# The AI Risk Mitigation Repository | Peter Slattery, PhD **By:** Peter Slattery, PhD **Published:** 2026-11-12T17:27:42.643Z **Source:** [LinkedIn](https://www.linkedin.com/posts/peterslattery1_the-ai-risk-mitigation-repository-ugcPost-7526697242187968513-rQ4u) --- # Peter Slattery, PhD’s Post [Peter Slattery, PhD](https://www.linkedin.com/in/peterslattery1?trk=public_post_feed-actor-name) MIT AI Risk Initiative | MIT FutureTech 📢 The [MIT AI Risk Initiative](https://www.linkedin.com/showcase/mit-ai-risk-initiative/?trk=public_post-text) just released the third version of our AI Risk Mitigation Taxonomy: 2,206 mitigation strategies for AI risks, extracted from 47 frameworks across technical, policy, and governance domains, shared in a database and interactive explorer. 🔍 What we did: We systematically reviewed the academic and gray literature for frameworks, taxonomies, and structured lists of AI risk mitigations. We extracted and classified each by who acts, when, against which risk, what it aims to achieve, and the lever used (e.g., technical artifacts, law, audits etc).  🧭 What we found: 1️⃣ Reducing AI risk extends far beyond improving the technology itself. While 55 percent of proposed mitigations focus directly on AI systems, researchers found that many others target organizational processes, governance structures, public policy, and broader societal systems 2️⃣ Responsibility is spread across the AI ecosystem. Developers account for 41% of mitigations, deployers 33%, and governance actors such as regulators 18%. 3️⃣ Mitigation coverage does not fully track expert priorities. Our Delphi study of 272 experts rated competitive dynamics and power centralization among the five risks most likely to cause severe or catastrophic harm by 2030, but only 15 mitigation records directly addressed them. 💡 Recommendation: We find that mitigations are difficult to compare because of inconsistent descriptions (only 3.4% of records shared names). To address this, we present a new tool called ‘PRIMER’, an eight-item reporting checklist that includes the mitigation, the risk it addresses, the implementer, and when, and how, it is implemented. 🔗 How can you engage? Please try our interactive explorer (see links in comments) which allows you to search for mitigations, browse categories, and examine mapping to AI risks, responsible actor, and stage of the AI lifecycle. The tool also enables users to compare mitigations across different taxonomies to identify patterns, areas of concentration, and gaps. We'd love developers and deployers to explore reporting mitigations with PRIMER and provide us with feedback.  🙏 Please feel free to share this work with anyone relevant. Thanks to the co-first authors [Alexander Saeri](https://au.linkedin.com/in/aksaeri?trk=public_post-text) and [Kun Zhao](https://au.linkedin.com/in/dr-kun-zhao?trk=public_post-text), and also [Michael Noetel](https://au.linkedin.com/in/mnoetel?trk=public_post-text), [Jess Graham](https://au.linkedin.com/in/jessica-jane-graham?trk=public_post-text), [Aileen Liang](https://www.linkedin.com/in/aileen-liang?trk=public_post-text), Catherine Simons, [Elisa Zhang](https://www.linkedin.com/in/elisa-zhangg?trk=public_post-text), [Fatimeh Al Ghannam](https://www.linkedin.com/in/fatimeh-al-ghannam?trk=public_post-text), [Soniya Agrawal](https://www.linkedin.com/in/soniya-agrawal-2a3001168?trk=public_post-text), [Spencer Michaels](https://www.linkedin.com/in/spencer-michaels?trk=public_post-text), [Neil Thompson](https://www.linkedin.com/in/neil-thompson-5724a614?trk=public_post-text), and the 40+ experts who helped. Tagging a few people who might be interested: [Kevin Fumai](https://www.linkedin.com/in/kevinfumai?trk=public_post-text) [Yonah Welker](https://ch.linkedin.com/in/welker?trk=public_post-text) [Sam Burrett](https://au.linkedin.com/in/sam-burrett?trk=public_post-text) [Chris Kraft](https://www.linkedin.com/in/ctkraft?trk=public_post-text) [Kuba Szarmach](https://pl.linkedin.com/in/jakubszarmach?trk=public_post-text) [Tony Moroney](https://ie.linkedin.com/in/tonymoroney?trk=public_post-text) [Nafis Alam](https://my.linkedin.com/in/nafisalam?trk=public_post-text) [Luiza Jarovsky, PhD](https://il.linkedin.com/in/luizajarovsky?trk=public_post-text), [Oliver Patel,](https://uk.linkedin.com/in/oliver-patel?trk=public_post-text) [Samuel Salzer](https://se.linkedin.com/in/samuelsalzer?trk=public_post-text) [Phil