Navigating Data Privacy

This Stanford study examined how six major AI companies (Anthropic, OpenAI, Google, Meta, Microsoft, and Amazon) handle user data from chatbot conversations.  Here are the main privacy concerns. 👀 All six companies use chat data for training by default, though some allow opt-out 👀 Data retention is often indefinite, with personal information stored long-term 👀 Cross-platform data merging occurs at multi-product companies (Google, Meta, Microsoft, Amazon) 👀 Children's data is handled inconsistently, with most companies not adequately protecting minors 👀 Limited transparency in privacy policies, which are complex and hard to understand and often lack crucial details about actual practices Practical Takeaways for Acceptable Use Policy and Training for nonprofits in using generative AI: ✅ Assume anything you share will be used for training - sensitive information, uploaded files, health details, biometric data, etc. ✅ Opt out when possible - proactively disable data collection for training (Meta is the one where you cannot) ✅ Information cascades through ecosystems - your inputs can lead to inferences that affect ads, recommendations, and potentially insurance or other third parties ✅ Special concern for children's data - age verification and consent protections are inconsistent Some questions to consider in acceptable use policies and to incorporate in any training. ❓ What types of sensitive information might your nonprofit staff  share with generative AI?  ❓ Does your nonprofit currently specifically identify what is considered “sensitive information” (beyond PID) and should not be shared with GenerativeAI ? Is this incorporated into training? ❓ Are you working with children, people with health conditions, or others whose data could be particularly harmful if leaked or misused? ❓ What would be the consequences if sensitive information or strategic organizational data ended up being used to train AI models? How might this affect trust, compliance, or your mission? How is this communicated in training and policy? Across the board, the Stanford research points that developers’ privacy policies lack essential information about their practices. They recommend policymakers and developers address data privacy challenges posed by LLM-powered chatbots through comprehensive federal privacy regulation, affirmative opt-in for model training, and filtering personal information from chat inputs by default. “We need to promote innovation in privacy-preserving AI, so that user privacy isn’t an afterthought." How are you advocating for privacy-preserving AI? How are you educating your staff to navigate this challenge? https://lnkd.in/g3RmbEwD

