We
read
about
it
every
day.
A
lawyer
uses
a
large
language
model
(LLM)
to
do
some
research.
They
copy
that
research
into
a
brief,
but
the
research
contains
cases
that
don’t
exist.
The
lawyer
is
busted,
the
judge
furious,
and
the
client
starts
looking
for
a
better
lawyer.
It
has
everyone
scratching
their
heads.
I
mean,
everyone
knows
the
AI
systems
will
do
this,
so
why
does
it
keep
happening?
A
new Cornell
University study
and
paper sheds
some
light
on
this,
the
problem
of
overreliance,
and
why
the
volcano
of
serious
AI
flaws
may
be
about
to
erupt.
Quite
simply,
the
cost
of
verifying
the
results
of
the
AI
tools
exceeds
any
savings
from
their
use.
It’s
a
paradox.
In
Part
I
of
an
examination
of
the why
a
volcano
of
AI
problems
may
be
about
to
erupt, I
looked
at
the
dangers
of
overreliance
on
AI
given
the
gaps
in
the
underlying
infrastructure.
But
there’s
more
to
the
story. The
simple
fact
is
that
AI tools
have
fundamental
reality
and
transparency
flaws
is
risky
and
downright
foolhardy.
Given
the
profound
and
breadth
of
the
impact
of
these
flaws
and
the
corresponding cost to
verify
outputs,
the
use and role
of
AI
in
legal
may end
up being more
limited
than
many
think.
The
Assumptions
As
pointed out
in
the
study,
the
assumption
fueling
the
explosion
of
AI
use
in
legal
is
that
will
save
gobs
of
time.
This
savings
will
inure
to
the
benefit
of
lawyers
and
clients,
will
lead
to
fairer
methods
of
billing
like
alternative
fee
structures,
will
get
better
results,
improve
access
to
justice,
and
lead
to
world
peace.
Well, maybe
even
the
vendors
would
not
go
so
far
as
to
guarantee
the
last
one.
But
vendors
do
seem
to
be
guaranteeing
everything
but
that. And
pundits
talk
as
if
AI
will
transform
legal
from
the
ground
up. Law
firms
are
buying
into
the
hype,
investing
in
expensive
systems
that
do
things
they
barely
understand.
But
not
so
fast.
All
this
hinges
on
the
assumption
that
the
time
saved
will
vastly
exceed
the
additional
steps
needed
to
verify
the
output
and
that
any
issues
of
AI
with
things
like
accuracy
will
soon
be
solved.
The
Cornell
study
throws
some
cold
water
on
all
these
assumptions and
challenges
them
head
on.
The
Cornell
Study
The
study
identifies
two fundamental
LLM flaws.
The
first
we
all
know
about:
the
propensity
of
the
systems
to
hallucinate
and
provide
inaccurate
information.
The
study
refers
to
this
flaw
as
a
reality
flaw.
It’s
a
big
problem
in
a
profession
like
law
where
being
wrong
can
have
severe
consequences.
The
second
flaw
identified by
the
study
it
calls a transparency one.
We
don’t
really
know
how
these
systems
work.
The
reality
flaw, says
the
study, stems
from
the
fact
that
generative
systems
“are
not
structurally
linked
to
reality:
namely
factual
accuracy…a
machine
learning
model
does
not
learn
the
facts
underlying
the
training
data
but
reduces
that
data
to
patterns
which
it
then
ingests
and
seeks
to
reproduce.”
And
the
study
notes
that
it’s
not
just
the
public
systems
like
ChatGPT
that
demonstrate
this
flaw,
it’s
also
the
ones
built
for
legal
as
well.
So,
the
study
concludes,
“any
output
generated
by
AI
must
be
verified
if
the
user
wishes
to
satisfy
themselves
as
to
the
accuracy
and
connection
to
reality,
of
that
output—especially
in
legal
practice.”
In
other
words,
check your cites.
The
second
flaw, one
of
transparency,
is the
black
box
problem.
It in
turn
creates
a trust issue, says
the
study.
If
you
don’t
know
how
a
decision
is
made
or
a
conclusion
is
reached,
how
can
you
trust
it?
For a
legal system that
depends
on
reasoning
and
logic, that’s a
big
issue. I
would
phrase
it
this
way: how
can
you
rely
on
something
when
you
don’t
know
how
it
works,
how
it
reached
the
decision
it
reached,
and
you
get
different
answers
to
the
same
questions.
Use
of AI in
legal
hinges
on
the
need to be
able
to
explain
how
a
decision
was
reached.
That’s
a
cornerstone
of
how
legal
processes
and
even
the
rule
of
law
is
based.
The
study
further
concludes
that
neither
of
these
flaws
will
be
overcome
anytime
soon.
What
Does
This
Mean?
The
study
goes
on
to
talk
about
what
this
means.
It
suggests
that
the
plethora
of
cases
where
lawyers have failed to
check
cites
and
end
up
having
a
hallucinated or
inaccurate case
or
facts
recited in
filings
means
lawyers are underestimating the
flaws.
Or have
been convinced
by
providers
that
the
risks
are
negligible.
These
lawyers have
simply
overrelied
on
a
tool
they
believed
or
were
led
or
lulled into believing was
more
accurate
than
what
it
is. The
result
so
far
has
been
a great
hue
and
cry
by
everyone
that
you have
to check
cites.
