The 2026 Stanford AI Index delivers a verdict that the research community had been accumulating evidence for years: 90% of notable AI models now originate from private companies, and the majority of them launch with almost no technical documentation attached. Of the 95 most notable models released last year, 80 shipped without training code — a figure that would have been implausible in 2020, when academic labs and semi-open organizations still shaped the field’s norms.
The collapse in AI private models transparency is not incidental. It reflects deliberate choices by the three labs that set norms for the entire industry: Google DeepMind, Anthropic, and OpenAI. All three have quietly abandoned the disclosure practices that characterized AI research from 2017 through roughly 2022. What they stopped sharing — dataset sizes, training duration, compute budgets — is exactly what independent researchers need to verify whether these systems are safe to deploy at scale.
AI Private Models Transparency: What the Stanford Data Actually Shows
The Stanford Foundation Model Transparency Index (FMTI) tracks disclosure across 100+ indicators. The trend since 2023 is unambiguous: private labs dominate AI model development and share progressively less information with each successive release.
Of 95 notable models catalogued in the most recent index period, 80 released no training code. The majority withheld dataset composition, training compute figures, and environmental impact data. Fewer than 15% disclosed anything substantive about evaluation methodology — the procedures that determine whether a model is safe to deploy in high-stakes environments like healthcare or legal systems.
The private sector’s 90% share of notable models represents a near-complete reversal from 2019, when academic institutions and semi-open research organizations contributed a meaningful fraction of the field’s foundational work. The Stanford AI Index reports that the shift accelerated sharply in 2023 and has not slowed since.
What Google, Anthropic, and OpenAI Stopped Disclosing
These three labs determine industry norms. What they withhold, every other lab withholds without consequence.
OpenAI stopped publishing dataset details after GPT-3. The GPT-4 technical report ran to 98 pages and deliberately excluded training data volume, compute budget, and hardware configuration — citing, verbatim, “competitive landscape and safety reasons.” Subsequent models including o3 and GPT-5 maintained this pattern. OpenAI’s pivot toward large-scale commercial licensing arrangements has aligned the company’s transparency policy with commercial interests rather than scientific norms.
Google DeepMind published the original Transformer paper in 2017 — the technical foundation of every major language model in existence. By 2024, Gemini Ultra’s technical report benchmarked performance extensively while omitting training data composition, pretraining corpus details, and fine-tuning procedures. Its successors have maintained the pattern.
Anthropic built its public identity on safety research: Constitutional AI, the Responsible Scaling Policy, mechanistic interpretability work. Claude 3 Opus, Claude 3.5 Sonnet, and subsequent releases launched without dataset disclosure, training compute figures, or model weights. When a developer accidentally published source code for a Claude agent SDK — an incident MegaOne AI covered — the accidental disclosure revealed more about Anthropic’s internal architecture than the company had officially shared in months of public communications. The irony is structural: a safety-first lab achieving more transparency through a mistake than through policy.
Three Reasons Labs Stopped Sharing — Only One Is Legitimate
Three forces converged to produce the current disclosure environment. Only one has genuine merit.
Competitive exposure is the most defensible argument. Training recipes contain real proprietary information: data sourcing strategies, filtering approaches, fine-tuning pipelines. A lab that discloses its pretraining corpus composition and compute budget hands competitors a meaningful roadmap. This argument holds at the margin. It does not explain why dataset licenses, evaluation methodology, environmental impact data, or basic compute order-of-magnitude figures need to remain confidential.
Regulatory uncertainty contributed more than labs publicly acknowledge. As the EU AI Act entered enforcement and US executive orders expanded AI oversight requirements, legal teams at major labs advised against disclosures that might constitute an implicit admission of high-risk system classification under emerging frameworks. Opacity became a compliance strategy — a way to minimize regulatory surface area before the regulatory landscape clarified.
Safety as pretext is the most consequential mechanism. OpenAI explicitly cited safety concerns when withholding GPT-4’s training details, arguing that adversaries who know exactly what data was used might exploit it. Researchers at MIT, Carnegie Mellon, and the Alignment Research Center have publicly rejected this framing. Dataset size and training duration provide no meaningful attack surface. What they do provide is accountability — which appears to be the actual concern.
Four Research Capabilities That Opacity Disabled
The consequences are concrete, not abstract. Withholding training details has disabled four categories of research that AI safety depends on.
