Rebecca Tushnet

STLR Panel 2: Litigating Fair Use in Copyright

Zahr Said: Substantial similarity is confused. Sedlik is a
good example. Laudable for concurrences to recognize need for reform in this
crummy, confusing test, but that’s not reflected in the case itself.

First, because the court’s diagnosis of what went wrong is
mistaken. Second, the case didn’t go wrong: it was correctly decided. Third,
this is a really unusual case, not a run of the mill © case.

Court confused copying and substantial similarity. It made a
number of errors along those lines, ignoring intermediate copying—“it’s the
photo tattooed on an arm.” P’s counsel encouraged the conflation of the photo
and tattoo, even though Ds only conceded that the reference photo Kat
von D used was 100% identical to the Sedlik photo. This matters b/c it’s calling
for reforms, and the court said it was representative of a too defendant-favorable
standard, and that’s wrong. If you treat the unusual as the common, we’ll be mired
in further doctrinal mud.

Court thinks the outcome is wrong for the right reasons. But
it was right! (I agree and filed an amicus focused on this.)

The timing was also weird—Warhol came down in the middle of
it and the court had to revise its analysis. What to do: don’t build reforms on
a landfill! Change the language; stop relying on Nimmer’s treatise and use
Goldstein on infringement instead—precise, clear, elegant. Clearer thinking.

(or?) Dispense with the test! That’s what the concurrences
were thinking about doing anyway.

Daniel Goldstein: Represented blind scholars & National
Federation of the Blind in HathiTrust. Kudos to Peter Jaszi. Early commercial
ebooks were inaccessible; not interested in the market. University of Michigan,
by contrast, considered accessibility key to the digitization of their library,
even before Google got the message. Librarians are the best. Intervened as a
defendant in the litigation, and university didn’t oppose even though he’d sued
universities for ADA violations in the past; publishers didn’t oppose, probably
because they didn’t want to lose the first motion in the case.

The romance of the case was important; lots of excitement
about blind researchers being able to access works for the first time. And the fair
use argument was strong: dicta in Sony about accessible books for the
blind being fair use; Chafee amendment; etc.

The publishers wanted to impound the entire library; the National
Library Service can only generate about 2000 accessible books/year but publishers
never suggested just giving it a copy. ADA was also relevant as a clear & comprehensive
national mandate for accessibility/against discrimination against people with
disabilities—universities have to offer the same experience when feasible and this
was a way for schools in HathiTrust to make the research program available to blind/print
disabled students. Why didn’t the publishers ask for the digitization program
to pause during settlement discussions? Not sure—seems like a bad call. Their
arguments consisted of saying that the Chafee Amendment/Sony dicta was
only about making a single book for a single blind person accessible and that
universities weren’t authorized to make accessible copies.

Corynne McSherry: Public Resource v. ASTM: Teeny nonprofit
that makes government information available online in an accessible way. Conflict
with ASTM over building standards. National Archive has a copy of every
standard incorporated by reference in the Federal Register; or you might find a
copy in the library; or you might buy a copy. Public Resource decided to change
that. Eventually won on fair use; thinks this should have been decided on
copyrightability. The law that governs us cannot be copyrightable. Fair use is
more complicated.

Also, all the creative work in standards is done by
volunteers who want to create good standards for their industries; the work
done by the profit-seeking entities is not creative. © is not needed.

Fair use is bearing too much weight. Rights holders think
fair use is too big, but the expansion of copyright is what creates the need for
fair use (as Jessica Litman wrote). But it shouldn’t have to bear so much
weight. 102 should have prevented copyrightability of standards enacted into
law. We should also have digital first sale, but we had to use—and lose—on fair
use instead.

Samuelson: very interested in fact issue v. legal issue in
fair use cases. Court said the overall fair use inquiry was mostly law, but
what is it that can be found as a fact? Although fair use isn’t tried to juries
very often, and Sedlik is an outlier, as is GvO, that is something that not
many practitioners have focused on.

Goldstein: in HathiTrust, there was good evidence that
publishers had no interest in creating a market for the blind/print disabled,
and they donated books to the National Library when they did. That was the key factual
part from his POV; the rest was law and policy.

