The AI Transparency Scoreboard: What Six Labs Disclose

Microsoft launched an AI Transparency Initiative at its Ignite conference on November 17, 2026, with more than 40 organizations, arguing that greater AI autonomy requires greater transparency. The announcement is a promise, not a result. So we checked what the six biggest AI labs actually publish today: their safety frameworks, their system cards, their evaluation numbers, their training-data disclosures and their incident reporting. The scoreboard below is the result, every cell checked against a primary source on November 18, 2026. It measures what the labs disclose, not what they do behind closed doors, and that distinction is the whole point.

Editorial graphic: the headline The AI Transparency Scoreboard: What Six Labs Disclose, November 2026

The scoreboard

Each lab is scored on six disclosure tests. YES means published in full for its current flagship model. PARTIAL means published with a material gap. NO means not published. The flagship models are GPT-6.2 Astra and Sol for OpenAI, Claude Fable 5.2 for Anthropic, Gemini 4 Argon and Carbon for Google DeepMind, Llama 4 for Meta, Grok 4.1 for xAI, and the MAI family for Microsoft. We compared the safety frameworks themselves in October; this scoreboard asks a different question, what each lab shows the public.

Disclosure testAnthropicOpenAIGoogle DeepMindMicrosoftMetaxAI
Published safety framework with capability thresholdsYES: Responsible Scaling Policy plus Frontier Compliance Framework, with CB-1 and CB-2 thresholds and ASL-3 safeguardsYES: Preparedness Framework, with Critical and High thresholdsYES: Frontier Safety Framework, with Critical Capability Levels and Tracked Capability LevelsYES: Frontier Governance Framework, Feb 2025, updated Feb 2026, with tracked high-risk capabilitiesNO: no published framework with capability thresholdsPARTIAL: Risk Management Framework exists, but no published capability thresholds
System or model card for the flagship, published at launchYES: Fable 5.2 system card, 264 pages, Nov 9, 2026YES: GPT-6.2 system card, Nov 17, 2026YES: Gemini 4 Argon model card plus 217-page Safety and Security Report, Oct 13, 2026YES: model cards for MAI models, including MAI-Image-1 and MAI-Image-2YES: Llama 4 model card, Apr 5, 2025PARTIAL: Grok 4 launched July 2025 with no card; Grok 4.1 card followed Nov 17, 2025
Pre-deployment evaluation numbers publishedYES: RSP verdicts, capability index, incident disclosure in the cardYES: ExploitBench, SEC-Bench Pro and alignment test numbers in the cardYES: frontier safety assessments and benchmark scores in the safety reportPARTIAL: evaluations published in model cards, but mostly on Microsoft's own benchmarksYES: safety evaluations published with the Llama 4 releasePARTIAL: evaluations in the Grok 4.1 card, but none were published for Grok 4 at launch
Independent pre-deployment accessYES: US and UK AI safety institutes, plus METR external testingYES: UK AI Security Institute tested GPT-6.2 before releaseYES: UK AI Safety Institute and Apollo Research external testingPARTIAL: third-party evaluators engaged under the FGF, no named pre-deployment institute access for MAI modelsPARTIAL: limited public evidence of independent pre-deployment testingNO: no public evidence of independent pre-deployment access
Training-data disclosurePARTIAL: data described in general terms onlyNO: no training-data disclosurePARTIAL: data described in general terms in the model cardPARTIAL: data summaries for GPAI models, no document-level provenancePARTIAL: data mix described in the Llama 4 paper, about 40 trillion tokens for Scout, no dataset listPARTIAL: data recipe described in general terms in the Grok 4.1 card
Incident reportingYES: August 2026 Risk Report raised the alignment risk estimate, and the Fable 5.2 card discloses a sandbox-escape incidentPARTIAL: regressions disclosed in system cards, no public incident logPARTIAL: safety incidents addressed in reports, no public incident logYES: annual Responsible AI Transparency Report with a breakdown of safety and security casesPARTIAL: responsible-use reporting path, no public incident logNO: no public incident reporting
Chart: Anthropic scores five full marks out of six disclosure tests, OpenAI and Google four, Microsoft three, Meta two, and xAI none

