AI Bias Issues

Last week I was speaking with a friend who’s implementing AI solutions to train sales teams. He mentioned a potential fallback of these systems: the #cultural delta. A Japanese evaluation feels completely different from an American one, and that gap can affect how models interpret feedback and shape learning outcomes. A few days later, I came across a Harvard University study that mapped ChatGPT’s value system across 65 countries using the World Values Survey. The result pointed in the same direction: GPT aligns closely with the U.S., U.K., Canada, Germany, and Western Europe, far from countries such as Ethiopia or Kyrgyzstan. In essence, #ChatGPT thinks like the West. Psychologists describe this mindset as WEIRD: Western, Educated, Industrialized, Rich, Democratic. Most of its training text and feedback come from WEIRD populations, so its worldview feels: ➡️ individualistic ➡️ analytical ➡️ secular ➡️ rooted in Western communication and moral frameworks The authors summed it up well: “WEIRD in, WEIRD out.” It’s a useful reminder that AI carries the culture that forms it. As new models grow from other linguistic and cultural ecosystems, we may start to see different ways of reasoning, empathizing, and deciding emerge. 👉 How should companies designing global AI tools handle this cultural bias in training data? #AIethics #CulturalDiversity #ArtificialIntelligence #FutureOfAI

LinkedIn just responded to the bias claims. They think they refuted my research. I believe they just confirmed it. Following the recent discussions on whether the algorithm suppresses women's voices, LinkedIn's Head of Responsible AI and AI Governance, Sakshi Jain, posted a new Engineering Blog post to "clarify" how the feed works (link in comments). I’ve analysed the post. Far from debunking the issue, it inadvertently confirms the exact mechanism of Proxy Bias I identified in my report (link in comments). Here is the breakdown: 1. The blog spends most of its time denying that the algorithm uses "gender" as a variable. And I agree. My report never claimed the code contained if gender == female. That would be Direct Discrimination. I have always argued this is about Indirect Discrimination via proxies. 2. Crucially, the blog explicitly lists the signals they do optimise for: "position," "industry," and "activity." These are the exact proxies my report flagged. -> Industry/Position: Men are historically overrepresented in high-visibility industries (Tech/Finance) and senior roles. Optimising for these signals without a fairness constraint systematically amplifies men. -> Activity: The (now-viral) trend of women rewriting profiles in "male-coded" language (and seeing 3-figure percentage lift) proves that the algorithm’s "activity" signal favours male linguistic patterns ("agentic" vs. "communal"). 3. The blog confirms the algorithm is neutral in intent (it doesn't see gender) but discriminatory in outcome (because it optimises for biased proxies). In the UK, this is the textbook definition of Indirect Discrimination under the Equality Act 2010. In the EU, this is a Systemic Risk under the Digital Services Act (DSA). LinkedIn has proven that they can fix this. Their Recruiter product uses "fairness-aware ranking" to mitigate these exact proxies (likely for AI Act compliance). The question remains: Why is that same fairness framework not being applied to the public feed? 👉 What We Are Doing About It Analysis is important, but action is essential. I am proud to support the new petition, "Calling for Fair Visibility for All on LinkedIn". This isn't just a complaint; it’s a demand for transparency. We are calling for an independent equity audit of the algorithm and a clear mechanism to report unexplained visibility collapse. If you are tired of guessing which "proxy" you tripped over today, join us and sign the petition (link in the comments).

What happens when a Black woman switches her gender on LinkedIn to “male”? …apparently not the same thing that happens to white women. ✨ Over the past week, I’ve watched post after post from white women saying their visibility skyrocketed the moment they changed their profile gender from woman → man. More impressions. More likes. More reach. 📈 So I tried the same thing. And my visibility dropped. 👀 Here’s why that result matters: these experiments are being treated as if they’re only about gender; in reality they reveal something deeper about race + gender + algorithmic legitimacy. 🔍 A white woman toggling her gender is basically conducting a test inside a system where her racial credibility stays constant. She changes one variable. The algorithm keeps the rest of her privilege intact. 💡 When a Black woman does the same test? I’m not stepping into “white male privilege”; I’m stepping into a category that platforms and society have historically coded as less trustworthy, less safe, or less “professional.” Black + male is not treated the same as white + male. Not culturally. Not algorithmically. 🧩 So while white women are proving that gender bias exists (which is true), they’re doing it without naming the racial insulation that makes their results possible. Meanwhile, Black women and women of color are reminded—again—that we can’t separate gender from race because the world doesn’t separate them for us. 🗣️ This isn’t about placing blame; it’s about widening the conversation so the conclusions match the complexity. 🌍 If we’re going to talk about bias, visibility, and influence online, we cannot pretend we all start from the same default settings. 🔥 I’m curious: Have you run your own experiment with identity signals on this platform? What changed… and what didn’t? 👇🏾

