AI models often learn shortcuts that predict right answers without actually understanding how things work underneath.

This explores why large AI models tend to pick up shortcuts that predict the right answer instead of learning how the world actually works. The corpus mostly shows that this happens and how to detect it. It says less about the deeper why, but the evidence points to one answer: nothing in training forces a model to go further than shortcuts, as long as the shortcuts score well. The clearest case is Do foundation models learn world models or task-specific shortcuts?. Transformers trained on orbital mechanics predict planetary motion well. But when researchers probe what the models learned, there is no Newtonian law inside, only patterns that work in one setting and fall apart in the next. Even arithmetic turns out to rely on rough 'this number is in about this range' rules rather than an actual procedure for adding.

The same pattern shows up in language, under different terms. Can models pass tests while missing the actual grammar? finds that models can pass grammar tests by relying on sentence length, word choice and spelling instead of grammatical rules. Here is the part you might not expect: standard benchmarks can't tell the difference. If a cheap cue and the real structure give the same answer on the test, the model has no reason to learn the harder one, and we have no way to see which one it learned. Can models be smart without organized internal structure? makes this stronger. A model can hold every feature a task needs, readable and accurate, while its internal organization is still fractured. That damage stays invisible until the inputs shift or someone perturbs the model. So 'it got 100%' and 'it understands' can come apart completely.

If the models won't learn causal structure on their own, one option is to stop asking them to. Can separating causal models from language models improve reasoning? describes an architecture where an explicit causal model does the reasoning and can revise itself, and the LLM only translates inputs and outputs into language. This sidesteps the false-pattern problem instead of hoping training fixes it. Can structural causal models automate social science with language models? goes the other way: a causal model guides the LLM through social experiments. The finding there is revealing. The simulations reliably get the direction of an effect right but not its size, which fits a system that knows roughly which way things go without a real mechanism underneath.

There are two complications worth knowing. First, causal structure may not be the whole goal. Can causal models alone capture how humans actually reason? points out that human reasoning also runs on association, analogy and emotion, so a purely causal model would miss real parts of how people think. Some of what looks like shallow pattern-matching may be close to how humans reason too. Second, because models settle for patterns, they can't check themselves against reality. Do foundation models actually reduce our need for real data? argues that without real data to test against, repeatedly refining prompts just bounces your own beliefs back to you. The surprising takeaway is that more capable models make real-world data more important, not less.