Intelligence is a relationship.
Design the connection.

The question is older than generative AI.

In 1960, J. C. R. Licklider imagined people and computers working in a close partnership. People would set goals, form hypotheses, choose criteria, and make judgments; computers could take on routinizable information work that prepares people to decide. It was a proposal about how to arrange work, not a claim that machines and people think in the same way.

Two years later, Douglas Engelbart described “augmenting human intellect”: increasing a person’s ability to approach complex problems and reach useful solutions. His frame included the person, tools, concepts, and methods around the work. The computer mattered as one part of that wider system.

These early ideas point away from the contest “human or machine.” A better question is what a person can do with a system, and which parts of that system make the person more capable of understanding and acting.

Read Licklider’s “Man-Computer Symbiosis” ↗
Read Engelbart’s “Augmenting Human Intellect” ↗

Collaboration begins when the roles are clear.

“Human in the loop” is too vague to guide a design. Does the person set the objective, provide missing context, compare options, correct a draft, approve an action, or take responsibility for the decision? Does the system retrieve, transform, draft, or recommend? Each answer creates a different relationship—and a different way it can fail.

Start with the actual task. Give the system a contribution it can perform and the person a meaningful way to shape or judge the result. Make handoffs explicit: what the system received, what it produced, what it could not establish, and what the person is being asked to do.

In its 2019 human–AI interaction work, Microsoft Research proposed 18 design guidelines across first use, regular interaction, system errors, and change over time. Its validation included a study in which 49 design practitioners applied the guidance to 20 AI-infused products. One durable idea is practical: when an AI feature is wrong, people need a clear route to correct it and continue their task.

Read the human–AI interaction guidelines ↗

A convincing reason can still lead to a worse decision.

An explanation can make a recommendation easier to inspect. It does not prove the recommendation is correct, and it does not guarantee that a person will use it well.

A 2020 study by Bansal and colleagues examined human–AI decisions across three datasets. In their experiments, the AI had accuracy comparable to the participants. AI assistance produced complementary improvements on the studied tasks, but adding explanations did not increase those improvements. The explanations did increase participants’ likelihood of accepting a recommendation, whether or not it was correct.

That finding belongs to the study’s tasks and setup; it is not a universal result about every explanation or product. It does make a useful design challenge hard to ignore: measure whether an explanation improves judgment, not only whether people find the answer persuasive. Give people evidence they can check and a low-friction way to disagree.

Read “Does the Whole Exceed its Parts?” ↗

Compare the joint result with a real baseline.

To find out whether AI helps people do a task, compare people working alone, the AI working alone where that comparison makes sense, and the people–AI workflow. Use the same task materials and success criteria. Include the time spent checking, correcting, and recovering—not just the time to first answer.

Test the ordinary path and the awkward ones: missing context, a plausible but wrong result, conflicting evidence, an unexpected system change, and a person trying to correct the output. Report where the combination helps, where it costs more, and where it fails. A result from one team or task is a starting point for a decision, not a universal ranking of people and models.

For a step-by-step evaluation method, see how to evaluate an AI agent ↗. For the wider question of what benchmark scores mean, read how to read AI benchmarks ↗. For a creative-work case, see what experiments say about creativity and idea diversity ↗.

Make the relationship answerable to the person using it.

A good interaction does not require people to admire the system or accept its account of itself. It lets them define what success means, see the material behind consequential outputs, correct a misunderstanding, and choose whether the process should continue.

• Show what the system can do in this task, and what remains outside its reach.
• Keep important sources, assumptions, and changes available for inspection.
• Make correction specific: let people change the part that is wrong.
• Make override and pause available before an action becomes difficult to reverse.
• Re-evaluate the workflow when the model, tools, or task changes.

This is the Symploka thread: intelligence takes shape through relationships, but relationships work best when the parts remain distinct and their influence can be examined. Read the Symplokē principle ↗, or explore our vision for connected intelligence ↗.