Venables](https://www.linkedin.com/in/philvenables?trk=public_post-text) [Renan Araujo](https://uk.linkedin.com/in/araujonrenan?trk=public_post-text) [Kevin Klyman](https://www.linkedin.com/in/kevin-klyman?trk=public_post-text) [Peter Slattery, PhD](https://www.linkedin.com/in/peterslattery1?trk=public_post_comment_actor-name) MIT AI Risk Initiative | MIT FutureTech See the tools and paper here: [https://airisk.mit.edu/ai-risk-mitigations](https://airisk.mit.edu/ai-risk-mitigations?trk=public_post_comment-text) 2,206 mitigations is a map nobody can execute. If a team can fund five this quarter, which five does the taxonomy point them to? The scale of this is impressive, [Peter](https://www.linkedin.com/in/peterslattery1?trk=public_post_comment-text). What stands out most is how clearly it shows that AI risk is not just a technical problem. Governance, ownership and organisational controls have to mature alongside the technology. This is interesting, but I believe the distinction between identifying risks and having clear ownership for addressing them is important, as mitigation only works when someone is accountable for implementing it. [Peter](https://www.linkedin.com/in/peterslattery1?trk=public_post_comment-text) Thank you for this — the finding that only 3.4% of mitigations are described consistently enough to compare is the one that matters most for practitioners. In our work on auditable AI, we see the same pattern: a mitigation that can't be described precisely enough to be checked claim-by-claim can't really be verified either. Does PRIMER score mitigations on the strength of evidence behind them, or mainly on how consistently they're described? I'd be curious whether adding an evidence-grading layer — what would have to be true, and shown, for a mitigation to count as working — would help deployers choose between the 2,206 options rather than just catalogue them. That's a great repository and it really helped me in shaping my research on critical infrastructure resilience ! Thank you very much for this great contribution ! Peter, this is impressive work, particularly the effort to create a common frame of reference across technical, operational, governance and accountability approaches. One question it raises for me is what comes after a mitigation is selected and implemented. As AI systems, permissions, integrations and capabilities change over time, how do we establish that a particular mitigation was actually in force, applicable to the system operating at that moment, and functioning as intended when a consequential action occurred? It seems that mitigation and verification ultimately have to meet somewhere: not only what safeguards were prescribed, but what can later be independently established about the system, the authority it possessed, the controls that applied, and what actually happened. Fascinating work. I look forward to exploring the taxonomy. Thank you for sharing this, I am still learning about AI governance. I was surprised that so much of the risk work is not just about the technology itself. Where would a newcomer start with the explorer? Now we just need an AI to read the whole repository for us. What could go wrong? - New work from the MIT AI Research Initiative: We just released the third version of our AI Risk Mitigation Taxonomy: 2,206 mitigation strategies for AI risks, extracted from 47 frameworks across technical, policy, and governance domains, shared in a database and interactive explorer. [Peter Slattery, PhD](https://www.linkedin.com/in/peterslattery1?trk=public_post_reshare_feed-actor-name) MIT AI Risk Initiative | MIT FutureTech 📢 The [MIT AI Risk Initiative](https://www.linkedin.com/showcase/mit-ai-risk-initiative/?trk=public_post_reshare-text) just released the third version of our AI Risk Mitigation Taxonomy: 2,206 mitigation strategies for AI risks, extracted from 47 frameworks across technical, policy, and governance domains, shared in a database and interactive explorer. 🔍 What we did: We systematically reviewed the academic and gray literature for frameworks, taxonomies, and structured lists of AI risk mitigations. We extracted and classified each by who acts, when, against which risk, what it aims to achieve, and the lever used (e.g., technical artifacts, law, audits etc).  