AI capability keeps scaling. Disclosure does not. Stanford University’s 2025 Foundation Model Transparency Index captures the gap. In one year, the average score across major developers fell from 58 to 40. Meta moved from 60 to 31. Mistral from 55 to 18. Fewer companies even showed up. In a single year, participation declined to 30%, down from 74%.
This decline in AI transparency coincides with the accelerating deployment of AI. Search systems rely on it. So do corporate forecasts, clinical screening, and financial models. Outcomes are visible. Design choices are not.
When systems shape decisions at the population scale without inspection, who actually holds power?
The Paradox of Progress
As foundation models become more sophisticated and integrated into critical infrastructure, the companies building them are pulling down the shutters. Stanford’s research team found companies especially opaque about training data, environmental impact, and downstream risks.
Consider the basics: Amazon, Google, Midjourney, Mistral, OpenAI, and xAI don’t even disclose fundamental information that includes input modalities, model size, or architecture. Ten companies disclose no information about their environmental footprint, despite data centers that strain energy grids. Training data remains a black box.
This opacity reflects strategic calculations about competitive advantage, legal exposure, and market positioning.

The Competitive Pressure Cooker
As the AI industry matures, the logic of secrecy intensifies. When GPT-4 or Claude 3 represented quantum leaps in capability, openness about their development didn’t risk much. Competitors couldn’t easily replicate the achievement. But as improvements become more incremental, every detail matters. Small technical choices now matter at the margin. A training method. A data selection rule. A fine-tuning decision. These differences can separate firms that lead markets from those that follow.
The competitive structure has changed. Seven AI-heavy companies represent more than 35% of the S&P 500. OpenAI and Anthropic rank among the highest-valued private companies on record. Capital exposure sits in the hundreds of billions. Under these conditions, transparency is treated as a cost rather than a virtue.
Legal risk reinforces that view. The New York Times has sued OpenAI and Microsoft, claiming unlawful exploitation of copyrighted articles. Similar allegations are now being heard in U.S. courts as judges investigate how existing copyright law applies to large-scale model training. Every disclosure about training data, every admission about what information was used and how, creates potential legal liability. Opacity becomes a form of legal defense.
Business model and disclosure context
Training advanced AI models requires capital outlays measured in the hundreds of millions of dollars. Deployment at scale depends on large infrastructure investments. Information about training data, model architecture, or environmental impact can expose companies to legal review, regulatory inquiry, or competitive disadvantage.
These conditions apply across the sector. Startups and established technology firms operate under similar cost structures and legal constraints. Disclosure practices vary by company. Some organizations provide audit access or governance documentation, while others limit public detail.
Industry behavior reflects existing cost, risk, and market conditions.

The Information Commons Under Siege
The transparency crisis threatens the information ecosystem that made AI possible. AI answer engines are disrupting the web’s economic model in fundamental ways. Consider searching for a lasagna recipe. A decade ago, queries surfaced food blogs with stories, photos, and techniques. Creators earned revenue through ads. Today, AI delivers a synthesized recipe instantly, no click required. The creator gets nothing.
According to the Reuters Institute for the Study of Journalism’s latest report, global Google search traffic to publishers declined by 33% between November 2024 and November 2025, with a 38% declinein the United States alone. This isn’t just affecting recipe bloggers. It’s unraveling the implicit contract that sustained the open web for three decades.
This operates through what researchers call “intent capture” and “substitution.” Platforms capture user intent with on-platform answers, substituting their synthesis for original content. Attribution gets buried in footnotes. Monetization flows to platforms. And critically, the learning loop breaks. As creators lose revenue, they retreat behind paywalls or stop publishing, shrinking the information commons that future AI systems need.
This creates a vicious cycle. As high-quality open information becomes scarce, AI trains on increasingly narrow, degraded datasets. This leads to “model collapse”, when synthetic content replaces real data, systems amplify errors and biases.
Power Without Accountability
The transparency deficit concentrates power in ways that should concern anyone who values democratic oversight. When companies do not disclose what data trains their models, how they mitigate risks, or the impacts their systems create, accountability becomes nearly impossible.
This matters profoundly in B2B contexts. A procurement officer evaluating fraud-detection platforms might once have clicked through to vendor websites, reading whitepapers and case studies. Today, an AI summary condenses the field into a few paragraphs. Vendors see dwindling click-driven sales and retreat from publishing open materials. Some lock content behind steep paywalls, others cut report production entirely, and still others sign exclusive data deals with platforms just to stay visible.
The once-diverse supply of open industry insight contracts into privatized silos. Information asymmetry increases. And the platforms that extract this value become gatekeepers, a potential “Data OPEC,” controlling access to high-quality information as OPEC once controlled oil.
This isn’t hypothetical speculation. We’re observing it happen in real time across industries: journalism, academic research, technical documentation, and specialized expertise. Each domain faces value extraction without compensation, attribution without traffic, citation without revenue.

