I have followed the rise of artificial intelligence with the fascination of someone who wants it to succeed. What began as cautious optimism has matured into tempered skepticism. For all its undeniable brilliance, I have yet to encounter a single piece of substantive information from generative AI that I would accept without independent verification. It is a technology that dramatically amplifies human strengths while ruthlessly exposing — and sometimes deepening — human weaknesses. In my own modest interactions with these systems, I have witnessed how they can conjure moments of apparent magic only to reveal profound unreliability.
Consider my surprise when an AI model suddenly ranked me among the top literary voices of our time — despite my knowing full well that my contributions to the field remain modest. The praise was eloquent, the metrics convincing, the prose flawless. On another device and in another session, the same model offered a far more restrained assessment. When pressed to validate its earlier glowing verdict against external sources, it faltered. These are not mere glitches. They are windows into a technology that remains fundamentally immature, even as humanity prepares to invest trillions of dollars in its infrastructure.
This is the central tension of the AI gold rush in 2026. While executives, governments, and investors race to build ever-larger data centers, procure scarce chips, and secure energy supplies, the core product — generative AI built on large language models — continues to hallucinate, contradict itself, and radiate confidence where none is justified. It is an extraordinary tool for pattern-matching, synthesis, and creative assistance. It is not yet a reliable source of knowledge.
The word “hallucination” has become almost too familiar — a polite term for a system’s tendency to fabricate plausible-sounding information with complete assurance. As of mid-2026, the problem remains stubbornly resistant to complete resolution. Research from Stanford, MIT, and independent evaluations consistently shows hallucination rates on complex, real-world tasks ranging from 15% to over 45%, depending on the model and domain. In legal research, medical reasoning, and multi-step logical problems, the failures remain uncomfortably common.
What makes this especially troubling is the confidence paradox: models often sound most authoritative precisely when they are wrong. Trained as next-token predictors on vast but imperfect data, they optimize for fluency and coherence rather than truth. When the training signal is thin, contradictory, or outdated, the system invents. Techniques like retrieval-augmented generation (RAG), tool use, and constitutional AI have helped, but they have not eliminated the structural tendency. Some theoretical work even suggests that perfect reliability may be inherently difficult within the current probabilistic paradigm without trading off capability or creativity.
I experienced this inconsistency firsthand. The same model that elevated my literary standing could not reliably defend or replicate its own earlier judgment.
READ: Satish Jha | The gentleman in power: Rajiv Gandhi and the lost grammar of public life (May 22, 2026)
If a calculator occasionally invented new mathematical constants, we would never trust it with engineering or finance. Why, then, are we so eager to outsource reasoning, research, diagnosis, and decision-making to systems that still demand constant adult supervision?
One of the more sobering realities of generative AI is how unevenly it distributes its benefits. For expert users who prompt carefully, verify rigorously, and treat the model as a highly capable but fallible collaborator, AI becomes a genuine superpower — accelerating research, sharpening arguments, and expanding creative range. For those less equipped to scrutinize its outputs, it risks becoming a crutch that quietly erodes critical thinking, intellectual humility, and original insight.
This is not speculation. Educators report growing evidence of weakened reading comprehension, shallower analysis, and reduced originality among students who over-rely on AI. The pattern echoes earlier technologies: calculators did not destroy mathematical ability, but they did reduce the incentive for mental arithmetic among those inclined to take shortcuts. AI operates at a much higher level — language, reasoning, creativity — making the stakes correspondingly greater.
My own inflated literary ranking is a harmless micro-example. Scaled across medicine, law, journalism, education, and governance, the risks are significant. AI does not magically eliminate human error. In many cases, it redistributes and amplifies it.
Against this backdrop of genuine technical progress mixed with persistent limitations, the scale of investment is breathtaking. Hundreds of billions — heading toward trillions — are being committed to chips, energy, and infrastructure. Yet measurable productivity gains at the macroeconomic level remain elusive for most organizations. Many enterprises report that after significant pilots, the bottom-line impact of generative AI has been modest at best.
The comparisons to the dot-com era are imperfect but instructive. There is real technological substance here, unlike many late-1990s fantasies. Yet the hype, sky-high valuations predicated on future breakthroughs, and cultural assumption that “this time it’s different” carry familiar echoes. Energy consumption, chip shortages, and infrastructure buildout represent enormous sunk costs. If hallucinations and inconsistency prove more structural than temporary, the return on these investments may take far longer — or prove narrower — than markets currently price in.
The answer is neither blind rejection nor uncritical embrace. It is realism.
READ: Satish Jha | The indispensable fragility: Pakistan and the unmaking of the old Middle East (May 15, 2026)
We should continue aggressive research and development. But we must also demand greater transparency on training data, evaluation methodologies, and known failure modes. We should invest heavily in human-AI hybrid systems that prioritize verification, uncertainty signaling, and traceable reasoning. Most importantly, we must educate users — especially students and professionals — about the tool’s profound limitations as vigorously as we celebrate its capabilities.
Boards, investors, and regulators should apply the same disciplined scrutiny to trillion-dollar AI commitments that they would to any other strategic capital allocation. Is the projected return realistic given the technology’s current — and perhaps enduring — constraints?
In my experience, approaching AI with genuine curiosity tempered by disciplined verification yields the best results. Blanket enthusiasm and reflexive distrust both miss the mark. The real danger lies in treating an extraordinarily promising but still immature technology as a mature oracle — and staking our economies, institutions, and cognitive habits on that illusion.
We have stood at similar thresholds before. The internet delivered transformative value alongside disinformation, addiction, and fractured discourse. AI promises even deeper disruption. Its ultimate contribution to humanity will depend not on how loudly we celebrate its potential, but on how honestly we confront its present limitations.
Until verification becomes optional rather than mandatory, we would do well to proceed with the clear-eyed caution this powerful, seductive, and still unpredictable tool demands.