Once again, AI was one of the biggest topics at the 2026 World Economic Forum.
For the first time this year, the tone was noticeably tense. Companies and journalists anxiously asked deep tech leaders questions about AI security, governance, the burden on infrastructure, whether the dreaded “AI bubble” is really a bubble, and when investments will start to yield economic returns. In other words, the stakes have never been higher.
VP of Engineering at LikelyAI.
Of the many AI leaders who spoke at Davos, Microsoft CEO Satya Nadella came closest to hitting the mark. He warned that AI can only avoid becoming a bubble if it produces real outcomes that are widely distributed, rather than concentrating value in a handful of companies and economies.
Unreliable AI (particularly the issue of hallucinations) further deepens the lack of trust in businesses and prevents positive economic impact.
Why today’s AI debate is starting in the wrong place
Much of the conversation at Davos reflected the reality that today’s dominant AI systems, large-scale language models (LLMs), are currently receiving the most power, attention, and investment.
Regulations, infrastructure plans, and economic models are all built on that reality. As a result, hallucinations are treated as regrettable but unavoidable risks that can be disclosed or mitigated.
LLMs are probabilistic systems that produce output by predicting what will happen next based on statistical patterns learned from large datasets. This is why they are linguistically fluent and flexible, but also why they hallucinate.
If LLM produces convincing but incorrect answers, that’s not a bug, but a result of how it’s designed.
Although hallucinations were frequently discussed at Davos as a governance and security issue, hallucinations are inherent in the probabilistic approach itself.
This distinction is important because it determines whether to treat hallucinations as a workaround or as a signal that a different system design may be required for a particular use case.
If hallucinations are treated as inevitable, the only countermeasures available are warnings, disclaimers, human oversight, and increasingly complex guardrails.
That’s why much of the conversation at Davos focused on disclosure, risk transfer, and regulation. All of these are necessary, but none of them can turn an unreliable system into a reliable infrastructure.
Combination of flexibility and reliability
So what are the alternatives? The fact that probabilistic models are not the only way to build AI systems. Long before generative AI gained public attention, symbolic reasoning systems were used to encode knowledge as explicit rules, facts, and constraints.
These systems make no guesses. Given the same input, it always produces the same output.
Most people interact with symbolic systems every day without thinking. Spreadsheets are just one example. When the spreadsheet calculates the results, the user doesn’t have to worry about hallucinating another answer that “sounds right.” Businesses want and need this same determinism from AI.
Most of the software in use today is symbolic systems. They don’t handle natural language processing very well, but that’s the beauty of LLMs. However, the choice between neural and symbolic is not an either/or.
There is now a growing class of hybrid systems known as neurosymbolic AI, which intentionally combine the best of both approaches.
Neural networks are used when flexibility is needed, such as when interpreting language or extracting information from documents, while symbolic reasoning layers apply explicit rules, constraints, and logic to determine outcomes.
Importantly, this means that the output does not depend solely on statistical validity. Neural symbolic systems can track how conclusions are reached, produce the same results for the same inputs, and clearly signal when a question cannot be answered confidently.
Such characteristics are essential in environments where decisions must be explained, audited, and defended.
Cost of losing alternatives
This narrow focus has real consequences. Much of the anxiety expressed at Davos stems from the true limitations of LLM: a system that offers extraordinary capabilities but grapples with unavoidable reliability challenges.
When these limitations become apparent, trust is lost, human oversight becomes mandatory, and productivity gains become difficult to achieve.
Many organizations find it difficult to scale pilot projects, especially when legal and compliance teams raise concerns about deliverables that cannot be reliably defended or audited.
While ROI results vary across industries, a recurring challenge is that the systems that provide the best functionality are not designed to justify their decisions.
Businesses at Davos were right to ask how AI should be managed, regulated, and integrated into the global economy. Neurosymbolic systems offer a natural solution to business concerns about implementation. In LLM-only systems, reliability and explainability introduce new risks.
But we can’t meaningfully answer these questions without first opening up the conversation about what AI actually is.
Most practitioners understand the limitations of LLM and are already discussing mitigation strategies. However, there is a difference between mitigating inherent limitations and choosing an architecture that avoids them for a particular use case.
The question is not whether to abandon LLM, but whether to default to LLM too easily, even when reliability requirements suggest a different approach.
If AI is to power the economy rather than just impress in demonstrations, reliability cannot be an afterthought. For a design to be auditable, compliant, and reliable from the beginning, it must become a standard.
Davos 2026 raised some pressing questions, and the answers already exist in the LLM’s approach, which combines flexibility and deterministic reasoning.
Much of the discussion still treats hallucinations as inevitable, rather than recognizing that hallucinations are inherent in probabilistic systems and that alternatives exist for use cases where reliability is paramount.
Reliable AI is not something we wait for us to invent. It already exists in the form of neuro-symbolic AI. Until that reality is reflected in mainstream developments, the gap between Davos’ ambitions and what organizations can safely rely on will continue to widen.
We’ve featured the best AI website builders.
This article is produced as part of TechRadarPro’s Expert Insights channel, featuring some of the brightest minds in technology today. The views expressed here are those of the author and not necessarily those of TechRadarPro or Future plc. If you’re interested in contributing, find out more here. https://www.techradar.com/news/submit-your-story-to-techradar-pro
