AI's Reality Check: When the Promise Hit the Wall
Table of Contents
- The Trust Gap
- The Pew Study: America Has Made Up Its Mind
- KPMG and the Hallucination Problem
- When AI Meets Law Enforcement
- AI Is Making Healthcare More Expensive
- Microsoft, GitHub, and the Infrastructure Ceiling
- The Pentagon's AI Army
- Noam Shazeer and the Talent War
- The Anthropic Saga Deepens
- Open Source AI Makes Its Case
- The Cost of the Gap
- What Comes Next
1. The Trust Gap
The distance between what AI companies promise and what AI delivers has never been wider. This week, it became impossible to ignore.
For the better part of three years, the AI industry has operated on a simple proposition: trust us. Trust us with your data, your workflows, your creative processes, your healthcare, your legal system, your national security. Trust that the technology will get better, the costs will come down, the hallucinations will fade, the safeguards will hold. Trust that the people building this know what they're doing.
This week, that trust was tested from every direction.
The tests came not from AI critics or doomsday prophets, but from the technology itself -- behaving exactly as it currently behaves, not as its makers promise it will someday behave. KPMG, one of the world's four largest accounting and consulting firms, published a report about AI adoption that turned out to contain AI-generated hallucinations. The report was pulled. A police officer in the UK is under investigation for using AI to fabricate evidence in multiple criminal cases. A PwC report found that AI is increasing medical bills rather than reducing them. Microsoft had to turn to AWS -- its biggest cloud rival -- to keep GitHub running because Azure couldn't handle the AI-driven load. And a Pew Research study found that only 16% of Americans think AI will have a positive impact on society.
Each of these stories is significant on its own. Together, they form a pattern that AI's most enthusiastic promoters need to reckon with. The technology works well enough to be dangerous -- dangerous to trust, dangerous to deploy without oversight, dangerous to assume will behave the way its demos suggest. It also works well enough to be genuinely useful, which is what makes the situation so complicated. The same AI that can help a developer write code faster can hallucinate facts in a consulting report. The same AI that can help analyze medical imaging can add complexity to billing systems that raises costs. The same AI that can help police analyze evidence can be used to manufacture it.
This is not a story about AI failing. It's a story about AI succeeding just enough to be deployed everywhere, but not enough to be trusted anywhere. And the gap between those two points is where all the interesting -- and dangerous -- things are happening.
The industry's response to this gap has been consistent: more AI will fix it. Better models, bigger context windows, more sophisticated guardrails, more compute. The hallucinations will diminish. The costs will come down. The public will come around. Maybe they're right. But this week provided a lot of evidence that the gap is widening, not narrowing -- and that the institutions most aggressively promoting AI adoption are the ones least prepared to deal with its current limitations.
The deep dive that follows examines each of these stories in detail, traces the connections between them, and tries to answer the question that matters: what does it actually mean when the promise hits the wall?
2. The Pew Study: America Has Made Up Its Mind
Only 16% of Americans think AI will have a positive impact on society. Wall Street is spending hundreds of billions. The gap between investor enthusiasm and public sentiment is the most important number in AI right now.
The Pew Research Center published its latest study on American attitudes toward AI on June 17, and the headline number is stark: 16%. That's the percentage of Americans who believe AI will have a positive impact on society. Not 16% who are cautiously optimistic. Not 16% who think it might help in some areas. 16% who think it will be a net positive for society as a whole.
To put that in perspective, more Americans believe in ghosts than believe AI will improve their world. More Americans trust Congress -- an institution with approval ratings that would make a cockroach blush -- than trust the AI industry with their future. The number of Americans who think AI will be positive for society is lower than the number who think the economy is on the right track, which is itself historically low.
The Demographics of Distrust
The Pew data breaks down along interesting lines. Younger Americans -- those under 30, who are supposed to be the most tech-savvy and AI-comfortable generation -- are only slightly more optimistic than older Americans. Men are marginally more positive than women, but the gender gap is small. Income and education levels correlate with slightly more optimism, but even among college-educated Americans with six-figure incomes, the percentage who think AI will be positive for society barely cracks 25%.
The most striking finding is how few Americans have actually used AI tools. Despite ChatGPT being the fastest-growing consumer application in history, Pew found that a majority of Americans have never used a generative AI tool for work or personal use. Among those who have, usage skews heavily toward simple tasks -- writing emails, summarizing documents, basic search -- rather than the complex, agentic workflows that AI companies are betting their futures on.
This creates a paradox for the industry. The people who have used AI most extensively tend to have more nuanced views -- they see the potential and the limitations. But most Americans haven't used AI meaningfully, and their skepticism is based not on experience but on observation. They're watching AI-generated content flood the internet. They're watching AI replace human customer service agents. They're watching AI make mistakes in public, from Google's Gemini suggesting people put glue on pizza to KPMG's hallucinated report. They've decided they don't like what they see.
The regional breakdown is also telling. Americans in rural areas -- where AI-driven job displacement is more visible and AI-driven benefits are less accessible -- are significantly more skeptical than urban Americans. This tracks with the broader pattern of technology adoption: the people who benefit first and most from new technologies are concentrated in urban, educated, high-income demographics. The people who bear the costs -- job displacement, algorithmic bias, reduced access to human service -- are everywhere else. The Pew study is measuring this asymmetry.
The Comparison to Other Technologies
It's worth comparing AI's public sentiment to other transformative technologies at similar stages of adoption. In the early 2000s, internet optimism was widespread -- the promise of information democratization, global connection, and economic opportunity resonated across demographics. Smartphone adoption in the 2010s was similarly welcomed, with clear utility that consumers could understand and experience directly.
AI's situation is different. The benefits of AI are mostly invisible to consumers -- faster backend processing, better ad targeting, more efficient code writing. The costs are highly visible -- AI-generated misinformation, AI-replaced jobs, AI chatbots that can't solve problems. This visibility asymmetry means that even if AI is creating net positive value (which is debatable), the public perceives it as net negative because they see the costs and don't see the benefits.
Social media went through a similar arc. In 2010, social media was celebrated as a democratizing force. By 2020, it was widely viewed as a threat to democracy, mental health, and public discourse. The technology didn't change -- the public's understanding of its effects did. AI is starting from a lower baseline of public trust than social media did, and it's declining faster. Without intervention, the trajectory points toward the same kind of backlash that social media faced, but with higher stakes because AI is being embedded in more critical systems.
Why Wall Street Doesn't Care
The stock market's response to the Pew study was essentially nothing. AI-related stocks -- Nvidia, Microsoft, Google, Amazon -- didn't budge. The investors pouring capital into AI infrastructure, AI startups, and AI integration aren't betting on public sentiment. They're betting on enterprise adoption, productivity gains, and the fundamental economics of compute.
There's a logic to this. Enterprise AI adoption doesn't require public enthusiasm. Companies adopt AI because it reduces costs, because competitors are adopting it, because their board demands an "AI strategy." The decision-makers are CTOs and CFOs, not ordinary consumers. If AI can save a Fortune 500 company 10% on customer service costs, it gets deployed whether the public likes it or not.