🧭 What we found: 1️⃣ Reducing AI risk extends far beyond improving the technology itself. While 55 percent of proposed mitigations focus directly on AI systems, researchers found that many others target organizational processes, governance structures, public policy, and broader societal systems 2️⃣ Responsibility is spread across the AI ecosystem. Developers account for 41% of mitigations, deployers 33%, and governance actors such as regulators 18%. 3️⃣ Mitigation coverage does not fully track expert priorities. Our Delphi study of 272 experts rated competitive dynamics and power centralization among the five risks most likely to cause severe or catastrophic harm by 2030, but only 15 mitigation records directly addressed them. 💡 Recommendation: We find that mitigations are difficult to compare because of inconsistent descriptions (only 3.4% of records shared names). To address this, we present a new tool called ‘PRIMER’, an eight-item reporting checklist that includes the mitigation, the risk it addresses, the implementer, and when, and how, it is implemented. 🔗 How can you engage? Please try our interactive explorer (see links in comments) which allows you to search for mitigations, browse categories, and examine mapping to AI risks, responsible actor, and stage of the AI lifecycle. The tool also enables users to compare mitigations across different taxonomies to identify patterns, areas of concentration, and gaps. We'd love developers and deployers to explore reporting mitigations with PRIMER and provide us with feedback.  🙏 Please feel free to share this work with anyone relevant. Thanks to the co-first authors [Alexander Saeri](https://au.linkedin.com/in/aksaeri?trk=public_post_reshare-text) and [Kun Zhao](https://au.linkedin.com/in/dr-kun-zhao?trk=public_post_reshare-text), and also [Michael Noetel](https://au.linkedin.com/in/mnoetel?trk=public_post_reshare-text), [Jess Graham](https://au.linkedin.com/in/jessica-jane-graham?trk=public_post_reshare-text), [Aileen Liang](https://www.linkedin.com/in/aileen-liang?trk=public_post_reshare-text), Catherine Simons, [Elisa Zhang](https://www.linkedin.com/in/elisa-zhangg?trk=public_post_reshare-text), [Fatimeh Al Ghannam](https://www.linkedin.com/in/fatimeh-al-ghannam?trk=public_post_reshare-text), [Soniya Agrawal](https://www.linkedin.com/in/soniya-agrawal-2a3001168?trk=public_post_reshare-text), [Spencer Michaels](https://www.linkedin.com/in/spencer-michaels?trk=public_post_reshare-text), [Neil Thompson](https://www.linkedin.com/in/neil-thompson-5724a614?trk=public_post_reshare-text), and the 40+ experts who helped. Tagging a few people who might be interested: [Kevin Fumai](https://www.linkedin.com/in/kevinfumai?trk=public_post_reshare-text) [Yonah Welker](https://ch.linkedin.com/in/welker?trk=public_post_reshare-text) [Sam Burrett](https://au.linkedin.com/in/sam-burrett?trk=public_post_reshare-text) [Chris Kraft](https://www.linkedin.com/in/ctkraft?trk=public_post_reshare-text) [Kuba Szarmach](https://pl.linkedin.com/in/jakubszarmach?trk=public_post_reshare-text) [Tony Moroney](https://ie.linkedin.com/in/tonymoroney?trk=public_post_reshare-text) [Nafis Alam](https://my.linkedin.com/in/nafisalam?trk=public_post_reshare-text) [Luiza Jarovsky, PhD](https://il.linkedin.com/in/luizajarovsky?trk=public_post_reshare-text), [Oliver Patel,](https://uk.linkedin.com/in/oliver-patel?trk=public_post_reshare-text) [Samuel Salzer](https://se.linkedin.com/in/samuelsalzer?trk=public_post_reshare-text) [Phil Venables](https://www.linkedin.com/in/philvenables?trk=public_post_reshare-text) [Renan Araujo](https://uk.linkedin.com/in/araujonrenan?trk=public_post_reshare-text) [Kevin Klyman](https://www.linkedin.com/in/kevin-klyman?trk=public_post_reshare-text) - [Chase P.](https://www.linkedin.com/in/petersonchase?trk=public_post_feed-actor-name) Stuff like this is absolute catnip to me. 🐱 Anyone can ship fast. Watching how different industries engineer control, containment, and safety when the dust settles? It’s easy to get swept up in the hype of what new tech can do. The difficulty and intriguing bit is to ask how we actually control and contain it. MIT’s risk mitigation breakdown is a masterclass in how different fields are solving that exact problem. I know my evening reading is squared away 😁 [#AIRisk](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fairisk&trk=public_post-text) [#TechLeadership](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Ftechleadership&trk=public_post-text) [#AISafety](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Faisafety&trk=public_post-text) [#RiskMitigation](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Friskmitigation&trk=public_post-text) [Peter Slattery, PhD](https://www.linkedin.com/in/peterslattery1?trk=public_post_reshare_feed-actor-name) MIT AI Risk Initiative | MIT FutureTech 📢 The [MIT AI Risk Initiative](https://www.linkedin.com/showcase/mit-ai-risk-initiative/?trk=public_post_reshare-text) just released the third version of our AI Risk Mitigation Taxonomy: 2,206 mitigation strategies for AI risks, extracted from 47 frameworks across technical, policy, and governance domains, shared in a database and interactive explorer. 