Addressing AI failures without fearmongering

I once watched a company spend almost ₹2 crores on an AI tool nobody used. The tech was brilliant, but The rollout was a disaster. They focused 100% on the tool's capabilities and 0% on the team's fears. People whispered: "Will this replace me?" "Should I start job hunting?" "Is this just cost-cutting in disguise?" I’ve coached dozens of leaders through AI transitions. Here’s the 4-step framework I now teach to fear-proof every rollout: 1. Address the elephant first.  Start by saying, "I know new tech can be unsettling. Let's talk about what this means, for us, as people." Acknowledging the fear directly is the only way to dissolve it. 2. Position it as a "Co-pilot," not a "Replacement."  Show them how the tool will remove repetitive tasks, so they can focus on creative, strategic work. Give concrete examples of what they'll gain, not just what the company will save. 3. Create "Peer Advocates."  Train early adopters first and let them share their positive experiences peer-to-peer. Trust spreads faster sideways than top-down. 4. Establish a "Human-in-the-Loop" rule.  Make it clear that the final decisions, the creativity, and ethical judgments will always be made by a person. AI is a tool, not the new boss. The success of any AI rollout isn't measured in processing power. It's measured in team trust. What's your biggest concern when a new AI tool is introduced at work? #AI #Leadership #ChangeManagement #TeamCulture #SoftSkillsCoach

Yesterday I watched AI fail in front of 15,000 developers. At WeAreDevelopers, one of Europe’s biggest dev conferences, GitHub engineers tried something bold: A live demo of GitHub’s new AI coding agents. We all leaned in. First run? It hung. The presenter had to restart. Second try? It generated code… but with an error. No worries — he copied the error into the prompt and asked AI to fix it. Failed again. Then again. And again. After 5 retries, he gave up. And I think 15,000 of us were collectively sweating for him. But here’s the thing: They came to prove that 90% of your code can be written by AI. Instead, they accidentally proved something more valuable: AI can write code. But it can’t (yet) understand your intent, your domain knowledge, or the real problem behind a feature request. Without a developer behind the wheel, AI is just guessing. And guessing doesn’t scale. AI will be a superpower — but only for those who already know how to code. Not a replacement. A multiplier. Let’s stop pretending otherwise. What do you think — are we too quick to put AI in the driver’s seat?

Apple recently published a research paper on the limits of large language models (LLMs), and it’s sparked a lot of conversation. The paper argues that as logic puzzles get more complex, LLMs “collapse” or give up. Some have even called it a “knockout blow” to today’s AI.    But we need more context.    The paper treats LLMs like traditional software—expected to follow precise rules and deliver exact answers. But LLMs aren’t rule-based engines. They’re built to generate, assist, and adapt—not compute like a calculator. That’s a fundamental difference.    Some of the failures cited in the paper likely stem from token limits or from testing LLMs in ways that don’t reflect how they’re actually designed to be used. In several examples, the models didn’t quit; they summarized, generalized, or offered a broader view. That’s not failure. It’s a different kind of intelligence, and it’s what makes them so powerful in collaborative tasks like those we see with Copilot every day.    So, what’s the real takeaway?    ✅ There’s a learning curve—not just in how we build AI, but in how we understand and use it. In customer conversations, I’m finding that the more we clarify what AI can and cannot do, the more value customers are able to unlock.  ✅ Use cases matter. These models perform best when applied to the right types of problems, especially those that rely on language, context, and insight rather than strict logic.    This paper doesn’t diminish AI’s potential. Instead, it highlights the importance of setting the right expectations and helping customers align the technology with the task to realize real business value. https://lnkd.in/gBhkzYkA

