When a person and AI judge together, the pair often lands below the AI alone, and people keep only about half of its gains.

This explores what happens to accuracy when a person and an AI model work on the same judgment, and whether the pair ends up better than either one alone. The most direct evidence in the collection is sobering. In a study of 535 participants, when the model got better on a given item, the people it was assisting kept only about half of that improvement. Their combined accuracy often landed below what the stronger partner, the AI, could have scored by itself Why does assisted accuracy capture only half the LLM gain?. The key distinction is that human and AI having different strengths (complementarity) only creates the possibility of a better team result (synergy). It doesn't guarantee one. Much of the AI's advantage gets lost when it passes through a human decision.

Why does it get lost? One answer is in how the AI's help is delivered. When the AI hands over a finished answer, people tend to anchor on it: they defer when they shouldn't, or they push back unevenly, and the hard cases still fall to the human without much support. One alternative flips the setup. Instead of the machine deciding and the human accepting or overriding, the machine points out which parts of the input deserve attention, and the human makes the call. This approach, called Learning to Guide, removes the anchoring problem and keeps responsibility with the person while still improving their judgment Can AI guidance reduce anchoring bias better than AI decisions?. The lesson is that the half-gain result may come less from people ignoring good advice and more from the way the advice is framed.

There is also a cost that per-answer accuracy scores don't show. AI suggestions can hurt reasoning even when they are correct, because they break a person's concentration and force them to rebuild their train of thought before continuing Does AI assistance always help reasoning or does it carry hidden costs?. So a correct suggestion can lower the team's overall performance. If you only measure whether each suggestion was right, you miss what it did to the rest of the task.

The AI side of the partnership has weak spots too. People can only calibrate how much to trust the model if its confidence means something. One line of work shows that a model's confidence becomes more reliable when it is grounded in its own record of past successes and failures on similar cases, rather than in the current answer alone Can past performance predict when a model will be right?. Assistants also have no built-in sense of what they don't know about the person they're helping. Adding an explicit list of these unknowns to the prompt cut sycophantic and harmful advice by 50–75% Do language models know what they don't know about users?. Both problems make it harder for a person to tell when to lean on the AI and when to override it.

The collection has only one study that directly measures human-AI team accuracy, so treat the "about half" figure as a striking single data point, not a settled law. Still, the takeaway holds up: the shortfall in a human-AI team usually traces back to the handoff. That covers how advice is framed, when it interrupts, and whether the AI's confidence can be trusted, more than either partner's raw ability.