Last
month,
a
developer
asked
me
if
ChatGPT
was
deleting
API
requests
immediately.
He’d
been
sending
customer
data
through
it
for
six
months
without
reading
the
terms.
It
wasn’t.
Anthropic
kept
his
conversations
in
a
storage
system
for
30
days
by
default.
OpenAI’s
policies
vary
depending
on
which
product
you’re
using.
The
gap
between
what
people
assume
happens
to
their
data
and
what
actually
happens
is
wide
enough
to
sink
a
production
system.
This
article
walks
through
the
exact
data
retention,
processing,
and
usage
policies
for
the
three
LLMs
most
people
use—ChatGPT
(OpenAI),
Claude
(Anthropic),
and
Gemini
(Google).
Not
marketing
speak.
Actual
terms,
practical
implications,
and
the
workflows
that
let
you
keep
sensitive
data
off
the
internet.
Why
LLM
Data
Policies
Matter
More
Than
You
Think
When
you
send
text
to
an
LLM,
two
things
happen
immediately:
the
model
processes
it,
and
the
company
keeping
the
model
logs
it.
Those
two
things
have
different
implications.
Processing
is
quick
and
invisible.
A
vendor
sends
your
request
to
a
server,
the
model
reads
it,
generates
a
response,
and
returns
the
output.
That’s
done
in
seconds.
Logging
is
what
creates
long-term
risk.
After
your
request
reaches
the
server,
the
company
can
choose
to:
- Retain
it
for
a
set
period.
Some
vendors
keep
conversations
for
weeks
or
months
to
improve
models
or
support
troubleshooting. - Use
it
to
train
future
versions
of
the
model.
This
was
the
default
for
OpenAI’s
ChatGPT
Web
until
November
2023,
when
they
added
an
opt-out. - Share
it
with
third
parties.
Less
common,
but
possible
in
enterprise
agreements. - Delete
it
immediately.
Only
certain
API
plans
guarantee
this.
The
risk
level
depends
entirely
on
what
data
you’re
sending.
A
customer
name
or
email
is
low
risk.
A
medical
record,
financial
statement,
or
proprietary
algorithm
is
not.
OpenAI
ChatGPT:
Web
vs.
API
vs.
Enterprise
OpenAI
runs
three
separate
products
with
three
separate
data
policies.
Most
people
don’t
realize
this.
ChatGPT
Web
(the
free
and
paid
tiers)
When
you
log
into
ChatGPT
on
the
web
and
have
a
conversation:
- OpenAI
retains
your
conversation
history
indefinitely
(unless
you
delete
it
manually). - Your
data
is
not
used
to
train
ChatGPT
by
default—but
only
if
you
have
a
Plus
subscription
or
free
trial
account
created
after
April
2023. - Free
accounts
created
before
April
2023:
conversations
were
used
for
training.
If
you
still
have
one,
assume
older
conversations
were
part
of
the
training
data. - Conversations
are
encrypted
in
transit
but
not
at
rest
on
OpenAI’s
servers
(they
control
the
encryption
keys).
Practical
impact:
You
can
use
ChatGPT
Web
for
brainstorming,
writing,
and
debugging.
Don’t
send
customer
data,
source
code,
or
anything
confidential.
If
you
need
training
opt-out
guarantees,
get
a
Plus
subscription
explicitly
for
that
reason,
or
use
the
API.
OpenAI
API
The
API
has
stricter
terms—but
only
if
you
know
to
use
them:
- Default
API
behavior
(pay-as-you-go):
Requests
are
retained
for
30
days
for
security
and
debugging.
They
are
not
used
for
training. - API
with
opt-out
(requires
contacting
OpenAI):
If
you’re
an
enterprise
customer
or
request
it
explicitly,
OpenAI
can
delete
logs
after
30
days
without
retention
for
training
research. - Data
residency
options:
If
you’re
EU-based
and
handle
sensitive
data,
you
can
request
EU
data
residency
through
the
dedicated
API.
Real
example:
A
fintech
company
I
worked
with
was
sending
anonymized
transaction
data
through
the
API
for
fraud
detection
patterns.
The
default
30-day
retention
was
unacceptable
for
their
compliance
team.
They
requested
the
extended
opt-out,
got
it,
and
now
logs
are
deleted
after
30
days
without
training
reuse.
OpenAI
Enterprise
Agreement
If
you’re
using
OpenAI
through
a
dedicated
enterprise
contract:
- Data
retention
is
negotiable.
Some
enterprises
get
0-day
retention
(logs
deleted
immediately
after
processing). - Training
opt-out
is
guaranteed. - Data
can
stay
in
your
region
or
within
a
VPC.
Cost:
Enterprise
plans
start
at
$30,000/year
and
go
up
from
there,
depending
on
usage
and
requirements.
Anthropic
Claude:
Clearer
by
Default
Claude’s
data
policy
is
more
straightforward,
which
is
one
reason
production
teams
are
switching
from
ChatGPT
to
Claude
for
sensitive
workflows.
Claude
Web
(Claude.ai)
- Conversations
are
retained
for
30
days
by
default
to
improve
the
model. - You
can
opt
out
of
training
usage
via
your
account
settings
(Settings
→
Data
→
Uncheck