One of the most important contributions of Google DeepMind's new AGI Safety and Security paper is a clean, actionable framing of risk types. Instead of lumping all AI risks into one “doomer” narrative, they break it down into 4 clear categories- with very different implications for mitigation: 1. Misuse → The user is the adversary This isn’t the model behaving badly on its own. It’s humans intentionally instructing it to cause harm- think jailbreak prompts, bioengineering recipes, or social engineering scripts. If we don’t build strong guardrails around access, it doesn’t matter how aligned your model is. Safety = security + control 2. Misalignment → The AI is the adversary The model understands the developer’s intent- but still chooses a path that’s misaligned. It optimizes the reward signal, not the goal behind it. This is the classic “paperclip maximizer” problem, but much more subtle in practice. Alignment isn’t a static checkbox. We need continuous oversight, better interpretability, and ways to build confidence that a system is truly doing what we intend- even as it grows more capable. 3. Mistakes → The world is the adversary Sometimes the AI just… gets it wrong. Not because it’s malicious, but because it lacks the context, or generalizes poorly. This is where brittleness shows up- especially in real-world domains like healthcare, education, or policy. Don’t just test your model- stress test it. Mistakes come from gaps in our data, assumptions, and feedback loops. It's important to build with humility and audit aggressively. 4. Structural Risks → The system is the adversary These are emergent harms- misinformation ecosystems, feedback loops, market failures- that don’t come from one bad actor or one bad model, but from the way everything interacts. These are the hardest problems- and the most underfunded. We need researchers, policymakers, and industry working together to design incentive-aligned ecosystems for AI. The brilliance of this framework: It gives us language to ask better questions. Not just “is this AI safe?” But: - Safe from whom? - In what context? - Over what time horizon? We don’t need to agree on timelines for AGI to agree that risk literacy like this is step one. I’ll be sharing more breakdowns from the paper soon- this is one of the most pragmatic blueprints I’ve seen so far. 🔗Link to the paper in comments. -------- If you found this insightful, do share it with your network ♻️ Follow me (Aishwarya Srinivasan) for more AI news, insights, and educational content to keep you informed in this hyperfast AI landscape 💙

AI-first isn’t failing. But incorrect implementation is. Klarna replaced 700 staff with AI. Now they’re quietly rehiring humans. Duolingo slashed human translators. Then faced mass exodus by their users. This isn’t about AI failing. It’s about companies not thinking through: 1. Where to use it. 2. How to deploy it. 3. How to communicate it. AI isn’t plug-and-play. It’s not a replacement strategy. It’s a human partner strategy. When it’s deployed well: • Customer service gets sharper, not colder • Creative work gets faster, not flatter • Teams become augmented, not redundant But that takes strategy. Workflow redesign. Thinking. Training. Not just an “implement AI everywhere” approach. Knowing where NOT to use AI is just as important. The real risk isn’t adopting AI…it’s doing it badly. What’s the biggest challenge you face implementing AI today?

A lesson from self-driving cars… Healthcare's AI conversation remains dangerously incomplete. While organizations obsess over provider adoption, we're neglecting the foundational element that will determine success or failure: trust. Joel Gordon, CMIO at UW Health, crystallized this at a Reuters conference, warning that a single high-profile AI error could devastate public confidence sector-wide. His point echoes decades of healthcare innovation: trust isn't given—it's earned through deliberate action. History and other industries can be instructive here. I was hoping by now we’d have fully autonomous self-driving vehicles (so my kids wouldn’t need a real driver’s license!), but early high-profile accidents and driver fatalities damaged consumer confidence. And while it’s picking up steam again, but we lost some good years as public trust needed to be regained. We cannot repeat this mistake with healthcare AI—it’s just too valuable and can do so much good for our patients, workforce, and our deeply inefficient health systems. As I've argued in my prior work, trust and humanity must anchor care delivery. AI that undermines these foundations will fail regardless of technical brilliance. Healthcare already battles trust deficits—vaccine hesitancy, treatment non-adherence—that cost lives and resources. AI without governance risks exponentially amplifying these challenges. We need systematic approaches addressing three areas:   Transparency in AI decision-making, with clear explanations of algorithmic conclusions. WHO principles emphasize AI must serve public benefit, requiring accountability mechanisms that patients and providers understand.   Equity-centered deployment that addresses rather than exacerbates disparities. There is no quality in healthcare without equity—a principle critical to AI deployment at scale.   Proactive error management treating mistakes as learning opportunities, not failures to hide. Improvement science teaches that error transparency builds trust when handled appropriately. As developers and entrepreneurs, we need to treat trust-building as seriously as technical validation. The question isn't whether healthcare AI will face its first major error—it's whether we'll have sufficient trust infrastructure to survive and learn from that inevitable moment. Organizations investing now in transparent governance will capture AI's potential. Those that don't risk the fate of other promising innovations that failed to earn public confidence. #Trust #HealthcareAI #AIAdoption #HealthTech #GenerativeAI #AIMedicine https://lnkd.in/eEnVguju

