Several model instances configured with shared instructions, goals, or constraints, run together on parts of a task so their outputs fit together.

Presented as a reliability feature, where agreement between the agents is offered as evidence of correctness, and the word "aligned" usually means they were given the same system prompt.

• Agreement as Proof: Instances of the same model sharing training also share their errors, so consensus confirms nothing.
• Alignment Undefined: The term covers everything from a shared prompt to a formal constraint, and the deck never says which.
• House-Cleaning Reality Check: Like asking four people who all read the same recipe whether the recipe is right, and treating the yes as verification.

12 Angry Men (1950s) where the jury begins in complete agreement, and that unanimity is worth nothing until one juror refuses to go along with it.

Several model instances given the same instructions and run together on one task.

“Multiple aligned AI agents cross-validate every output for enterprise-grade accuracy.”

“We run four copies of the model on different sections of the document using the same instructions, then combine the results.”