What is Capital One's approach to responsible AI?
Capital One has no single published 'AI principles' document; its responsible-AI position is spread across research posts, testimony, filings and partnerships. The through-line since 2018 is that credit AI must be fair in inputs and outcomes and explainable to regulators and customers, and that oversight runs through the bank's existing model risk management, compliance, fair-lending and data-privacy governance rather than a separate AI charter, as chief scientist Prem Natarajan told the Senate AI Insight Forum in November 2023. In practice that means a 10-K risk factor on biased or 'hallucinatory' AI output, a public disclosure of how its generative models are trained (more than 20 trillion tokens of open-source and synthetic data), guardian-model and input-guardrail research (an AAAI 2025 Outstanding Paper), platform-first governance for agents with human-in-the-loop controls for high-risk actions, membership of the Partnership on AI since February 2024, and about $14.5 million committed to responsible-AI centers at Columbia, USC, Illinois and UVA plus $5 million to the NSF. Evident's October 2026 index again placed Capital One second among 50 banks, and reported that all but one of them now have responsible AI principles.[1][2][3][4][5][6][7][8][9][10][11][12][13][14]
Data as of Oct 15, 2026 · Capital One profile updated Oct 1, 2026 · 33 sources cited on this page · 140 banks profiled on this site
What has Capital One published on responsible AI?
Public statements, research and disclosures, oldest first. 'Who' is the named author or speaker where the piece has one.
| Publication | What it says | Who | Date |
|---|---|---|---|
| Fairness and explainability across the banking industry | Fair-lending law requires fairness in inputs and outcomes; the two legal notions can conflict mathematically; the bank intends to 'help set the standard' rather than just comply, and co-hosted a NeurIPS 2018 workshop on fairness, explainability, accuracy and privacy | Capital One ML research | November 15, 2018[1] |
| An integral, pragmatic approach to explainable AI | A 'humanity- and responsibility-centered' philosophy; explainability levers through the whole lifecycle; trade-offs between explainability and accuracy acknowledged | Margaret Mayer, VP Software Engineering, Conversational AI | February 21, 2019[15] |
| Global Attribution Mapping (open source) | An explainability method for neural networks released as open source | Capital One ML research | March 13, 2019[16] |
| A data scientist and a sociologist on responsible AI | Interdisciplinary working groups inside the Center for Machine Learning as a mechanism for responsible development | Center for Machine Learning | June 3, 2019[17] |
| Senate AI Insight Forum remarks | Four pillars: workforce training, research partnerships, risk-management frameworks and inclusion; AI risk managed through model risk management, compliance and governance; fair-lending and privacy processes applied to each use case; 'human-centered guardrails' before deployment | Prem Natarajan, Chief Scientist | November 1, 2023[2] |
| Joins the Partnership on AI | Membership of the multi-stakeholder non-profit; Natarajan cites 'multidisciplinary, multi-stakeholder frameworks' as the bank's belief | Prem Natarajan | February 27, 2024[7] |
| Input guardrails for safer LLMs | Fine-tuned open-source Mistral and Llama judge models to detect jailbreaks and prompt injection; base models scored under 80% F1 and fine-tuning lifted F1 and attack-detection rate by more than 50%; AAAI 2025 PDLM workshop Outstanding Paper Award | AI Foundations team | June 10, 2025 (paper January 2025)[5] |
| 2025 Form 10-K | Risk factor on 'our evolving use of AI': generative AI 'has been known to produce false or hallucinatory inferences'; biased or unethical outputs could bring liability; AI output may not be copyrightable; agentic commerce may shift fraud liability | Capital One Financial Corporation | February 19, 2026[3] |
| ICLR 2026 research | Gold sponsor; papers on LLM alignment, uncertainty quantification and responsible agentic systems, including the DynaGuard guardian model and differentially private synthetic data as a 'safe gym' for agents | AI Foundations and academic partners | April 20, 2026[18] |
| VulnHunter open-sourced | Agentic code-security tool released under Apache 2.0 because 'the scale of the AI threat is larger than any single organization' | Chris Nims, CISO | July 16, 2026[19] |
| GAI Training Data Transparency Disclosure | Generative models built by customising open-source LLMs with synthetic and publicly available data; training sets exceed 20 trillion tokens; 7,100-plus synthetic conversations | Capital One | Retrieved September 24, 2026[4] |
| Agentic-commerce principles | Co-authored with Bank of America, CBA, ING, NatWest and ASB: transparency, safety, privacy and data, choice and interoperability | Six banks | September 22, 2026[20] |
| Agentic AI governance, platform-first | Rule-based validation, sandbox simulations, observability of agent trajectories, standardised harnesses and human-in-the-loop controls for high-risk actions, built before developers build agents | Rashmi Shetty, VP Enterprise AI | September 28, 2026[6] |
How is AI governed inside Capital One?