The Regulatory Response Gap
Governments have introduced regulatory frameworks that include transparency and data access requirements. The European Union’s AI Act and the United Kingdom’s Online Safety Act both address disclosure and researcher access. Implementation and enforcement vary across jurisdictions, particularly where governments also pursue investment from AI companies.
Regulatory objectives differ within the same policy environments. Some governments seek to reduce AI-related risks while supporting domestic growth and international partnerships. These approaches coexist with business models that limit public disclosure.
Industry benchmarking initiatives show mixed results. The World Benchmarking Alliance reports that 71 of 80 participating companies publish AI-related principles. Adoption and execution differ. A significant share of companies did not respond to investor engagement efforts, and fewer than half demonstrate consistent implementation of published commitments.
Past regulatory efforts illustrate potential side effects. The FOSTA and SESTA laws in the United States, which were introduced to combat sex trafficking, resulted in more extensive limitations on internet material. Information on sex work, educational resources, reproductive health materials, and LGBTQ+ content were among the affected areas.
What Artificial Integrity Demands
Moving from band-aid solutions to a framework based on three pillars of what can be referred to as “Artificial Integrity” is necessary to address this crisis:
- Information Provenance Integrity ensures that sources are visible, traceable, and appropriately credited using machine-readable, verifiable metadata. Citations must carry active provenance signatures, not just cosmetic links. This safeguards against erasure and ensures creators receive recognition even when users don’t click through.
- Economic Integrity of Information Flows requires rethinking how citations create value. Today, a link matters only if clicked. In an integrity-based model, being cited in an AI-generated answer would carry economic weight automatically. Platforms would compensate not only for traffic delivered but for information cited, shifting incentives from click-chasing to knowledge contribution.
- The integrity of the Shared Information Commons requires that AI platforms reinvest revenues to sustain open datasets. This isn’t voluntary philanthropy but mandatory reinvestment, architecturally embedded into model development pipelines. Each improvement cycle would refresh and maintain open-access resources while also tuning performance.
Verifiable transparency standards, independent auditing, and real enforcement, combining advantages (tax credits, lower legal risk) with penalties (increasing fines, possible limits) for platforms that disregard integrity requirements, would be necessary for implementation.
Oversight at the Point of Deployment
AI systems are increasingly used across commercial, public, and institutional settings. Information about training data, model structure, and risk controls is not uniformly available. Access to such data varies depending on the business, the legal system, and the terms of the contract.
Dependence on these systems increases as deployment expands across domains such as content mediation, procurement, compliance, and decision support. Disclosure practices have not followed a single pattern. Some organizations provide limited technical documentation or governance statements. Others restrict access or provide only high-level descriptions.
Current conditions reflect a mismatch between system reach and external visibility. Oversight, evaluation, and verification depend on the availability of underlying information. Where access is limited, evaluation is based on secondary indicators rather than direct review.
Oversight ultimately takes shape at the point of deployment. What can be reviewed depends on the regulatory context, contractual access, and the position of the party requesting the information. As a result, similar systems can be subject to very different levels of scrutiny once they enter use.