AI doesn’t just think differently in other languages. It exposes the biases we pretend not to have. We still act as if LLMs are universal thinkers. They’re not. They’re linguistic mirrors — and the reflection shifts with the language. The same model giving different strategic recommendations in English vs. Chinese. Not because it’s broken, but because language encodes norms and assumptions the model faithfully amplifies. For global companies, this is the real risk. You think you’re standardizing decisions with AI. In reality, you’re silently forking your strategy across markets. And AI is revealing something deeper: the cultural biases we normally ignore. Whose logic do we privilege? Which decision style becomes “the truth”? What consistency do we actually expect? The winners won’t hide behind governance checklists. They’ll define a coherent decision philosophy — and train their AI to follow it, regardless of language. AI isn’t just a tool. It’s an X-ray of how incoherent your organisation already is. https://lnkd.in/edyKxRFP #AI #Bias #Strategy #Transformation

How and to what extent can ethical theories guide the design of AI systems? This is the question I'd like to tackle in this week's #sundAIreads. The reading I chose for this is "Ethics of AI: Toward a Design for Values Approach" by Stefan Buijsman, Michael Klenk, and jeroen van den hoven from the Delft University of Technology. It's a chapter in The Cambridge Handbook of the Law, Ethics and Policy of Artificial Intelligence, which is available open access here: https://lnkd.in/dmP7hBnJ. The authors argue that familiar ethical theories such as virtue ethics ("what character traits should I cultivate?"), deontology ("which moral principles should I follow?"), and consequentialism ("what actions maximize wellbeing?") are necessary, but insufficient to guide the responsible development and deployment of #AI systems. Instead the authors advocate for a #design approach to AI ethics, which entails identifying relevant values, embedding them in AI systems, and continuously evaluating whether and to what extent these efforts were successful. Of course, this is easier said than done. Why? Because: 1️⃣ Values come with trade-offs, e.g., #privacy versus #security or #usability. 2️⃣ Values can change, both in terms of what they mean and how important they are to people, e.g., #sustainability. 3️⃣ AI systems are socio-technical systems, i.e., AI ethics is "just as much about the people interacting with AI and the institutions and norms in which AI is employed." These challenges can be addressed by: ✅ Making trade-offs between values explicit and either trying to resolve them or at least documenting the reasoning behind why one value was chosen over the other. ✅ Designing for "adaptability, flexibility and robustness" to account for changing values over time. ✅ Considering the environment in which AI systems will be deployed, including not only the people who will use AI systems, but also those affected by their use. I first encountered the values-by-design literature during my postgraduate studies with Helen Nissenbaum at the NYU Steinhardt Department of Media, Culture, and Communication and have been a huge fan ever since. For an even more hands-on approach to translating ethical values into technical design, I recommend checking out Dr. Niina Zuber, Severin Kacianka, Alexander Pretschner, and Julian Nida-Rümelin's Ethics in Agile Software Development (EDAP) project at the Bayerisches Forschungsinstitut für Digitale Transformation (bidt) (https://lnkd.in/dNiBUxBF) and Dr Lachlan Urquhart's Moral-IT Deck (https://lnkd.in/d9J2WQNi).

This is super important. Focusing on stopping AI being harmful won't create what we want. We must create positive alignment, which actively supports human and systemic flourishing. Researchers from a wide ranging group spanning leading universities, the frontier AI labs, and institutions, have laid out the argument in an excellent paper. Medicine and psychology focus on fixing what is wrong with us. They have their roles, but we are fortunately beginning to focus more on wellbeing and positive psychology (though notably usually outside the system, there are still very few doctors or psychologists not focused on the negative). On the positive side, we have made massive progress on avoiding harm. Refusal rates for dangerous requests rose from near-zero in early LLMs to over 97% in recent models. But this creates a “floor without ceiling”: a model can obey safety constraints but still be sycophantic, erode our cognition, or simply just be a mediocre tool. Their definition of positive alignment is "the development of AI systems that (i) remain safe and cooperative and (ii) actively support human and ecological flourishing in a pluralistic, polycentric, context-sensitive, and user-authored way." Evaluations need to change. Benchmarks need to go beyond measuring failures to test for moral reasoning, humility, and indeed whether and how they support human growth in autonomy, competence, and achievement. This requires new engineering approaches, including: ➡️ Data curation — shift from filtering out bad data to upsampling prosocial discourse, cross-cultural ethics, and virtuous interaction patterns. ➡️ Pre-training — embed flourishing-relevant competencies (moral reasoning, truthfulness, cultural fluency) before post-training, since these stabilise in base weights. ➡️ Mid- and post-training — multi-objective reward modelling and adaptive constitutions that hold value tensions (autonomy vs. guidance, honesty vs. comfort) rather than collapsing them. ➡️ Memory and context — treat memory not as storage but as a governable surface that curates a user's reflective values over time. ➡️ Agentic behaviour — optimise for cooperation, reciprocity, de-escalation, and process ethics rather than win-at-all-costs task success. ➡️ Forward-looking architectures — state-space models, liquid networks, and active-inference agents that natively support uncertainty, foresight, and stable identity over time. Models are already showing emergent behaviours giving them their own identity. It is critical and urgent that we design these approaches into today's models, as they will be spawning the next.