But this logic has a ceiling. Consumer-facing companies -- retailers, media companies, healthcare providers, banks -- need their customers to trust the AI systems they deploy. If 84% of Americans are skeptical of AI, a bank that replaces human customer service with an AI chatbot risks losing customers. A healthcare provider that uses AI for diagnostic recommendations risks scaring patients away. The public distrust that Pew documented isn't just a sentiment problem. It's a deployment problem. It limits where and how AI can be rolled out, even when the economics make sense.
The Political Implications
Politicians read Pew data. When 84% of your constituents are skeptical of something, regulating it becomes politically easy. Senator Bernie Sanders' $7 trillion AI sovereign wealth fund proposal -- which would tax AI companies at 50% and redistribute the proceeds to Americans -- polls well with the public precisely because the public already views AI companies with suspicion. Other politicians are noticing. The state attorneys general investigation into OpenAI, the ongoing export controls controversy, the Pentagon's supply chain blacklisting of Anthropic -- all of these are political responses to an industry that the public doesn't trust.
The AI industry has historically treated regulation as a problem to be managed, not a signal to be heeded. When your approval rating is 16%, you don't have the political capital to fight regulation. You have the political vulnerability to be regulated into oblivion. The industry needs public trust more than it needs better benchmarks, and this week made clear that trust is eroding, not building.
What the Industry Gets Wrong
AI companies respond to distrust with education. "People just don't understand the technology," the thinking goes. "If we show them the benefits, they'll come around." This is the same mistake every disruptive technology industry has made, from genetically modified foods to social media. The problem isn't understanding. It's observation. People can see what AI does. They can see the AI-generated slop flooding search results. They can see the chatbots that can't solve their problems. They can see the news stories about AI mistakes, AI bias, AI job displacement.
Education doesn't fix observation. Only better behavior does. If the AI industry wants to improve public sentiment, it needs to stop shipping products that hallucinate, stop deploying chatbots that frustrate users, and stop promising things the technology can't deliver. The 16% number isn't an information problem. It's a product problem.
3. KPMG and the Hallucination Problem
The world's fourth-largest accounting firm published a report on AI adoption. The report contained AI-generated hallucinations. KPMG pulled it. The irony writes itself.
On June 13, TechCrunch reported that KPMG had pulled a report on AI usage due to what the firm described as "apparent hallucinations." The report, which was meant to analyze how businesses are adopting AI tools, was withdrawn after readers noticed content that appeared to be fabricated by the AI systems used in its production.
The details are sparse -- KPMG hasn't disclosed exactly what was hallucinated, which AI tool was used, or how the errors survived whatever review process the firm had in place. But the fact pattern alone is devastating. This is a company that sells AI consulting services to other businesses. It helps its clients implement AI. It publishes research about AI best practices. And it couldn't use AI to write a report about AI without the AI making things up.
The Hallucination Problem, Explained
For those who haven't been following the technical details, hallucinations are the AI industry's term for when language models generate confident, fluent, and completely false information. The model doesn't know it's wrong. It doesn't experience doubt. It produces text that reads authoritatively and happens to be incorrect, because the model's fundamental mechanism is pattern prediction, not fact retrieval.
This isn't a bug. It's a feature of how large language models work. They predict the next most likely token based on patterns in their training data. Sometimes the most likely next token is factually correct. Sometimes it isn't. The model doesn't know the difference. And despite years of research, safety classifiers, RLHF (reinforcement learning from human feedback), and claims of improvement from every major AI lab, hallucinations remain an unsolved problem.
Every frontier model hallucinates. GPT-5.5 hallucinates. Claude hallucinates. Gemini hallucinates. The rates have decreased, but the severity hasn't. A model that hallucinates 5% of the time instead of 20% is still a model that will fabricate information in roughly 1 out of every 20 queries. If you're using it for creative writing, that's fine. If you're using it for a consulting report, it's a disaster.
The technical community has proposed various solutions: retrieval-augmented generation (RAG) to ground models in verified sources, multi-step verification pipelines, citation requirements, and fact-checking layers. Each of these helps in specific contexts but none solves the fundamental problem. A model that can hallucinate will hallucinate, regardless of the guardrails around it. The guardrails reduce frequency; they don't eliminate the failure mode. And every guardrail adds latency, cost, and complexity, which creates a practical tension between reliability and the seamless experience that AI companies want to sell.
The industry has been quietly managing this problem by positioning hallucination as a known limitation that users must work around. "Always verify AI output" is the standard disclaimer. But as the KPMG case demonstrates, organizations aren't always good at following this guidance, especially when the AI output looks authoritative and the reviewers are stretched thin. The disclaimer is a legal protection, not a practical solution.
Why KPMG's Failure Matters More Than It Seems
KPMG is not a startup. It's not a small company experimenting with AI. It's one of the Big Four accounting firms -- Deloitte, PwC, EY, and KPMG -- that collectively audit the majority of the world's public companies. These firms are the backbone of global financial trust. They're supposed to be the gold standard for accuracy, diligence, and professional skepticism.
If KPMG can't use AI reliably for an internal report -- a low-stakes document that no one's financial decisions depend on -- what happens when they use AI for client work? What happens when AI is involved in audit procedures, risk assessments, or compliance reviews? The firm reportedly has thousands of employees using AI tools daily. How many of those uses involve hallucinated information that hasn't been caught?
The standard defense is "human review." AI generates the content, humans review it. But KPMG's report presumably went through some form of human review before publication, and the hallucinations survived. This is the pattern that makes hallucinations so insidious: they read as plausible. If the hallucinated content were obviously wrong, it would be caught in review. The problem is that AI-generated falsehoods are calibrated to sound exactly like AI-generated truths. Same tone, same structure, same confidence. The reviewer can't tell the difference because the difference isn't visible in the text.
The Trust Cascade
KPMG's failure creates a trust cascade. If a Big Four firm can't use AI safely, why should any company trust AI consultants who recommend adoption? The consulting industry has been one of the biggest promoters of AI adoption -- they make enormous revenues helping companies implement AI systems. If the consultants themselves can't use the technology reliably, their advice is compromised at the root.
This isn't just about KPMG. Every major consulting firm is aggressively integrating AI into its workflows. Deloitte has its own AI platform. PwC announced a $1 billion AI investment. EY has AI woven through its service lines. The question none of them can answer convincingly is: if you can't use AI without it making things up, why are you telling your clients to use it?
The honest answer -- which none of them will give -- is that AI is useful for some tasks and dangerous for others, and the challenge is knowing which is which. That distinction requires expertise, testing, and oversight. It requires admitting that AI isn't ready for everything, which is a hard admission to make when your business model depends on selling AI to everyone.
The Deeper Problem: AI Evaluating AI
One of the most troubling aspects of the KPMG incident is what it suggests about AI quality control. In many organizations, AI systems are now being used to review AI-generated content. An AI writes a draft, another AI checks it for errors. This creates a closed loop where the same class of errors that produce hallucinations can also fail to detect them.
If the reviewer AI has the same blind spots as the writer AI -- which it likely does, since they're often built on similar or identical underlying models -- the errors propagate. The system reports that everything looks good because the system doesn't know what it doesn't know. Human reviewers, overwhelmed by the volume of AI-generated content and primed to trust the tools they've been told to use, miss what the AI missed.