Usually
this
is
delivered
with
a
wry
grin
that
says
it’s
just
dumb
and
lazy
lawyers
to
blame. But
the
fact
is
the
problem
is
not
going
away.
In fact, it
seems
to
be
getting
worse.
It
may
be
that
the
guilty
lawyers
are
dumb
or
lazy, although
as
I
have written
before,
that’s
not
the
whole
story. But
what’s
left unsaid is
something
the
study
points
out:
“the
net
value
of
an
AI
model
in
legal
practice
can
only
be
assessed
once
the
efficiency
gain
(savings
on
time,
salary
costs,
firm
resource
costs,
etc.)
is
offset
by
the
corresponding
verification
cost
(cost
to
manually
verify
AI
outputs
for
accuracy,
completeness,
relevance,
etc.). Those
caught
with
hallucinated
cases
in
their
papers
simply
didn’t
take
the
time
to
verify
relying
on
the
AI
tool.
Because
the
demand
for
accuracy
in
legal
is
so
high,
the
study
notes,
the
verification
cost
for
many
actions
in
legal
is
too
high
to
offset
the
savings. The
study
also
concludes
that
this
cost
is
not
ameliorated
by
automated
systems
since
the
reality
and
transparency
risks
may
still
exist.
Hence what the study
calls
a verification
paradox.
And
we
see the
impact
of
this
paradox already
with
fines
imposed
by
courts
for
hallucinated
cases.
We
will
no
doubt
see
malpractice
and
ethical
violation
claims.
The
cost
of
being
wrong
in
law
is
just
too
great
to
not
verify
and
verify
thoroughly.
Granted,
AI
can
do
lots
of
things
well
where
the
risks
of
being
wrong
are
not
that
great.
It
will
have
an
enormous
impact
in
business
and
maybe
other
professions.
But
for
law,
not
so
much:
“The
more
important
the
output,
the
more
important
it
is
to
verify
its
accuracy.”
The
study
concludes:
The
verification-value
paradox
suggests
the
net
value
of
AI
to
legal
practice
is
grossly overestimated,
due
to
an
underestimation
of
the
verification
cost.
A
proper
understanding of the costly
and
essential
nature
of
verification
leads
to
the
conclusion
that
AI’s
net
value will
often
be negligible
in
legal
practice:
that
is,
in
most
cases,
the
value
added
will
not
be
sufficient
to
justify the
corresponding
verification
cost.
The
Reality
It’s
easy
to
see
the
economic
impact
of
the
verification
paradox
when
you
compare
the
cost
of
getting
a
piece
of
work
done
by
an
LLM
with
that
done
by
a
human.
Let’s
assume
you
ask
an
LLM
to
do
some
legal
research
that
would
normally
take
you
10
hours.
You
get
the
result,
but
it’s
got
some
25
case
citations.
Now
you
have
to
a)
check
to
make
sure
every
case
exists
and
b)
make
sure
that
the
case
stands
for
the
proposition
the
LLM
says
it
does.
By
the
time
you
do
that,
you
could
very
well
spend
the
eight
hours,
if
not
more.
A Volcano
About
to
Erupt?
It
may
be
too
late
to
completely
put AI back
in
the
bottle.
But
where
it
takes
just
as
long
if
not
longer
to
verify
the
results
of
an
AI
tool
you’ve
spent
thousands
of
dollars
on,
you’re
not
going
be
predisposed
to
buy
more.
Certainly,
your
clients
aren’t
going
to
be
wild
about
your
use
of
a
tool
that
not
only
fails
to
save
them
money
but
costs
them
more
and
exposes
them
to
risk.
It’s
easy
to
envision
the
fundamental
conclusion
that
using
AI
for
many
things
is
not
worth
the
risk
and
the
cost
of
validating
its
result.
It’s
easy
to see
how this
fact will temper the enthusiasm
and
reliance
on
AI.
We may
rapidly
conclude the
costs
and
risks
of
doing
so
are
too
high and
simply
not
worth
it
in
the
long-
and
perhaps
even
the
short-run.
When
that
happens,
a
lot
of
lawyers
are
going
to
be
caught
with
expensive
systems
that
they
don’t
need.
A
lot
of
vendors
may
have
to
go
in
other
directions.
A
lot
of
venture
capital
may
go
down
the
drain.
The
proverbial
volcano
may
be
about
to
erupt.
That’s
something
worth
considering
before
you
buy
the
next
shiny
new
AI
toy and before
you
use
AI
shortcuts
to do the
hard
work,
before
you
blindly
expect
people
you
supervise
to
do
the
right
thing
and
before
you
accept
without
question
their
work.
In
the
meantime,
check
your
citations.
Please.
Stephen
Embry
is
a
lawyer,
speaker,
blogger,
and
writer.
He
publishes TechLaw
Crossroads,
a
blog
devoted
to
the
examination
of
the
tension
between
technology,
the
law,
and
the
practice
of
law.
Melissa
Rogozinski is
CEO
of
the
RPC
Round
Table
and
RPC
Strategies,
LLC,
a
marketing
and
advertising
firm
in
Miami,
FL.