raise questions about model reliability, data quality, or output verification get quietly sidelined. Not fired, usually. Just deprioritized. Left off the next meeting invite. Labeled as resistant to change. This is the most dangerous signature because it actively removes the people most capable of catching errors before they compound.
The Architecture of How It Spreads
Understanding why this spreads so fast matters as much as recognizing it.
The divergence in that chart is the story. Executive confidence in AI is climbing steeply. Verified, measurable ROI is barely moving. That gap is the psychosis spreading in real time.
The mechanism works like this:
graph TD
A[Compelling AI Demo] --> B[Executive Buys In]
B --> C[Strategy Deck Updated — AI-First]
C --> D[Roadmap Locked to AI Features]
D --> E[Engineers Begin Building]
E --> F{Reality Check?}
F -->|Skeptic raises concern| G[Labeled as Blocker / Resistant]
F -->|No challenge| H[Demo-Quality Output Ships]
G --> I[Skeptic Sidelined]
I --> H
H --> J[AI Output Enters Decision Loop]
J --> K[Confident Wrong Answers Become Org Facts]
K --> L[Strategy Doubles Down]
L --> C
style G fill:#ef4444,color:#fff
style K fill:#f59e0b,color:#fff
style L fill:#6366f1,color:#fff
It’s a closed loop. Every cycle makes the organization less capable of self-correction. The feedback mechanisms that would normally catch errors — engineering pushback, measurement rigor, honest retrospectives — get systematically disabled.
The demo is the original sin. AI demos are extraordinarily compelling. I’ve built demos of my own tools that looked far more capable than the production reality. That’s not dishonesty — it’s a structural feature of how LLMs work. Curated prompts, clean data, the right questions in the right order. It’s going to look magical.
But organizations are making multi-year strategic bets based on demo performance. That’s the psychosis accelerant. By the time production reality diverges from demo promise, the roadmap is committed, the press release is drafted, and the engineers who saw it coming have been reassigned.
Who’s Most at Risk
Not every company is equally vulnerable. The pattern I observe is that mid-market enterprises (500–5,000 employees) are the highest-risk cohort right now.
Large enterprises have enough governance infrastructure and enough cynical senior engineers that some skepticism survives. Early-stage startups iterate fast enough to hit reality quickly and pivot. The mid-market is where the danger concentrates: enough budget to make big bets, enough hierarchy to suppress dissent, not enough engineering rigor to catch the gap before it becomes structural.
Industries processing unstructured language at scale — legal, financial services, healthcare, consulting — are particularly exposed. These are exactly the domains where LLM hallucination rates are highest (domain-specific terminology, regulatory precision, factual accuracy requirements), and also where the demos look most compelling because the before-state (manual document review) is so obviously painful.
The vendor ecosystem isn’t helping. Every AI software vendor’s incentive structure rewards demos, not production reliability. Every consulting firm is staffed up on “AI transformation” engagements. Every conference speaker has a case study that shows the highlight reel. Nobody’s keynote is titled “Here’s How Our AI Initiative Failed Quietly Over 14 Months.”
What Breaks the Loop
I’m not writing this to be doom-and-gloom. I’m writing it because I think the loop is breakable — but it requires a specific kind of organizational courage.
Treat AI outputs as first drafts, not conclusions. This sounds simple. It almost never happens in practice because it slows things down and the demo made it look so fast.
Version control your prompts like you version control your code. If a text string is making decisions in your system, it deserves a PR, a review, and a blame history. I do this now on every project I build. It’s not glamorous. It’s how you stop the prompt-as-architecture problem before it bites you.
Protect your skeptics. The engineer who says “I’m not confident this is reliable enough to put in front of customers” is not a blocker. They are your most valuable person in the room. The organizational immune system only works if it’s allowed to do its job.
Measure from production, not demos. What’s the hallucination rate on your actual use case? What’s the failure mode? What happens when the model encounters something it wasn’t tuned for? These aren’t anti-AI questions. They’re the questions that let you actually use AI safely at scale.
Slow down the roadmap commitment. Specifically: don’t lock in an AI-dependent roadmap until you’ve run the feature in production conditions for at least 60 days. Not a pilot with selected users. Not a demo. Production conditions. Real edge cases. Real failure modes.
The Bottom Line
We’ve watched companies destroy themselves chasing SOA, microservices, blockchain, and the metaverse. Each cycle had true believers, sunk costs, and suppressed skeptics.
AI is genuinely more capable than all of those. That’s not the argument. The argument is this: a powerful tool used under a psychotic organizational relationship with it is still dangerous. Maybe more dangerous, because the capability makes the belief feel more justified.
The companies that come out of this cycle with real competitive advantage won’t be the ones who went “AI-first” the loudest in 2024. They’ll be the ones who stayed epistemically honest long enough to figure out what actually worked — and built on that, not on the demo.
The most expensive AI implementation isn’t the one that failed at launch. It’s the one that succeeded in the demo, convinced everyone it was working, and silently hallucinated its way into your strategy for two years before anyone was allowed to say so.
That’s AI psychosis. And the antidote isn’t skepticism of AI. It’s restoring the organization’s basic right to ask: does this actually work?