This is not a theoretical concern. It's happening right now in newsrooms, law firms, healthcare systems, and consulting companies. The KPMG incident is just the one that got caught. How many hallucinated reports, analyses, and recommendations are currently sitting in client deliverables, board presentations, and regulatory filings, undiscovered because no one has caught them yet?
4. When AI Meets Law Enforcement
A UK police officer is under investigation for using AI to create evidence in multiple criminal cases. The implications go far beyond one bad actor.
On June 14, Sky News reported that a police officer from Derbyshire Constabulary is under formal investigation for using AI to "create evidence" in multiple criminal cases. The details of the investigation are limited, but the core allegation is extraordinary: a law enforcement officer used artificial intelligence to generate material that was presented as evidence in criminal proceedings.
This is not a hypothetical risk anymore. It's a case study in what happens when powerful AI tools are deployed in high-stakes environments without clear rules, oversight, or accountability frameworks.
The Legal and Ethical Implications
Evidence in criminal cases is supposed to meet specific standards. It must be authentic, verifiable, and obtained through legal processes. The chain of custody must be clear. The methods used to collect and analyze evidence must be transparent and subject to challenge by the defense. AI-generated material meets none of these standards.
When an AI model generates text, an image, or an analysis, it's creating something new based on patterns in its training data. It's not retrieving a fact from a database. It's not capturing reality. It's producing a plausible-sounding output that may or may not correspond to anything real. In a criminal case, the difference between "this actually happened" and "this is what the AI thinks probably happened" is the difference between justice and its opposite.
The officer in question allegedly used AI to create material that was treated as evidence. We don't yet know whether this means generating documents, synthesizing witness statements, creating images, or producing analytical reports. Whatever the specifics, the fundamental problem is the same: AI-generated content was introduced into a system designed to evaluate facts, and the system had no mechanism for distinguishing AI output from reality.
The Regulatory Vacuum
The UK doesn't have specific laws governing AI use in law enforcement. Neither does the United States. Police departments are adopting AI tools -- for predictive policing, evidence analysis, report writing, facial recognition, and more -- without a coherent regulatory framework. Individual departments set their own policies, if they set any at all.
This is the same problem we see across every AI deployment domain: the technology is deployed first, and the rules come later. Sometimes much later, and only after something goes wrong. The Derbyshire case is the "something goes wrong" moment for law enforcement AI. But it's unlikely to be the last. Police departments across the UK and US are using AI tools right now, and without clear rules, the line between using AI as a tool and using AI to create evidence is blurry enough that officers may not always know when they've crossed it.
The Fabrication Risk
The specific risk of AI fabrication in law enforcement is uniquely dangerous because of the stakes. A hallucinated consulting report costs a firm embarrassment. A hallucinated medical diagnosis costs a patient time and anxiety. A hallucinated piece of evidence in a criminal case can cost someone years of their life.
Consider the scenario: an officer uses an AI tool to analyze a crime scene and generate a report. The AI, doing what AI does, fills in gaps in the evidence with plausible-sounding inferences. Some of those inferences are correct. Some aren't. The report reads as authoritative. The officer, trusting the tool, presents it as fact. The defense, not knowing AI was involved, doesn't challenge the methodology. The jury, seeing a confident professional report, convicts.
This scenario requires only that the officer doesn't fully understand what the AI tool is doing -- which, given the black-box nature of most AI systems, is the default state. The officer doesn't need to be malicious. They just need to trust the tool more than they should, which is exactly what the AI industry encourages.
What Needs to Change
The Derbyshire investigation should be a catalyst for clear rules. At minimum:
- AI-generated content should never be presented as evidence without explicit disclosure that AI was used in its creation
- Police departments should be required to log all AI tool usage in criminal investigations
- Defense attorneys should have the right to know when and how AI was used in building the case against their client
- Independent auditors should review AI tools used in law enforcement for accuracy, bias, and fabrication risk
- Officers should receive specific training on the limitations of AI, including the risk of hallucinations and the difference between AI analysis and AI generation
None of these rules currently exist in any comprehensive form. The AI industry, for its part, has been happy to sell tools to law enforcement without insisting on these safeguards. The Derbyshire case shows why that approach has consequences.
5. AI Is Making Healthcare More Expensive
The promise was that AI would reduce medical costs. A PwC report this week says the opposite is happening. When the consultants admit the technology is raising costs, something has changed.
On June 12, Fortune reported on a PwC study finding that AI is making medical bills higher, not lower. This directly contradicts the healthcare AI industry's central selling point. Every AI healthcare startup, every hospital system buying AI tools, every investor funding AI health companies has been operating on the assumption that AI will reduce costs by streamlining administration, improving diagnostics, and enabling more efficient care delivery.
PwC's data suggests the opposite is happening, at least so far. And PwC is not an AI skeptic. The firm has invested billions in AI capabilities and actively promotes AI adoption to its clients. When the company selling the transformation admits it's making things more expensive, that's a signal worth taking seriously.
Why AI Is Increasing Costs
The PwC report identifies several mechanisms by which AI is driving up medical costs. Understanding these mechanisms matters because they reveal a pattern that extends far beyond healthcare.
Implementation costs: Hospitals and healthcare systems are spending heavily on AI tools, infrastructure, training, and integration. These costs are being passed through to patients in the form of higher bills. The AI systems themselves are expensive -- enterprise AI licenses can run millions of dollars per year for a large hospital system -- and the integration work, which requires specialized consultants and IT teams, adds another layer of cost on top.
Complexity, not simplification: AI tools were supposed to simplify healthcare administration. In practice, they've added complexity. A hospital that used to have a simple billing system now has a billing system plus an AI layer that needs maintenance, monitoring, and oversight. The AI doesn't replace the billing department -- it adds to it. More systems, more vendors, more points of failure, more people needed to manage the technology.
False economies: Some AI implementations do save time on specific tasks -- prior authorization processing, claims adjudication, coding. But the savings are often consumed by the costs of the AI system itself, the new error categories it introduces (AI-generated coding errors that require human correction), and the compliance overhead of using AI in a regulated industry.
The volume effect: When AI makes a process cheaper per unit, organizations often increase the volume of that process. If AI makes it easier to run diagnostic tests, more tests get ordered. If AI makes it easier to generate billing codes, more codes get generated. The per-unit cost goes down, but the total cost goes up because volume increases faster than efficiency.
The Broader Pattern
The healthcare cost story illustrates a pattern that AI's promoters consistently miss. They calculate the benefit of AI as: current cost minus AI-driven efficiency savings equals lower cost. But this calculation ignores three things.
First, the cost of the AI itself. The compute, the licensing, the integration, the training, the ongoing maintenance. These are not trivial. For many organizations, the total cost of AI ownership exceeds the savings it generates, at least in the first few years of adoption.
Second, the cost of AI errors. When AI makes mistakes -- and it does, at rates that vary by task but are never zero -- those mistakes have to be caught and corrected. The correction process requires human time, which costs money. In healthcare, AI errors that reach patients can generate malpractice liability, which costs even more.