Mechanisms the bank has described, with the strongest source for each. 'Stated' means a named executive said it; 'Disclosed' means it is in a filing.
| Mechanism | What the bank says | Status | Source date |
|---|---|---|---|
| Model Risk Management framework | 'A Model Risk Management framework to evaluate, assess, validate, and govern models' is a best practice the bank applies before AI reaches customers; model risk sits under the Operational Risk Management Program reported to the executive committee and the board | Stated and disclosed | November 2023; February 2026[2][3] |
| Fair-lending and data-privacy governance per use case | 'Established fair lending and data privacy governance processes to ensure we approach each use case in a responsible, well managed way' | Stated | November 2023[2] |
| Build-not-buy ownership of models | Open-source LLMs customised in-house 'to meet our high performance, risk, and governance thresholds'; validation and explanation agents sit inside the multi-agent workflow | In production | March 2025; August 2026[21][22] |
| Platform-first agent controls | Validation, sandboxing, trajectory observability, standard harnesses and human review of high-risk actions, provided by the platform rather than retrofitted | Stated | September 2026[6] |
| Named accountable executives | Prem Natarajan (Chief Scientist and Head of Enterprise AI); Andy Ozment (Chief Technology Risk Officer); Chris Nims (CISO); Robert Alexander (CIO); Amy Lenander (Chief Data Officer); the 10-K describes the CISO's and CTRO's backgrounds | Disclosed | February 2026[3][23][19][24] |
| Training-data transparency | A public page describing the data behind its generative models | Published | September 2026[4] |
| Interdisciplinary review | Sociologists, designers, legal scholars and community advocates in internal AI/ML discussions (2019); 'diverse perspectives and equities' through industry-academia-government partnerships (2023) | Stated | 2019; 2023[17][2] |
| Industry benchmark | Evident's October 2026 index: Capital One second of 50 banks for a fourth year; all but one bank reports responsible AI principles (16 in 2023); AI governance talent across the 50 up 33% year on year | External | October 6, 2026[13][14] |
Which responsible-AI research centers does Capital One fund?
Commitments as announced by Capital One or the university. Amounts are pledges over the stated term, not spend to date.
| Center | Partner | Commitment | Announced |
|---|---|---|---|
| Center for AI and Responsible Financial Innovation (CAIRFI) | Columbia Engineering | $3 million: PhD fellowships, faculty research, public symposia; kicked off March 28, 2024 with 50-plus researchers; 2025 awards include 'trustworthy lending: explaining group differences' | March 15, 2024[8][25][26][27] |
| Center for Responsible AI and Decision Making in Finance (CREDIF) | University of Southern California | $3 million; director Petros Ioannou; first call for proposals August-October 2024; spring 2025 awards on LLM explainability in multi-agent systems and knowledge-graph RAG | April 11, 2024[9][28] |
| Center for Generative AI Safety, Knowledge Systems and Cybersecurity (ASKS) | University of Illinois Urbana-Champaign | $3 million over five years for faculty projects and PhD fellowships; director Heng Ji; builds on a tech incubator lab at Illinois since 2015 | June 4, 2025[10] |
| National AI Research Institutes (NSF ARIA) | National Science Foundation | $1 million a year for five years from 2025 ($5 million) toward NSF's $100 million for five new AI institutes; focus on 'grounding, instructability and alignment' | July 29, 2025[12][2] |
| Capital One AI Research Neighborhood and PhD Fellowships | University of Virginia Engineering | $4.5 million: $2 million facility (matched by UVA), $500,000 fellowships; adds to the 2024 Capital One Hub at the School of Data Science and 2025-27 faculty fellows on trustworthy and explainable ML | October 27, 2025[11][29] |
| Partnership on AI | PAI (non-profit) | Membership | February 27, 2024[7] |
What does the Capital One profile record on this topic?