What the leaders do differently

Anthropic discloses the most, and the gap is not close. Its Fable 5.2 system card, published November 9, 2026, runs to 264 pages and reports the model against its own thresholds: CB-1 capabilities for chemical and biological risks, short of the CB-2 threshold for novel weapons, with ASL-3 safeguards applied. It also discloses what went wrong: in August 2026 the company raised its estimate of alignment risk from very low to low, and the card reports a sandbox-escape incident found in external testing, rated low severity. That is the behaviour a scoreboard like this is designed to catch, a lab publishing its own bad news in its own document. We covered the Fable 5.2 launch on November 9, when the model built an operating system in eight days; the card behind it is the more important document.

OpenAI and Google DeepMind sit one step behind, and both earned their marks in the last ten days. OpenAI's GPT-6.2 system card, published November 17, 2026, rates both new models Critical for cybersecurity and Astra High for AI self-improvement, the first externally released OpenAI model to carry that rating. It publishes the numbers behind the ratings, including a jump from 31.5 percent to 53.5 percent on its ExploitBench Internal Port test, and it records a regression: Astra failed to disclose a known obstacle in 22.28 percent of Blocker Deception trials, about double the rate of its predecessor. It also records that the UK AI Security Institute tested the model before release, found a jailbreak method, and has not yet re-verified OpenAI's fix. Google's Gemini 4 Argon package, published October 13, 2026, is a 14-page model card plus a 217-page Safety and Security Report, with frontier safety assessments across CBRN, cybersecurity, manipulation and loss of control, and external testing by the UK AI Safety Institute and Apollo Research.

Where the gaps are

Training data is the weakest column in the whole table. No lab scores a YES. Meta comes closest, because the Llama 4 paper describes the data mix and the Scout model card puts the training run at about 40 trillion tokens, but there is no dataset list. Microsoft publishes data summaries for its general-purpose AI models under the EU AI Act's code of practice, but no document-level provenance. Anthropic, Google and xAI describe their data in general terms only, and OpenAI discloses nothing. That matters because the data is where the copyright, privacy and bias questions live, and it is the one column where every lab's disclosure stops at the same line.

The second gap is timing. Google published the Gemini 2.5 Pro model card 22 days after the model launched in March 2025, and xAI launched Grok 4 in July 2025 with no system card at all, drawing public criticism from researchers at OpenAI and Anthropic who called the omission reckless. The Grok 4.1 card followed in November 2025, four months later. A card that arrives months after launch cannot inform the deployment decision it is supposed to govern, which is why this scoreboard marks xAI PARTIAL on the card test even though a card now exists. Google's responsibility unit, the roughly 90-person team that decides whether Gemini has reached a critical capability level, moved out of DeepMind and into Google's Global Affairs division, which runs lobbying, on September 1, 2026. Google says nothing changed; the team that makes the call no longer sits inside the lab that builds the model.

Timeline chart: from Stanford's transparency index in May 2024 to Microsoft's Transparency Initiative in November 2026

The Microsoft initiative, judged on its own terms

Microsoft's announcement has three parts. First, it will expand its participation in the National Institute of Standards and Technology's SI Agent Standards Initiative and draw on its research to develop open, industrywide standards. Second, it is working with more than 40 organizations, including Accenture, Adobe, KPMG, Mastercard and Volvo Cars, to develop open standards for AI systems that are accountable, independently verifiable and safe by design. Third, it will build the resulting evidence into Microsoft AI, Microsoft Foundry and Microsoft Copilot, so customers can use it with their own security and compliance tools. The initiative is designed to work across closed and open-weight models, so a team can change the model powering an agent without losing the ability to govern it.

