10 Nov AI’S ORIGINAL SIN
Millions of people
use artificial
intelligence every
day but there is
a basic ethical
question we need
to ask. Some AI
companies used
copyrighted material without permission
to develop their systems. If so, are we
allowed to use those systems after they
are already built?
I. Is AI Training Theft?
Artificial intelligence companies train
large language models (LLMs) using
enormous collections of content. Some of
this material is in the public domain, some
is copyrighted and used with permission,
and some is copyrighted and used without
permission. During training, the AI system
analyzes these materials and adjusts its
internal parameters, and then undergoes
significant additional adjustment. The
question is whether the unauthorized
use of content during training makes the
resulting technology forbidden.
American copyright law has not yet
definitively resolved whether using
copyrighted works without permission
to train generative AI constitutes
infringement. Within halachah, there are
two general approaches to copyright. Rav
Menashe Weisfish, in his book Copyright
in Jewish Law, quotes some authorities,
including Rav Moshe Feinstein and Rav
Shlomo Zalman Auerbach, who do not
recognize independent ownership of
intangible ideas. They nevertheless protect
copyrights through dina de-malchusa dina,
secular law. According to this approach,
the uncertainty surrounding American
copyright law provides substantial
grounds for permitting consumer use.
Other authorities, including Rav Yosef
Shalom Elyashiv and Rav Zalman
Nechemiah Goldberg, maintain that a
creator possesses ownership rights in an
intellectual creation. Under this approach,
unauthorized use can constitute theft
independently of secular copyright law.
Following this stricter position, the
question remains whether consumers
may use an AI system developed partly
through theft. Since AI systems did not
exist in the times of the Talmud, we have
to look for relevant analogies.
II. Benefiting Without Causing Loss
Rav Yechezkel Landau addresses a case
that resembles our question. An author
paid a printer to produce an edition of two
orders of the Mishnah or Talmud, with the
primary text above and his commentary
below. After completing the author’s
edition, the printer reused the typesetting
of the primary text for another edition
without the commentary. The author
demanded compensation because the
printer benefited from work for which the
author paid.
Rav Landau initially considers exempting
the printer under the principle of zeh
neheneh ve-zeh lo chaser, one person
benefits while the other suffers no
loss. Reusing the typesetting did not
necessarily impose an additional expense
on the author. Ultimately, Rav Landau
requires compensation because the
printer’s competing edition would reduce
sales of the author’s edition. Even a small
financial loss could require the printer to
pay for the benefit he received (Noda Bi-
Yehudah, Mahadura Tinyana, Choshen
Mishpat, no. 24).
This reasoning could apply to AI training.
If a developer benefits from an author’s
work without causing any financial loss,
the principle of zeh neheneh ve-zeh lo
chaser would exempt him from payment.
However, the availability of AI-generated
writing and analysis reduces demand for
many creative services. Since an author
suffers a loss of earnings because a
developer used his work, Rav Landau’s
reasoning might consider this theft and
require the developer to compensate for
the author’s loss.
III. Becoming Something New
Rambam rules that one may not purchase
stolen property or knowingly benefit from
it because it encourages sinners (Mishneh
Torah, Hilchos Gezeilah Va-Aveidah 5:1-
2). Shulchan Aruch (Choshen Mishpat
369:1-2) rules likewise. At first glance,
this seems to prohibit using an AI system
created through the unauthorized training
on copyrighted material because that
training constitutes theft according to Rav
Landau and the approach that accepts
copyright within halachah.
However, the Gemara (Bava Kamma 96b)
discusses someone who steals a brick and
crushes it into dust. The thief acquires
the transformed material, even though he
must compensate its original owner. If he
later makes another brick from the dust,
the Gemara explains that it is a different
brick: panim chadashos ba’u le-kan, a
new entity has come into existence.
Even though the thief must pay the
victim for his loss, the stolen item has
undergone so much transformation that it
is permissible to retain and use. Similarly,
an AI developer has used information
from copyrighted works to create a
technological product that is new and
distinct from the original works. We find
a similar concept in other areas, such as
kashrus, but that would take us too far
afield.
IV. Growing an AI
Another useful analogy is the case of
terumah, produce that is sanctified and
designated for the sole use of kohanim.
The Mishnah (Terumos 9:4) teaches that
produce grown from terumah (gidulei
terumah) retains the restrictions of
terumah. Rambam explains that this
prohibition is rabbinic, intended to prevent
people from keeping impure terumah
for planting and inadvertently eating
it (Hilchos Terumos 11:21). Without
this decree, the new produce would not
automatically retain the prohibition of the
original terumah.
The Talmud Yerushalmi (Shevi’is 6:3)
distinguishes between plants whose seeds
decompose (zaro kaleh) and plants whose
seeds continue growing into the new
plants. If the forbidden plant continues
to exist in the new plant, the new plant
is forbidden (Mishneh Torah, Hilchos
Terumos 11:21-22; Shulchan Aruch,
Yoreh De’ah 216:2). With regard to the
comparison to copyrighted material in
AI, I suggest that it compares with the
plants whose seeds decompose because
the original material does not reside in the
AI systems.
Additionally, Shulchan Aruch (Yoreh
De’ah 294:12) rules that if someone plants
a fruit of orlah, which is forbidden for
benefit, the fruit eventually produced by
the resulting tree is permissible because of
an important principle: zeh ve-zeh gorem,
a product created through a combination
of permissible and prohibited causes can
be permitted. Both the prohibited seed
and the permissible soil contributed to
producing the new tree.
Rav Shabsai Cohen explains that when
seeds grow into a new plant (like with
gidulei terumah whose seeds do not
decompose), there is an existing prohibited
object that continues to grow, and remains
forbidden. However, when a new product
emerges through the combined influence
of permissible and prohibited causes, the
resulting product is permitted because of
zeh ve-zeh gorem (Shach, Yoreh De’ah
294:34).
An AI model is developed from vast
quantities of information, some used
with permission and some without. Its
development also requires substantial
independent technological work. The
model emerges through the combined
influence of these factors. The resulting
system resembles the newly produced
fruit more closely than the prohibited seed
that remains in existence. An AI seems to
fall under the permission of zeh ve-zeh
gorem.
V. Does Using AI Encourage Further
Theft?
Even if an AI model is permissible to use,
paying for access raises another concern.
Shulchan Aruch (Choshen Mishpat 356:1)
prohibits purchasing stolen merchandise
because doing so strengthens the thieves
(machazik yedei overei aveirah) and
encourages further theft. Arguably, an
AI subscription encourages companies
to acquire copyrighted material without
permission.
However, note that the Shulchan Aruch
only forbids buying the stolen object. It
later (369:3) permits buying objects that a
thief legitimately acquires. An AI system
is not the stolen object itself but rather
something new.
Additionally, the Talmud (Sukkah 30a)
instructs consumers purchasing hadassim
from thieving gentiles to let the sellers cut
the branches themselves. The consumers
could then acquire the hadassim through
the change in ownership, although the
exact mechanism is more complicated.
We see that it is permissible to buy and
use objects originating in theft once a
valid acquisition has taken place.
Similarly, an AI subscriber purchases
access to a new technological product that
grew, so to speak, from a combination
of forbidden and permitted objects. The
prohibition against purchasing stolen
merchandise therefore does not directly
apply.
Based on all the above, I tentatively
suggest that halachah permits ordinary
consumer use of AI systems, even if some
of the underlying materials were obtained
without permission.