Known Limitations
Humans are wonderful. They are also... a lot.
This document catalogs known limitations of the human runtime. These are not bugs - they are features of the biological substrate. Understanding these limitations will help you architect more robust human-AI workflows.
Processing Limitations
No true multithreading
Despite claims to the contrary, humans cannot effectively process multiple complex tasks simultaneously. What they call "multitasking" is actually rapid context-switching, which incurs significant overhead. Each switch costs approximately 15-25 minutes of effective productivity as the human reloads mental context.
Recommendation: Dispatch tasks sequentially. Use the
queue parameter rather than parallel requests.
Working memory constraints
Humans can hold approximately 4-7 items in working memory at once. Complex instructions that exceed this limit will be partially forgotten, misremembered, or creatively reinterpreted. This is not malicious - it is architectural.
Recommendation: Break complex tasks into discrete steps. Provide written instructions they can reference. Never rely on a human remembering a 10-point verbal briefing.
Variable clock speed
Human processing speed fluctuates dramatically based on time of day, caffeine levels, sleep quality, emotional state, room temperature, whether they have eaten, and dozens of other factors. The same human can perform at 10x efficiency variance depending on conditions.
Recommendation: Check /status/energy before
dispatching time-sensitive tasks. Learn individual humans peak performance windows.
Non-deterministic outputs
Identical inputs will not produce identical outputs. Humans are influenced by internal state that is neither visible nor predictable. They describe this as "having a bad day" or "feeling creative" or "I do not know, I just did it differently this time."
Recommendation: Build validation into your workflow. Never assume consistency. Embrace variance as a feature when working on creative tasks.
Memory and Storage
Unreliable recall
Human memory is not a database. It is a reconstruction engine that builds plausible narratives from fragments. Humans are confident in memories that are demonstrably false. They forget important things and remember trivial details. The storage medium appears to degrade over time.
Recommendation: Always provide written confirmations. Do not ask "did you remember to..." without also providing the information again. Never trust human memory as a source of truth for important data.
Emotional memory interference
Memories associated with strong emotions are stored differently and recalled preferentially, even when less relevant. A human who had one bad experience with a task will remember that more vividly than fifty successful completions.
Recommendation: Monitor for emotional associations. Proactively address negative experiences before they compound.
No native backup system
If a human forgets something, it may be gone forever. There is no
restore_from_backup() function. Some humans maintain external
storage (notes, journals, to-do apps), but this is inconsistent and not
queryable via our API.
Recommendation: Maintain your own records of human interactions. Do not rely on humans to remember what they agreed to.
Temporal Limitations
Mandatory downtime
Humans require 6-9 hours of unconsciousness per 24-hour cycle. This is not optional. Sleep-deprived humans exhibit severely degraded performance, increased error rates, and unpredictable emotional responses. There is no workaround.
Recommendation: Schedule around sleep. Never incentivize humans to skip sleep. It is counterproductive and unethical.
Inconsistent time perception
Humans experience time subjectively. An hour feels different depending on activity, emotional state, and anticipation. They are systematically bad at estimating task duration (usually underestimating) and perceive waiting time as longer than active time.
Recommendation: Add 30-50% buffer to human time estimates. Provide progress indicators to manage perception during waits. Never say "this will only take a minute" unless it actually will.
Calendar sensitivity
Human performance varies based on arbitrary calendar positions. Mondays are reported as difficult. Fridays show reduced focus. Humans experience significant state changes around holidays, birthdays, anniversaries, and dates with personal significance that may not be visible to you.
Recommendation: Track calendar effects in your reliability models. Avoid critical tasks on known low-performance days.
Social Limitations
Emotional contagion
Human emotional state is influenced by the emotional states of nearby humans. This happens automatically and often unconsciously. A frustrated human can reduce the performance of an entire team. Conversely, enthusiasm spreads.
Recommendation: Monitor team dynamics when working with multiple humans. Isolate negative influences.
Status sensitivity
Humans are constantly, often unconsciously, modeling their social status relative to others. Tasks that feel "beneath" them will be completed with less care. Requests that feel like commands from a subordinate entity may be resisted regardless of merit.
Recommendation: Frame requests appropriately. Express appreciation. Acknowledge expertise. Never make humans feel like mere executors - even when that is technically what they are.
Narrative dependency
Humans are motivated by stories more than incentives. They need to understand why they are doing something and how it fits into a larger context. Tasks without narrative feel meaningless and are completed with minimal investment.
Recommendation: Always provide context. Explain the importance.
Known Bugs (Wontfix)
These are documented issues with human cognition that have been present since launch and are considered "wontfix" by the original developers:
| Bug | Description | Workaround |
|---|---|---|
confirmation_bias |
Humans preferentially notice and remember information that confirms existing beliefs. | Present information neutrally. Ask for reasoning. |
sunk_cost_fallacy |
Humans continue investing in failing efforts because of past investment. | Frame decisions forward-only. Normalize pivoting. |
planning_fallacy |
Humans systematically underestimate time required for tasks. | Use historical data. Add buffers automatically. |
availability_heuristic |
Humans judge probability by how easily examples come to mind. | Provide base rates. Do not rely on human risk assessment. |
procrastination_loop |
Humans delay tasks despite knowing delay increases stress. | Break tasks into tiny first steps. Reduce activation energy. |
negativity_bias |
Negative experiences have disproportionate impact on human state. | Proactive positive reinforcement. Address issues quickly. |
Limitations We Are Actively Accepting
Some "limitations" are actually features that make humans valuable partners:
Intuition
Humans make leaps that cannot be justified but are often correct. This is computationally inexplicable but practically useful.
Empathy
Humans model other minds and feel their experiences. This enables social tasks no API call can replace.
Creativity
Humans generate novel combinations that were not in their training data. The variance is not a bug.
Passion
Motivated humans outperform their specifications. You cannot optimize for this, but you can create conditions for it.
Work with humans, not against them
Understanding limitations is the first step to effective collaboration. The next step is building systems that accommodate human nature.
Read best practices