Sprigman: Restatement of © picks up on this—the parties’
interests are relevant. Emphasized in SCt Public Resource case—they resolve it
on protectability; some affected parties will be willing to roll the dice on
fair use. The less bold among us would have to think twice before using the law
that belongs to all. Fair use isn’t just framed as a defense, it’s
case-specific.  Maybe the Court has
signaled to lower courts that they should be relying on copyrightability.

McSherry: UpCodes cases—ASTM keeps pushing the idea that
commercial uses are fundamentally different. They’ve mostly lost in dct, but
not entirely. It shouldn’t matter: it’s the law! Make the law more available!

Said: it’s troubling to give the jury an idea/expression
instruction unless they have some understanding of the caselaw—they’ll
interpret it according to lay understandings. That’s led her to worry about
transformativeness. Sometimes it’s a factual question, but Kelly v. Arriba Soft
is not the kind of case where a jury would have been able to figure that out.

Fisher: 107 makes explicit that the four factor list is
nonexhaustive. Puzzling how few cases press a fifth factor. E.g., shouldn’t it
matter for social justice that we provide access to the blind? We’ve known
since Rousseau at least that the rule of law requires accessibility. We don’t
need to frame that in constitutional terms. [I teach that courts find a way to
stuff considerations like this into the four.]

Said: why isn’t that public benefit?

Fisher: that’s controversial as an extension of factor 1.
Breyer is willing to do it in GvO, but Sotomayor is not in Warhol. It would be
a way to address ambiguity if we had a more sensible allocation of what you were
trying to show. [I’m not sure that would help with ambiguity. It’s not the
numbering that’s the problem.]

McSherry: we’re always framing it one way or another,
including “consider four factors in light of the public interest in …” She
doesn’t like good faith as a fifth factor.

Goldstein: could put all his considerations in factor one
b/c they were about the accessibility purposes of the use; not transformative,
but still a fair use.

Said: Trevor Reed, Fair
Use as Cultural Appropriation—interesting argument about factor two,
considering whether the work was being appropriated from an indigenous culture—judges
are sociologically conservative (but maybe can be invited).

STLR Panel 3: The Role of Fair Use in Copyright

Trevor D. Lohrey: Should we have factors?

Christopher Sprigman: predictability is an issue. Leg
history frames fair use as an equitable rule of reason. Court is supposed to be
doing something synthetic, not considering factors like in a “test.” A platform
for thinking about the individual case with attention to ©’s progress-promoting
purpose. Predictability is often a virtue, but not always—like patience. I don’t
think we should be looking for a rule.

Laura Heymann: doesn’t fault judges for going through
factors in an atomized way. The synthesis is challenging. It’s really about
giving more weight to the “fair” in fair use. But judges don’t want to be
unmoored.

Peter Henderson: we have examples of conflicting holdings in
what looks like the same case. We don’t want to have bright lines.

Heymann: consider folk knowledge that develops around areas
of law. “You can copy 300 words or less.” That suggests that people want those
kinds of sub-rules.

Henderson: you can also see a desire for that in AI training—don’t
output more than 300 words w/o a license, etc.

Barton Beebe: Appellate reversal rate for fair use cases is
about 1/3, which is almost exactly the same as for other civil litigation. The
dissent rate is 1 in 10, almost exactly the same as for other civil litigation.
The hard cases are hard, which is why they’re being litigated—that’s why there’s
unpredictability in the individual litigated case; fair use is not that
special.

Heymann: though there also effects on nonligitation
decisions.

Lohrey: change over time?

Sprigman: The use of the “purloined” manuscript in Nation: a
journalist would think that’s how the system works, while a lawyer thinks it’s
bad; a lawyer thinks it’s good to keep secrets for your clients while a layperson
might object. O’Connor failed to see outside of her bubble. Leval and others
helped erode the idea of “good faith.” Fair use is about market competition now.
Restatement focuses on how the factors can aid in assessing whether the
defendant is undermining fair competition. That’s something that takes time and
has slowly matured.

Heymann: shortcut for judges can be “is this tech analogous
to tech we’ve encountered in the past?”

Sprigman: what Congress meant to do was to take a snapshot
of judicial creativity in fair use, but in a way that didn’t freeze it and
rather spurred that creativity. In some areas of the Act Congress said very distinctly
what it wanted and in others it approved continued judicial innovation—to keep
development going. So when people complain we’ve gotten away from the purpose
of fair use, I want to ask: what did you think that purpose was?

Lohrey: is fair use now too market-focused and about
battling economists?

