In the standard evaluation of artificial intelligence, success is measured by accuracy: how often does the model get the right answer? However, as AI moves from isolated benchmarks into real-world workflows, a different metric is emerging as the true bottleneck for productivity: interaction cost.
When a human and an AI work together, the AI’s contribution is not "free." Every suggestion, prompt, and automated action imposes a cognitive tax on the human partner. If the effort required to prompt the system, verify its output, and reconcile disagreements exceeds the time saved by the automation, the collaboration becomes a net negative.
The Productivity Paradox and the Collaboration Tax
We often assume that a "general-purpose" model is inherently useful. However, Pablo A. M. Casares and Bao Sheng Loe argue in How General-Purpose Is a Language Model? Usefulness and Safety with Human Prompters in the Wild that the practical utility of these models is frequently diminished by the cumulative human effort required to use them. This includes the "hidden" costs of devising prompts, formatting inputs, and extracting usable information.
This phenomenon is often described as a collaboration tax. As Haotian Li and Q. Vera Liao found in their study of data workers, experts will often revert to manual execution if the cognitive or temporal cost of managing the AI—what they call the "overhead of managing the AI"—exceeds the efficiency gains. In high-stakes environments, this overhead is even more pronounced. Maytal Saar-Tsechansky and Tong Wang identify a specific reconciliation cost: the mental energy an expert spends processing and resolving a conflict between their own judgment and an algorithmic suggestion.
The Cognitive Tax of Proactivity
To make AI more helpful, developers often make it more proactive—offering autocompletions, suggestions, or unsolicited advice. But proactivity can backfire.
In Evaluations at Work: Measuring the Capabilities of GenAI in Use, Pamela Mishkin demonstrates that "noisy" AI proactivity creates an extraneous load that can be three times more damaging to performance than the actual difficulty of the task. This happens because the human must constantly triage and integrate unsolicited information, which disrupts their natural workflow.
This "triage" cost is particularly visible in text interfaces. Shani Goren and Oren Kalinsky note in their work on ChaI-TeA that providing suggestions at every character input offers marginal typing savings while significantly increasing the cognitive load of reviewing those suggestions. Similarly, Yijiang River Dong and Tiancheng Hu propose using the Value of Information (VoI) to help agents decide whether to ask a question or act autonomously, weighing the benefit of the answer against the "cognitive tax" imposed on the user.
Strategic Overreliance as a Rational Choice
If the cost of verifying an AI's work is too high, humans don't just work harder; they stop verifying. Helena Vasconcelos and Michael S. Bernstein argue that overreliance on AI is often an economically rational choice. When the stakes are low or the effort to verify a suggestion is high, users adopt a low-effort strategy of simply accepting the AI's output to save cognitive resources.
This creates a dangerous feedback loop:
- The AI provides a suggestion.
- The human finds it mentally taxing to verify the suggestion.
- The human accepts the suggestion without verification to save energy.
- Errors propagate, but the "interaction cost" remains low.
To break this loop, researchers like David Solans suggest that we must design systems that lower the verification cost—making it easier for humans to see why a suggestion was made—rather than just making the AI more accurate.
Designing for Empowerment
The goal of modern human-AI interaction research is shifting from "maximizing AI autonomy" to "minimizing interaction friction."
One promising direction is the Empower framework by Evan Ellis and Sergey Levine. Instead of trying to complete a whole task (which requires the human to review the entire output for errors), the agent learns to complete only the predictable, "boilerplate" parts. By ceding control at critical decision points, the agent reduces the volume of suggestions, which paradoxically increases user efficiency and acceptance rates.
Similarly, the REV-AI framework by Maytal Saar-Tsechansky focuses on creating advisors that know when to stay silent. By accounting for the reconciliation cost of disagreements, these systems only speak up when the potential improvement in decision quality outweighs the mental effort the human must spend to process the advice.
Ultimately, as Wout Schellaert and Fernando Martínez‐Plumed suggest, we should view AI not as an autonomous agent, but as a cognitive extender. Its value is not in its independent capability, but in how effectively it can be integrated into the human mind's existing workflows without overwhelming them.