🔍 What we did: We systematically reviewed the academic and gray literature for frameworks, taxonomies, and structured lists of AI risk mitigations. We extracted and classified each by who acts, when, against which risk, what it aims to achieve, and the lever used (e.g., technical artifacts, law, audits etc).  🧭 What we found: 1️⃣ Reducing AI risk extends far beyond improving the technology itself. While 55 percent of proposed mitigations focus directly on AI systems, researchers found that many others target organizational processes, governance structures, public policy, and broader societal systems 2️⃣ Responsibility is spread across the AI ecosystem. Developers account for 41% of mitigations, deployers 33%, and governance actors such as regulators 18%. 3️⃣ Mitigation coverage does not fully track expert priorities. Our Delphi study of 272 experts rated competitive dynamics and power centralization among the five risks most likely to cause severe or catastrophic harm by 2030, but only 15 mitigation records directly addressed them. 💡 Recommendation: We find that mitigations are difficult to compare because of inconsistent descriptions (only 3.4% of records shared names). To address this, we present a new tool called ‘PRIMER’, an eight-item reporting checklist that includes the mitigation, the risk it addresses, the implementer, and when, and how, it is implemented. 🔗 How can you engage? Please try our interactive explorer (see links in comments) which allows you to search for mitigations, browse categories, and examine mapping to AI risks, responsible actor, and stage of the AI lifecycle. The tool also enables users to compare mitigations across different taxonomies to identify patterns, areas of concentration, and gaps. We'd love developers and deployers to explore reporting mitigations with PRIMER and provide us with feedback.  🙏 Please feel free to share this work with anyone relevant. Thanks to the co-first authors [Alexander Saeri](https://au.linkedin.com/in/aksaeri?trk=public_post_reshare-text) and [Kun Zhao](https://au.linkedin.com/in/dr-kun-zhao?trk=public_post_reshare-text), and also [Michael Noetel](https://au.linkedin.com/in/mnoetel?trk=public_post_reshare-text), [Jess Graham](https://au.linkedin.com/in/jessica-jane-graham?trk=public_post_reshare-text), [Aileen Liang](https://www.linkedin.com/in/aileen-liang?trk=public_post_reshare-text), Catherine Simons, [Elisa Zhang](https://www.linkedin.com/in/elisa-zhangg?trk=public_post_reshare-text), [Fatimeh Al Ghannam](https://www.linkedin.com/in/fatimeh-al-ghannam?trk=public_post_reshare-text), [Soniya Agrawal](https://www.linkedin.com/in/soniya-agrawal-2a3001168?trk=public_post_reshare-text), [Spencer Michaels](https://www.linkedin.com/in/spencer-michaels?trk=public_post_reshare-text), [Neil Thompson](https://www.linkedin.com/in/neil-thompson-5724a614?trk=public_post_reshare-text), and the 40+ experts who helped. Tagging a few people who might be interested: [Kevin Fumai](https://www.linkedin.com/in/kevinfumai?trk=public_post_reshare-text) [Yonah Welker](https://ch.linkedin.com/in/welker?trk=public_post_reshare-text) [Sam Burrett](https://au.linkedin.com/in/sam-burrett?trk=public_post_reshare-text) [Chris Kraft](https://www.linkedin.com/in/ctkraft?trk=public_post_reshare-text) [Kuba Szarmach](https://pl.linkedin.com/in/jakubszarmach?trk=public_post_reshare-text) [Tony Moroney](https://ie.linkedin.com/in/tonymoroney?trk=public_post_reshare-text) [Nafis Alam](https://my.linkedin.com/in/nafisalam?trk=public_post_reshare-text) [Luiza Jarovsky, PhD](https://il.linkedin.com/in/luizajarovsky?trk=public_post_reshare-text), [Oliver Patel,](https://uk.linkedin.com/in/oliver-patel?trk=public_post_reshare-text) [Samuel Salzer](https://se.linkedin.com/in/samuelsalzer?trk=public_post_reshare-text) [Phil Venables](https://www.linkedin.com/in/philvenables?trk=public_post_reshare-text) [Renan