If your due-diligence framework for AI vendors assumes that transparency improves over time as the industry matures, the 2025 data shows the opposite. The Foundation Model Transparency Index average fell from 58 in 2024 to 40 in 2025. Most frontier model providers got less transparent in 2025, not more.

Foundation Model Transparency Index scores by major dimensions, 2025:

The Foundation Model Transparency Index (FMTI), now in its third year, scores model developers across three stages of the model lifecycle. Upstream covers what goes into building a model: training data, labour, and compute. Model covers what is disclosed about the system itself: architecture, capabilities, mitigations. Downstream covers what happens after release: distribution, usage policies, post-deployment monitoring, impact reporting.

The 2025 results: IBM leads the index at 95 (Granite 3.3). Writer follows at 72 (Palmyra X5). Anthropic Claude 4 scored 46. Google Gemini 2.5 scored 41. OpenAI o3 scored 38. xAI Grok 3 and Midjourney V7 score 14 each. The weakest area across the index is Upstream: training data and compute disclosure, where most providers score near zero.

The Artificial Analysis Openness Index, a separate measure focused on model weights, methodology disclosure, and training data transparency, paints the same picture from a different angle. Most leading models score between 2 and 16 out of 100. K2 Think and Olmo 3 32B Think score 16, and they are the only two models in the index that scored any points for pre-training data transparency. Every other model scores zero in that category.

Openness Index by components (top frontier models):

The two indices measure different things. FMTI scores developers across the full lifecycle; the Openness Index scores models on weights and data access. Both converge on the same finding. The amount of information frontier AI developers disclose about their models, training data, compute, and post-deployment behaviour declined in 2025 across most providers, with a few outliers (IBM, Writer, K2 Think, Olmo 3) operating substantially above the field.

The structural reasons for the decline are visible in the field.

The first is competitive pressure. As frontier-model capability converged across vendors (see Ch 2), the differentiation moved from "what can the model do?" to "what training data and methodology produced these capabilities?" Disclosing training data composition and post-training methodology now exposes competitive advantage in a way it did not in 2023. The incentive to publish detailed model cards is weaker than it was when capability was the headline differentiator.

The second is regulatory exposure. Detailed disclosure of training data invites legal questions about copyright, fair use, and data provenance. Detailed disclosure of post-training methodology invites legal questions about safety claims and liability. The legal environment of 2025–2026 makes disclosure costlier than it was in 2023, partly because the litigation surface around AI has expanded substantially.

The third is the lack of structural pressure for disclosure. There is no equivalent of GAAP for AI model disclosure. No regulator currently requires frontier AI developers to publish training data composition or methodology details. Without mandatory disclosure, the trajectory is set by voluntary publication, which is dropping as competitive and legal pressures rise.

For procurement and due-diligence teams in 2026, the declining transparency creates a specific problem. Vendor evaluation that depends on disclosed model characteristics (training data composition, fine-tuning methodology, safety evaluations, post-deployment monitoring practices) is now operating against a thinner information layer than the equivalent evaluation did 12–24 months ago. The procurement question has not changed: "is this vendor's model appropriate for our deployment?" The information available to answer it has shrunk.

Three implications for due-diligence frameworks.

The first: due-diligence frameworks should now require disclosure beyond what vendors voluntarily publish. Many enterprises have leverage in procurement conversations to request, through contracts, security questionnaires, or vendor management processes, disclosure that does not appear in public model cards. Vendors that decline to disclose under those conditions are signalling something about the trade-offs they have made. The procurement framework should treat that signal as information.

The second: third-party transparency assessments (FMTI, Openness Index, and emerging equivalents) should be primary inputs to due-diligence, not supplementary. These assessments standardise comparison across vendors in a way that no individual vendor will provide. A procurement framework that anchors on third-party transparency scores is operating on better information than one that relies on vendor-supplied disclosure alone.

The third: the small number of vendors with substantially higher transparency scores (IBM at 95, Writer at 72, K2 Think and Olmo 3 at the top of the Openness Index) are differentiating themselves through disclosure. For enterprises that value due-diligence-ready vendors, the transparency-leader cohort is a meaningful procurement signal. The capability differences between these vendors and the lower-disclosure frontier-tier vendors are smaller than the transparency differences, and the transparency differences are now consequential for governance, compliance, and risk management.

The prescription: rebuild due-diligence frameworks around third-party transparency assessments as primary inputs, contractual disclosure requirements beyond what vendors voluntarily publish, and a procurement preference (where capability allows) for the transparency-leader cohort. The trajectory of declining voluntary disclosure across the field will continue unless regulatory mandates force the trend in the other direction. Due-diligence frameworks that build for the declining-transparency reality will produce better outcomes than ones built on the assumption that the disclosure environment improves over time.

There is a longer view. If voluntary transparency continues to decline through 2026–2027, regulatory pressure will probably increase. The EU AI Act's transparency provisions for general-purpose AI models will phase in. Similar provisions are likely to follow in other jurisdictions. The current trajectory is unstable. It cannot continue indefinitely without regulatory intervention forcing the trend. The defensible planning anchor is that voluntary disclosure declines further before mandatory disclosure forces a reversal. Due-diligence frameworks should be designed to operate well in both regimes, with the contractual-disclosure mechanism as the bridge.

Sources