Read More
Generative AI’s greatest strength is range: the ability to generate options, structures, framings, and drafts at scale. But range without limits is a dangerous illusion.
ADVERTISEMENT
SCROLL TO CONTINUE WITH CONTENT
GenAI fails for the same reason it works. It predicts. It does not verify. It produces plausible continuations, not guaranteed truths. Some failures are not accidental; they’re structural.
The first category is intrinsic limits: boundaries built into the model itself. AI has no built-in mechanism for truth. It can fabricate a source, misstate a number, or present an incomplete answer with complete confidence. It can also struggle with exact calculation because numbers are processed as tokens, not mathematical objects. Better prompting can reduce these failures, but not remove the limit.
The second category is interaction limits: boundaries that arise in use. AI only works with what is visible in the conversation. Context can be missing, buried, or poorly framed. Small changes in wording can shift the answer. A long prompt can look comprehensive while the model quietly underweights the middle. These limits are managed through better specification, structure, and review.
The third category is governance limits: boundaries built around the model. The AI you use is not raw capability. It comes inside a product, wrapped in guardrails set by its creator: safety rules, access permissions, memory settings, retrieval tools, interface design, and commercial choices. These guardrails decide what the system can see, refuse, remember, and produce. Sometimes the limit is not the engine. It is the road, the rules, and the fence around it.
This is why “AI made a mistake” is often the wrong diagnosis. The better question is: which limit did we fail to account for?
AI’s limitations do not make it useless. They make human judgment necessary. Mature AI use begins when we stop asking the system to be unlimited and start designing work around what it cannot do.
Frank Ng is a retired NASDAQ CEO, who co-authors this column with his son Ryan after publishing their book Hey AI, Let’s Talk!
𝗗𝗼𝘄𝗻𝗹𝗼𝗮𝗱 𝗧𝗵𝗲 𝗦𝘁𝗮𝗻𝗱𝗮𝗿𝗱 𝗔𝗽𝗽 ↓