𝟔𝟔% 𝐨𝐟 𝐀𝐈 𝐮𝐬𝐞𝐫𝐬 𝐬𝐚𝐲 𝐝𝐚𝐭𝐚 𝐩𝐫𝐢𝐯𝐚𝐜𝐲 𝐢𝐬 𝐭𝐡𝐞𝐢𝐫 𝐭𝐨𝐩 𝐜𝐨𝐧𝐜𝐞𝐫𝐧. What does that tell us? Trust isn’t just a feature - it’s the foundation of AI’s future. When breaches happen, the cost isn’t measured in fines or headlines alone - it’s measured in lost trust. I recently spoke with a healthcare executive who shared a haunting story: after a data breach, patients stopped using their app - not because they didn’t need the service, but because they no longer felt safe. 𝐓𝐡𝐢𝐬 𝐢𝐬𝐧’𝐭 𝐣𝐮𝐬𝐭 𝐚𝐛𝐨𝐮𝐭 𝐝𝐚𝐭𝐚. 𝐈𝐭’𝐬 𝐚𝐛𝐨𝐮𝐭 𝐩𝐞𝐨𝐩𝐥𝐞’𝐬 𝐥𝐢𝐯𝐞𝐬 - 𝐭𝐫𝐮𝐬𝐭 𝐛𝐫𝐨𝐤𝐞𝐧, 𝐜𝐨𝐧𝐟𝐢𝐝𝐞𝐧𝐜𝐞 𝐬𝐡𝐚𝐭𝐭𝐞𝐫𝐞𝐝. Consider the October 2023 incident at 23andMe: unauthorized access exposed the genetic and personal information of 6.9 million users. Imagine seeing your most private data compromised. At Deloitte, we’ve helped organizations turn privacy challenges into opportunities by embedding trust into their AI strategies. For example, we recently partnered with a global financial institution to design a privacy-by-design framework that not only met regulatory requirements but also restored customer confidence. The result? A 15% increase in customer engagement within six months. 𝐇𝐨𝐰 𝐜𝐚𝐧 𝐥𝐞𝐚𝐝𝐞𝐫𝐬 𝐫𝐞𝐛𝐮𝐢𝐥𝐝 𝐭𝐫𝐮𝐬𝐭 𝐰𝐡𝐞𝐧 𝐢𝐭’𝐬 𝐥𝐨𝐬𝐭? ✔️ 𝐓𝐮𝐫𝐧 𝐏𝐫𝐢𝐯𝐚𝐜𝐲 𝐢𝐧𝐭𝐨 𝐄𝐦𝐩𝐨𝐰𝐞𝐫𝐦𝐞𝐧𝐭: Privacy isn’t just about compliance. It’s about empowering customers to own their data. When people feel in control, they trust more. ✔️ 𝐏𝐫𝐨𝐚𝐜𝐭𝐢𝐯𝐞𝐥𝐲 𝐏𝐫𝐨𝐭𝐞𝐜𝐭 𝐏𝐫𝐢𝐯𝐚𝐜𝐲: AI can do more than process data, it can safeguard it. Predictive privacy models can spot risks before they become problems, demonstrating your commitment to trust and innovation. ✔️ 𝐋𝐞𝐚𝐝 𝐰𝐢𝐭𝐡 𝐄𝐭𝐡𝐢𝐜𝐬, 𝐍𝐨𝐭 𝐉𝐮𝐬𝐭 𝐂𝐨𝐦𝐩𝐥𝐢𝐚𝐧𝐜𝐞: Collaborate with peers, regulators, and even competitors to set new privacy standards. Customers notice when you lead the charge for their protection. ✔️ 𝐃𝐞𝐬𝐢𝐠𝐧 𝐟𝐨𝐫 𝐀𝐧𝐨𝐧𝐲𝐦𝐢𝐭𝐲: Techniques like differential privacy ensure sensitive data remains safe while enabling innovation. Your customers shouldn’t have to trade their privacy for progress. Trust is fragile, but it’s also resilient when leaders take responsibility. AI without trust isn’t just limited - it’s destined to fail. 𝐇𝐨𝐰 𝐰𝐨𝐮𝐥𝐝 𝐲𝐨𝐮 𝐫𝐞𝐠𝐚𝐢𝐧 𝐭𝐫𝐮𝐬𝐭 𝐢𝐧 𝐭𝐡𝐢𝐬 𝐬𝐢𝐭𝐮𝐚𝐭𝐢𝐨𝐧? 𝐋𝐞𝐭’𝐬 𝐬𝐡𝐚𝐫𝐞 𝐚𝐧𝐝 𝐢𝐧𝐬𝐩𝐢𝐫𝐞 𝐞𝐚𝐜𝐡 𝐨𝐭𝐡𝐞𝐫 👇 #AI #DataPrivacy #Leadership #CustomerTrust #Ethics

Passing the CIPP doesn't make you a privacy professional. It makes you someone who can pass the CIPP. The gap between certification and real privacy work hits different when you're sitting across from a frustrated marketing director who needs a yes-or-no decision they can act on, not an explanation of state privacy law nuances. If you're a privacy officer feeling like your certification didn't quite prepare you for the day-to-day reality, you're not alone. But here's what I learned after years of figuring this out the hard way: The path forward isn't more legal theory. It's operational mastery. Stop waiting for perfect conditions. Start solving real problems. → Find something broken and fix it. You'll learn more from streamlining one data subject request process than from reading ten compliance guides. Pick the thing that's driving everyone crazy and make it work better. → Measure what actually matters. Forget policies written or training sessions conducted. The main metric that counts: Are people making better privacy decisions without you in the room? → Build relationships, not barriers. That marketing team's campaign deadline isn't your enemy, it's your reality. Learn their world. Speak their language. Become the privacy person they actually want to work with. → Embrace "good enough" as your superpower. Perfect compliance is a myth in dynamic organizations.Effective privacy professionals make good decisions fast, document their reasoning, and improve over time. Paralysis isn't prudence. → Master practical risk assessment. Stop trying to eliminate all privacy risk. Start communicating trade-offs like a business partner. "What's the worst realistic outcome? How likely? What will it cost to prevent versus fix?" This is how you earn a seat at the table. → Build systems that work without you. Your goal isn't to review every decision. It's to create templates, practical procedures, and processes that make the right choice the easy choice. Scale yourself through systems. The confidence breakthrough: Privacy expertise isn't about having perfect legal knowledge. It's about developing unshakeable operational judgment and trusting yourself to apply it. You don't need more certifications. You need more practice making decisions. Start with lower-stakes decisions. Document your reasoning. Learn from outcomes. Your judgment will get better. Because it has to. What's the biggest operational gap you've found between privacy law and practice?

