Nine AI Paradoxes That Must Be Addressed

The future with AI was the theme. The setting was Brunel University of London's Research Festival 2026, the university's 60th anniversary, and the occasion was the first live, livestreamed episode of The AI Adoption Podcast.

Five panellists joined me to discuss five topics: opportunities and risks, societal readiness, the agentic organisation, responsible and regulated AI, and the impact on people and jobs. Zahra Bahrololoumi CBE, President and CEO of Salesforce UK and Ireland; Lord Tim Clement-Jones, Liberal Democrat Spokesperson in the Lords on Science, Innovation and Technology and Founder and Co-Chair of the APPG AI; Alex Dalman, Managing Partner at faith; Janusz Marecki, CEO and Co-Founder of Fractal Brain; and Maggie Sarfo, CEO of Meres Consulting. 

Their arguments did not converge. They sharpened into nine paradoxes that resist resolution and demand honest confrontation. I am indebted to the panel for the frankness and openness with which they presented their views, backed by data, real-world examples, and thoughtful analogies. Their generosity in doing so made it possible to identify the cross-cutting tensions that lie at the heart of each paradox. I invite you to listen to the full panel discussion on Spotify, Apple Podcasts or You Tube, (open.spotify.com/show/296zibtjU3w4ANuUsPSu2D), Apple Podcasts (podcasts.apple.com/gb/podcast/the-ai-adoption-podcast/id1811897501), and YouTube (youtube.com/@aiadoption-conversations) and form your own view on where these tensions lead.

These are not philosophical puzzles to be resolved through clever argument. They are structural tensions at the heart of how AI is being built, deployed, governed, and experienced. Each carry real stakes. Each forces a choice and trade-off about who bears the cost when it is resolved in a particular direction. None of them can be settled by assertion, whether from a technology optimist, pessimist, regulator, academic or a boardroom.

AI Is Unreliable by Design. People Are Compensating More Capably Than the Industry Acknowledges.

The underlying architecture of the large language models that dominate public and commercial AI today samples from a distribution over logic rather than reasoning from ground truth. Janusz Marecki made this argument with precision: these systems are pre-trained on internet-scale data that contains inaccurate, toxic, and harmful information, and the technology is non-reliable by construction. And yet the evidence from practice is that people are already discerning, already experimenting, already seeing through the outputs that do not hold up. They are not passive recipients of whatever AI produces. The paradox is that the industry continues to push adoption while relying on human judgment to catch what the technology gets wrong, without acknowledging that this is the arrangement it has put in place.

Regulation Is Necessary. It Will Be Designed for Today's AI Architecture and Overlook Tomorrow's Entirely.

Clarity, consistency, and international standards are the foundation of public trust in AI. That case is well made and I find it persuasive. Transparency requirements, continuous monitoring, and training standards require no bureaucratic leap to adopt and would create a better platform for both enterprise adoption and citizen trust. And yet the counter-argument is one that no regulator has yet answered satisfactorily. Janusz Marecki drew the analogy directly: you could have regulated the jet engine in 1941 and missed the rocket entirely, because the next generation of AI systems may use none of the constructs that today's regulatory frameworks are being built around. The paradox is that the regulation needed to build public trust today may entrench the assumptions that prevent the responsible development of what comes next.

Jobs Are Being Displaced Now. New Jobs May Not Arrive in Time to Absorb That Displacement, Societal Disruption and Inequalities.

Lord Tim Clement-Jones offered an image the famous photograph of Fifth Avenue, one motor car in 1903 and one horse and carriage in 1913. The jobs created by the automobile industry came, eventually, but the pain was real and it took a decade. The view that an economy based on AI will create new jobs quickly enough in the near term to offset the displacement already underway remains untested.  This is particularly the case in service industries where tasks rather than whole jobs are being eliminated, and the risk of compounding existing inequalities is real. Against that sits the evidence from creative industries that junior staff who have embraced AI are leapfrogging, hiring is increasing, and the appetite for AI-native talent is growing. Both observations are accurate. The paradox is that they are happening simultaneously in different sectors, different firms, and different geographies, and the people bearing the displacement cost are not the same people capturing the productivity gain.

Many Organisations Are Unprepared and Muddling Through AI Adoption. Some Are Already Transforming at Scale and Saving Millions a Year.

Organisations do not understand the effort required to set up AI agents correctly, train them on the right data, or govern them appropriately. Alex Dalman was direct: the early rhetoric around AI being easy, cheap, and quick has done genuine damage to realistic expectations, and too many organisations are reaching for the latest shiny toy without asking the business problem they are actually solving. Against that sits a live example of agentic technology deployed at scale on Salesforce’s platform: 82 to 84 percent of complex queries resolved without human intervention, no staff displaced, and millions saved annually, with agents and humans working alongside each other in a way that expands rather than diminishes human capacity. The paradox is not that one picture is wrong. It is that the distance between these two realities is growing rather than closing, and the organisations in the first picture largely do not know it.

Hallucination Is a Failure to Be Eliminated. It Is Also a Commercial Design Variable Some Organisations May Want.