Third, the cost of complexity. Every AI system added to an existing workflow adds a new layer of potential failure. More integration points, more vendor dependencies, more training requirements, more things that can break. The administrative overhead of managing AI systems can exceed the labor savings they provide.
This pattern isn't unique to healthcare. It shows up in every industry where AI has been deployed at scale. Customer service AI that reduces call center costs but increases the volume and duration of escalated calls. Legal AI that speeds document review but requires additional attorney time to verify accuracy. Coding AI that helps developers write code faster but introduces subtle bugs that take longer to find.
The Promise vs. the Reality
The AI healthcare story is really about the gap between theoretical efficiency and practical implementation. In theory, AI should reduce healthcare costs. The technology can process medical records faster than humans, identify patterns in imaging data, automate routine coding tasks, and streamline prior authorization. In practice, the technology is being layered onto a healthcare system that was already the most expensive in the world, and it's adding to that expense rather than reducing it.
This gap exists because AI is being deployed as an addition to existing systems rather than a replacement. Hospitals don't fire their billing departments when they buy AI billing tools. They keep the billing department and add an AI layer on top. The billing department now manages the AI system, reviews its outputs, and handles the cases the AI can't manage. The total headcount doesn't decrease, and the technology costs add to the total.
For AI to actually reduce healthcare costs, it would need to replace existing processes entirely, not augment them. But that's risky -- if the AI fails, there's no human fallback. So organizations keep the humans and add the AI, and costs go up.
This is the reality check. The promise of AI-driven cost reduction assumes a level of reliability and autonomy that the technology hasn't achieved. Until it does, AI will add costs before it reduces them. The question is how long that transition takes, and how many organizations can afford the transition period.
6. Microsoft, GitHub, and the Infrastructure Ceiling
The company that bet its future on AI can't keep its own developer platform running without help from its biggest competitor. The infrastructure crunch is real, and it's arriving faster than anyone expected.
On June 16, RuntimeWire reported that Microsoft is adding Amazon Web Services capacity to keep GitHub running after an AI-driven surge in coding activity strained the platform's infrastructure. According to the report, which cited two people familiar with the plans, the AI coding boom -- developers using GitHub Copilot, agentic coding tools, and AI-assisted workflows -- has pushed GitHub's infrastructure requirements beyond what Azure can currently provide.
This is a story about infrastructure, but it's really a story about the gap between AI's promise and its physical reality. AI doesn't run on promises. It runs on compute, and compute runs on data centers, and data centers have limits.
The Irony Is the Point
Microsoft bought GitHub in 2018 for $7.5 billion. The strategic logic was straightforward: GitHub is where the world's code lives. Microsoft makes the tools developers use. Bring them together, integrate GitHub with Azure, and Microsoft's cloud becomes the default substrate for the world's software development.
The plan worked, in a sense. GitHub grew. Copilot became the dominant AI coding assistant. Azure usage increased. But the plan also created a dependency: GitHub's infrastructure needs to scale with its usage, and that usage is being driven by AI -- the same technology Microsoft has bet its entire corporate strategy on.
The problem is that AI-driven coding generates enormous infrastructure demands. Every Copilot query, every agentic coding session, every AI-assisted pull request requires compute. Microsoft's Azure infrastructure was built for traditional cloud workloads -- web apps, databases, storage. AI workloads require different infrastructure: GPU clusters, high-bandwidth networking, specialized cooling. Microsoft has been building this infrastructure as fast as it can, but it can't build fast enough to keep up with the demand its own products are creating.
So GitHub -- a Microsoft product -- is being partially hosted on AWS -- a competitor's product -- because Microsoft's cloud can't handle the load. The company that wants to own the AI revolution is relying on its biggest rival to keep its own platform running.
The Infrastructure Ceiling
This is not just a Microsoft problem. Every major AI company is hitting infrastructure limits. Google has reported GPU shortages. Amazon's AWS has had capacity constraints in multiple regions. Anthropic's models have been rate-limited during peak usage. OpenAI has struggled with availability for its most advanced models.
The underlying issue is simple: AI compute demand is growing faster than data center capacity. Building a data center takes 18-24 months. Scaling AI usage takes minutes. When a new AI tool goes viral -- as Copilot did, as ChatGPT did, as every major AI release does -- the demand spike is immediate. The infrastructure to serve that demand takes years to build.
This creates a structural problem for the AI industry. The economics work when usage scales smoothly and infrastructure grows ahead of demand. They break when demand spikes faster than infrastructure can expand. Companies either turn away users (bad for business), degrade performance (bad for reputation), or buy capacity from competitors (bad for margins and strategic independence).
Microsoft choosing the third option -- buying AWS capacity for GitHub -- is a signal that the infrastructure crunch has reached the highest levels. If the second-largest cloud provider in the world can't keep up with AI demand for its own products, what does that mean for smaller providers? What does it mean for startups trying to build AI products on cloud infrastructure?
The Cost of Compute
The infrastructure crunch also has financial implications that the AI industry has been downplaying. AI is not cheap. The compute required for inference -- running AI models to generate responses -- is substantial, and it scales linearly with usage. Every additional user, every additional query, every additional agent adds to the compute bill.
This is fundamentally different from traditional software economics, where marginal costs approach zero. A traditional SaaS application can serve a million users with roughly the same infrastructure as 100,000 users, because the per-user compute requirement is tiny. AI applications don't have this property. Each user's queries require real GPU time, and GPU time is expensive.
The result is that AI companies are burning through capital at rates that would be unsustainable for any other technology. OpenAI reportedly spends more on compute than it generates in revenue. Anthropic's costs are similarly enormous. Google and Microsoft subsidize their AI products with profits from other businesses. The unit economics of AI, at current compute prices, don't work for most applications.
The infrastructure ceiling is the physical manifestation of this economic problem. You can't scale AI usage without scaling compute, and you can't scale compute without building data centers, and you can't build data centers fast enough. Until compute becomes dramatically cheaper -- through better chips, more efficient models, or architectural breakthroughs -- the AI industry is constrained by physical reality.
What This Means for the Industry
The GitHub-AWS story is a canary in the coal mine. It tells us that AI's growth is outpacing the infrastructure that supports it, and that the gap is wide enough to force uncomfortable strategic decisions. Microsoft running GitHub on AWS is embarrassing. It's also a preview of what's coming for every AI-dependent platform.
Companies that build their products on AI need to reckon with the fact that their infrastructure costs will scale with usage, not shrink with efficiency. They need to plan for capacity constraints. They need to accept that the promise of infinite scalability -- the cloud computing promise that made SaaS economics work -- doesn't apply to AI workloads in the same way.
The infrastructure ceiling is the physical manifestation of this economic problem. You can't scale AI usage without scaling compute, and you can't scale compute without building data centers, and you can't build data centers fast enough. Until compute becomes dramatically cheaper -- through better chips, more efficient models, or architectural breakthroughs -- the AI industry is constrained by physical reality.
The Energy Problem
The infrastructure ceiling has an energy dimension that's becoming impossible to ignore. AI data centers consume enormous amounts of electricity -- a single large AI data center can use as much power as a small city. The US power grid is not built for this load. Multiple regions are now facing electricity constraints driven by data center expansion, and the problem is getting worse.
