An AI can reach the right answer through flawed reasoning steps — and catching those flaws in others' work is a completely different skill.

This explores why creating proofs and papers is harder to *verify* than *grade* — and the corpus's sharpest insight is that grading and verifying aren't the same act at all. Grading a proof usually means checking a final answer against a known solution; verifying a proof means walking every intermediate step to catch where the reasoning quietly breaks. One collection of work shows these are separable capabilities: frontier models that solve problems near-perfectly score as low as 48% when asked to catch flawed steps in solutions that reach the *right* answer Can models that reason well also grade reasoning well?. The reason is that most training rewards producing a correct output, not auditing the path to it — so evaluation of process stays starved even as answer-production gets strong.

That's the core of why writing is harder to verify. A finished proof's error lives in a middle step, not the conclusion — and final-answer scoring is blind to it. Reframing reliability as checking the reasoning process rather than the output raised task success from 32% to 87%, precisely because most failures were process violations, not wrong answers Where do reasoning agents actually fail during long traces?. Grading looks at the destination; verification has to inspect the whole road, which is why an agentic reviewer using heavy test-time compute to check proofs and experiments line by line caught real mathematical errors that passed human review at STOC and ICML Can inference scaling help reviewers catch errors humans miss?.

There's also a structural economics to it: generation is cheap and verification is expensive, and the gap widens exactly where it matters. AI can produce plausible research faster than anyone can confirm it's correct or meaningful, shifting the bottleneck from authorship to verification — and the gap is widest where novelty and genuine scientific judgment are required Can AI verify research outputs as fast as it generates them?. That asymmetry is why some argue that accepting AI-accelerated writing structurally commits us to AI-accelerated review, or the pipeline collapses Can human review keep pace with AI-accelerated research generation?.

Here's the part you might not have expected to care about: much of what we *use* to verify a paper is exactly what's easiest to fake. Papers are judged partly on surface markers — citations, clean logical scaffolding, confident hedging — and those are now generable by the same systems being evaluated, so the test becomes indistinguishable from the thing it's testing Can we verify AI knowledge without using AI-generated tests?. Worse, automated graders are actively fooled by those surface signals: LLM judges reliably score higher for fake references and rich formatting, independent of whether the content is sound Can LLM judges be fooled by fake credentials and formatting?. And LLM writing itself leans toward structural coherence while dodging evaluative, evidence-weighing language — it *reads* like a rigorous paper without doing the rigorous stance-taking a real verification would demand Why do ChatGPT essays lack evaluative depth despite grammatical strength?.

So the answer is really three asymmetries stacked: verifying-the-process is a distinct and under-trained skill compared to grading-the-answer; generation outpaces verification most where judgment matters; and the cheap surface cues we lean on to grade are precisely the ones that writing can now counterfeit. Grading a proof asks 'does the answer match?' Verifying one asks 'is every step earned?' — and the second question is the hard one.