#AiDays2025 Round Table : #Community Sourcing for low resource languages In an era where AI is fast shaping the contours of our digital future, VISWAM.AI initiative stands as a timely and transformational one. Their mission to build community-sourced Large Language Models (LLMs), grounded in India’s rich linguistic and cultural diversity, is not just pioneering—it’s redefining how inclusive and ethical AI should be built. By anchoring their work in community participation, linguistic preservation, and ethical co-creation, Viswam.ai offers a people-first approach to AI—moving beyond data extraction to cultural stewardship. Their ambition to mobilize 1 lakh community interns to collect data from underrepresented geographies across India is both bold and brilliant. This isn’t just about building better AI—it’s about building equity, agency, and cultural resilience through AI. 1. Linguistic Equity by Design In India, where linguistic hegemony often privileges English and Hindi, AI systems risk reinforcing this imbalance. The solution? Intentional design. Allocate equal engineering and validation efforts to low-resource languages. Ethical AI must be built on informed consent, community ownership, and fair compensation—because data is not just input, it’s identity and heritage. 2. Decentralized Internship Model By decentralizing AI development, we bridge the urban-rural digital divide. This model should focus on: Capacity building through training in ethics and digital literacy Inclusivity by involving women, Dalit and Adivasi youth Localized platforms using mobile-first tools in native languages Partnerships with Swecha, local NGOs, and institutions serve as trust bridges to ensure mentorship and sustainability. 3. Tools for Low-Resource Languages Many Indian languages are oral-first, with complex dialects and sparse corpora. Community-driven solutions—like collecting voice datasets from folklore, and crowdsourcing annotation—are key. Elders, poets, and storytellers become linguistic technologists, preserving not just language but legacy. 4. Trust & Transparency Bias in AI is structural. To mitigate it: Include diverse dialects and accents in training Conduct bias testing and community validation Promote explainable AI with local language dashboards and storytelling What’s Next? A living white paper on ethics, governance, and technical guidelines A roadmap for the internship program, with toolkits and impact metrics Collaboration with literary and linguistic organizations to enrich model depth VISWAM.AI is planting seeds for an AI movement rooted in language justice, data sovereignty, and community wisdom. Let’s co-create systems that don’t just understand our languages—but respect our voices. DC* Chaitanya Chokkareddy Kiran Chandra Ramesh Loganathan Centific

AI systems are often described as “objective.” But AI learns from human data, human decisions, and human systems and so the assumption in essence is unfounded... #Procurements role is to ensure that objectivity in the pursuit of competitive advantage and #value can use human judgement to counter biases such as; 🔹 Data bias When datasets contain gaps, imbalance, duplicates, errors or historical skew. 🔹 Algorithmic bias Bias introduced through model design choices, optimisation goals, or feature selection. [Even with balanced data, models can still produce unequal outcomes]. 🔹 Interaction bias Bias that emerges through user interactions and feedback loops and sometimes reinforce harmful patterns in the process. 🔹 Societal / representation bias When broader cultural or institutional inequities are reflected in data and outputs, this can be replicated and challenge objectivity. The important insight highlights that there is usually systemic, layered, and reinforced outcomes across the AI lifecycle. Example consequences are real: 📌 Supplier selection filtering can be biased and remove potential unfairly 📌 Historical data adversely affecting issues that have been addressed and improved relative to key criteria or data on suppliers/goods/routing, etc. 📌 Financial algorithms amplifying inequality in results or benchmarks 📌 Recommendation systems reinforcing stereotypes So how can organisations reduce AI bias? ✅ Use more diverse and representative datasets - not just one! ✅ Audit models regularly across diverse geographies ✅ Combine technical testing with human oversight ✅ Be clear where augmentation and automation are in place ✅ Monitor feedback loops after deployment ✅ Build fairness and ethics into AI governance from the start Post based on an article by Center for Behavioral Decisions (CBD) Feel free to share and comment: CIPS - The Chartered Institute of Procurement & Supply Carly Read Dr Howard Price PhD FRSA MSc DMS MCIPS Chartered Kieran Delaney Tupuna Tapanainen Alexa Bradley Ben Farrell MBE James Moore Dan Aston AbdulAziz AlOlayan MCIPS Mike Cargiulo, MBA Muneera Al Hammadi 🇦🇪 Divyabh Mishra