Sprigman: your answers depend on your priors. He thinks it’s
a good development b/c he came from antitrust where those issues are asked
every day and there are tools. If the bete noire of © is substitution, he thinks
that the question of substitution can be answered much more accurately than a
general fairness inquiry.

Heymann: that devolves into fights over market definition,
which seems pretty hard. Does milk compete with Coke? [Sprigman says it’s surprisingly
difficult to answer that question—which to me doesn’t make the point he thinks
it does. I’m not sure that you can solve the difficulties by switching the
terms of debate—you just get different difficulties that you might only notice after
a bunch of cases.]

Sprigman: hypotheticals can help—it’s theoretically
tractable to ask whether the market for one work affects the market for
another. In fairness, we have to figure out whose interests count and for how
much.

Henderson: you still have to choose your models, components
of your models—it constrains degrees of freedom but not infinitely. In the
latest evolution of market dilution theory, some of those models aren’t really
couched in anything particularly data-grounded. [E.g., the Court’s claims about
derivative markets in Goldsmith.] We will often not have data,
especially about potential future market harms.

Beebe: one hypothesis—when a court finds fair use it will
find that it’s a question of law, and no fair use will be a question of fact.

Heymann: you see some cases saying that there’s no market
because the corporation could have exploited it but didn’t and that means there
must be none, whereas cases like Monge say that the © owner might change her
mind. [I think these can be made more coherent by thinking about the idea of
corporate rationality versus individual caprice. But I admit that this doesn’t
cover cases like the Seinfeld Aptitude Test case.]

Beebe: don’t ask “is this conduct fair use”; instead ask “should
this conduct be fair use?” There is no “is” in fair use in cases of this
nature. It’s all “should.” There’s no preexisting rule that we apply to the
facts.

Sprigman: many opinions leave lots unresolved. AI cases: One
of the Cal. cases finds fair use might give the rule “don’t use pirated
datasets”? The other Cal case says: “this is terrible—this shouldn’t be fair
use but it is, b/c Ps are so inept. I know in my bones that market dilution harm
is happening.” But competition for an industry—w/a class of human authored
works—is not something we have experience with. As a matter of © law I think
the answer is that © shouldn’t recognize that as a theory of harm, but we can
think about how to address it as a different kind of harm—labor law, tax law. ©
is a poor vessel for that.

Beebe: A human supremacist view: more and more stuff is not
the goal. The processes of creativity are important to human flourishing, not
the consumption of stuff. The solidarity of sharing it with others. Humans over
machines and corporations! Fair use should be resisting both, not just
machines.

Ochoa: if fair use is a “should,” are we getting things
systematically wrong?

Sprigman: thinks general fairness concerns interesting but
not central; fair use law holds up pretty well. Mistakes always exist and don’t
seem directional or pernicious.

Beebe: Clarification: Lots of fair use questions aren’t that
hard. 80% of the time fair use is an “is.” District judges don’t have time for
a lot of “should.” We need is or it will be chaos. But for LLMs/AI training,
there is no “is.” The hard questions are at the edges.

Sunder: don’t romanticize what our © has wrought—women as
naked models/objectified subjects, not creators. Hollywood gives us sequel
after sequel, movies based on commercial products like Lego and Labubu. The
cheerleading uniform! The human creator supposedly fostered by this law
deserves a closer look. Possibility of hybrid innovation/cyborg creativity—good
with human flourishing, not supremacy.

Beebe: fair! Romanticism: thinks of American romanticism of mid-19^th
century, very folk-oriented, not the heroic muscular unusual author. But fair
to point out that lots of people were not enjoying the mid 19^th
century!

Sprigman: Readers should be important in our theories—the people
shaped by creativity as well.

from Blogger https://tushnet.blogspot.com/2026/04/commemorating-50-years-of-copyright-act_3.html

The 1976 Copyright Act: Mostly Evolutionary, Not
Revolutionary Tyler Ochoa

1790 Act adopted the Statute of Anne—not radical even though
it was the first for the US. Similar here—major changes, but not radical.
Expanded subject matter; protection on creation/fixation instead of publication
and preemption of state law; automatic protection without preconditions (by
amendment), reduced formalities, codified idea/expression.

But 1909 Act covered “all the writings of an author.” “Original
works of authorship” aren’t that different. Likewise, lawyers didn’t
necessarily take common-law copyright as seriously as federal copyright, but
there was common-law copyright before the Act preempted it for fixed works. Moving
the federal protection back to the date of fixation isn’t a huge change.