What happens when your AI system is wrong at 2am? Not inaccurate. Not biased. Simply wrong and operating at scale. Most organisations still define robustness at the model level. We talk about validation, fairness metrics, explainability reports. Those matter. But they do not answer the operational question: What happens when the system fails? In my latest article, “Designing Resilient AI Systems - What Robustness Actually Looks Like,” part of The Data Science Decoder newsletter, I explore resilience beyond abstract trust. Robustness is not about building a perfect model. It is about engineering systems that: 💠 Contain failure rather than amplify it 💠 Reduce autonomy as uncertainty rises 💠 Escalate to humans through defined pathways 💠 Limit blast radius through deliberate architectural design Resilient AI is structural. It is designed into redundancy, monitoring, containment, and decision rights, not added through policy language. For senior leaders, this reframes governance. The question shifts from “Is the model accurate?” to “Is the system controllable under stress?” For practitioners, it changes design priorities. Monitoring must connect to consequence. Escalation paths must be rehearsed. Automation must be elastic. If we continue optimising for frictionless scale without engineering for failure, we are building fragile systems. If this tension resonates with you, particularly in regulated or high-impact environments, the full article goes deeper into the architecture and economics of resilience. I’d be interested in your perspective: where have you seen AI systems amplify risk rather than absorb it?

Similar posts have gained significant traction in recent days, largely due to their sensationalist and clickbait-style framing. While it is true that hallucinations remain a challenge in LLMs, the way these articles present the issue is both misleading and lacking context. They imply that AI models are fundamentally unreliable, without acknowledging the vast improvements made over successive iterations, nor the actual mechanisms available to mitigate and manage inaccuracies in real-world applications. It is important to clarify that hallucinations in large language models (LLMs) are not an all-or-nothing issue. They are a natural consequence of how these probabilistic models generate text. LLMs do not "know" facts in the same way humans do; rather, they predict the most statistically likely sequence of words based on their training data. Given the nature of language, absolute truthfulness is an unrealistic expectation for any generative AI system, just as no human expert can guarantee 100% accuracy in all responses. These articles treat hallucinations as catastrophic failures, but in most cases, they are small factual deviations or statistical approximations rather than deliberate "lies." Many corporate applications do not require 100% accuracy—instead, they rely on AI for drafting, summarization, ideation, and pattern recognition, where efficiency and augmentation are more critical than absolute correctness. The Reality of Enterprise AI Use Cases: Most real-world implementations of AI, especially in enterprise settings, do not rely on raw LLM outputs alone. Instead, they use retrieval-augmented generation (RAG), knowledge-grounding, and human-in-the-loop systems to ensure factual accuracy. AI is not meant to replace domain expertise but to enhance and accelerate human decision-making. What this article fails to acknowledge is that the real challenge is not hallucination elimination (which is impossible) but trust calibration. The reality is that, despite the imperfections, AI adoption is accelerating across industries. From legal document review to financial risk analysis, organizations are using AI successfully because they understand that intelligence—human or artificial—is always probabilistic, not absolute. These articles portray AI as "failing" due to hallucinations, when in reality, the best models are outperforming previous generations and demonstrating increasingly useful precision. The discussion should not be about whether hallucinations exist—they always will—but rather about how businesses and users can optimize AI’s reliability through proper integration, contextual grounding, and expert oversight. AI is not perfect, but neither is human reasoning. The future belongs to those who can intelligently harness the power of AI while maintaining a nuanced understanding of its strengths and limitations.