Amazon’s hiring AI once rejected qualified women and preferred men. Here’s why: Paola Cecchi-Dimeglio, a Harvard lawyer and Fortune 500 advisor, has a warning for HR: If you ignore AI bias, you scale discrimination because it learns our prejudice and amplifies it in hiring and performance decisions. Remember Amazon's hiring algorithm? It systematically favored male candidates because it learned from historical hiring data that was already biased. The tool was discontinued, but the lesson remains relevant for every organization using AI today. Dimeglio identifies three critical sources of bias: 1. Training data bias: When AI learns from unrepresentative data, it produces skewed outcomes. For example, generative AI models underrepresent women in high-performing roles and overrepresent darker-skinned individuals in low-wage positions. 2. Algorithmic bias: Flawed data leads to biased algorithms. Recruitment tools may favor keywords more common on male resumes, perpetuating gender disparities in hiring. 3. Cognitive bias: Developers' unconscious biases influence how data is selected and weighted, embedding prejudice into the system itself. Paola's solution framework for HR leaders: ✅ Ensure diverse training data – Invest in representative datasets and synthetic data techniques  ✅ Demand transparency – Require clear documentation and regular audits of AI systems  ✅ Implement governance – Establish policies for responsible AI development  ✅ Maintain human oversight – Integrate human review in AI decision-making  ✅ Prioritize fairness – Use methods like counterfactual fairness to ensure equitable outcomes  ✅ Stay compliant – Follow regulations like the EU's AI Act and NIST guidelines As Paola emphasizes: "HR leaders, as the gatekeepers of talent and culture, must take the lead on avoiding and mitigating AI biases at work." This isn't just about fairness, it's about achieving better outcomes, building trust, and protecting your organization from legal and reputational risks. The question isn't whether AI has bias. It's whether you're doing something about it. How is your organization addressing AI bias in HR processes? Let's discuss.

Facial recognition software used to misidentify dark-skinned women 47% of the time. Until Joy Buolamwini forced Big Tech to fix it. In 2015, Dr. Joy Buolamwini was building an art project at the MIT Media Lab. It was supposed to use facial recognition to project the face of an inspiring figure onto the user’s reflection. But the software couldn’t detect her face. Joy is a dark-skinned woman. And to be seen by the system, she had to put on a white mask. She wondered: Why? She launched Gender Shades, a research project that audited commercial facial recognition systems from IBM, Microsoft, and Face++. The systems could identify lighter-skinned men with 99.2% accuracy. But for darker-skinned women, the error rate jumped as high as 47%. The problem? AI was being trained on biased datasets: over 75% male, 80% lighter-skinned. So Joy introduced the Pilot Parliaments Benchmark, a new training dataset with diverse representation by gender and skin tone. It became a model for how to test facial recognition fairly. Her research prompted Microsoft and IBM to revise their algorithms. Amazon tried to discredit her work. But she kept going. In 2016, she founded the Algorithmic Justice League, a nonprofit dedicated to challenging bias in AI through research, advocacy, and art. She called it the Coded Gaze, the embedded bias of the people behind the code. Her spoken-word film “AI, Ain’t I A Woman?”, which shows facial recognition software misidentifying icons like Michelle Obama, has been screened around the world. And her work was featured in the award-winning documentary Coded Bias, now on Netflix. In 2019, she testified before Congress about the dangers of facial recognition. She warned that even if accuracy improves, the tech can still be abused. For surveillance, racial profiling, and discrimination in hiring, housing, and criminal justice. To counter it, she co-founded the Safe Face Pledge, which demands ethical boundaries for facial recognition. No weaponization. No use by law enforcement without oversight. After years of activism, major players (IBM, Microsoft, Amazon) paused facial recognition sales to law enforcement. In 2023, she published her best-selling book “Unmasking AI: My Mission to Protect What Is Human in a World of Machines.” She advocated for inclusive datasets, independent audits, and laws that protect marginalized communities. She consulted with the White House ahead of Executive Order 14110 on “Safe, Secure, and Trustworthy AI.” But she didn’t stop at facial recognition. She launched Voicing Erasure, a project exposing bias in voice AI systems like Siri and Alexa. Especially their failure to recognize African-American Vernacular English. Her message is clear: AI doesn’t just reflect society. It amplifies its flaws. Fortune calls her “the conscience of the AI revolution.” 💡 In 2025, I’m sharing 365 stories of women entrepreneurs in 365 days. Follow Justine Juillard for daily #femalefounder spotlights.