Araujo](https://uk.linkedin.com/in/araujonrenan?trk=public_post_reshare-text) [Kevin Klyman](https://www.linkedin.com/in/kevin-klyman?trk=public_post_reshare-text) - More than 2,200 proposed AI safeguards. How do organizations decide which ones to use? On November 12, MIT Sloan announced a new AI Risk Mitigation Repository: a free public database cataloging 2,206 safeguards drawn from 47 AI governance and risk-management frameworks. The resource organizes safeguards by the risks they address, who should act, when action should occur during the AI lifecycle, and how each measure is intended to reduce risk. That structure brings an important governance challenge into focus: Having access to safeguards is not the same as knowing which safeguards your organization needs. At the Institute for AI Governance and Ethics (IAGE), we believe the practical value of a framework lies in the decisions it helps people make. For boards, executives, and governance teams, the next questions should be: - Which risks are most relevant to this specific AI use? - Who has the authority and resources to address them? - How will we test whether the selected safeguards work? - What would trigger a change, pause, or withdrawal? The repository catalogs proposed safeguards. Inclusion should not be mistaken for proof that a measure will work in every setting. The opportunity is to move from collecting frameworks to making deliberate, documented choices about risk. In your experience, where does AI governance get stuck most often: choosing the right safeguards, assigning ownership, or measuring effectiveness? What has helped your team move forward? Read the announcement: [https://lnkd.in/gCX2bQ\_B](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FgCX2bQ_B&urlhash=pUkJ&trk=public_post-text) [#AIGovernance](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Faigovernance&trk=public_post-text) [#ResponsibleAI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fresponsibleai&trk=public_post-text) [#RiskManagement](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Friskmanagement&trk=public_post-text) [#IAGE](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fiage&trk=public_post-text) [#AIEthics](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Faiethics&trk=public_post-text) [ MIT researchers create roadmap to help organizations mitigate AI risks globenewswire.com ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fwww%2Eglobenewswire%2Ecom%2Fnews-release%2F2026%2F11%2F12%2F3401754%2F0%2Fen%2Fmit-researchers-create-roadmap-to-help-organizations-mitigate-ai-risks%2Ehtml&urlhash=z5dg&trk=public_post_feed-article-content) - We are getting better at measuring what AI can do. But are we getting equally good at measuring the harm it can cause? I recently came across the MIT AI Risk Initiative’s AI Incident Tracker – Harm Taxonomy, and I think it highlights an important gap in how we approach AI governance and risk management. The framework classifies AI-related incidents across 13 harm categories, each with five severity levels, ranging from Negligible to Catastrophic. What I find particularly interesting is that it goes beyond the traditional focus on technical failures and brings different dimensions of harm into the picture, including: \* Physical and psychological harm. \* Economic and reputational damage. \* Privacy and security violations. \* Discrimination and harmful bias. \* Epistemic harm — the erosion of our ability to distinguish reliable information from misinformation. \* Child sexual exploitation and abuse. A few technical observations stand out to me: 1. Model accuracy is not a proxy for safety. An AI system can achieve excellent benchmark results and still create significant harm when deployed in a real-world environment. We need to evaluate not only model performance, but also its operational context, users, dependencies and potential failure modes. 2. Risk assessment needs to be multidimensional. A single risk score is rarely sufficient. The same incident can generate several types of harm simultaneously. This makes it essential to assess severity, likelihood, exposure, affected stakeholders and downstream consequences independently. 3. AI incidents should become engineering feedback loops. Incident databases should not simply serve as retrospective reporting mechanisms. They should feed directly into model evaluation, red teaming, threat modelling, system design and continuous monitoring. In my view, this is where AI governance becomes genuinely valuable: when it translates observed failures into measurable engineering controls. 