Against: where the combination fails

The evidence that a person plus a model is not always better.

Shape: against

The meta-analysis

• Vaccaro 2024 pooled 106 experiments and 370 effect sizes. Human-AI combinations did worse than the best of human or AI alone: Hedges' g of -0.23, confidence interval -0.39 to -0.07 (secondary, read in the abstract).
• Decision tasks lost; content-creation tasks gained.
• When the human beat the AI alone, the combination gained. When the AI beat the human alone, it lost.
• The authors name possible publication bias and varied designs.

The jagged frontier

• Dell'Acqua 2023: outside the model's reach, workers using it were 19 percentage points less likely to produce correct solutions (secondary; the paper's abstract gives the figure in points, not percent).

The paradox

• Raisch 2021: augmentation cannot be separated from automation; stressing either one creates reinforcing cycles with bad outcomes.

What it does to the family

• A rule that the person decides is not a guarantee of a better result.
• A claim of a gain must compare human alone, model alone and the combination.
• The task type matters: the family works on content creation, which is the side that gained.