What AI Still Cannot Do

Current AI systems cannot reliably tell you when they are wrong, cannot know what happened after training ended unless something looks it up for them, cannot be counted on to repeat a success once conditions change, and cannot be the party held responsible for an outcome. Most are engineering problems being narrowed year by year. The last is not an engineering problem at all.

The limits are not where intuition puts them

The usual answer is a list of human virtues: empathy, creativity, judgment. Comforting, and mostly untested. The measured picture is stranger.

In a randomised field experiment with Boston Consulting Group, published in Organization Science in 2026, 758 consultants worked with and without GPT-4. On tasks inside the model's competence, the consultants using AI finished more of them and worked faster. On one task deliberately placed just outside it, they were markedly less likely to reach the correct answer than colleagues working without AI at all. The authors called the pattern the jagged technological frontier: the boundary between what AI does well and badly is irregular, and it does not track human difficulty. The size of each of those effects is set out in the evidence on AI and productivity.

Stanford's 2026 AI Index illustrates it by putting two different systems side by side. Gemini Deep Think took a gold medal at the 2025 International Mathematical Olympiad. On ClockBench, an unrelated test of reading analogue clocks, the strongest model measured was GPT-5.4 High, correct 50.6% of the time in March 2026 against 90.1% for humans on correctly formatted clocks. Separate systems, separate tests, which is what makes the pairing useful: the field can take the olympiad medal and still lose to the clock. The International AI Safety Report 2026, chaired by Yoshua Bengio, puts it plainly: leading systems "may excel at some difficult tasks while failing at other, simpler ones." Unevenness on that scale is also why the question of how close AGI is has no clean answer.

It cannot reliably tell you when it is wrong

A language model produces the next likely piece of text. Nothing in that process separately checks whether the result is true, and the model's confidence is a poor guide to its accuracy.

Researchers at OpenAI and Georgia Tech argued in September 2025 that this is partly a scoring problem. Most evaluations award a point for a correct answer and nothing for admitting uncertainty, so a system optimised against them learns that guessing beats abstaining, as a student does on a multiple-choice exam. Their remedy is to change what leaderboards reward. That is a testable claim about incentives from people with an interest in the field, not a settled finding.

Looking information up helps and does not fix it. Stanford's RegLab tested three purpose-built legal research tools that retrieve real legal sources before answering, across 202 queries, as those products stood in 2024. Lexis+ AI hallucinated on about 17% of them, Westlaw AI-Assisted Research on about a third, and Ask Practical Law AI on about 19% while declining to answer properly on 62%. Accurate answers ran from 65% down to 19%. These are products sold to lawyers on the strength of grounding answers in sources. The mechanism is covered in depth in why AI makes things up.

It does not know what happened after it stopped learning

A model's factual knowledge is fixed when its training data ends. That date, the knowledge cutoff, is published by most developers and is softer than it sounds. Researchers at Johns Hopkins found in 2024 that the effective cutoff, the point where knowledge of a given source actually stops, often differs from the reported one and varies by topic, because large web crawls carry older material inside newer collections and deduplication misses near-copies. A stated cutoff is an approximation, not a line.

Search and retrieval close most of the gap, which is what RAG does. Retrieval changes where the answer comes from. It does not give the system a way to judge whether the page it found is any good.

A benchmark score is not a job reference

Benchmark results are the main public evidence about AI capability, and weaker than their precision suggests. Three documented problems sit underneath them.

Tests get used up. Humanity's Last Exam was designed to stay difficult and still gained about 30 percentage points between 2024 and 2025. That is benchmark saturation, and it shortens the window in which a test can separate systems.

Tests contain errors. A 2025 review cited by the AI Index found invalid question rates from 2% on MMLU Math to 42% on GSM8K, one of the most quoted maths benchmarks in the field.

Tests leak. When benchmark items reach training data, a score measures recall rather than capability, and contamination is hard to rule out from outside. The AI Index cites the 2025 dispute over Meta's Llama 4, which the company denied.

The gap also opens between two tests of the same kind. Robots reached 89.4% success on RLBench, eighteen short manipulation tasks in a controlled simulation, as of January 2026. On BEHAVIOR-1K, which is also a simulation but sets long household activities drawn from what people said they wanted help with, the best team in the 2025 challenge finished 12.4% of tasks outright. Both figures come from simulators. Lengthening the task, rather than leaving the simulator, is what moved one number to the other.

It does not hold up when the conditions change

A model learns the statistical regularities of the data it trained on. Move it where those regularities differ and performance falls, usually with no warning in the output.

The clearest documented case is medical. The Epic Sepsis Model was running in hundreds of US hospitals. Researchers at Michigan Medicine validated it externally across 27,697 patients and 38,455 hospitalisations and published in JAMA Internal Medicine in 2021. It scored an area under the curve of 0.63, against the 0.76 to 0.83 the developer had reported, and at the recommended alert threshold caught 33% of sepsis cases, missing 1,709 of 2,552 patients who had sepsis while firing alerts on 18% of all hospitalisations.

The model worked where it was built and did not transfer. That is why generalization and concept drift are first-order concerns for anyone deploying these systems, and why a vendor's validation figures describe their setting rather than yours.

It cannot be handed long work without checkpoints

METR, a nonprofit evaluation organization, measures task difficulty in units of human time: the length, in human expert working time, of a task a model completes at a given success rate. It is not a measure of how long an agent can run unattended, and METR says so directly. In its March 2025 paper that figure had doubled roughly every seven months across six years, and it has kept climbing. On METR's live results page, last updated on 8 May 2026, the strongest released model measured, Claude Opus 4.6, sits near twelve hours at 50% success and a little over an hour at 80%. A preview build, listed as Claude Mythos Preview (early), is charted higher again at roughly seventeen hours at 50%, past the point where METR says its own task set stops measuring reliably.

Two things matter. The number is rising fast, so any figure quoted here will date. And 50% is not a working standard: a coin flip on whether the job got finished is not delegation. METR notes that measurements above 16 hours are unreliable with its current task set, that the tasks are mostly software, machine learning and cybersecurity, and that they represent what someone with no prior context could do rather than skilled work inside an organization. For now, AI agents need checkpoints rather than a brief and an afternoon.

It cannot be the one who is accountable

This limit is different in kind. It does not move when models improve, because it is not about capability.

In February 2024 the British Columbia Civil Resolution Tribunal decided Moffatt v. Air Canada, after the airline's chatbot gave a customer wrong information about bereavement fares. Air Canada argued, among other things, that the chatbot was a separate legal entity responsible for its own actions. The tribunal rejected that and ordered the airline to pay. Responsibility stayed where it started.

Regulators have written the same principle into law. Article 14 of the EU AI Act requires high-risk systems to be designed so they can be "effectively overseen by natural persons" while in use, and that oversight includes overriding an output, declining to use the system at all, and staying alert to automation bias. Where a decision needs someone answerable for it, better models do not remove the need for that person, which is part of why the jobs question rarely resolves into simple replacement.

What would move each of these

Five of the six are open engineering problems where progress is real and uneven. The International AI Safety Report 2026 puts the pace this way: the trajectory to 2030 is uncertain, current trends are consistent with continued improvement, and progress could plateau, hold steady, or accelerate dramatically. Date every claim about a limitation, including the ones here, and check whether it still holds when you need it to.