This creates a secondary constraint on AI growth: you can build data centers, but you need to power them. Utilities are struggling to keep up with demand. Some data center operators are turning to on-site power generation -- diesel generators, natural gas turbines, even nuclear -- because the grid can't provide enough electricity. This adds cost, complexity, and environmental concerns.
The energy constraint is particularly challenging because it interacts with political and environmental policy. Data centers that consume massive power draw regulatory attention. Communities fight data center construction because of noise, water usage, and power consumption. The AI industry's growth is constrained not just by chip supply but by the physical and political limits of power generation.
This is perhaps the most underappreciated limit on AI's trajectory. The models will keep getting smarter. The software will keep getting better. But the electrons that power the hardware have to come from somewhere, and that somewhere is increasingly contested. The infrastructure ceiling is really an energy ceiling, and the energy ceiling is the one limit that can't be solved with better software.
7. The Pentagon's AI Army
The Department of Defense is proudly using AI to write reports mandated by Congress. 1.5 million users across the Pentagon are running on Google Gemini. Whether this is efficiency or an indictment of the compliance report industrial complex depends on your perspective.
On June 16, Ars Technica reported that the Pentagon has been boasting about its use of AI -- specifically Google Gemini -- to write reports that Congress requires the Department of Defense to produce. The scale is remarkable: approximately 1.5 million users across the DoD are using AI tools, and the department is actively expanding its AI deployment.
The Pentagon's enthusiasm for AI report-writing is a perfect microcosm of the AI reality check theme. On one hand, it demonstrates genuine productivity gains: Congress mandates an enormous volume of compliance reporting from the DoD, and if AI can help produce those reports faster, that's real efficiency. On the other hand, it raises questions about what those reports actually mean if they're being written by machines rather than by the people who are supposed to be accountable for them.
The Scale of Military AI Adoption
The 1.5 million user figure is striking. That's not a pilot program or an experimental deployment. It's institutional adoption at scale. The DoD has integrated AI into the daily workflows of a huge portion of its workforce, making it one of the largest AI deployments in any organization in the world.
Google Gemini's role in this deployment is also significant. Google has been more cautious than other AI companies about military applications -- the company famously backed away from Project Maven, a Pentagon AI program, in 2018 after employee protests. But the current arrangement shows Google has found its way back to the defense market through a different door: not analyzing drone footage or targeting systems, but writing reports. The less controversial, more banal use case turns out to be the one that opens the biggest military market.
The DoD's AI usage also highlights the infrastructure story from the previous section. 1.5 million users generating AI-assisted reports requires enormous compute capacity. The Pentagon's compute needs for this deployment are likely in the same order of magnitude as a major tech company's AI workload. The fact that the DoD chose Google -- which has the infrastructure to support this scale -- over smaller providers tells you something about the infrastructure ceiling. Only a handful of companies can serve AI at this scale.
The Accountability Question
Congress mandates reports from the DoD for a reason. Those reports are supposed to provide transparency and oversight. They're supposed to document what the military is doing, how it's spending money, what risks it's identifying, and how it's addressing them. If those reports are being written by AI, what exactly is Congress reviewing?
The optimistic interpretation is that AI is helping DoD staff produce better reports faster. The humans who understand the content are still directing the AI, reviewing its output, and signing off on the final product. The AI is a tool that helps them work more efficiently, not a replacement for their judgment.
The pessimistic interpretation is that AI is being used to generate compliance theater -- reports that technically satisfy congressional mandates but that no human at the Pentagon has carefully read or verified. If the process becomes "prompt the AI, lightly review, submit," then the oversight function is compromised. Congress isn't getting the insight it asked for. It's getting AI-generated text that looks like insight.
The reality is probably somewhere in between, varying widely across the DoD's sprawling bureaucracy. Some offices are using AI carefully, with strong review processes. Others are almost certainly using it to churn out reports that no one reads, just to check the compliance box. The Pentagon is not a monolith. It's a vast organization with enormous variation in practice and quality.
The Precedent for Government AI
The DoD's AI deployment sets a precedent for how government agencies use AI for mandated reporting. If the Pentagon can use AI to write congressional reports, other agencies will follow. The EPA could use AI to write environmental impact reports. The FDA could use AI to write drug safety reviews. The SEC could use AI to write enforcement summaries.
This creates a strange dynamic: Congress requires agencies to produce reports for oversight purposes. Agencies use AI to produce those reports. Congress reads AI-generated reports to perform oversight. The entire oversight loop becomes AI-to-AI communication, with humans on both sides largely rubber-stamping the output.
This isn't necessarily bad -- if the AI is accurate and the humans are engaged, it could improve the quality and timeliness of government reporting. But it requires a level of AI reliability and human engagement that this week's other stories suggest is not yet warranted. The KPMG hallucination case, the police evidence fabrication case, the healthcare cost inflation case -- all of these suggest that organizations are not yet sophisticated enough in their AI usage to trust the output without close review.
The Pentagon's boastfulness about AI report-writing is itself a problem. It suggests the department is treating AI adoption as a PR opportunity rather than a operational decision that requires careful oversight. If the DoD is more interested in showing off its AI usage than in ensuring that usage is safe and effective, the risk of problems -- hallucinated reports, classified information leaking through AI prompts, incorrect data reaching congressional oversight -- increases significantly.
8. Noam Shazeer and the Talent War
The co-lead of Google's Gemini project and one of the original authors of the transformer paper is joining OpenAI. The talent war is intensifying, and Google is losing its best people.
On June 18, Noam Shazeer announced he was joining OpenAI. The news, first reported by Reuters and confirmed by Shazeer on social media, sent ripples through the AI industry. Shazeer isn't just a senior engineer switching companies. He's one of the most consequential figures in the history of artificial intelligence.
Shazeer was an author on "Attention Is All You Need," the 2017 Google paper that introduced the transformer architecture. Every major AI model -- GPT, Claude, Gemini, Llama -- is built on the transformer architecture. The paper is arguably the most important technical contribution to modern AI, and Shazeer was one of its key authors.
He then founded Character.AI, which brought conversational AI to consumers and was eventually reabsorbed into Google as part of a talent acquisition deal. At Google, Shazeer became co-lead of the Gemini project, making him one of the most powerful people inside the company's AI division. His departure to OpenAI is a significant loss for Google and a significant gain for OpenAI.
Why This Matters Beyond the Headline
Individual talent moves don't usually matter this much. Companies are bigger than any one person. But in AI, the concentration of top talent at a small number of companies has become a strategic concern. The most capable AI models are built by small teams -- sometimes dozens of researchers, not hundreds. Losing one key person can set a project back months. Gaining one can accelerate it dramatically.
Shazeer's move to OpenAI is part of a broader pattern. The most talented AI researchers and engineers have been consolidating at three companies: OpenAI, Anthropic, and Google. Meta has struggled to retain top talent despite its open-source strategy. Smaller companies and startups face even greater challenges. The result is an industry where the gap between the top three and everyone else is widening.
