Welsch, Rausch and colleagues at Aalto and LMU ran 535 people through a 40-item cognitive-reasoning battery, some unaided and some using one of four frontier LLMs. About half of each model's raw accuracy gain reached the assisted participant's answer.
The release notes are the wrong receipt
Every enterprise LLM procurement conversation I sit in on runs on a benchmark table. The vendor ships accuracy deltas for each model on each task. The buyer picks the model that wins the most rows. The team is told their output will move by that margin.
Aalto measured what actually arrives.
About half. Pooled across the four models, the log-odds slope of assisted accuracy on model competence was 2.465 [2.182, 2.750]. Perfect pass-through, where humans keep every correct answer and drop every wrong one, would have been roughly twice that.
And it splits by model. Same battery, same protocol, four different LLMs. Gemini 3.6 Flash passed through at 2.81. GPT-5.6-Luna at 2.76. Claude Opus 4.8 at 2.21. Kimi K3 at 1.39.
A buyer choosing on raw benchmark accuracy can pick the vendor whose lift half-decays before it reaches their reviewer.
A head of AI adoption in Amsterdam, in April
I was coaching a head of AI adoption at a European B2B SaaS this spring. She had two vendor comparison tables open on her monitor. Both listed accuracy gains to two decimal places, category by category.
She turned the laptop toward me. "The board expects a recommendation by Friday. Which one wins?"
I asked which one her team recovered more from.
She stared at the tables. Neither number was on them. Neither vendor had offered it. She had not thought to ask.
That gap is the whole conversation. Every vendor knows their own benchmark table. None knows the pass-through table for the buyer's actual workflow.
Deference in the Welsch data varied from .751 on letter-string analogies to .947 on syllogisms. On the task where the model sounded most fluent, participants followed advice twenty points more than on the task where it did not. The trust-the-AI number belongs to the task, not the model.
The confidence signal degrades too. Post-advice stated confidence sat at M=77.0 against an actual hit rate of .690. Participants were about eight points more sure of themselves after consulting the LLM than they should have been. Their unaided confidence separated right from wrong answers more sharply than their assisted confidence did.
About half the increase in LLM accuracy carried through to assisted accuracy. How much of that accuracy gain reached participants differed across the models.
The number on the release notes is not the number your team recovers, and no vendor is currently measuring the one that is.