This new white paper by Stanford Institute for Human-Centered Artificial Intelligence (HAI) titled "Rethinking Privacy in the AI Era" addresses the intersection of data privacy and AI development, highlighting the challenges and proposing solutions for mitigating privacy risks. It outlines the current data protection landscape, including the Fair Information Practice Principles, GDPR, and U.S. state privacy laws, and discusses the distinction and regulatory implications between predictive and generative AI. The paper argues that AI's reliance on extensive data collection presents unique privacy risks at both individual and societal levels, noting that existing laws are inadequate for the emerging challenges posed by AI systems, because they don't fully tackle the shortcomings of the Fair Information Practice Principles (FIPs) framework or concentrate adequately on the comprehensive data governance measures necessary for regulating data used in AI development. According to the paper, FIPs are outdated and not well-suited for modern data and AI complexities, because: - They do not address the power imbalance between data collectors and individuals. - FIPs fail to enforce data minimization and purpose limitation effectively. - The framework places too much responsibility on individuals for privacy management. - Allows for data collection by default, putting the onus on individuals to opt out. - Focuses on procedural rather than substantive protections. - Struggles with the concepts of consent and legitimate interest, complicating privacy management. It emphasizes the need for new regulatory approaches that go beyond current privacy legislation to effectively manage the risks associated with AI-driven data acquisition and processing. The paper suggests three key strategies to mitigate the privacy harms of AI: 1.) Denormalize Data Collection by Default: Shift from opt-out to opt-in data collection models to facilitate true data minimization. This approach emphasizes "privacy by default" and the need for technical standards and infrastructure that enable meaningful consent mechanisms. 2.) Focus on the AI Data Supply Chain: Enhance privacy and data protection by ensuring dataset transparency and accountability throughout the entire lifecycle of data. This includes a call for regulatory frameworks that address data privacy comprehensively across the data supply chain. 3.) Flip the Script on Personal Data Management: Encourage the development of new governance mechanisms and technical infrastructures, such as data intermediaries and data permissioning systems, to automate and support the exercise of individual data rights and preferences. This strategy aims to empower individuals by facilitating easier management and control of their personal data in the context of AI. by Dr. Jennifer King Caroline Meinhardt Link: https://lnkd.in/dniktn3V

In a discussion about synthetic data generation and privacy last week, Alexandra Ebert had this fascinating metaphor: « Would you rather sit in a car that's "theoretically" safe/safe on paper or one where the car manufacturer's new car model had to actually undergo (empirical) crash tests? » 🚘 In this metaphor, theoretical safety is differential privacy, and crash tests are empirical privacy tests. The former gives you formal, proven guarantees about the level of risk; the latter empirically compares the synthetic data to the true data to check that it doesn’t seem too revealing 📊 This got me wondering: is there something that anonymization practitioners can learn from industries with mature safety programs? 🤔 When automotive engineers start on a new design, they know all the properties of the materials they work with: elasticity, hardness, shear strength, and so on. They account for all this information in the design & manufacturing process, to predict what would happen during a crash: when will collision detectors activate, how materials will deform, how fast airbags will inflate… Those are all “theoretical” properties — calculations and simulations run before the car even gets built. Standards like ISO 26262 require all this — car manufacturers must be able to demonstrate the safety impact of every step of the design and manufacturing process, and quantify all the potential risks 🦺 So why do crash tests? Because they provide a final verification that the practice matches the theory. Sometimes, defects pop up in surprising places, or system components interact in an unforeseen way. This should not happen: by the time the car is actually thrown against a wall, engineers know how the car will behave, and the crash test is only there to check that nothing unexpected is happening. No car manufacturer would use this kind of final verification as a primary mechanism to ensure safety! 🤨 This is an excellent metaphor to draw parallels with the anonymization industry. Just like car manufacturers, vendors of anonymization tech should adopt a safety-first approach to building their products. They should fully understand their privacy guarantees, grounding those in a solid theoretical foundation, like — you guessed it — differential privacy. They should also be able to demonstrate that their implementation actually achieves the theoretical guarantees. There, end-to-end empirical tests can be useful¹, along with good security practices like publishing open-source code, building for auditability, writing unit tests, hiring third-party auditors, and so on 💡 Just using empirical tests as the core mechanism to provide privacy guarantees, though? Hmm. I know I wouldn’t want to climb in a car whose safety story boils down to “we smashed it against the wall 5 times and our dummies were fine every time” 😬 ¹ Though not all tests are created equal, and there’s more to say about that… but that’s a hot take for another time. #syntheticdata #privacy