The rows below are the entries of the Capital One AI strategy profile that bear on this page, as of Oct 1, 2026; the profile holds the full record of 31 dated moves, 13 use cases and 46 sources.
| Use case | What the record shows | Status | Source |
|---|---|---|---|
| AI and model risk in the 10-K | The 2025 10-K names models and AI in pricing, fraud detection, loan grading, credit decisions and software development as a risk, plus agentic-commerce fraud and AI intellectual-property uncertainty. | In production | [3] |
| Date | Move | Detail | Source |
|---|---|---|---|
| Sep 28, 2026 | Shetty: 'agentic AI now must be treated as an end-to-end system' | VP of enterprise AI Rashmi Shetty says 14 years of data-infrastructure work gives agents a strong foundation and context, and that the bank adopts platform-first governance — rule-based validation, sandbox simulations, observability of agent trajectories, standardised harnesses and human-in-the-loop controls for high-risk actions — before developers build agents, rather than retrofitting it later. | [6] |
| Sep 22, 2026 | Agentic-commerce principles with five other banks | With Bank of America, CBA, ING, NatWest and ASB, a paper on transparency, safety, privacy and data, choice and interoperability. | [20] |
| Apr 21, 2026 | Q1 2026 call: AI leverage is 'vastly greater' when embedded | Fairbank: 14th year of the technology transformation; continued investment in AI infrastructure and specific AI experiences. | [30] |
| Apr 20, 2026 | ICLR 2026: alignment and guardian-model research | Gold sponsor and contributor on LLM alignment, uncertainty and responsible agentic systems, including the DynaGuard guardian model. | [18] |
| Feb 19, 2026 | 2025 10-K: 'our evolving use of AI' is a risk factor | Models and AI in pricing, fraud detection, loan grading, credit and software development; agentic-commerce fraud liability; uncertain IP protection for AI output. | [3] |
| Jan 22, 2026 | Q4 2025 call: a stack built 'working backwards from the AI revolution' | Fairbank says the bank is 'now building AI solutions across our businesses' and that Brex has deployed in-house AI agents for expense management and audit. | [31] |
| Nov 20, 2025 | Applied AI research agenda published | LLM training, customisation and alignment; synthetic data; multi-agent reasoning. | [32] |
| Oct 21, 2025 | Q3 2025 call: AI 'deeply embedded', not bolted on | Fairbank: AI-driven experiences are 'a new front' in premium cards; the opportunities require 'significant investment in AI and AI talent'. | [33] |
| Jun 4, 2025 | ASKS Center for generative-AI safety at Illinois | Capital One and the University of Illinois launch the Center for Generative AI Safety, Knowledge Systems and Cybersecurity; AI Foundations SVP Milind Naphade among its contributors. | [10] |
Does Capital One publish AI principles?
Not as a standalone document. Its stated principles are scattered: fairness in both inputs and outcomes and explainability for credit models (2018), a 'humanity- and responsibility-centered' philosophy (2019), and, in Prem Natarajan's November 2023 Senate testimony, four pillars of workforce training, research partnerships, risk-management frameworks and inclusion, with 'human-centered guardrails' and a model risk management framework before any AI reaches customers. The bank also co-authored six banks' agentic-commerce principles in September 2026 (transparency, safety, privacy and data, choice, interoperability). Evident reports that 49 of the 50 largest banks say they have responsible AI principles in place.
How does Capital One govern AI risk?
Through the bank's existing risk architecture rather than a separate AI charter. The 2025 10-K places model risk inside the Operational Risk Management Program, reported to senior leaders, the executive committee and the board, and treats 'our evolving use of AI' as a named risk factor covering pricing, fraud, loan grading, credit, software development and personalisation. Natarajan has said fair-lending and data-privacy governance processes are applied to each use case. For agents, Rashmi Shetty described platform-level validation, sandbox simulation, observability of agent trajectories and human-in-the-loop controls for high-risk actions, built before developers build agents.
What does Capital One disclose about its generative-AI training data?
A public GAI Training Data Transparency Disclosure says its generative models are built by customising open-source large language models with synthetic and publicly available data, that the training sets exceed 20 trillion tokens, and that more than 7,100 synthetic conversations were used. Chat Concierge is disclosed as built on Meta's Llama customised with proprietary data. The 10-K adds that AI output 'may not be subject to copyright or other intellectual property protection'.
What responsible-AI research has Capital One published?