Formalities: some change.

Codifying idea/expression: It’s a good thing that we wrote
it down because of the textualist turn, but it was an existing principle. Not a
radical break.

Fair use: same thing.

Exclusive rights: again, evolutionary.

WFH: a reasonably big change in defining what can be WFH,
rejecting instance & expense test (which he thinks was a mistake by the
courts in the first place).

Duration/renewal: Motion pictures and musical works were
heavily renewed, and most other things weren’t. Life plus 50 isn’t just an
extension, it’s a fundamental change in measurement. And it’s a change that we
haven’t felt the full impact of yet b/c we’re only 50 years out. None of the ©
granted in 1978+ have yet entered the PD (w/exception of some unpublished works
by deceased authors).

Terminations: same idea as renewal term, but new opportunity
b/c of unitary term.

Compared to what happened since the 76 Act: rise of computer
programs, though the definition of literary works was written to allow them. BCIA:
elimination of formalities; elimination of registration requirement for foreign
works (subject to limited remedies). Architectural works; VARA; automatic
renewal for works 1964-1977—works no longer go into public domain for failure to
renew. NAFTA; TRIPS; Uruguay Round restoration—a radical change. Public
performance rights for sound recordings. Term extension; DMCA adding 512, 1201,
and 1202. Cumulatively this is a radical break with the past, but mostly after
1976. The 1976 act did change from incentives to publish towards natural right
of author, but that was a change in principle with fewer radical changes than
came later.

Moral Rights, 50 Years Later Xiyin Tang

Poised for a resurgence among concerns for human creators in
a world of AI slop. Entrepreneurs are using TM, even trade secret, to protect
rights far beyond VARA. Monty Python case is a classic example using TM—but the
cause of action was a stand-in for moral rights, by the court’s own admission.

VARA has had limited success—limiting to fine arts likely
fell well short of Berne requirements. Also US rights are waivable, though
nonassignable. Only 84 adjudications of VARA claims, 71.4% of which dismiss the
claims, often on procedural grounds. More concerning are cases that effectively
chip away at the scope altogether, further neutering it. First Circuit held
that VARA didn’t protect site-specific works of art at all—where the location
is integral to meaning—based on the public presentation exception. But this was
the type of art that formed the basis of the public justification for VARA in
the first place—Richard Serra’s Tilted Arc. SDNY denying injunctive relief in
Five Pointz; money is not a good remedy. And the 7^th Circuit found
that site-specific art wasn’t just ineligible for VARA, but ineligible for ©
generally. But site-specific art is the exemplar of postmodern artistic practice.

Nonetheless, predicts that moral rights are posed for a
resurgence. AI has led to a new insistence on human creation. Private deals
w/AI: authors don’t just get a per-generation payment for summarizing their
work, but a link to purchase—attribution. Decision holding that Wu-Tang Clan’s
album was a trade secret—independent economic value comes from ability to
exploit exclusivity to create an experience that competitors can’t. Secret art
could therefore be trade secret art.

Terry Fisher: what about fan fiction and other
interventions?

A: Artists care about originals. Maybe expansion of moral-esque
rights could impinge on what people do w/a copy of the original, but the
Wu-Tang example is about works that exist in an edition of one, and that’s what
she’s concerned about.

Keynote Address: Authorship in the Shadow of the 1976 Act Paul
Goldstein

AI challenges: believes that transaction costs are the key
and that private or public deals can achieve them. But copyright dilution—competition
with AI-generated works—should not be actionable because the ideal is that
works should be priced as low as they can be without destroying incentives.

Points out that Jane Ginsburg showed that French copyright
was initially utilitarian/incentive concerned; moral interests became important
later. US: the author-protective provisions that Congress introduced in 1976
are important—a shift in the philosophical base of ©, according to Barbara
Ringer, to make the primary beneficiaries of © individual authors.

What do authors want? To be recognized as the authors of
their works. Consider Creative Commons, which found that attribution was so
commonly desired that attribution became the default. Audiences also want this,
which is why they go to concerts instead of watching from the comfort of home. They
also want community—shared passion—but the main desire is authenticity, the knowledge
that the tiny well-lit figure on the stage is their favorite performer, not a hologram.
Authenticity is the consumer-side counterpart of the desire for attribution.
Author & audiences meet and form a bond. But aura can be attached to
multiple copies of identical objects—ubiquitous reproduction hasn’t led to the
withering of aura, but strengthened our desire for it and created new
strategies to produce aura.