Guardrails in enterprise AI are not about constraint; they are about safety. The ability to use any model while masking personally identifiable data before it enters that model, applying toxicity and bias filters when information returns into the flow of work, and setting a parameter around hallucination calibrated to the specific use case: this is what responsible enterprise AI deployment looks like in practice. And yet the hallucination parameter is not set to zero across all deployments, because some organisations want a degree of generative uncertainty in their context and some simply cannot afford any. Zahra Bahrololoumi made the point with precision: it is not about constraining AI, it is about applying it safely, and that distinction matters enormously for how governance frameworks are designed and implemented. The paradox sits in the gap between those two positions with the industry simultaneously trying to eliminate hallucination as a defect and preserve it as a feature.  The frameworks being built do not yet distinguish clearly between the two.

AI Is Invisible Unlike Previous Technology Revolutions. That Invisibility Is Simultaneously Its Power and the Source of Public Distrust.

The mobile phone, the computer, the motor car have something in common. People could see them, physically touch them, understand what they were interacting with. Alex Dalman made an observation that I think is one of the most important in the entire conversation: AI is the scary thing over here that no one really understands, precisely because it has no physical form to point to. It is embedded in decisions, workflows, and recommendations that arrive without visible authorship. That invisibility is also what makes AI extraordinarily effective, as it enters systems without friction, operates at scale without announcement, and influences outcomes without requiring the user to understand the mechanism. The paradox is that the feature most responsible for AI's transformative power is the same feature most responsible for public fear and resistance. An invisible technology cannot easily demonstrate that it is trustworthy, and trust built on invisibility is fragile by nature.

Enterprise AI Built on Curated, Governed Data Is Trustworthy. Consumer AI Trained on Unfettered, Unchecked, Biased Internet Data Is Eroding the Public Trust That Enterprise AI Depends On.

Successful enterprise AI is categorically different from consumer AI. For enterprise AI to be valuable and trusted, it needs to be grounded in data that has veracity, has been curated, and is engaged with appropriate guardrails: toxicity filters, governance, auditability, traceability, and observability. Zahra Bahrololoumi drew this distinction with force: consumer AI trained on unfettered, uncurated, potentially manipulated data is a categorically different thing from enterprise AI built on governed data, and conflating the two does serious damage to public understanding of what AI actually is and what it can responsibly do. The paradox is that public trust in AI as a category is being shaped primarily by consumer AI experiences, and that erosion of trust threatens the very conditions under which enterprise AI can be adopted and scaled. The two are not separate markets with separate reputations. They share one.

The C-Suite Must Lead AI Transformation. The C-Suite Is the Least Equipped Layer to Do So and Is Delegating Accountability Downward.

Boards and C-suite leaders who are not embracing AI are delegating the responsibility to a tech person, bolting it on, and wasting shareholders' money in the process. Maggie Sarfo's readiness assessment work makes this visible with uncomfortable precision: when organisations are tested on what they are actually doing with AI across culture, strategy, and governance, rather than on whether they are aware of it, many score one or two on a five-point maturity scale. The fears that are not talked about, including the fear that AI will take jobs, are among the factors holding that score down. C-suite leaders need to brief themselves far more rigorously on the possibilities and risks of agentic AI and take personal accountability for the systems they put in place, including decisions about where a human must remain in the loop. The paradox is that the layer of the organisation with the most responsibility for AI transformation is the layer with the least preparation to discharge that responsibility, and the delegation of accountability that results is precisely the condition that makes bad outcomes more likely.

AI Should Make Us More Human. AI Is Removing Human Contact from the Most Consequential Moments in People's Working Lives.

The case for human-centred AI adoption rests on a genuine argument: that AI, properly deployed, releases people from routine work and frees them to apply emotional intelligence, judgment, and the kind of meaning-making that machines cannot replicate. Maggie Sarfo has built her consulting practice around this principle, and I find it persuasive. The analyst role has not disappeared; it has been transformed, with linear processes compressed into real-time prototyping and a much shorter feedback loop. The nature of the work changes; the need does not. And yet Lord Tim Clement-Jones described the experience of young people applying for jobs in service industries: multiple AI assessment stages, no human contact, no feedback, no empathy, no acknowledgment that a person applied. That is soul-destroying, and it is happening at scale. The paradox is that the technology being used to argue that humans will be freed for more human work is, in its current deployment across hiring, customer service, and operations, removing human contact from exactly the moments where it matters most.

These nine paradoxes will not be resolved in a single conversation, a policy paper, or an earnings call. They are the terrain that leaders, policymakers, academics and practitioners have to address without a map that makes the contradictions disappear. The panel at Brunel's Research Festival did not offer false resolution. It offered something more honest: a clear account of what is actually at stake in each tension, and a refusal to pretend that assertion is the same as answer. The question of in whose favour each of these tensions is resolved, and who bears the cost when it is not, is the central unresolved question of the AI era.

Listen to the full panel discussion on Spotify: open.spotify.com/show/296zibtjU3w4ANuUsPSu2D | Apple Podcasts: podcasts.apple.com/gb/podcast/the-ai-adoption-podcast/id1811897501 | YouTube: youtube.com/@aiadoption-conversations