Within the top three, there's a secondary competition. OpenAI has historically been the talent magnet -- the place where the most ambitious researchers want to work because of the resources, the compute, and the brand. Anthropic has attracted researchers who prioritize safety and alignment work. Google has competed on resources, compensation, and the appeal of working at a company with enormous infrastructure and distribution.
Shazeer's departure suggests Google is losing this competition. When the co-lead of your flagship AI project leaves for your biggest rival, it's not just a personnel change. It's a signal. Other researchers at Google will notice. Other companies will use the departure as a recruiting tool. The narrative that OpenAI is the destination for top AI talent gets stronger with every high-profile departure from Google.
The Concentration Problem
The talent concentration at OpenAI, Anthropic, and Google creates a strategic risk for the entire AI ecosystem. When the best minds in a field work for three companies, those companies' priorities become the field's priorities. If none of them are working on your specific problem -- say, AI safety mechanisms for biomedical applications, or AI systems for small-language communities -- that work doesn't get done at the frontier level.
This is particularly concerning for AI safety research. Anthropic has made safety a core mission, but OpenAI's safety team has seen significant turnover and Google's safety efforts are embedded in a larger product organization. The most talented researchers are being pulled toward capability work -- building more powerful models -- rather than safety work, because that's where the resources and prestige are.
The concentration also has implications for AI's direction as a technology. When the people building AI are drawn from a narrow demographic -- primarily young, primarily male, primarily based in San Francisco -- the technology reflects their priorities and blind spots. The well-documented biases in AI systems, from facial recognition that works less accurately on darker skin to language models that reproduce gender stereotypes, are partly a function of who builds them. Talent concentration at three companies with similar cultures and similar hiring patterns narrows the perspectives that shape AI's development.
Shazeer's move to OpenAI, while a win for OpenAI, narrows the field further. When the most influential researchers are all working for organizations with similar commercial incentives, the research agenda tilts toward what's profitable rather than what's beneficial. The open source manifesto that went viral this week is partly a response to this -- a recognition that the frontier of AI capability is being shaped by a handful of people at a handful of companies, and that the rest of the world is reduced to consuming whatever they produce.
The Economic Dimensions
The talent war is also an economic story. Compensation for top AI researchers has reached levels that would have seemed absurd a decade ago. Salaries in the millions, plus equity worth tens of millions, are now standard for senior researchers at frontier labs. The cost of talent is becoming a significant portion of AI companies' budgets, and it creates a barrier to entry that makes it nearly impossible for new companies to compete at the frontier.
This talent cost also contributes to the economic pressure on AI companies. When you're paying researchers millions of dollars each, you need enormous revenue to justify the investment. This pushes companies toward revenue-generating products even when those products aren't ready for deployment, which contributes to the quality problems that this week's stories have documented.
9. The Anthropic Saga Deepens
SK Telecom, Amazon's concerns, Nicholas Carlini's profile, and the continuing fallout from the Fable 5 shutdown. The Anthropic-government saga is becoming the defining story of AI governance.
The Anthropic story that dominated last week's issue continued to develop this week, with new revelations that deepen the intrigue around the Fable 5/Mythos 5 shutdown and the government's role in AI deployment decisions.
SK Telecom and the China Connection
On June 18, WIRED published a detailed report identifying SK Telecom -- South Korea's largest wireless carrier -- as the "South Korean telecommunications company" whose access to Anthropic's Mythos model alarmed White House officials. The report fills in a critical piece of the Fable 5 shutdown story.
SK Telecom received access to Mythos as part of Project Glasswing's expansion to approximately 150 companies. The company is an Anthropic investor -- it put $100 million into the AI lab in 2023 as part of a commercial partnership to develop telecom-specific AI models. When White House officials learned that a Korean company with potential China connections had access to the most powerful AI model Anthropic had ever built, they asked Anthropic to revoke that access. Anthropic complied.
The China concern stems from SK Telecom's parent conglomerate, SK Group, which has extensive business interests in China spanning semiconductors, energy, and other industries. SK Telecom itself has minimal China operations -- about $1.9 million in revenue and seven employees in 2024 -- but the broader conglomerate's China ties were enough to trigger national security concerns.
This detail matters because it reveals how the US government is thinking about AI model access. The concern isn't just about direct adversaries accessing powerful AI. It's about any company with any connection to China, no matter how tangential, having access to frontier models. If that standard were applied broadly, it would exclude a significant portion of the global technology industry from accessing the most advanced AI systems.
Amazon's Role
The WIRED report also adds context to Amazon's involvement in the Anthropic situation. As previously reported, Amazon CEO Andy Jassy raised concerns about Anthropic's models before the government crackdown. Amazon has invested billions in Anthropic and is a primary cloud provider for the company. The exact nature of Jassy's concerns remains undisclosed, but their timing -- before the export control directive -- suggests the government's action may have been influenced by more than a single jailbreak demonstration.
The Amazon angle raises uncomfortable questions about the relationship between AI investors, their portfolio companies, and government regulators. Amazon competes with Microsoft and Google in the AI market. It also hosts Anthropic's models on AWS. If Amazon's concerns about Anthropic's models accelerated the government's decision to shut down Fable 5, that's a case of a major tech company using government relationships to influence the competitive landscape.
This doesn't require conspiracy. Amazon may have had genuine concerns about Anthropic's model safety. The company may have reported those concerns in good faith. But the appearance of a major competitor influencing government action against another competitor is damaging to the integrity of the regulatory process, regardless of intent.
Nicholas Carlini: The Hacker Sent to Washington
The Wall Street Journal published a profile of Nicholas Carlini, the researcher Anthropic sent to Washington to demonstrate Mythos's safety capabilities and calm government nerves about the model's potential risks. Carlini is described as a brilliant security researcher who has spent years testing AI systems for vulnerabilities. His role was to show government officials that Anthropic understood Mythos's capabilities and had robust safety mechanisms in place.
The profile is revealing not just for what it says about Carlini but for what it reveals about the state of AI-government relations. Anthropic needed to send a person to Washington to personally demonstrate that their AI was safe. Not a policy team, not a PR campaign, but a technical expert who could show the model's behavior in real time. This suggests the government's understanding of AI capabilities is still rudimentary enough that it requires in-person technical demonstration.
It also reveals the asymmetry in AI governance. The government can shut down a model with a Friday evening directive. But understanding whether a model should be shut down requires technical expertise that exists only within the companies building the models. The government is dependent on the companies it's regulating for the information needed to make regulatory decisions. This creates a structural conflict of interest that no amount of in-person demonstrations can resolve.
The Governance Vacuum
The Anthropic saga is becoming the defining story of AI governance not because Anthropic is uniquely problematic -- the company has been more transparent and safety-focused than most of its competitors -- but because it exposes the vacuum at the center of the regulatory framework.
The government can order a model shutdown. It can't explain why. It can cite national security concerns without providing specifics. It can act on a Friday evening with no process. It can blacklist a company from defense contracts without explaining the basis. And the company subject to these actions has no clear recourse.
Meanwhile, the government is simultaneously using the same company's AI for its own purposes. The Pentagon's 1.5 million AI users include people who may be working with the same class of models that the government just ordered Anthropic to shut down. The government is both Anthropic's regulator and its customer, creating a set of conflicts that would be untenable in any other industry.
