A technology arrives that the profession does not know how
to think about, and the response follows a reliable sequence. First the
anecdotes, then the sanctions, then the committees, and, at last, the question
of whether a new body of law is required.
Hallucinated citations have now carried us to the fourth
stage, and bar associations, courts, and firms are all busy, in committees,
orders, and blog posts, drafting proposals on what should be done.
This sequence is not new, and the profession already
witnessed it unfold, with the same sound and fury, after the arrival of the
world wide web.
In 1996, at a conference convened to consider the law of “cyberspace,”
Frank Easterbrook advised the assembled participants that their subject did not
warrant a body of law of its own. Invited to speak on property in the age of
the internet, Easterbrook reported that his first instinct had been to wonder
whether this was simply the law of the horse in new dress. As Easterbrook
noted, the formulation came to him from Gerhard Casper. As dean of the
University of Chicago Law School, Casper had boasted that his school taught no
such course.[1] Equine disputes scatter across sale, tort, licensing, and
professional negligence, and, consequently, any attempt to gather them into a
subject of their own is, in Easterbrook's words, "doomed to be shallow and
to miss unifying principles."[2] Accordingly, Easterbrook provided the
prescription: "Develop a sound law of intellectual property, then apply it
to computer networks."[3]
There is no law of the hallucination
Let's set memories of ‘90s dial-up aside for a moment
to point out what a lawyer has actually done when a brief goes out bearing
citations to cases that do not exist. Essentially, that lawyer has signed a
document they did not read.
The profession has rules for that, and they are neither
obscure nor recent. A lawyer who signs a filing certifies that the legal
contentions it advances are warranted by existing law.[4] Competence has always
been understood to require the thoroughness and preparation necessary for the
representation.[5] Candor toward a tribunal forbids a knowing false statement
of law and imposes an affirmative duty to correct one already made.[6] And the
rules governing supervision extend the same logic to work a lawyer has
delegated to someone (or something) else.[7]
None of these duties contains an exception for text the
signing lawyer did not personally compose, and none ever has. The profession
has never accepted "the summer associate drafted it" as an answer
(regardless of whether that summer associate was a younger version of this
author getting lost on the subway on the way to the office, on one end of the
spectrum, or an associate on the masthead of the Yale Law Journal, on the other
end of the spectrum), and nothing about the drafter being a machine rather than
a second-year alters the structure of the obligation. A rule addressed
specifically to generative artificial intelligence would add nothing the
signature block does not already supply.
Easterbrook had made the same observation about copying.
Cheaper copying, he allowed, "may make violations of the law more
attractive," and may warrant spending more to catch the violators,
"but movement along a cost continuum does not call for change in legal
substance."[8]
If this is the case, then, why are courts and bar
associations frantically scribbling guidance for practitioners, if the practice
already has a strong layer of rules that apply to hallucinations?
What Lessig saw
Lawrence Lessig published his answer to Easterbrook
years afterward, and the striking feature of that answer is that it does not
proceed by defending the law of the horse. Lessig concedes most of the frame at
the outset, and concedes it in terms that leave him very little room: "I
am not defending the law of the horse. My claim is specific to
cyberspace."[9]
The claim he does make is that "[b]ehavior... is
regulated by four kinds of constraints" rather than one: law, social
norms, the market, and what he calls architecture, meaning the structure of the
space in which the behavior occurs.[10] Law is the modality lawyers are trained
to notice, because law is the modality we think about continuously, being in
the profession. However, the other constraints of social norms, market, and
architecture also constitute levers of control. For the “cyberspace” of the ‘90s,
and for the large language model boom of the 2020s, the design and code of the
technologies provide such architecture.
Lessig's own example is worth following, because its shape
recurs. He asks the reader to imagine walking into a store where a guard
records your name, cameras track which items you look at and which you ignore,
and an employee follows you down each aisle timing you. If that happened in a
shop, he observes, "you would notice." You might not object. But you
would know enough to choose, if you possessed preferences regarding such
surveillance.[11]
In “cyberspace” you would not, and the reason is
architectural. On the internet, "[s]urveillance is not
self-authenticating," Lessig writes; "[n]othing reveals whether you
are being watched."[12] No privacy rule was repealed to bring this about.
What changed is that the (cyber) space, due to its architecture, stopped
supplying (for free) facts about the user's privacy (or lack thereof), which
were otherwise provided by nature of browsing a physical store in person. The
ordinary condition of the architecture does the regulatory work, independent
from any laws.[13]
What Lessig draws out of this is the following
proposition: "Cyberspace has no nature; it has no particular architecture
that cannot be changed."[14] He means it narrowly. It's not that the
present design has no features, but that other designs are possible.