4. We should be careful about what the data actually tells us. The MIT framework itself acknowledges important limitations: voluntary reporting creates sampling bias, incident descriptions vary in quality, and LLM-based classification can introduce errors. There is another important limitation: assessing individual incidents does not necessarily capture cumulative or systemic harm, such as long-term labour displacement or the gradual erosion of trust in information. This distinction matters. What we can measure is not necessarily everything we need to manage. My main takeaway: The next stage of responsible AI will not be defined simply by building more accurate, powerful or efficient models. It will be defined by our ability to anticipate, detect, quantify and mitigate the consequences of deploying these systems at scale. I believe AI risk management needs to become an integral part of the engineering lifecycle — not a compliance exercise added just before deployment. Because ultimately, the real measure of a safe AI system is not just how well it performs, but what happens when it operates in the real world. What do you think: are current AI governance frameworks sufficiently connected to actual engineering and operational risk management? Source: MIT AI Risk Initiative – AI Incident Tracker / Harm Taxonomy [https://lnkd.in/eHnS6-kw](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FeHnS6-kw&urlhash=0uVj&trk=public_post-text) [#ArtificialIntelligence](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fartificialintelligence&trk=public_post-text) [#AIRiskManagement](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fairiskmanagement&trk=public_post-text) [#ResponsibleAI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fresponsibleai&trk=public_post-text) [#AIGovernance](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Faigovernance&trk=public_post-text) [#AIEngineering](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Faiengineering&trk=public_post-text) [#Cybersecurity](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fcybersecurity&trk=public_post-text) [#MachineLearning](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fmachinelearning&trk=public_post-text) [ AI Incident Tracker airisk.mit.edu ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fairisk%2Emit%2Eedu%2Fai-incident-tracker%2Fharm-taxonomy&urlhash=AMKd&trk=public_post_feed-article-content) - As [#AI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) moves into more consequential and business-critical processes, organizations recognize that [#trust](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Ftrust&trk=public_post-text) must be a prerequisite for progress. This [EY](https://uk.linkedin.com/company/ernstandyoung?trk=public_post-text) article explores why [#governance](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fgovernance&trk=public_post-text), [#monitoring](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fmonitoring&trk=public_post-text), [#accountability](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Faccountability&trk=public_post-text) and clear decision-making frameworks are non-negotiable as organizations expand their use of AI. Organizations making the most progress aren't treating trust as an add-on. They're building it in from the start. [ Trustworthy AI: five steps to scale with confidence eyconsulting.smh.re ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Feyconsulting%2Esmh%2Ere%2F0_U&urlhash=WTB2&trk=public_post_feed-article-content) - Your vendors are using AI. The bigger question is: do you know what it's doing with your data?   Great perspective from [Aravo Solutions](https://www.linkedin.com/company/aravo-solutions?trk=public_post-text) [Dean Alms](https://www.linkedin.com/in/deanalms?trk=public_post-text) in [Procurement Magazine](https://uk.linkedin.com/company/procurement-magazine-bc?trk=public_post-text) on why AI risk can't be a one-and-done checkbox at onboarding. As AI use evolves, your third-party oversight needs to keep up.   