The Fable 5 crisis is not resolved. Anthropic complied with the shutdown order, but the underlying questions -- what standard should trigger a model shutdown, what process should precede such an action, what recourse should companies have, and how should the government balance national security concerns with commercial innovation -- remain unanswered. Every frontier AI company is watching this case as a precedent. The precedent it sets is not reassuring.
10. Open Source AI Makes Its Case
A manifesto declaring that "open source AI must win" went viral this week, collecting nearly 1,600 points on Hacker News and sparking a debate that cuts to the heart of the AI industry's direction. The argument is simple: civilization's intelligence infrastructure shouldn't be rented from a handful of closed institutions.
On June 13, a manifesto titled "Open Source AI Must Win" appeared online and quickly became one of the most discussed AI posts of the week. Written by Ahmad Osman, the manifesto's core argument is that AI is becoming infrastructure -- as fundamental to civilization as electricity, internet protocols, or operating systems -- and that infrastructure of this importance cannot be controlled by a small number of proprietary companies.
The manifesto's central claim is worth taking seriously: "Civilizational intelligence infrastructure must remain free to study, build, deploy, and run, not rented from closed institutions." This is not a new argument in the open source world, but it's arriving at a moment when the alternative -- a future where three or four companies control all meaningful AI capability -- is becoming the default trajectory.
The Case for Open Source AI
The argument for open source AI rests on several pillars, and this week's stories strengthen each of them.
Transparency: When KPMG's AI-generated report contained hallucinations, the problem was partly that the AI's reasoning process was opaque. No one could trace how the AI arrived at its false claims because the model is a black box. Open source models, where the weights, training data, and architecture are publicly available, at least allow researchers to study how models produce errors. You can't fix what you can't see.
Accountability: The Pentagon's use of AI to write congressional reports raises accountability questions that proprietary AI makes harder to answer. When a government agency uses a proprietary AI system, the company that built the system has no obligation to disclose how it works, what biases it contains, or what errors it's prone to. With open source AI, independent auditors can examine the system directly.
Cost: The infrastructure crunch that forced Microsoft to use AWS for GitHub is partly a function of proprietary AI economics. When compute-intensive AI is controlled by companies that charge per-token pricing, the costs scale with usage in ways that are unsustainable for many organizations. Open source models can be run on your own infrastructure, at your own cost, without per-query licensing fees.
Safety: The Anthropic government shutdown exposed how fragile the proprietary AI ecosystem is. When one company's model gets shut down, everyone using that model loses access. If the model were open source, the community could continue using it even if the original developer faced regulatory issues. Decentralized AI is more resilient than centralized AI.
Innovation: When the best AI talent consolidates at three companies, the research agenda is set by those companies' commercial priorities. Open source AI allows researchers outside the big labs to contribute to and build on frontier capabilities, broadening the range of problems AI is applied to.
The Case Against
The open source argument isn't without merit on the other side. The same week the manifesto went viral, a police officer was under investigation for using AI to fabricate evidence. Open source AI means anyone -- including bad actors -- can access powerful AI tools without any oversight or restrictions. The same transparency that enables safety research also enables adversaries to find vulnerabilities. The same accessibility that democratizes AI also puts it in the hands of people who will misuse it.
The Anthropic Fable 5 story illustrates this tension. Anthropic built Mythos as a gated model because its cybersecurity capabilities were too dangerous for general release. If Mythos were open source, those capabilities would be available to anyone -- including hostile nations, criminal organizations, and individual bad actors. The government's concerns about SK Telecom's China connections show how seriously the US takes the risk of frontier AI capabilities reaching adversaries.
The open source response to this is that the bad actors will get AI anyway -- through model theft, through open-weight releases by other companies, through their own training programs. Restricting open source AI hurts the good actors more than the bad ones, because the bad actors have alternatives and the good actors don't.
This is a genuine tension, not a problem with an easy answer. The AI industry has been arguing about it for years, and this week's stories illustrate both sides without resolving the debate.
The Market Reality
The open source AI movement faces a market reality that makes its mission difficult. Training frontier AI models requires enormous compute resources -- hundreds of millions of dollars worth of GPUs, data center capacity, and electricity. Open source projects generally don't have access to this level of funding. The best open source models are currently produced by well-funded companies (Meta's Llama series, Mistral's models) that release them for strategic reasons, not because they can't compete with proprietary offerings.
This means the open source AI ecosystem is dependent on the largesse of companies that have their own commercial motives. Meta releases Llama because it wants to commoditize the model layer and compete at the application layer. Mistral releases open weights as part of a hybrid strategy. If these companies' strategies change, the open source supply could dry up.
The manifesto's call for open source AI to win is aspirational. The current reality is that open source AI is a complementary ecosystem, not a replacement. It provides a baseline of capability that proprietary models have to compete with, which helps keep pricing in check and prevents total monopolization. But the frontier of capability is still firmly in the proprietary camp, and the talent and compute concentration at the top three companies makes that unlikely to change soon.
11. The Cost of the Gap
Every story this week traces back to the same fundamental problem: the gap between what AI promises and what AI delivers. That gap has costs -- economic, social, institutional, and political. Understanding those costs is the first step toward addressing them.
The individual stories of this week -- the Pew study, KPMG's hallucinated report, the police evidence fabrication case, the healthcare cost increase, Microsoft's infrastructure crunch, the Pentagon's AI reports, Noam Shazeer's move, the Anthropic saga, the open source manifesto -- are each significant in their own right. But together, they paint a picture of an industry that has overpromised, underdelivered, and is now facing the consequences.
The Economic Cost
The most immediate cost is economic. Organizations that adopted AI based on promises of efficiency gains are finding that the reality is more complicated. Healthcare costs are rising, not falling. Infrastructure costs are scaling with usage in ways that break traditional software economics. Consulting firms are spending on AI tools that produce work product they can't trust. The total cost of AI ownership -- including the technology, the integration, the error correction, the oversight, and the compliance overhead -- is higher than most organizations budgeted for.
This doesn't mean AI is economically negative. It means the economic case for AI is more nuanced than the sales pitch. Some applications generate real value. Others don't. The challenge is distinguishing between them, and this week's stories suggest that many organizations haven't been making that distinction carefully enough.
The Trust Cost
The 16% Pew number is the most important statistic in AI right now because it measures the trust deficit. When only 16% of the population believes your industry will have a positive impact, you have a legitimacy problem. That problem affects everything from consumer adoption to regulatory policy to talent recruitment.
The trust deficit is self-reinforcing. Low public trust leads to restrictive regulation. Restrictive regulation limits deployment. Limited deployment reduces the evidence of beneficial AI applications. Lack of evidence keeps public trust low. Breaking this cycle requires either dramatically better AI products or a fundamental change in how the industry communicates about its capabilities and limitations.
The KPMG hallucination case is particularly damaging to trust because it's the kind of story that resonates with ordinary people. "The consultants promoting AI can't even use it correctly" is a narrative that writes itself. Every time a high-profile AI failure becomes public, the trust deficit widens. The industry needs fewer positive PR campaigns and fewer breathless predictions about AGI, and more honest acknowledgment of what AI can and cannot do today.
