7 Forbidden AI Secrets That Big Tech Won’t Admit

This isn’t a theory anymore. Bartz v. Anthropic settled in 2025 for $1.5 billion — the largest copyright settlement in U.S. history. The case centered on Anthropic downloading millions of pirated book copies to train Claude. The New York Times is still going. Authors Guild is still going. Universal Music Publishing Group filed a $3.1 billion lawsuit against Anthropic in January 2026.

There are now 166 active AI copyright lawsuits across 15 jurisdictions. The Munich Regional Court ruled in November 2025 that OpenAI’s use of German song lyrics to train GPT-4 violates German copyright law — the first major European ruling of its kind. In January 2026, Judge Sidney Stein ordered OpenAI to produce 20 million anonymized ChatGPT logs to copyright plaintiffs. OpenAI fought it. Lost.

The industry’s public defense — “fair use, it’s transformative” — is genuine, but unresolved. No U.S. court has ruled on AI training fair use at scale yet. That ruling comes in 2026. If it goes against the labs, the repricing of everything downstream will be brutal.

What this means practically: Companies fine-tuning open models on their own proprietary data are building moats that this legal wave cannot touch. Companies running everything through closed APIs face potential “retroactive licensing” demands — and no exit strategy.

A single ChatGPT query uses roughly 10× the electricity of a Google search. Training a GPT-5-class model consumes more than 100 GWh — the annual electricity use of 10,000 U.S. households. The IEA World Energy Outlook 2025 warned that under high-AI scenarios, data centers alone could consume 1,000 TWh by 2026.

Goldman Sachs Research forecasts a 175% rise in power demand from data centers by 2030.

Here’s how seriously the hyperscalers are taking it: Microsoft signed a nuclear deal with Helion (delivery 2028). Amazon purchased an entire nuclear plant in Pennsylvania. Epoch AI’s September 2025 analysis pointed out that Meta’s >1.5GW Hyperion cluster isn’t expected until late 2029 or 2030 — because building infrastructure at that scale takes 4–5 years. The grid cannot support current AI expansion plans on the timelines being promised.

Environmental groups are already suing. In December 2025, one filed against the Public Service Commission of Wisconsin over Meta’s data center. In 2026, surging electricity rates are on local ballots in multiple states. “We’re buying carbon offsets” is the public line. The private line — visible in every major hyperscaler’s energy procurement strategy — is: we need gigawatts of new nuclear generation we don’t yet have.

Every frontier model in 2025 still hallucinates 12–28% of the time on hard tasks (Stanford AI Index 2025, HELM Safety benchmark). Not some models. All of them.

This isn’t because labs haven’t tried to fix it. It’s because the architecture — next-token prediction at scale — is probabilistic by design. It doesn’t know what it doesn’t know. It generates confident-sounding text because confident-sounding text dominated its training data. Google’s own researchers published papers documenting that LLMs cannot reliably reason or retrieve information beyond statistical patterns.

The industry’s current answer is Retrieval-Augmented Generation (RAG). Bolt a database onto the LLM; let it look things up instead of hallucinating them. This works — but it’s not the “general intelligence” being sold in the keynotes. It’s a workaround for a fundamental limitation that most enterprise AI vendors don’t advertise.

The practical takeaway is this: Never trust raw LLM output for anything where being wrong has consequences. Medical, legal, financial, compliance. Any company selling “AI agents” without heavy RAG and verification guardrails is selling 2024 technology with 2026 pricing.

HEC Paris published an analysis in late 2025 noting what it called “a well-kept secret in the AI industry: for over a year now, frontier models appear to have reached their ceiling.” TechCrunch reported in November 2024 that “diminishing returns set in” and that “AI labs are rethinking strategies.” The Economist’s September 2025 piece described investor enthusiasm cooling as LLMs “fall short of their promise to become digital gods.”

The data is fairly clear. OpenAI’s o3 reasoning model required roughly 30× more compute than o1 for incremental gains — that’s the definition of diminishing returns. Epoch AI’s September 2025 analysis flagged that massive compute scaling is likely to slow due to increasing economic uncertainty and longer build cycles: constructing a very large data center takes 1–3 years; Meta’s planned 1.5GW Hyperion cluster isn’t expected until late 2029.

The honest version of the insider view: the next breakthrough in capability will come from architecture innovation — mixture-of-experts, test-time scaling, synthetic data distillation — not from throwing more compute at the same approach. DeepSeek proved this in early 2025 by matching GPT-4-class performance at a fraction of the training cost.

The benchmark data as of early 2026 is not ambiguous on this.

The gap in price-performance is now negative — open models are better value by most measures. The remaining closed-model advantages are plugin ecosystems and the deliberate crippling of public versions (see Secret #6). If you’re paying premium API prices for tasks where Llama or DeepSeek performs equivalently, you’re funding someone else’s moat.

This one is harder to document cleanly, because the evidence is by definition not in press releases. But the pattern is consistent enough that it’s worth naming directly.

Employees at major hyperscalers — routinely, across platforms — describe the public-facing models as “the 2023–2024 versions.” The leaked benchmark comparisons that surface periodically suggest internal models are significantly more capable than what’s available commercially. This isn’t surprising if you think about it: these companies have enormous incentive to release just enough capability to generate revenue without releasing their full competitive advantage.

The practical implication is that waiting for “the next GPT” to solve your AI limitations is a losing strategy. The capability you need may already exist internally. It won’t be released until the competitive and legal landscape makes it advantageous.

Klarna’s story is the most important case study in enterprise AI that almost nobody tells correctly.

Goldman Sachs estimated in August 2025 that 6–7% of the U.S. workforce — roughly 10 million workers — is economically displaceable today. IBM froze hiring in roles where AI can substitute 90% of work (internal memo, May 2025). Salesforce, Duolingo, Dropbox, and BT Group all announced 2025 layoffs explicitly tied to AI productivity gains.

The Klarna lesson isn’t that AI doesn’t displace jobs. It does. The lesson is: full replacement fails; hybrid beats replacement; the companies that figure this out first win, and the ones that follow the hype cycle hurt their customers and then reverse quietly.

The pattern will repeat. It’s already repeating.