Three problems AI models may never fully solve: hallucination, reasoning limits, and the data wall

01 The uncomfortable question

For two years the public conversation about AI has been dominated by capability announcements: longer contexts, better benchmarks, more convincing demos. A physics-trained commentator pushing in the opposite direction — that current AI models face problems that scale and money may not fix — reached over a million viewers this year. That audience exists because the question is real. A large language model, as Wikipedia defines it, is an AI model trained on vast amounts of text for language generation — the basis of ChatGPT, Claude, Gemini, Grok, and DeepSeek. What that definition quietly contains is the seed of each structural problem: the models learn from text, and text is not truth.

None of these problems is secret. Labs publish about them, researchers measure them, and experienced practitioners engineer around them daily. But the gap between how the technology is marketed and how it behaves under load is where most real-world disappointment lives. Three problems in particular keep resurfacing: hallucination, brittle reasoning, and the tightening supply of training data. Each has a mitigation. None has a cure.

02 Problem one: hallucination is structural, not a bug

A language model generates the most plausible next token given its context. Plausibility and truth overlap often enough to be useful, and diverge often enough to be dangerous. When they diverge, the model does not experience uncertainty — it produces fluent, confident, fabricated detail. This is not a defect that patching removes; it is the direct consequence of training on next-token prediction. The model optimizes for what sounds right, and most falsehoods sound perfectly right.

The mitigations are real but partial. Retrieval augmentation grounds answers in retrieved documents. Citation systems make claims checkable. Calibration research tries to make confidence scores honest. All help, and all fail at the edges: the failure mode returns precisely in the situations where users least expect it, because fluent output feels verified. The engineering answer is to treat every model output as a draft requiring review — an answer that satisfies professionals and frustrates anyone hoping for automation without oversight.

03 Problem two: reasoning that bends under pressure

The second problem is generalization. Models demonstrably solve reasoning problems similar to their training distribution and demonstrably fail on small rearrangements of the same problems. Research on arithmetic, logic puzzles, and planning tasks repeatedly finds this brittleness: performance collapses when surface features change while the underlying structure does not. The model has learned patterns that correlate with correct answers, not algorithms that generate them.

Reasoning-extended systems — models that generate long chains of intermediate steps before answering — have materially improved measured performance on hard problems. But they convert one failure mode into another: longer reasoning chains multiply the places an error can enter, and confident multi-step derivations that end wrong are harder to audit than a single wrong sentence. The improvement is genuine; the residual risk just changes shape.

04 Problem three: the data wall

The third problem is arithmetic. Model capability has scaled with training compute, and compute has grown far faster than the supply of high-quality public text. Frontier labs have consumed on the order of the well-curated public web; the remainder is lower quality, duplicated, or machine-generated — and training on machine-generated text risks degrading the very abilities being scaled. Synthetic data, licensed archives, and human-annotated material all extend the runway, and all cost more per token than scraping did.

This is why the industry's enthusiasm for multimodal and embodied data is not only about capability. Video, audio, and interaction logs are the remaining large pools of human-generated signal. Whether they teach the same lessons that text does is an open research question, which is precisely the point: the scaling recipe that produced the current generation is running out of its original ingredient.

05 What the labs are actually doing

The response to all three problems is visible in product decisions. Hallucination pressure produced citations, search grounding, and refusal behavior. Reasoning brittleness produced the reasoning-model class and heavier investment in verification at inference time. The data wall produced a quiet licensing race — publishers, archives, and platforms signing training deals — plus synthetic data pipelines with quality filters. Stanford's AI Index series tracks these capability-and-limits trends annually and is the best single reference for the current state of play.

06 Living with the limits

The practical conclusion is neither hype nor dismissal. These systems compress and deploy human knowledge at a scale no previous tool approached, and they fabricate, stumble, and consume data at a scale no previous tool approached. Both facts are true simultaneously, and the second does not cancel the first.

For users, the discipline is matching task to failure mode. Ideation and drafting tolerate hallucination; medical, legal, and financial claims do not. Verification effort should concentrate where being wrong is expensive. For the industry, the honest framing is that the next capability jump must come from architecture, data quality, or verification — not from repeating the last jump louder. The models are extraordinary. They are also, in three specific and well-documented ways, unfinished.