THE TURN WE NO LONGER REMEMBER
ALM No.96, December 2026
ESSAYS


A small blue arrow knows how to get almost anywhere. It turns before
we notice the street sign, adjusts when a lane is closed, and rarely
needs to ask what the neighborhood used to be called. When it says to
turn left in four hundred feet, the instruction seems too modest to
change anything about the person following it. One turn, then another.
A journey accumulates without requiring a map in the traveler's mind.
Arriving is a useful result. Anyone who has been late for an
appointment, or lost in an unfamiliar city after dark, understands the
value of being told where to go. Still, arrival and orientation
describe different achievements. A person can reach an address and
remain unable to explain where it sits in relation to the river, the
train station, or the place they started. The device has completed the
journey. The traveler may have learned little about the landscape.
There is something strangely intimate about knowing a place. It is
possible to recognize a corner before remembering its name, to guess
where a street will emerge, to notice that the land slopes toward
water even when the water is hidden by buildings. Knowledge grows from
repeated encounters and occasional mistakes. A wrong turn can be
irritating precisely because it insists on being remembered. Next
time, the bakery on the corner means something. It is no longer
decoration.
The small blue arrow is an example of a much larger bargain. We give a
tool the work of finding a path, and it returns certainty at the next
intersection. A new generation of machines offers a similar bargain
with arguments, decisions, and words. The question is how much of our
own capacity to navigate survives after we accept the directions.
Psychologists Louisa Dahmani and Veronique Bohbot studied the
relationship between habitual GPS use and spatial memory. In a 2020
paper in Scientific Reports, they examined fifty regular drivers and
found that greater past GPS use was associated with poorer performance
on several measures of self-guided spatial memory. Thirteen
participants returned three years later, and increased GPS use over
that period was associated with declining performance on a memory
measure. The small follow-up group and observational design do not
justify treating the device as the proven cause of every deficit. They
do show a relationship worth taking seriously. The study is available
at https://doi.org/10.1038/s41598-020-62877-0.
A separate study by Alexis Topete and colleagues, published in 2024,
explored different ways of using GPS and their relationships to
navigation ability. It highlights a useful distinction: an aid may
supplement a person's mental map or replace the effort of building
one. The full study is at https://doi.org/10.1186/s41235-024-00545-x.
The difference is easy to overlook because both users may reach
exactly the same destination.
We have been making such trades for a long time. Printed tables
relieved people of routine calculation. The calculator made arithmetic
faster and less error-prone. Search engines removed the necessity of
knowing which shelf held a fact. Few of us would choose to restore
every slow method they displaced. Some skills ought to become less
central as tools improve. There is no virtue in memorizing a telephone
number simply because a phone will remember it for us.
But not all forgotten details have the same consequence. Losing the
habit of remembering telephone numbers is usually harmless. Losing the
ability to tell whether a numerical answer is reasonable can be
expensive. Losing the names of streets is often fine; losing the
ability to recognize that a suggested route crosses a closed bridge is
another matter. The important skill may move upward, from performing
each operation to deciding whether the operation belongs in this
situation.
That movement sounds simple until a tool becomes so smooth that its
mistakes no longer feel like mistakes. A paper map forces a user to
orient the page, compare landmarks, and decide which road matters. A
turn-by-turn voice allows attention to drift. An AI-generated
recommendation can do something similar. It offers a plan already
sorted into headings, alternatives, and a reassuring conclusion. If
the reader sees only the conclusion, what happens to the small doubts
that would have arisen while building the plan?
Consider a manager asked to choose between two vendors. An AI tool can
summarize proposals, compare prices, and produce a recommendation.
Those are legitimate uses. Yet a recommendation depends on assumptions
about delivery, service, and what the organization values. One
contract may be cheaper because it leaves an important exception
undefined. Another may promise faster service but exclude weekends.
The issue is not whether the machine can read those clauses. Often it
can. The issue is whether the person accepting its summary still knows
which question to ask next.
The old map offers an unexpectedly good model for that question. It
puts the surrounding terrain on the page, including the roads a
traveler will not take. The blue arrow gives the next instruction.
Both representations have uses. When the bridge closes, though, the
broader map makes it easier to improvise. The traveler has an account
of the world that survives a failed instruction.
A competent use of AI therefore requires more than checking a few
facts at the end. We have to preserve an independent representation of
the problem. What outcome are we seeking? Which constraint would make
the proposed solution unacceptable? What evidence would change our
mind? These questions sound procedural, but they are really questions
about how knowledge is formed. A person who never develops an account
of the terrain has no way to distinguish an excellent route from a
disastrous one that merely looks plausible.
For a learner, this raises a harder problem. We can ask an experienced
accountant to inspect an AI-generated financial analysis because the
accountant already has a sense of what numbers should resemble. The
apprentice is supposed to acquire that sense by working through cases,
making estimates, and discovering the consequences of an overlooked
assumption. If the machine performs every early attempt, the
apprentice may inherit the power to generate answers without the
capacity to recognize why an answer is wrong.
Research on AI-assisted work does not support an easy conclusion that
technology must impoverish novices. A study of customer-support agents
by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond found that AI
assistance improved productivity and could help less-experienced
workers benefit from effective patterns used by stronger colleagues.
The working paper is at https://www.nber.org/papers/w31161. That is an
important possibility: the tool can become a means of learning,
especially when it reveals the reasoning and examples behind an
effective response.
The difficulty begins when we confuse observing a polished answer with
developing the capacity that produced it. Reading an elegant solution
to a problem is a form of instruction. Attempting a problem, being
wrong, and locating the source of the error can teach something
different. The first gives us a path. The second helps us recognize
the landscape.
No one needs to get lost on purpose in dangerous conditions to make
this point. In daily life, a navigational app is often the sensible
choice. Likewise, there is little reason to make employees write
routine summaries from scratch merely to demonstrate discipline. We
can delegate the repetitive work while preserving chances for people
to exercise judgment. A trainee might estimate an answer before seeing
the generated one, identify the evidence that would overturn it, or
defend a different approach with the tool temporarily set aside.
These exercises carry a cost in time. It would be misleading to
pretend otherwise. The short-term efficiency of automated guidance can
conflict with the long-term task of developing capable people. An
organization that counts only minutes saved may miss that trade. The
same is true of a school that counts assignments completed but cannot
tell whether students have learned how to assess a claim without
assistance.
A deeper form of competence is needed to decide where delegation helps
and where it deprives us of necessary practice. I think of it as
metacompetence: the ability to recognize what sort of understanding a
task requires, which work may be handed to a tool, and what ability
must remain available when the tool is mistaken. It is a moving skill.
The details of the tool change. The obligation to judge the boundary
remains.
Sometimes the most sensible use of an AI assistant will be to ask it
for a first draft. Sometimes it will be to request an objection rather
than a recommendation. Sometimes it will be to set the assistant aside
and work through the problem unaided, so the next assisted decision is
actually informed. Knowing which of those moments one has reached is
itself a kind of practiced judgment.
The blue arrow still has an excellent job to do. It can guide us
through an unfamiliar city, around a crash, and toward the place we
need to be. We should keep it. But every so often it is worth looking
past the glowing screen at the streets on either side. There is a
difference between reaching a destination and knowing where we are.
The next generation of tools will make that difference easier to
forget, and more important to remember.
Gleb Tsipursky, PhD, behavioral scientist, CEO of Disaster Avoidance Experts, and author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026). https://disasteravoidanceexperts.com/aibook
Subscribe to FREE digital flip copy of the Adelaide Magazine printed edition.