Every time we share data, we walk a tightrope between utility and privacy. I have seen how the desire to extract value from data can easily collide with the need to protect it. Yet this is not a zero-sum game. Advances in cryptography and privacy-enhancing technologies are making it possible to reconcile these two goals in ways that were unthinkable just a few years ago. My infographic highlights six privacy-preserving techniques that are helping to reshape how we think about secure data sharing. From fully homomorphic encryption, which allows computations on encrypted data, to differential privacy, which injects noise into datasets to hide individual traces, each method reflects a different strategy to maintain control without losing analytical power. Others, like federated analysis and secure multiparty computation, show how collaboration can thrive even when data is never centralized or fully revealed. The underlying message is simple: privacy does not have to be an obstacle to innovation. On the contrary, it can be a design principle that unlocks new forms of responsible collaboration. #Privacy #DataSharing #Cybersecurity #Encryption #DigitalTrust #DataProtection

Yesterday, Anthropic quietly dropped a bombshell. Unless users explicitly opt out by September 28, it will use consumer chat data to train future AI models. This is a stunning reversal from Anthropic’s previous position as the privacy-first alternative to ChatGPT. Previously, Anthropic automatically deleted user conversations after 30 days. Under the new policy, conversations from users who don’t opt out will be retained for five years. The new policy affects all consumer tiers: Claude Free, Pro, and Max users, plus those using Claude Code. Importantly, business customers using Claude for Work, Claude Gov, Claude for Education, or API access through services like Amazon Bedrock remain unaffected. This creates a clear two-tiered privacy system where enterprise customers get protection while consumers become training data. Anthropic frames the change around improving “model safety” and helping future Claude models “improve at skills like coding, analysis, and reasoning.” The company emphasizes user choice and the ability to change settings at any time. This is total nonsense, of course. In reality, training AI models requires vast amounts of high-quality conversational data, and accessing millions of Claude interactions will provide exactly the kind of real-world content that can improve Anthropic’s competitive positioning against rivals like OpenAI and Google. This isn’t happening in isolation. Google recently announced a similar opt-out policy for Gemini, set to take effect on September 2. That policy is similarly broad, covering user-uploaded files, photos, videos, and even screenshots that users ask questions about. The entire industry is converging on the same strategy: make data collection the default and require users to actively opt out. If your company uses Claude, review your access method immediately. Consumer accounts now default to data sharing. Enterprise accounts maintain privacy protections, but at significantly higher cost. And you’ll probably want to let your workforce know that they have to properly configure their personal AI accounts if they are likely to accidentally input sensitive company data while using their personal devices. To opt-out today, go to Settings>Privacy. Under the Privacy settings area, you’ll see “Help improve Claude.” Toggle it off. Accept the terms. You’re done. The deadline is September 28, 2025. After that date, users must make their selection to continue using Claude. I think we should consider this a preview of coming industry standards. Privacy-by-default will quickly transition to privacy-by-choice, with the burden shifting to users to protect their own data.