Easterbrook's essay is remembered for its opening about
equine law. However, the second half of his essay, about costs, is also
relevant to discussions of regulatory architecture. Having declined to
prescribe a law of “cyberspace,” he turns to the economics of copying:
"Cyberspace reduces the effective cost of copying."[15] Easterbrook
noted that copyright already prohibited what the photocopier and then the
network made easy. The difficulty is that a right becomes unenforceable when
transacting over it costs more than the transaction is worth.[16]
His remedy was design-driven, as opposed to the imposition
of law. Each file would carry its own instructions on copying; programs would
read those instructions and negotiate automatically; a standards body would
publish a menu of options from which parties could choose at low cost.[17] The
model he reached for was the Uniform Commercial Code and Karl Llewellyn's
vision of giving merchants defaults they could adopt without bargaining for
them, Llewellyn being the man who coined "the law of the horse" in
the first place, for the amateur trade the UCC was written to leave behind.[18]
However, the means by which Easterbrook imagines effectuating his remedy was
architectural: protocols built into the code which, by design, make a market
for intellectual property.
Accordingly, the judge who told the cyberlaw conference
that its subject did not deserve a course spent the rest of his remarks
designing an architecture for it! The distance between Easterbrook and Lessig
is smaller than the citation history suggests.
Returning to Lessig's view that "Cyberspace has no
nature; it has no particular architecture that cannot be changed," the
architecture of legal research has no nature either. It has a present design,
which is recent and malleable.
The price of an old duty
What was the architecture of a legal citation before
large language models arrived?
Until quite recently, the existence of a cited authority
was very nearly self-authenticating (using Lessig's language), and the
profession had no particular occasion to notice that strange reality. A
citation reached a lawyer through a reporter, a treatise, a citator, or a
commercial database, and whatever delivered it also vouched for it. To reach
print at all, the citation had to pass through someone who had found and read
the case, or through a deterministic process overseen by such a person. To reach
a brief, from print, it had to pass through a lawyer who located the entry in
one of those volumes and copied it out by hand. Human-linked checks therefore
occurred, automatically, as part of a lawyer's normal practice, even though
those checks were not designed as checks. Additionally, these checks seemed, to
the practitioner, free, since the costs of such checks (a publisher sourcing an
opinion for inclusion in their product or a lawyer finding and copying a
reference to an opinion from such a product) occurred incidentally with regular
lawyering (instead of requiring supplementary effort).
A language model severs these automatic checks from the
lawyerly day-to-day. A citation can now enter a draft without any publisher
having collected it and without any lawyer having found it, and where the model
sits inside the word processor, the citation can enter the draft without anyone
having looked at the citation, at all. Worse, the genuine and the invented
citations arrive in identical dress (plausibly formatted, plausibly numbered,
plausibly named), so a cursory read of the citations will not separate them.
It's worth being careful about what has changed here (and
what has not). Today's professional responsibilities regarding citations are
largely the same responsibilities that regulated practitioners in 1995.
However, in 1995, the additional costs of fulfilling those responsibilities, in
proving the existence of a citation, were something near zero, since human
oversight occurred automatically as a byproduct of a lawyer's use of a citation
in work product. These costs were already bundled in the publishing costs of
the reporter, a treatise, a citator, or commercial database, and in the cost of
the lawyer's own exertion in reading and copying from those publications.
Flash forward to 2026: the lawyer using a large language
model can produce written work product, citations included, while avoiding the
previously bundled costs of the publisher and their own toil in reading and
transcribing the citations. Because of this unbundling, the additional costs of
fulfilling the responsibilities related to the verification of any citation are
now substantive, and entirely borne by the lawyer.
That single displacement accounts for a good deal of
institutional behavior that otherwise looks reactionary. When firms, courts,
and bar committees reach for moratoria, disclosure regimes, and outright
prohibitions, they are not responding to any ambiguity in what the profession
requires; what the profession mandates, in terms of rules, is largely the same.
Instead, they are responding to the economic shift described above: a duty that
was seemingly cheap (because the costs were bundled in the nature of production)
has become expensive, and one way to manage the now-expensive duty is to forbid
the activity that requires it.
Two failures under one name
There is a second difficulty here, which emerges from
the purpose of citations in legal writing. Damien Charlotin's database of
judicial decisions addressing hallucinated material listed 2,006 decisions as
of September 1, 2026, of which 1,376 came from United States courts, sorted
into categories. A single decision can carry more than one tag, so the
categories overlap (and the totals do not sum). Across the database’s
American-tagged items, fabricated authorities appeared in 1,134 decisions.
Misrepresented authorities appeared in 586, and false quotations in 462.[19]
Those latter categories describe a different animal
altogether. There the case is real, the citation is accurate, and the reporter
volume checks out on inspection. However, the proposition attached to the
citation has been invented, with language placed in the mouth of a court that
never articulated such a position.