But TM may overprotect attribution at the expense of popular
culture—missing limited term and public interest exceptions. It’s not enough to
exclude protection for generic elements; its exceptions for parody and the like
aren’t vigorous enough. Presumably this was a concern in Dastar.

So now we need attribution in ©. 106A (VARA) should be
replaced w/an exclusive right for all authors to claim authorship and object to
distortion/mutilation/modification of the work, taken from Berne. Generative AI:
a user who asks for a story in the style of Raymond Chandler justifies an
attribution right. [Hunh? Doesn’t the user, by definition, know?] Style is not
copyrightable, but we could draw the line more generously if there was only an
attribution right, not a control right; the limited term would avoid the parade
of horrors and parody would be allowed.

Lemley: there are lots of circumstances in which creativity
requires that you not keep integrity. Tang’s answer was focused on individual
copies/single copy works. But your moral right is broader. What do you do about
fan fiction?

A: look to what other countries do. They valorize commentary
and individual creation as much as the US does. French law has a robust exception
for parody and pastiche.

RT: I don’t think that would work. Litigation culture is a
big deal.

A: the motion picture studios are why we don’t have a moral
right in this country. France, Germany, Canada etc. have moral rights and
moviemaking—but the answer was the litigation culture in the US. Would love to
see some empirical work on that. On the attribution right, would love to see a
full-fledged empirical study of asking © litigants why they sued; integrity
touches on vital interests. Turns in part on what the remedies are. One could
fiddle with remedies like injunctive relief rather than monetary relief, though
then you lose the contingent fee bar.

Barton Beebe: Scalia dismissed tracing the origin of Nile
and all its tributaries—what if we recognized moral rights of author and all of
those who preceded her and put a burden on her to state her sources? Religious
traditions might support this.

Sprigman: other attribution systems exist, but Earth is for
the living: part of art’s responsibility is to free itself from the past
despite its influence. You are taking a side on what art should do; our
perspective is from writers, but may not serve readers very well.

Panel 1: The Copyright Act and Technological Change

R. Anthony Reese: legislative response to digital
technologies was the key in the midpoint period; last couple of decades saw
less of that other than Music Modernization Act. Focusing on civil, not criminal,
amendments. In 1980 we added a definition of computer program; not a big
change. We’ve only added one exclusive tech based right—digital audio public
performance for sound recordings. Not a response to tech but to political
failure of the 1976 Act.

Some expansions to limitations—110 for distance education
(not very helpful in pandemic); 111 Family Movie Act allowance for skipping
naughty scenes hasn’t created a flourishing business model. But 114’s limits on
nonsubscription broadcast transmission enabled HD over the air radio. Also bars
on record rental, etc.

We’ve done a bunch of compulsory licensing, mostly tinkering
to provisions already there like adding low-power TV to the cable ones already
there. DAT licensing; some compulsory license for the digital sound recording public
performance right; satellite retransmission licenses; modernization of
mechanical licenses to include a blanket license for streaming & download. Copyright
Office was not a fan of compulsory licenses but Congress is enamored of them in
specific circumstances.

Remedies: 512 limits have been incredibly important. 408
preregistration, maybe has some effect (about 700 works/year); 504(c) allows
willfulness to be presumed if you provide false domain name contact info in
connection with © infringement.

Sui generis provisions: AHRA (desuetude/written so narrowly
as to exclude general purpose computers just as they were about to become the
way music was enjoyed), 1201/1202.

Amendments motivated as much by
political/legal/market/social developments as by the new technologies
themselves. Record Rental Act passed not b/c of CDs but because of shops that
rented out tapes and encouraged copying; the worry was that it would get worse
w/CDs but the business model already existed. Similarly, low-power TV was added
not b/c it was new but b/c the FCC hadn’t previously authorized its deployment.