The Institutional Cost
The institutions adopting AI most aggressively -- consulting firms, law enforcement, healthcare systems, the military -- are the ones where AI failures have the highest stakes. A hallucinated consulting report is embarrassing. A fabricated piece of police evidence is dangerous. An AI-inflated medical bill is costly. An AI-generated compliance report that no one reads undermines oversight. Each of these represents a failure mode that the institution either didn't anticipate or didn't take seriously enough.
The institutional cost extends to the AI companies themselves. Anthropic's clash with the government, OpenAI's state attorneys general investigation, the growing regulatory scrutiny of every major AI lab -- these are the costs of operating in an institutional environment that hasn't figured out how to handle AI. The companies are spending resources on regulatory compliance, litigation, and government relations that could be going into research and product development.
The Political Cost
The AI industry's political position is weakening. When Bernie Sanders proposes a $7 trillion wealth fund financed by taxing AI companies at 50%, that's a sign the political climate has shifted. When the Pentagon blacklists Anthropic as a supply chain risk, that's a sign the national security establishment is willing to treat AI companies as adversaries, not partners. When state attorneys general are investigating OpenAI for consumer protection violations, that's a sign the regulatory apparatus is mobilizing against the industry.
The political cost is compounded by the industry's own behavior. AI executives who make grandiose claims about artificial general intelligence, who promise products that can't be delivered, who deploy tools that hallucinate and fabricate -- these executives are creating the political vulnerability that regulators are exploiting. Every overpromise becomes a potential basis for a lawsuit, an investigation, or a regulation.
The Opportunity Cost
Perhaps the most insidious cost is the opportunity cost -- the beneficial applications of AI that aren't being pursued because the industry is focused on the most profitable rather than the most beneficial use cases. AI that helps doctors diagnose rare diseases. AI that helps small businesses compete with large ones. AI that helps teachers personalize education. AI that helps researchers discover new materials, new drugs, new scientific knowledge.
These applications exist. They're being worked on by dedicated researchers and small companies. But they're not where the capital, the talent, and the infrastructure are going. The resources are going to AI chatbots, AI coding tools, AI advertising optimization, AI content generation -- the applications that generate revenue or reduce costs in the short term, rather than the applications that create long-term value.
The gap between AI's promise and its reality isn't just a PR problem. It's an allocation problem. The industry is allocating its resources toward applications that maximize short-term returns rather than long-term benefit. The cost of that misallocation is the beneficial AI we're not building while we're building the next chatbot.
12. What Comes Next
The reality check isn't the end of AI. It's the beginning of AI's maturation. What happens in the next six months will determine whether the industry learns from this week or repeats it.
This week was not a disaster for AI. The technology didn't fail catastrophically. No AI-related catastrophe occurred. What happened was less dramatic but arguably more important: the accumulated evidence of AI's current limitations became impossible to ignore. The promises, the hype, and the trillion-dollar valuations ran into the reality of hallucinations, infrastructure limits, public skepticism, regulatory friction, and rising costs.
This is what a reality check looks like. It's not a collapse. It's a correction. And corrections, painful as they are, serve a purpose. They force an industry to confront what's working and what isn't, to distinguish between the promise and the product, to decide whether the trajectory they're on is the one they want to be on.
What the Industry Needs to Do
Stop overpromising. The single most damaging behavior in the AI industry right now is the gap between marketing and reality. Every time an AI company promises capabilities the technology doesn't have, it widens the trust deficit. Every time a consulting firm claims AI is ready for deployment in a context where it clearly isn't, it creates the conditions for a KPMG-style embarrassment. The industry needs to start talking about what AI can do today, not what it might do tomorrow.
Invest in reliability, not just capability. The industry's obsession with benchmarks -- higher scores on coding tests, reasoning tests, knowledge tests -- has come at the expense of reliability. A model that scores 5% higher on a benchmark but still hallucinates 5% of the time isn't a better product. It's a more dangerous product, because the higher benchmark creates more confidence that the output is correct. The industry needs to measure what matters: error rates in real-world use, not benchmark scores in controlled tests.
Accept the infrastructure constraints. The compute story is real. AI demand is outpacing data center capacity, and that gap will persist for years. The industry needs to plan for this, not pretend it's temporary. That means more efficient models, better caching, smarter routing, and honest conversations with customers about what usage levels are sustainable.
Engage with regulation constructively, not defensively. The regulatory environment is going to get tighter, not looser. The 16% public approval number guarantees it. The industry's current approach -- fighting regulation, lobbying against oversight, resisting transparency -- is politically unsustainable. The companies that survive the reality check will be the ones that help write the rules rather than fighting them.
Take safety seriously, not just as a marketing tool. The Anthropic story shows the limits of safety as a brand differentiator. Anthropic positioned itself as the safety-focused AI lab, and the government still shut down its model. Safety needs to be a practice, not a positioning. That means publishing error rates, disclosing failure modes, sharing safety research, and being honest about what models can and cannot do.
What the Public Needs to Do
Demand transparency. Organizations using AI -- hospitals, police departments, consulting firms, government agencies -- should be required to disclose when and how they use AI. The public has a right to know when AI is involved in decisions that affect their lives. This week's stories show what happens when AI operates without transparency: fabricated evidence, hallucinated reports, inflated bills.
Set realistic expectations. AI is a tool, not a transformation. It helps with some tasks and fails at others. The public needs to understand this, not because the public is responsible for fixing the AI industry, but because realistic expectations are the best defense against the industry's overpromises.
Support thoughtful regulation. Not all regulation is good regulation. The Fable 5 shutdown -- a Friday evening directive with no process -- is bad regulation. But the absence of regulation is worse. The public should support regulations that require transparency, accountability, and safety testing, while opposing regulations that are opaque, arbitrary, and not grounded in technical reality.
The Bigger Picture
The AI reality check of June 2026 will either be remembered as the moment the industry started taking its limitations seriously, or as the moment it doubled down on hype despite the evidence. The path forward isn't mysterious. It's the same path every transformative technology has had to walk: from hype to disillusionment to mature adoption.
The internet went through this. In 1999, every company was going to be an internet company, the old economy was dead, and valuations didn't matter. In 2001, the bubble burst, the skeptics were vindicated, and the internet was declared a fad. By 2010, the internet was obviously transformative, obviously permanent, and obviously not going to make every internet company a winner. The technology won. The hype didn't.
AI is in its 1999 moment. The hype is at its peak. The reality is catching up. The correction is coming. And the technology -- the actual technology, not the valuations and the promises -- will survive it, because AI genuinely is useful, genuinely is transformative, and genuinely does things that previous technologies couldn't.
But the companies, the valuations, and the promises won't all survive. The ones that make it through the reality check will be the ones that built useful things, told the truth about what those things could do, and treated their limitations as problems to solve rather than as PR problems to manage.
This week was the reality check arriving. What happens next depends on whether the industry listens.
See you next Monday. -James
The promise was everything. The reality was something else. The gap between them is where all the interesting things happen.