Why should that distinction matter so much?
Because whether a case exists is a question of lookup, and
lookup is a (relatively) easier category of problem (that's why it was also
largely solved by the "automatic" publisher and lawyer checks
described above). On the other hand, whether a case holds what it has been said
to hold is a question of reasoning (something difficult for a lookup against an
index).
This is why a tool that only checks existence and stops
there can be more dangerous than a tool that does nothing. Having disposed of
the easier half of the problem, it gives the lawyer a state of confidence, even
though hallucinations regarding the reasoning of the court may still lurk in
the shadows.
Rebuilding the architecture
If the law is already relevant to the problem of
hallucinations, let's turn back to Lessig and Easterbrook to see if we can also
use architecture to control this emergent problem. For an architecture to
regulate citational hallucinations effectively, we must consider questions on
how the architecture affects the costs borne by the lawyer to discharge their
professional duties.
The work we have been doing at Filevine divides into two
layers that together alter these costs.
The first addresses existence. Every citation to a case
produced by LOIS is automatically checked against source databases of
reporter-based citations. Verified citations are rendered as checkmarked links
that open the opinion beside the answer. Unverifiable citations, in contrast,
are flagged with caution highlighting and left unclickable.
The second layer addresses the harder failure:
hallucinations in legal reasoning. Where the tool relies on a given passage
from an opinion, it opens the opinion to that passage and highlights it for the
user, so that confirming an assertion becomes an immediate read rather than a
lost expedition through seventy pages of judicial writing. Additionally, using
this second layer, an attorney can put questions to a single opinion directly.
They can ask the opinion what the procedural posture was, how the court reached
its conclusion, how far the holding actually extends, or what the dissent was
arguing, for example. Those answers come directly from the text of the court
opinion, likewise highlighted.
What our architecture changes is the number of steps
(i.e., costs) for a lawyer to discharge their professional obligations: the
architecture gives the lawyer an amber gloss on top of a problematic citation
rather than plain text that masks a hallucination, a click in a panel rather
than a new lookup session on another platform, and a jumped-to, highlighted
paragraph rather than seventy unmarked pages.
The architecture that once made these duties seemingly
free was not designed by anyone; it was a byproduct of how citations happened
to reach, and be used by, lawyers. However, because LLM-driven legal research
no longer supplies such verification as a byproduct, anti-hallucination
architecture must be built on purpose, which we have sought to do.
An old kind of answer
Easterbrook, still on copying, described different
modes of regulating reproduction in the 90s internet era. He pointed at the
weakness of regulating behavior by fiat, regardless of the economic costs. The
Copyright Clearance Center printed a line at the foot of articles directing the
copier to remit fifty cents a page. He asked whether anyone actually remits:
"No, you don't, because the extent of the fair-use privilege is so hard to
grasp, and because remitting would be such a bother—do you tape two quarters to
a post card and mail it?"[20] Discharging the obligation was a bother, so
it went undischarged, and the Center "barely raises enough money to cover
the costs of its operations."[21]
Every prohibition now being drafted for generative
research tools is a slug line at the foot of an article. It states an
obligation that already existed but neglects the economic costs of the
obligation, given the technological change. Some lawyers will pay the fifty
cents (i.e., closely read and
double-check each citation and claim by hand). Some may not, and hallucinations
will thereby cause the chaos observed currently. Rather than more laws and
written guidance (slug lines), we propose new architecture, which, by design
and function, reduces the costs of discharging the professional obligation so
that the lawyer actually does so.
Notes
[1] Frank H. Easterbrook, Cyberspace and the Law of the
Horse, 1996 U. Chi. Legal F. 207.
[2] Id. at 207.
[3] Id. at 208.
[4] Fed. R. Civ. P. 11(b)(2).
[5] Model Rules of Professional Conduct Rule 1.1.
[6] Id. Rule 3.3(a)(1).
[7] Id. Rules 5.1, 5.3.
[8] Easterbrook, supra note 1, at 210.
[9] Lawrence Lessig, The Law of the Horse: What Cyberlaw
Might Teach, 113 Harv. L. Rev. 501, 502 (1999).
[10] Id. at 506.
[11] Id. at 504.
[12] Id. at 505.
[13] See id. at 506–08.
[14] Id. at 505.
[15] Easterbrook, supra note 1, at 208.
[16] Id. at 208–10.
[17] Id. at 213–14.
[18] Id. at 214 (citing Karl N. Llewellyn, Across Sales on
Horseback, 52 Harv. L. Rev. 725 (1939)).
[19] AI Hallucination Cases,
https://www.damiencharlotin.com/hallucinations (last visited Sept. 1, 2026).
[20] Easterbrook, supra note 1, at 209.
[21] Id.