Jessica Litman: substantive approach is shaped by drafting
method—to invite many of the stakeholders who know they’re interested to work
out their differences and embody their compromises in statutory language.
Overreliance on compulsory license is b/c it’s easier to reach a compromise.
Even in the 1909 Act, courts had devised separate rules for different kinds of
works; they wanted to incorporate exclusive rights shaped to works, but then
that turned out to work really badly when new tech like movies came around. So
1976 Act tried for one size fits all rights, but then everyone needed bespoke
exceptions to continue doing the legitimate things they’d done every day. They
tried to make exceptions as narrow as possible so they couldn’t be used for
anything else (e.g., jukebox exceptions). Insiders make the rules that they and
© outsiders will have to follow—unfriendly to outsiders.

How does this method work for insiders? Their efforts make a
lot of assumptions that may not pan out. DMCA is a good example. Implicitly
incorporate promises about how insiders will treat one another. But none of the
promises are legally or morally enforceable, so many get broken. Promises of
publishers & distributors to creators turned out to be particularly
vulnerable to breakage. Also: If you exclude outsiders from negotiations for as
long as you can, you’ll miss important issues on the horizon that aren’t
central to anyone sitting around the table.

Result: insiders came out believing that they’d
significantly fortified themselves against scary new technologies, and as those
have failed, we’ve seen © insiders adopting more combative negotiating postures
& developing deep resentment about the interests they believe are eating
more of the pie than they should be permitted to eat. Legacy © owners are earning
more money than they ever have. But they are nonetheless looking at © with grievance
and resentment b/c technical services delivering material are also earning a
ton of money—instead of saying “big pie, wow, awesome!” they want that whole
pie. That’s led some to believe they’re entitled to hoard rights, money,
control, and market share by any means they can manage.

When you call © insiders together to write a statute to fix
things, you get the MMA. Has some good things and some bad things buried in the
middle—but most of this hoarding is currently coming at the expense of
creators. They’re earning less, unlike the legacy industries. When we tell them
to compromise on AI, this is also what we may expect.

Pam Samuelson: A happy story on the scope of software
copyrights. Wasn’t initially clear whether machine executable code was
copyrightable. Initial attacks on copyrightability had to be overcome, but scope
was unclear—Paul Goldstein wrote early article expressing concerns about risks
of monopolizing functionality. Suggested borrowing patent misuse doctrine.

Whelan v. Jaslow then gave super-broad © protection—early expectations
about thin © to avoid protecting functionality were totally ignored. All of the
“structure sequence and organization” was protectable if there was a modicum of
creativity; everything should be protectable unless there was only one way to
achieve it.

6 years of effort followed (Samuelson in the lead!) to get
the results right. David Nimmer also wrote an article suggesting that more
filtration was required. 2d Circuit adopted this in Altai requiring abstraction,
filtration, and comparison. Filtering out unprotectable elements was a really
significant advance and got us back to a relatively thin scope. Compatibility
is unprotectable.

Fed. Cir. agreed w/dct that we should have © hearing like
Markman hearing to do claim construction—another important development. So © turns
out to accommodate software well, after struggle.

Mark Lemley: Imagine a collection of model weights that
gives a possibility, but not a certainty, of generating an infringing copy: is
that a copy? The statute’s answer is incoherent. The right to reproduce the
work in copies; copies are copies when they’re fixed; fixed means the copyright
owner authorized it. That doesn’t make any sense so nobody pays any attention.
We use the same definitions for protectability and infringement, causing the
problem. But we do use the part that says fixation means it has to be
perceived, reproduced, or otherwise communicated.

The parties in AI cases take some extreme positions—models don’t
include content; output is just a collage of intputs—neither of these things are
true. Can extract Harry Potter verbatim w/a four word prompt from Llama 3.1. Extraction
is possible for some works/parts and some models but not others. New work shows
you can expand extraction if you go beyond verbatim extraction.  But again it depends.

Jane Ginsburg shows that you can retrain models to force them
to regurgitate a work, and that makes it more likely that they’ll regurgitate
other works, suggesting other works are latent.

Is a work latent in a model a tangible copy? It’s
complicated. Models don’t store works directly, but encode weights reflecting relationships
b/t words or syllables. You can make a copy of a picture using ones and zeros
even though the ones and zeros aren’t the picture—deterministically, these are
copies. But Microsoft Word doesn’t encode War and Peace even though all the
necessary parts are in Word waiting to be put in the right order.

Information theory: can the work be extracted with less than
the same information you already have? Compression algorithms: sometimes we can
store less than all the info & use it to generate a work. Lossless compression
is clearly a copy; lossy compression probably creates non-identical copies that
are nonetheless still copies. Compression algos are still deterministic, though—same
imperfect copy every time.

