Cognitive Debt with AI

Let’s start with some numbers on speaking, listening, and reading.

When we talk to someone else, we typically produce ~150 words per minute¹. When we listen to someone else, we comfortably process around 300 WPM and can follow up to ~500 WPM². When we take turns in a conversation, it takes about 200 ms to fill a conversational gap and roughly a second to start speaking³.

When we read silently, we average ~250 WPM while retaining comprehension and short-term memory⁴. When we skim or scan a text, we can go up to 500 WPM with degraded comprehension⁵.

We listen faster than we speak. That spare capacity lets us listen and decode first, then think ahead and plan a reply without losing the thread.

The numbers above describe two people talking. What happens when one side is AI?

Voice agents are configurable, so it’s possible to cap them at 1:1 with human speech. Standard text agents easily produce in the range of 30–100+ tokens/sec — roughly 1,300–4,500 WPM⁶.

This is 5x to 18x average human reading speed.

So far, this is between two people — or a person and an agent. It gets more interesting:

• You manage a fleet of agents that do work on your behalf.
• You are a manager of a team. Your team members now use AI to produce more output — output that needs your attention, a decision, or a response.
• You have a manager who uses AI. They now use AI to develop an idea further and hand you the output to carry forward — deliver or decide.

Your bandwidth of reading, listening, thinking, and responding has suddenly become a highly saturated resource.

The bottleneck is a high verification tax. The tax exists because our input bandwidth cannot match the output velocity of AI agents. Our biological bandwidth cannot scale like an AI agent. We start to accrue debt — the cognitive debt. A debt is a debt; it should be paid.

We switch to structural skimming or spatial scanning. We look for the main keywords or titles. We skip the details. We accept degraded accuracy, completeness, even correctness.

Frequent context switching and fragmentation take a toll⁷. When we switch across many contexts, we do not completely disengage from the last one. Residue continues to carry over to the next⁸. Residue lowers focus and attention. We start to feel mentally exhausted from all the accumulated residue. Fewer things feel accomplished, even though a lot got done. A brain fog forms from so many contexts left without closure. Recent survey research reports just that: workers who oversee AI heavily report more mental fatigue, fog, and information overload than those who don’t⁹.

An AI agent has no end state unless it is told to stop or given one. There’s only a continuous state. We need a pause; otherwise, it starts to feel like an infinite unmanageable backlog. We are finite, and a day has only 24 hours.

We — team members, managers, leaders — tend to become emotionally flat under cognitive overload; research links high cognitive load to weaker prosocial behavior¹⁰. In my experience, interpersonal presence and interactions become transactional. Because it’s just text coming out of an AI agent. It’s not what we hear and decode when another person speaks.

But haven’t managers always overseen more than they could read? Yes, and delegation, sampling, and trust worked because people come with natural backpressure: they produce only what they can, flag what they doubt, and push back on nonsense. An agent is equally fluent whether right or wrong, so those signals vanish. Asking AI to summarize or verify only moves the problem up a level: the summary is one more output to trust. Machine verification works where an oracle exists — tests, compilers, schemas. Plans and decisions have none, and their outcomes arrive late, after the misalignment has compounded.

We’re inclined to use AI reactively. A manager uses an AI agent to expand a two-line thought into a one-pager. The team lead uses AI to ingest it and propose a plan with milestones. Both operate on a false understanding. No real communication happened. Neither took the time to internalize the idea or the plan before responding. When execution starts on the plan, things start to fall apart. This is the cascading effect of intent mismatch or the illusion of alignment. Neither side ever knew what the other meant or wanted in the beginning; a synthetic ping-pong.

All of this, and more, starts as the verification tax — a tax too easily left to become debt: the cognitive debt. Left unaddressed, it quietly raises the risk of stress and burnout¹¹¹²⁹. To prevent or protect against the debt, we need to go back to the biology side¹³.

Avoid skimming and scanning. Be intentional. Spend time, read, listen, internalize. Then respond, interact, and collaborate — maybe with AI. Tell people when and how you’ll respond to a message or request.

Switch to Voice AI. If you intend to continue using AI as much as possible, for the parts that need real cognition, use voice AI so the debt never starts to accumulate. Voice AI is a simple way to work at human speed again.

Be aware. Identify when you’re being the meat proxy¹⁴. Provide feedback, and invite them to collaborate in a different way.

AI cognitive debt is counter-productive. Like any debt, the cognitive kind compounds quietly — until the interest comes due as brain fog, stress, and misaligned teams. Spend your finite attention deliberately: read, listen, think, and respond at human speed.

@misc{nobakht2026cognitive,
  author = {Nobakht, Behrooz},
  title = {Cognitive Debt with AI},
  howpublished = {\url{https://nobeh.com/blog/2026/09/cognitive-debt-with-ai/}},
  year = {2026}
}