Your Brain Is Not a Prompt Queue

Why faster AI does not make a faster human

We have achieved something remarkable.
We can now produce information faster than we can possibly understand it.
Well done, everyone.

AI can research, prospect, compare, analyse, brainstorm, rewrite, translate, summarize and produce another 37 alternatives before the human has finished drinking the coffee that started the whole thing.
This sounds like productivity.
And sometimes it is.

But there is a small biological detail we may have overlooked:
the human being.

Humans have limited attention. Limited working memory. Limited time.
We get hungry. We get distracted. We become emotionally involved. We forget things. We need sleep.
Frankly, from a computing perspective, the specifications are embarrassing.

And yet this slightly inconvenient organism remains the person who eventually has to understand what matters, judge what is trustworthy, decide what to do, and live with the consequences.

That may be one of the central problems of our new relationship with AI.
AI can generate at machine speed.
Humans assimilate at human speed.

The brain is not a storage folder either

Cognitive science has known for a long time that working memory is severely limited. Researchers still debate exactly why it is limited—resources, interference and other mechanisms all play roles—but the limitation itself is not controversial.

We also pay a price when we constantly jump between things. Task-switching research finds reliable “switch costs”: after changing tasks, people tend to become slower and more error-prone.
Unfinished work can leave what researchers call attention residue—part of your mind is still hanging around in the previous problem while you have supposedly moved on to the next one.

Which makes the modern working day rather interesting.
Email.
Meeting.
AI summary.
Slack.
Another AI tool everyone apparently started using yesterday.
A report.
Six tabs about how to use that AI tool.
A notification explaining that the AI tool has already been replaced by a better AI tool.
Lunch, theoretically.

This is not simply an information problem.
It is an assimilation problem.

A 2024 review of information overload research linked overload with cognitive pressure, poorer decision-making and reduced productivity. A 2026 systematic review found that useful responses include filtering, prioritisation, simplification and—importantly—technological and organisational solutions, rather than expecting individuals simply to become better at absorbing infinity.

So perhaps what I have been calling AI congestion is not a scientific diagnosis, but the experience is recognizable:
I have learned an enormous amount today and have absolutely no idea what I am supposed to do tomorrow.

Knowledge can of course be valuable without being “actionable.” Curiosity feeds us too.
But there is a difference between being enriched by an idea and having another 47 vaguely important things floating around your head.

And unlike a hard drive, your brain does something peculiar with information.
It sleeps on it.
Literally.

Sleep plays an important role in learning and memory consolidation. What we experience and encode while awake continues to be processed and reorganized; a large scientific literature supports sleep’s role in memory, although the precise effects vary by type of memory and task.

So yes, perhaps one day someone will put a very clever chip in our heads.
It may make communication with machines dramatically faster.

But:
you can upgrade the port. The person still needs lunch.
And sleep.

AI can carry things for us. That does not mean we have learned them.

Humans have always outsourced cognition.
We write shopping lists. Set alarms. Keep calendars. Leave something beside the front door so we remember to take it with us.

Psychologists call some of this cognitive offloading, and research shows that external tools can be extremely effective at helping us remember intentions and reducing demands on internal memory.
AI massively expands this possibility.
Wonderful.

But offloading and understanding are not the same thing.

A system can produce an excellent summary of a book you have not understood, a beautiful strategy you have not internalised, and an immaculate action plan nobody will execute.

Recent research with knowledge workers using generative AI offers an interesting warning. In a 2025 study, greater confidence in GenAI was associated with less self-reported critical-thinking effort. The researchers also found that AI did not simply eliminate critical thinking—it shifted it toward activities such as verification, integration and stewardship of the result.

That distinction matters.
Maybe the future skill is not “knowing everything.”
It is knowing when to delegate cognition and when to remain cognitively present.

AI can prospect.
AI can analyse.
AI can compress.
But somebody still needs to assimilate, judge, choose, act and maintain.
Hello again, human.

And no, “human emotion + AI logic” is not quite the formula

There is another attractive idea: particularly in a crisis, humans become emotional while AI remains logical.

There is some truth hiding inside it. Acute stress can change human decision-making, sometimes impairing processes that good decisions depend on.
But neuroscience gives us a more interesting picture.

Emotion is not simply interference with rational thought. Research increasingly treats emotion and cognition as interacting systems. Emotions can distort judgment, certainly—but they also guide attention, signal importance, influence valuation and contribute to adaptive decisions.

Fear can make us overreact.
It can also tell us that something matters.
Empathy can complicate the neatest solution.
Sometimes it should.

Meanwhile AI can produce calm, beautifully organised reasoning while beginning from a false assumption.
Calm is not the same as correct.
And logic is not the same as wisdom.