🔗 [https://lnkd.in/eftzYf8i](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FeftzYf8i&urlhash=mNjZ&trk=public_post-text)   [#AI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) [#TPRM](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Ftprm&trk=public_post-text) [#ThirdPartyRisk](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fthirdpartyrisk&trk=public_post-text) [#AIGovernance](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Faigovernance&trk=public_post-text) [ Aravo: Responsible Third-Party AI Risk Governance procurementmag.com ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fprocurementmag%2Ecom%2Fnews%2Faravo-responsible-third-party-ai-risk-governance&urlhash=H8NA&trk=public_post_feed-article-content) - Sources of AI Risk: The Real Challenge Is Not Just Identifying Risks, but Understanding How They Interact. Artificial Intelligence can create tremendous business value. But have organisations fully understood where AI risks originate and how a seemingly minor weakness can become an enterprise-wide problem? AI risk doesn't begin or end with the algorithm. It can come from multiple sources. But the real challenge lies in how these sources connect. Consider an insurance company using AI to assess claims. Poor-quality data affects model outputs. Employees trust those outputs without adequate verification. Automated workflows propagate the resulting errors across thousands of claims. Customer complaints increase, regulatory concerns emerge, and the organisation faces financial and reputational consequences. What began as a data or model weakness has become an enterprise risk. Traditional risk management already recognises correlations, dependencies and common causes. AI adds further complexity through the speed, scale, automation and interconnectedness of decisions. A single weakness can travel across systems, weaken controls, influence human judgement and affect multiple business functions. Feedback loops and shared technology dependencies can make the problem even harder to detect. For Chief Risk Officers, the challenge is to move beyond the risk register. We must ask: ✔ Where does the risk originate? ✔ How can it spread across the organisation? ✔ What could amplify its consequences? ✔ Where can we detect, interrupt and correct the problem? ✔ Who has the authority and accountability to act? The objective is not to eliminate every AI risk. The objective is to understand how risks develop and ensure governance, monitoring, and human intervention keep pace with AI adoption. My central message: AI risk management must examine the entire decision chain—not merely the technology or the individual risk category. That is how organisations can move from risk identification to genuine risk intelligence. What do you believe is the most underestimated source of AI risk today: data, models, human judgement, third-party dependency or governance? [#AIRiskManagement](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fairiskmanagement&trk=public_post-text), [#AIGovernance](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Faigovernance&trk=public_post-text), [#EnterpriseRiskManagement](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fenterpriseriskmanagement&trk=public_post-text), [#RiskManagement](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Friskmanagement&trk=public_post-text), [#ResponsibleAI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fresponsibleai&trk=public_post-text), [#RiskIntelligence](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Friskintelligence&trk=public_post-text), [#ModelRisk](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fmodelrisk&trk=public_post-text), [#DataRisk](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fdatarisk&trk=public_post-text), [#CyberRisk](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fcyberrisk&trk=public_post-text), [#ThirdPartyRisk](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fthirdpartyrisk&trk=public_post-text), [#OperationalResilience](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Foperationalresilience&trk=public_post-text), [#ChiefRiskOfficer](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fchiefriskofficer&trk=public_post-text), [#ArtificialIntelligence](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fartificialintelligence&trk=public_post-text), [#GenerativeAI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fgenerativeai&trk=public_post-text), [#AICompliance](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Faicompliance&trk=public_post-text), [#CorporateGovernance](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fcorporategovernance&trk=public_post-text), [#AlgorithmicBias](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Falgorithmicbias&trk=public_post-text), [#TechnologyRisk](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Ftechnologyrisk&trk=public_post-text), [#InsuranceRiskManagement](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Finsuranceriskmanagement&trk=public_post-text), [#InsurTech](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Finsurtech&trk=public_post-text), [#FinancialRisk](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Ffinancialrisk&trk=public_post-text), [#RiskLeadership](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Friskleadership&trk=public_post-text), [#AI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) Ethics, [#DigitalTransformation](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fdigitaltransformation&trk=public_post-text), [#RiskAssessment](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Friskassessment&trk=public_post-text) [#AIRiskManagement](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fairiskmanagement&trk=public_post-text) [#ArtificialIntelligence](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fartificialintelligence&trk=public_post-text) [#EnterpriseRiskManagement](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fenterpriseriskmanagement&trk=public_post-text) [#AIGovernance](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Faigovernance&trk=public_post-text) [#ChiefRiskOfficer](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fchiefriskofficer&trk=public_post-text) [#ResponsibleAI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fresponsibleai&trk=public_post-text) [#RiskIntelligence](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Friskintelligence&trk=public_post-text) [#Insurance](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Finsurance&trk=public_post-text) - "Trust but verify" AI I've always disliked this phrase. I consider it oxymoronic (or at least self‑negating). Trust requires you to be vulnerable; verification reduces vulnerability. If you trusted me to enter your house and not steal anything, would you verify all your possessions were still there when I left? Would that really be trusting? So what do people mean when they say this? It means that you must manage risk when using AI. You can (but maybe shouldn't) assume that AI has the intent to execute the task that you have requested. Take a simple example. If I wanted the best Basque recipe, would it matter if the LLM provided a recipe from one of the many "top 10" webpages (you know what I'm talking about)? The risk is that you create a suboptimal cheesecake. There is a high probability that the cheesecake will taste good (I hope). Now consider when misinformation could result in something bad (not just a cheesecake no one will eat). If the stakes are high enough, don't use AI or come up with your own answer (you can validate or proofread it with AI). Are there other ways to manage risk when using AI? Here is a prompt that I want you to remember: "How did you get that information?" Here is why. I heard a sound bite and immediately knew it came from an episode of show X I have watched. I wanted to view the scene, so I asked AI what episode the sound bite came from. AI gave me an episode of another TV show with similar lines. It told me I was mistaken and went as far as gaslighting me with, "It's funny how memory works." Then I asked it, "How did you get that information?" It provided YouTube videos of the scene from the other TV show. Do you see the problem here? My follow‑up prompt was: "Do you have access to videos from show X?" and "Do you have access to the scripts from show X?" The answer to both was "no." Let's circle back to risk management. Does the information need to be 100% accurate? I prompted AI with "Wikipedia accuracy". The overall factual accuracy is about 80%, according to Reference Services Review, a peer‑reviewed academic study published in 2008 (the result of asking for the source). So in closing, there are parallels between checking the footnotes in a Wikipedia article and prompting for the source material. Chances are, your LLM was trained on Wikipedia articles. You decide if 80% accuracy is ok. - [ ### What Influences the consumption of Animal-Products? A Meta-Review ](https://www.linkedin.com/pulse/what-influences-consumption-animal-products-peter-slattery-phd?trk=public_post)