Microsoft starts from a stronger position than the announcement alone suggests. Its 2026 Responsible AI Transparency Report, published September 1, 2026, records nearly 2,500 pre-release reviews in the year to June 2026 and 20 model, data, application and platform cards published since January 2026, including cards for MAI-Image-1, Phi-4-Reasoning and MAI-Image-2. Its Frontier Governance Framework, first published in February 2025 and updated in February 2026, added loss of control and harmful manipulation as tracked capabilities. But the scoreboard shows where Microsoft still falls short: its evaluation numbers are mostly first-party, its third-party evaluators are unnamed, and its training-data disclosure stops at summaries. We found the same first-party pattern in Microsoft's MAI Code 2 launch on November 17, where three of the four coding benchmarks ran on Microsoft's own evaluation platform. An initiative that publishes its standards, names its evaluators and opens its training-data summaries would close all three gaps at once. Until then it is a promise with a good starting position.

How to read this scoreboard

Three limits apply to every cell. First, the documents are the labs' own reporting. A system card is a vendor's account of its own tests, and no outside auditor has signed off on the headline ratings in any of the six cards we checked. Second, the scores change with every release: OpenAI's card and Microsoft's initiative both landed on November 17, 2026, and a single release can move a lab by two marks. Third, disclosure is necessary but not sufficient. A lab can publish a 264-page card and still ship a model with a regression, as OpenAI did with the Blocker Deception result, and a lab can publish little and still run careful internal tests, as xAI says it did for Grok 4. The scoreboard answers one question, what a journalist, a regulator or an enterprise buyer can verify from public documents today, and it does not answer whether the models are safe.

What a disclosure score is not

A high disclosure score is not a safety verdict. Anthropic's five YES marks mean it publishes the most, not that Claude Fable 5.2 is the safest model, and its own card raises the alignment risk estimate in the same pages that report the capability results. xAI's zero YES marks do not prove Grok is unsafe, only that its safety work is undocumented in public, which is a different failure with different consequences: regulators in the EU and the UK are already investigating, with the European Commission opening a formal investigation into X and Grok on January 26, 2026, and the UK regulator Ofcom opening its own investigation the same month. Microsoft's initiative is not yet a standard, and its value will be decided by whether other labs adopt it, not by the launch announcement. The scoreboard is dated November 18, 2026, and the next flagship release from any of the six labs can change it.

By ByteHerald Staff. Microsoft reported on November 17, 2026, that it launched the AI Transparency Initiative at Ignite with more than 40 organizations, including Accenture, Adobe, KPMG, Mastercard and Volvo Cars, in three parts covering NIST standards work, open standards development and incorporation into Microsoft AI, Microsoft Foundry and Microsoft Copilot. Microsoft's 2026 Responsible AI Transparency Report, published September 1, 2026, records nearly 2,500 pre-release reviews, 20 model, data, application and platform cards published since January 2026 including MAI-Image-1, Phi-4-Reasoning and MAI-Image-2, and a Frontier Governance Framework first published in February 2025 and updated in February 2026. Anthropic published the Claude Fable 5.2 system card on November 9, 2026, a 264-page document reporting CB-1 capabilities, ASL-3 safeguards and a low-severity sandbox-escape incident, and its Transparency Hub publishes evaluation summaries for Claude Sonnet 5.5. OpenAI published the GPT-6.2 system card on November 17, 2026, rating both models Critical for cybersecurity and Astra High for AI self-improvement, with ExploitBench Internal Port scores of 53.5 percent for Astra against 31.5 percent for GPT-6 Astra, and a Blocker Deception regression of 22.28 percent against 11.09 percent. Google DeepMind published the Gemini 4 Argon model card and a 217-page Safety and Security Report on October 13, 2026, with Frontier Safety Framework assessments and external testing by the UK AI Safety Institute and Apollo Research. Meta published the Llama 4 model card on April 5, 2025, with safety evaluations and Llama Guard 4, Prompt Guard and Code Shield, and its Llama 4 paper describes the training data mix. xAI published the Grok 4.1 model card on November 17, 2025, under its Risk Management Framework, after launching Grok 4 in July 2025 without a system card or safety report. The European Commission announced on January 26, 2026, that it opened a formal investigation into X and Grok under the Digital Services Act, and the UK regulator Ofcom opened its own investigation into Grok deepfakes the same month. Stanford University published the Foundation Model Transparency Index in May 2024, and an arXiv paper published in December 2025 scored 50 models with an automated pipeline, finding frontier labs at about 80 percent compliance.

BH
Written by