- Benchmark contamination detection: Without training data disclosure, researchers cannot determine whether evaluation datasets appeared in pretraining. A model scoring 94% on MMLU after training on MMLU-derived data is not equivalent to a clean 94%. Every major public leaderboard is now potentially compromised by unverifiable contamination — and labs are under no obligation to say whether contamination occurred.
- Energy and compute accounting: The International Energy Agency estimated AI data centers consumed 415 TWh in 2024. Without per-model compute disclosure, attributing that footprint to specific systems requires guesswork that labs could immediately correct if they chose to. Facilities like Nebius’s planned $10 billion data center in Finland are expanding to meet demand from labs that won’t disclose how much compute their models require.
- Reproducibility: Without training code, independent researchers cannot verify results, audit methodology, or build credibly on published work. The scientific chain that made 2017–2022 AI research so unusually productive — each paper extending the last — has been severed at the foundation.
- Capability red-teaming: Third-party evaluators assessing dangerous capabilities — biological synthesis assistance, deceptive reasoning, cyberattack support — cannot run meaningful evaluations without understanding what data shaped model behavior and through what fine-tuning process.
The Open-Source Counterargument Doesn’t Survive Scrutiny
The standard defense invokes Meta’s Llama series and Mistral: open-source is thriving, so the transparency problem is overstated. This conflates two distinct concepts.
Llama 4 and Mistral Large are open-weight — model parameters are downloadable. Neither lab discloses full training data composition, training code, or compute budgets. Open-weight is a distribution decision. Transparency is a disclosure decision. They are not the same thing.
MegaOne AI tracks 139+ AI tools across 17 categories, and the pattern holds consistently: even models marketed as “open” typically withhold three or more of the six major transparency indicators tracked by the FMTI. Releasing weights without documentation is equivalent to publishing a pharmaceutical compound’s molecular structure while withholding dosage trials, synthesis methods, and what the compound was tested against.
The Self-Reinforcing Accountability Deficit
The disclosure blackout is self-reinforcing. When OpenAI established the precedent of withholding GPT-4’s training details in 2023, it reduced the reputational cost for every subsequent lab to do the same. Google and Anthropic followed within months. Smaller labs had no incentive to lead on transparency when the three dominant players had normalized opacity — and faced no commercial penalty for doing so.
The Humans First movement and allied advocacy groups have pushed for mandatory disclosure requirements for AI systems with societal-scale deployment across healthcare, legal, education, and government procurement sectors. As of April 2026, labs have successfully resisted statutory disclosure requirements in every major jurisdiction. The Frontier Model Forum’s voluntary transparency commitments, signed by several major labs in 2023, have been selectively honored and selectively ignored without consequence.
OpenAI’s commercial expansion has accelerated this dynamic. Once a lab’s valuation crosses $300 billion and its products are embedded in enterprise software used by hundreds of millions, technical transparency stops being a scientific virtue and starts being a liability — a source of competitive exposure, regulatory surface area, and litigation risk.
What Would Actually Change This
Voluntary frameworks have failed. Structural change requires one of three interventions:
- Mandatory disclosure as a deployment condition: Regulators require training data provenance and compute disclosure before models can be integrated into regulated sectors — finance, healthcare, legal, and government systems. This exists in draft form under the EU AI Act’s high-risk category requirements but has been applied so narrowly as to cover almost nothing at the frontier.
- Confidential third-party audit: Labs share training details under NDA with accredited safety evaluators, analogous to FDA drug trial oversight. This requires building a credible independent auditor class that does not yet exist at the scale needed to cover frontier models on reasonable timelines.
- Tiered disclosure by capability threshold: Models above a defined compute or parameter threshold face mandatory disclosure requirements; smaller models operate under lighter rules. Labs will contest any specific threshold, but the structural approach is workable and precedented in nuclear and pharmaceutical regulation.
The most actionable near-term lever requires no new legislation: procurement. Governments and large enterprises that license AI services can contractually require training data provenance, compute disclosure, and evaluation methodology as conditions of purchase agreements. That lever exists today and is not being systematically applied by any major government buyer.
The 90% private-sector figure is not a trend to monitor — it describes a completed transition. Labs that built their reputations on safety research are now running the least transparent development processes in the field’s history. The accountability gap is widening, and no voluntary commitment has closed it yet. It will not close without external pressure, whether from regulators, customers, or both.