Bosch has just published a very interesting assessment regarding cutting-edge privacy-preserving technology. A group of their researchers (such as Sven Trieflinger and Hossein Yalame) have worked with independent lawyers to determine if and when Secure Multiparty Computation (MPC) can be used to process private data cross-border according to the GDPR. Their conclusions make sense, also from my technical perspective: the data stored inside the MPC is not considered private, unless someone has the (legal) means to reconstruct it. You can't use MPC if everyone participating is controlled by the same legal entity (e.g. same company in one country). Once you go cross-border, however, things are different as subsidiaries must follow local privacy laws and thus have an obligation not to break privacy. However, they additionally recommend to put contractual agreements in place regarding the honest participation in MPC. This is particularly interesting when outsourcing the computation (or management thereof) to third-party technology providers. The full report can be found here: https://lnkd.in/dEwdtecZ

A hairdresser and a marketer came into the bar. Hold on… Haircuts and marketing? 🤔 Here's the reality: Consumers are more aware than ever of how their data is used. User privacy is no longer a checkbox – It is a trust-building cornerstone for any online business. 88% of consumers say they won’t share personal information unless they trust a brand. Think about it: Every time a user visits your website, they’re making an active choice to trust you or not. They want to feel heard and respected. If you're not prioritizing their privacy preferences, you're risking their data AND loyalty. We’ve all been there – Asked for a quick trim and got VERY short hair instead. Using consumers’ data without consent is just like cutting the hair you shouldn’t cut. That horrible bad haircut ruined our mood for weeks. And a poor data privacy experience can drive customers straight to your competitors, leaving your shopping carts empty. How do you avoid this pitfall? - Listen to your users. Use consent and preference management tools such as Usercentrics to allow customers full control of their data. - Be transparent. Clearly communicate how you use their information and respect their choices. - Build trust: When users feel secure about their data, they’re more likely to engage with your brand. Make sure your website isn’t alienating users with poor data practices. Start by evaluating your current approach to data privacy by scanning your website for trackers. Remember, respecting consumer choices isn’t just an ethical practice. It’s essential for long-term success in e-commerce. Focus on creating a digital environment where consumers feel valued and secure. Trust me, it will pay off! 💰

🚨 AI Privacy Risks & Mitigations Large Language Models (LLMs), by Isabel Barberá, is the 107-page report about AI & Privacy you were waiting for! [Bookmark & share below]. Topics covered: - Background "This section introduces Large Language Models, how they work, and their common applications. It also discusses performance evaluation measures, helping readers understand the foundational aspects of LLM systems." - Data Flow and Associated Privacy Risks in LLM Systems "Here, we explore how privacy risks emerge across different LLM service models, emphasizing the importance of understanding data flows throughout the AI lifecycle. This section also identifies risks and mitigations and examines roles and responsibilities under the AI Act and the GDPR." - Data Protection and Privacy Risk Assessment: Risk Identification "This section outlines criteria for identifying risks and provides examples of privacy risks specific to LLM systems. Developers and users can use this section as a starting point for identifying risks in their own systems." - Data Protection and Privacy Risk Assessment: Risk Estimation & Evaluation "Guidance on how to analyse, classify and assess privacy risks is provided here, with criteria for evaluating both the probability and severity of risks. This section explains how to derive a final risk evaluation to prioritize mitigation efforts effectively." - Data Protection and Privacy Risk Control "This section details risk treatment strategies, offering practical mitigation measures for common privacy risks in LLM systems. It also discusses residual risk acceptance and the iterative nature of risk management in AI systems." - Residual Risk Evaluation "Evaluating residual risks after mitigation is essential to ensure risks fall within acceptable thresholds and do not require further action. This section outlines how residual risks are evaluated to determine whether additional mitigation is needed or if the model or LLM system is ready for deployment." - Review & Monitor "This section covers the importance of reviewing risk management activities and maintaining a risk register. It also highlights the importance of continuous monitoring to detect emerging risks, assess real-world impact, and refine mitigation strategies." - Examples of LLM Systems’ Risk Assessments "Three detailed use cases are provided to demonstrate the application of the risk management framework in real-world scenarios. These examples illustrate how risks can be identified, assessed, and mitigated across various contexts." - Reference to Tools, Methodologies, Benchmarks, and Guidance "The final section compiles tools, evaluation metrics, benchmarks, methodologies, and standards to support developers and users in managing risks and evaluating the performance of LLM systems." 👉 Download it below. 👉 NEVER MISS my AI governance updates: join my newsletter's 58,500+ subscribers (below). #AI #AIGovernance #Privacy #DataProtection #AIRegulation #EDPB