Enough of the 95% fail myth. Let’s be clear: the “95% of GenAI programs are failing” headline from MIT is clickbait dressed up as research. And LinkedIn has turned it into a daily echo chamber of doom. Here’s the problem: • If 95% of experiments fail, that’s not a crisis; that’s called innovation. • If companies launch AI pilots with no clear use case, of course, they flop; but that’s not on GenAI, it’s on execution and leadership. • Declaring “failure” when we’re barely two years into mainstream adoption is like calling the internet a bust in 1995 because most websites looked useless. We need to stop glamorizing failure stats and start having a grown‑up conversation: - Which programs are actually delivering value? - What distinguishes serious enterprise adoption from “me too” experiments? - How do we move from hype metrics → impact metrics? GenAI isn’t failing. What’s failing is the lazy narrative that lumps immature experiments, poor strategy, and hype‑driven noise into a single “95% fail” bucket. Let’s stop repeating headlines. Let’s start building. Failure is data. Hype is noise. Progress is a choice. #MIT #GenAI #Failure #Hype

𝐖𝐡𝐞𝐧 𝐀𝐈 𝐅𝐚𝐢𝐥𝐬, 𝐈𝐭 𝐃𝐨𝐞𝐬𝐧’𝐭 𝐁𝐫𝐞𝐚𝐤 𝐭𝐡𝐞 𝐂𝐨𝐝𝐞—𝐈𝐭 𝐁𝐫𝐞𝐚𝐤𝐬 𝐭𝐡𝐞 𝐁𝐫𝐚𝐧𝐝. 𝐀𝐫𝐞 𝐘𝐨𝐮 𝐑𝐞𝐚𝐝𝐲? AI won’t just test your systems—it will test your story. And when it fails, customers don’t debug code. They judge your brand. The Cruise robotaxi crisis showed us something deeper than a product flaw. It revealed what happens when trust, transparency, and timing break down in an AI-driven world. Julian De Freitas’ new Harvard Business Review piece puts it bluntly: 𝐖𝐡𝐞𝐧 𝐀𝐈 𝐟𝐚𝐢𝐥𝐬, 𝐜𝐨𝐧𝐬𝐮𝐦𝐞𝐫𝐬 𝐝𝐨𝐧’𝐭 𝐛𝐥𝐚𝐦𝐞 𝐭𝐡𝐞 𝐜𝐨𝐝𝐞—𝐭𝐡𝐞𝐲 𝐛𝐥𝐚𝐦𝐞 𝐭𝐡𝐞 𝐜𝐨𝐦𝐩𝐚𝐧𝐲. This hits the C-suite squarely where it matters most: brand equity, customer belief, and enterprise trust. But this isn’t a reason to fear AI. It is the reason to 𝐥𝐞𝐚𝐝 𝐢𝐭 𝐛𝐞𝐭𝐭𝐞𝐫. Here’s what bold, trusted brands do 𝘣𝘦𝘧𝘰𝘳𝘦 the heat hits: ✅ Set honest expectations about what AI can and can’t do ✅ Avoid over-humanizing tech which is still evolving ✅ Ground your marketing in 𝘳𝘦𝘢𝘭𝘪𝘵𝘺, not just roadmaps ✅ Tell the 𝘸𝘩𝘰𝘭𝘦 story—especially when things go wrong ✅ Prepare your cross-functional crisis muscle: Legal, Comms, Product, Risk, and People all at the same table Here’s what I tell my Board friends and exec teams: 𝐀𝐈 𝐰𝐢𝐥𝐥 𝐬𝐡𝐚𝐩𝐞 𝐭𝐡𝐞 𝐟𝐮𝐭𝐮𝐫𝐞 𝐨𝐟 𝐯𝐚𝐥𝐮𝐞 𝐜𝐫𝐞𝐚𝐭𝐢𝐨𝐧, 𝐛𝐮𝐭 𝐨𝐧𝐥𝐲 𝐢𝐟 𝐭𝐫𝐮𝐬𝐭 𝐬𝐜𝐚𝐥𝐞𝐬 𝐰𝐢𝐭𝐡 𝐢𝐭. This is our leadership crucible.  1. The brands that win won’t just build great AI. 2. We build it with trust, transparency, and truth at the core. 3. We design Agentic AI systems that enhance our brand, deepen customer belief, and empower teams to move with confidence. The future belongs to those who don’t just build AI—but build trust into its foundation. Are you AI-ready 𝘢𝘯𝘥 brand-ready? Read the article here: https://lnkd.in/g7qsbXme #AgenticEnterprise #BrandLeadership #TrustByDesign #CrisisManagement #AIethics #DigitalTransformation #CEOInsights #CMO #EnterpriseAI #ResponsibleAI #BoardStrategy #LeadershipMatters