Perhaps the interesting partnership is therefore not:
emotional human + logical machine

but something closer to:
human meaning + machine perspective.

The human brings lived context, values, intuition, responsibility and the knowledge that somebody might actually get hurt.
AI can bring breadth, consistency, comparison, alternative hypotheses and the wonderfully useful absence of a racing heartbeat.

Each can compensate for weaknesses in the other.
But only if we design the partnership carefully.

A large 2024 meta-analysis provides an excellent dose of humility here. Across 106 experimental studies, human–AI combinations were not, on average, better than whichever performed best alone. Combination was particularly problematic in decision tasks, while creative tasks appeared more promising for collaboration.

Apparently putting two kinds of intelligence in the same room does not automatically make a genius.
Anyone who has attended a committee meeting will find this unsurprising.

There is also a darker reason judgment matters

Much of our conversation about AI focuses on what the technology can do for people.
Unfortunately, people have also noticed what it can do to other people.

Fraud. Impersonation. Deepfakes. Manipulation. Industrial-scale deception.

And “just teach people to spot it” is unlikely to be enough.

A 2024 meta-analysis covering 56 studies and more than 86,000 participants found that human deepfake detection performance was, overall, not reliably above chance; pooled raw accuracy was about 56%. Training and technological assistance helped.

Meanwhile, experiments have demonstrated that language models can be effective persuaders. In one 2025 controlled study, GPT-4 using personal information was significantly more persuasive than human debate opponents under the study’s conditions.

So the exhausted human cannot also be expected to serve as the last firewall against every sophisticated misuse of AI.
Safety has to live in the system too.

That means better safeguards, provenance, fraud detection, risk management, appropriate human oversight and systems designed to notice when capability is drifting toward harm. Frameworks such as NIST’s Generative AI Risk Management Profile explicitly treat safety and trustworthiness as properties that have to be designed and managed throughout the AI lifecycle.

Perhaps this is where “moral AI” becomes a more practical question.

Not:
Can we upload morality into the machine?

But:
Can we build systems that notice consequences, respect human agency, resist harmful uses, expose uncertainty and know when capability should be restrained?

Because more capability should require more judgment.
Not less.

Perhaps we are asking AI the wrong question

For several years the dominant question has been:
“What else can AI do?”

It is an exciting question.
It is also an infinite one.

For the human being sitting in front of it, a better question may be:
“What kind of cognition would actually help me here?”

Do I need to explore?
Analyse?
Challenge an assumption?
Compress?
Connect?
Decide?
Or have I already thought enough and need to do something?

Sometimes AI should give us more.
Sometimes it should give us less.
Sometimes the most intelligent response might be:
You already have enough information to make this decision.

That would be quite an achievement.

Not artificial intelligence that fills every available cognitive space.
But intelligence adapted to the person using it: their goal, their knowledge, their attention, their vulnerability, their available time and their capacity to act.

AI that understands that the objective is not maximum output.
It is useful human understanding.

Because our brains are not prompt queues.
They are not storage folders.

They are living systems that forget, associate, feel, wander, integrate, sleep, reconsider, make mistakes and occasionally produce something no prompt would have predicted.

That is not necessarily poor engineering.
It may be part of what makes judgment possible.

So by all means, make AI faster.
Make it smarter.
Connect the systems. Improve the memory. Increase the bandwidth.

But perhaps the more interesting challenge is to make all that intelligence increasingly sensitive to the small, slow, distracted, emotional creature on the other side of the interface.
The one who still has to decide what matters.
And who, despite several impressive technological advances, still needs lunch.

Agents Brain

Agents Brain is a personal network of specialized GPTs, tools, knowledge sources, and ways of thinking that extend my own cognition.

Agents brain

Selected scientific sources

Oberauer et al. (2016), What limits working memory capacity?, Psychological Bulletin.
Egner (2023), Principles of cognitive control over task focus and task switching, Nature Reviews Psychology; Leroy (2009), work on attention residue.
Hoedlmoser, Peigneux & Rauchs (2022), review of sleep, learning and memory consolidation; Mason et al. (2024), summary of systematic reviews and meta-analyses.
Gilbert et al. (2023), review of intention offloading; Risko and colleagues’ distributed-cognition perspective on technology.
Lee et al. (CHI 2025), The Impact of Generative AI on Critical Thinking.
Vaccaro, Almaatouq & Malone (2024), systematic review and meta-analysis of human–AI combinations, Nature Human Behaviour.
Lerner et al. (2015) and Phelps et al. (2014), reviews of emotion and decision-making.
Salvi et al. (2025), experimental work on AI persuasion; a 2024 meta-analysis of human deepfake detection.
NIST (2024, updated 2026), Generative Artificial Intelligence Profile for the AI Risk Management Framework.