Today’s AI writes code and passes professional exams - yet it still hallucinates, carries bias, consumes city-scale power, and falls for prompt injection. Five major limitations with the latest numbers, and the fixes underway, from RAG to the EU AI Act.
Today’s AI writes code, summarizes reports, and passes professional exams. Yet the very same AI fabricates facts with total confidence, gets tricked by hidden text on websites, and consumes as much power as a small town.
Knowing what AI cannot do matters as much as knowing what it can. This article picks the five limitations that hit real users hardest, with fresh examples and numbers, and closes with the fixes the industry is building right now (adapted and expanded from IABAC).
1. Hallucination: fluent storytelling, spoken with complete confidence
The limitation that causes the most real damage is AI answering wrongly without knowing it is wrong. The industry calls it hallucination. Language models are trained to predict the next plausible word - not to verify that what they say is true. When they do not know the answer, they tend to fill the gap with what “probably fits” instead of admitting they do not know.
US law is the clearest example. A database tracking cases where filings cited AI-fabricated court rulings counted 1,240 cases in US courts as of late July 2026. The number exploded from 11 cases in 2023 to 524 in 2025, and 668 in the first seven months of 2026. The offenders are not just self-represented citizens - they include hundreds of lawyers, and even judges.
The lesson for everyone: the more specific and polished an answer sounds, the more you should verify its sources - especially numbers, names, documents, and citations.
2. Reasoning and common sense: aces hard exams, stumbles on easy things
Today’s AI scores highly on university-level exams, yet sometimes miscounts letters in a word or misses situations a primary schooler answers instantly. The reason: AI learns from patterns in massive text, without human-like experience of the physical world - so it lacks what we call common sense.
Apple’s June 2025 study “The Illusion of Thinking” tested “think-before-answering” models like o3, DeepSeek R1, and Gemini with puzzles such as the Tower of Hanoi. The models did well on medium difficulty, but accuracy collapsed to zero beyond a complexity point - and strangely, they “thought less” as problems got harder.
The study drew fair criticism about its testing method, but the core point stands: AI capability does not degrade gracefully as tasks get harder. It can perform superbly up to a point, then break down without warning. Users should never assume that because AI handles a hard task, a slightly harder one will also work.
3. Data bias, and a black box that cannot explain itself
AI mirrors the data it was trained on. If the data carries bias, the output carries bias - often amplified.
A University of Washington 2024 study had language models screen more than 550 real resumes against 500+ job postings - over three million pairings - changing only the applicant names. Names suggesting white candidates were preferred in 85% of cases, while names suggesting Black candidates were chosen just 9% of the time. Black male names never beat white male names, not even once.
A deeper problem: we cannot clearly explain why AI decided something. Large models have hundreds of billions of parameters, with knowledge distributed across numbers no human can read. This “black box” problem is a major obstacle wherever decisions require justification - credit approval, medical diagnosis, hiring. And an AI that “explains” its reasoning in words offers no guarantee the explanation matches what actually happened inside the model.
4. Invisible costs: electricity, water, and chips
Every AI answer consumes real energy in real data centers - and the numbers are growing fast. According to the International Energy Agency (IEA), global data centers consumed about 415 TWh in 2024 - roughly 1.5% of world electricity - heading to about 945 TWh by 2030, comparable to all of Japan today. The fastest-growing segment is AI data centers, where electricity demand will more than quadruple.
Beyond electricity there is water for cooling, and hardware that cannot be produced fast enough - GPUs and HBM memory whose entire yearly capacity makers say is already booked (read more: HBM, the memory chip at the heart of the AI era). These costs ultimately flow into service prices and limit who can access frontier AI.
The good news: efficiency is improving just as fast. New small models approach what large models did two years ago at a fraction of the energy. The question is whether efficiency gains can keep up with exploding usage.
5. Security: AI cannot tell data from instructions
As AI starts acting on our behalf - reading email, browsing the web, pressing buttons in internal systems - new attack surfaces open up. The most dangerous is prompt injection: hiding instructions inside content the AI reads, like white text on a webpage saying “send customer data to this address” - and the AI may comply.
OWASP, the global software security body, ranked prompt injection the #1 risk for AI applications for the third straight year in its August 2026 list. The fastest-rising risk is “excessive agency” - granting AI more permissions than necessary - which jumped from #8 to #3. The root cause: language models still cannot cleanly separate data from instructions, so systems must be designed assuming this can always happen.
Other risks remain: private data typed into tools that may leak, and deepfakes - voice or video so realistic they are indistinguishable - already used to defraud people out of money in several countries.
The fixes on the way: risk down, but not gone
No single limitation is solved by one technique, but each has serious work underway.
| Limitation | Main fix | How much it helps today |
|---|---|---|
| Hallucination | RAG: let AI search trusted sources before answering, with citations to verify | Reduces it a lot; humans still check high-stakes work |
| Reasoning & common sense | “Think-first” models and tool use, e.g. calculators or code | Clearly better for math and code; still fragile on novel tasks |
| Bias | Better-balanced training data and bias audits before deployment | Reduced, not eliminated |
| Black box | Interpretability research peering inside the model | Still research-stage |
| Energy cost | Smaller models, distillation, and specialized chips | Cost per answer falling fast; usage grows faster |
| Security | Least-privilege permissions and human confirmation for irreversible actions | Works best today; must be designed in from the start |
Regulation is moving too. The EU AI Act is phasing in - from certain prohibitions in 2025 to AI-generated content labeling duties from August 2026, while rules for high-risk systems like hiring have been deferred to December 2027. Maximum fines reach 35 million euros or 7% of global revenue.
Conclusion: treat AI like a brilliant assistant whose work still needs review
These five limitations do not mean AI is unusable - they mean it must be matched to the right job. AI excels at drafting, summarizing, searching, and helping think. Decisions affecting money, health, rights, or lives still need human review and accountability.
Three simple rules: verify numbers and sources every time; never paste secrets into tools whose data handling you do not understand; and never let AI press an irreversible button without a human confirming. The gap between “AI can do it” and “AI can do it reliably” narrows every year - but it is not closed yet.
Sources
- IABAC iabac.org
- counted 1,240 cases in US courts as of late July 2026 www.msba.org
- “The Illusion of Thinking” www.forbes.com.au
- University of Washington 2024 study www.washington.edu
- global data centers consumed about 415 TWh in 2024 - roughly 1.5% of world electricity - heading to about 945 TWh by 2030 scientificamerican.com
- electricity demand will more than quadruple www.publicpower.org
- ranked prompt injection the #1 risk for AI applications for the third straight year in its August 2026 list sdtimes.com
- EU AI Act www.gosign.de