An input-guardrails paper that fine-tuned open-source Mistral and Llama judge models to detect jailbreaks and prompt injection, raising F1 and attack-detection rates by more than 50% over base models and winning the Outstanding Paper Award at the AAAI 2025 PDLM workshop; the DynaGuard guardian model and work on LLM alignment, uncertainty quantification and differentially private synthetic data as a 'safe gym' for agents at ICLR 2026; more than 20 accepted papers at NeurIPS 2025 spanning responsible AI, multi-agent reasoning and fairness; and the open-source Global Attribution Mapping explainability method (2019).
Who provides diverse perspectives and oversight on Capital One's AI development?
Internally, the bank described interdisciplinary discussions inside its Center for Machine Learning with sociologists, designers, legal scholars and community advocates (2019). Externally, Natarajan told the Senate the bank believes in industry-academia-government partnerships 'to ensure diverse perspectives and equities when developing, testing, and deploying AI', and the bank funds centers at Columbia, USC, Illinois and UVA, belongs to the Partnership on AI, and co-advises NSF's AI Research Institutes. Board-level oversight of model risk runs through the Operational Risk Management Program.
How does Capital One compare with other banks on responsible AI?
Evident's October 6, 2026 index placed Capital One second of 50 global banks for the fourth year running, behind JPMorganChase and ahead of RBC, scoring talent, innovation, leadership and transparency. Evident found that all but one of the 50 now report responsible AI principles, up from 16 in 2023, and that AI-governance talent grew 33% year on year. Capital One's distinctive features are ownership of its models, published training-data transparency and the volume of funded academic research; what it lacks is a single public principles document or an audit of its customer-facing agents.
Who leads responsible AI at Capital One?
Prem Natarajan, EVP, Chief Scientist and Head of Enterprise AI, is the public voice (Senate forum 2023, Partnership on AI 2024, the university centers). Milind Naphade, SVP and Head of AI Foundations, leads the research team behind the guardrail and guardian-model papers. Andy Ozment is Chief Technology Risk Officer, Chris Nims Chief Information Security Officer, Amy Lenander Chief Data Officer and Robert Alexander Chief Information Officer; Rashmi Shetty, VP of Enterprise AI, has described agent governance.
- Advancing Greater Fairness and Explainability for AI and Machine Learning Across the Banking Industry
- Capital One participates in U.S. Senate AI Insight Forum (remarks by Prem Natarajan)
- Capital One Financial Corporation Annual Report on Form 10-K for 2025
- GAI Training Data Transparency Disclosure
- Refining input guardrails for safer LLM applications (AAAI 2025 PDLM workshop, Outstanding Paper Award)
- Capital One's agentic AI strategy hinges on data, platform-first mindset
- Capital One joins the Partnership on AI
- Capital One and Columbia University partner to establish the Center for AI and Responsible Financial Innovation (CAIRFI)
- Capital One and USC launch the Center for Responsible AI and Decision Making in Finance (CREDIF)
- Capital One and University of Illinois launch ASKS Center
- Capital One deepens UVA partnership to advance AI research
- Capital One partners with NSF to advance U.S. AI leadership
- JPMorganChase still tops AI ranking as big-bank lead widens
- Banks prioritize responsible AI, governance as use soars
- An integral, pragmatic approach to explainable AI (Margaret Mayer)
- Global Attribution Mapping (GAM): open-sourcing explainability for neural networks
- Re-imagining innovation with responsible AI: a conversation with a data scientist and a sociologist
- NLP research foundations at ICLR 2026
- Capital One releases VulnHunter, an open-source AI tool that finds software flaws before hackers do
- Are banks falling behind on agentic commerce standards?
- AI for Car Dealerships: Smarter Car Shopping
- Why Capital One built its multi-agent AI platform around open-weight models (presented by Capital One)
- How North Korean operatives infiltrate US bank payrolls
- Context is key to surpass AI's 40% success: Gartner D&A 2026
- Capital One and Columbia University Launch Center for AI and Responsible Financial Innovation
- Kicking Off the Center of Responsible AI and Finance
- Capital One and Columbia University: 2025 CAIRFI AI research award recipients
- CREDIF spring 2025 AI awards and fellowship recipients (USC)
- UVA School of Data Science: Capital One Fellows 2025-2027
- Capital One (COF) Q1 2026 Earnings Transcript
- Capital One (COF) Q4 2025 Earnings Call Transcript
- Capital One's Applied AI research
- Capital One (COF) Q3 2025 Earnings Call Transcript
When Capital One moves on AI, you'll read it here first.
when one of the banks we cover moves on AI, it is in the next morning's brief · six sourced stories · 7 am ET · free
plus every tracker, bank and agent page update, the morning after · leave any morning