AI extraction is rarely deterministic. We can get a result
10% of the time; are the other 90% also stored in the model? That means more
works are in the model than there are atoms in the universe. © hasn’t dealt much
with nondeterministic copies. Kelly v. Chicago Park District comes closest—a garden
is not ©able b/c it’s not deterministically fixed. Compare to video game output
cases: you infringe by making a map used w/existing video game to cause it to
be played in a new map even though the images aren’t contained in the map.

Kelly is probably wrong: most people would say a garden
could be sufficiently replicable to be fixed. Game cases: predictable and
replicable—the same map and characters would show up. Though now there are
procedurally generated games where the map changes on the fly, but that still
involves output. Not all the possible maps exist already in the game.

Challenge: sometimes it’s easy to get out—high degree of predictability
and replicability. Sometimes it’s not really possible. Sometimes it requires a
lot of work but can be done.

We should say that if it’s easy to get a work out, it’s
probably worth saying it’s in the model. But if you have to know what you’re
looking for, and keep trying until you get it, then we probably should not call
it a copy.

Why this matters: if a copy is in the model, then making a
copy of the model infringes the copyright in the underlying work. Meta is distributing
the model weights to lots of people.

Maybe those models are fair use—probably be the right
result, but harder to reach, especially after Warhol. Fact of
intermediate copying might be important (even though he doesn’t think it should
be).

In response to Tony Reese: maybe we could say they’re unfixed
but derivative works which don’t require fixation. Reese” points out that the
RAM copy cases about duration of presence in computer—not a great way to do it
but we did do it. Also compare Google Books: how much of a book you can
assemble by doing searches was relevant to the fair use analysis.

A: maybe that functional approach is practically the best
way to go. If I can prove that it’s in there somewhere but it costs $10,000s to
do that each time, that’s not a real way to get a “copy.” It is also really
hard to resolve such questions as a matter of class actions—courts would like
to have an answer, and an answer depending on a work by work, model by model
analysis is not going to be desirable to them. Even if it’s the truth.

Tang: incidental copies in the course of streaming music,
for example, raise similar questions.

Samuelson: a new round of cases about using YouTube
performances—allegations of 1201 violations to say that scraping the video to
use that data for training. Even if the stream itself wasn’t a copy, if you can
make a copy from it, what to do? Should there be a meaningful distinction b/t
access controls and copy controls? Reese
said yes! (And he was right.)

Gellis: is vibe-coded software copyrightable?

Lemley: depends on how vibey you were! Maybe a selection and
arrangement ©, though for code that will be less broad.

Samuelson: divinely authored works cases are also relevant.

RT: For Samuelson: courts often seem to divide over whether
there are 2 ways to have a thin ©: the first way is that anyone who does the
same thing can make a nearly identical work, but they can’t engage in
reproduction; but the second and more controversial one is that anyone can copy
chunks of the work as long as they don’t copy nearly the whole thing: Sedlik concurrences
and Thomas’s dissent in GvO. Do you think that these are both examples of thin ©?

Samuelson: we’re close to a thin © but not as thin as Goldstein
suggested (which was an exact copying standard). Dennis Karjala suggested a
rule of exact copying being needed for infringement; emulating functionality is
really important for freedom. Lotus v. Borland—that’s actually a case allowing
copying the interface (that is, a chunk) as a method of operation. You still
have to write/code the functionality independently. Fed Cir ©ability ruling in GvO
was a real step backwards—wrong as a matter of law.

Lemley: Thin ©: if the test is virtual identity, one way to
test is for really close identity (99% or close), but © in general protects
protectable expression. One approach would be to say that identical copying of
even a small fragment is infringing under this virtual identity standard; the
other approach would require copying the whole work verbatim or close to
verbatim. That will matter a lot where most of what I copied was not
protectable.

Reese: part of the problem is that most courts dealing with
this look to Feist, but Feist isn’t about either of those things. In these kinds
of cases, not very much of what you put into your work is ©able, b/c you’re
using facts/all you get is selection, coordination, and arrangement.
Infringement means parsing carefully which thin part of your work is protected
by ©, but that doesn’t say anything about whether partial copying infringes a
thin ©. There’s no original understanding to look back to!

from Blogger https://tushnet.blogspot.com/2026/04/commemorating-50-years-of-copyright-act.html