Perplexity’s movement to dismiss in its protection in opposition to the New York Instances and Chicago Tribune tries to slim the instances, not finish them. (Chicago Tribune Co. LLC v. Perplexity AI Inc. (1:25-cv-10094) and The New York Instances Co. v. Perplexity AI Inc. (1:25-cv-10106).). It asks the courtroom to dismiss the publishers’ direct-infringement claims based mostly on allegedly infringing solutions, the contributory and vicarious infringement claims, and, within the Instances case, the trademark claims. Perplexity says these counts fail on the pleading stage as a result of the challenged outputs are generated solely after a consumer immediate, so the related “volitional” act is a consumer’s request slightly than Perplexity’s personal conduct; as a result of plaintiffs allegedly present solely a small variety of output examples slightly than work-by-work allegations throughout the total asserted catalog; as a result of merely offering an automatic software is just not sufficient to plead contributory legal responsibility; and since the Instances’ trademark allegations don’t plausibly present actionable use past nominative reference. Perplexity additionally says trimming these claims would depart the “core challenge of truthful use” for later levels.
As Perplexity describes its platform, it’s an “reply engine” that interprets a consumer’s query, searches the web in actual time, gathers data from “top-tier” or different authoritative sources, and returns a concise, conversational abstract with hyperlinks to the underlying sources and associated questions. It characterizes the method as automated retrieval-augmented era: retrieve related paperwork, mix them with the immediate, and have an LLM produce a natural-language response. Perplexity insists the system is designed to synthesize information, not merely regurgitate verbatim supply textual content.
In its briefing and public framing, Perplexity additionally positions the case as a dispute about the way forward for search and entry to data, not merely about copying. The corporate repeatedly characterizes itself as a next-generation search engine whose operate is to assist customers find and perceive data scattered throughout the open internet. In that framing, the publishers’ lawsuit is portrayed as an effort by massive information organizations to increase copyright past expression and into management over factual data and reporting subjects. Perplexity means that permitting such claims to proceed would danger turning copyright right into a mechanism for gatekeeping information—successfully letting main publishers assert quasi-exclusive management over the informational worth of their journalism.
Perplexity’s narrative echoes a well-recognized technology-sector theme that goes again to early web debates about linking and search indexing: information should not copyrightable, and instruments that assist customers uncover and synthesize data are a part of the infrastructure of the open internet. From that perspective, the platform describes its summaries as transformative instruments that assist customers navigate the web extra effectively whereas nonetheless pointing readers to the underlying sources by hyperlinks and citations. The implication is that the publishers’ principle would undermine the standard function of search engines like google and different data intermediaries by treating automated summarization as illegal appropriation slightly than as a authentic technique of organizing and speaking publicly obtainable data.
As a result of, you already know–data needs to be free.
Now I’m only a nation lawyer from Texas and I freely admit that I’m not as sensible as these metropolis fellers who work for Perplexity. However this simply feels like an enormous heaping rasher of crap to me.
I requested Perplexity to do one thing that’s distinctly unsearch-like: Write me a narrative about Philip Dru Administrator within the fashion of James Fenimore Cooper. That story doesn’t exist. Right here’s what I bought again, not a blinking cursor, however slightly a “fascinating combine”:
Perplexity responded by producing an unique artistic narrative — not by retrieving or synthesizing search outcomes. . This output raises vital stress with a number of of Perplexity’s core arguments in its movement to dismiss.
Perplexity’s Characterization of Its Personal Expertise
All through the movement, Perplexity characterizes its “reply engine” as a software that “searches for, acquires, and synthesizes data from a variety of related sources” and offers “environment friendly entry to concepts and information that nobody owns.” . Perplexity emphasizes that its system makes use of retrieval-augmented era (“RAG”) to “cull information in actual time from quite a few sources” after which “condenses and summarizes these information utilizing an LLM.” . In its telling of its personal tall story, the reply engine “mechanically generates an easy-to-understand, natural-language response, incorporating information and concepts gathered from sources on the web.” . It stresses that customers depend on Perplexity “to acquire up-to-date, succinct, factual solutions to an almost limitless variety of questions.” .
Perplexity additionally argues it isn’t straight answerable for copyright infringement as a result of it lacks “volitional conduct” — that it’s merely an automatic system that passively responds to consumer prompts, analogous to Cablevision’s distant DVR, the place “the copies at challenge have been ‘made mechanically upon [a] buyer’s command.’” . The movement frames Perplexity as akin to a passive middleman that “mechanically undertakes actions that end in” outputs and argues that “the mere possession, development, or supervision of the machine or system is not going to set up volitional conduct.” .
Why This Screenshot Is Probably Inconsistent With Perplexity’s Movement
The Philip Dru output undermines Perplexity’s framing in a number of key respects:
This isn’t search or data retrieval. I didn’t ask Perplexity to search out or summarize present data. I requested Perplexity to write an unique artistic story in a particular literary fashion. The ensuing output — a multi-paragraph fictional narrative with invented scenes, dialogue, and descriptive prose — is just not “information and concepts gathered from sources on the web.” . It’s generative artistic content material, which is basically completely different from the search-and-summarize operate Perplexity describes in its movement. .
It contradicts the MTD’s “factual solutions” framing. Perplexity’s legal professionals persistently describe its product as offering “factual solutions” and “entry to concepts and information.” . However writing a narrative in one other writer’s fashion is just not answering a factual query — it’s an train of artistic era that goes effectively past the retrieval-and-synthesis paradigm Perplexity presents to the Courtroom. It’s nearly…whatchamacallit…misleading.
It complicates the “passive conduit” / Cablevision analogy. Perplexity’s volitional conduct argument rests on the premise that it is sort of a DVR system that merely copies pre-existing content material at a consumer’s path. However within the Philip Dru instance, Perplexity is just not copying or retrieving any pre-existing content material — it’s creating new expressive content material. The truth is, it’s creating new expressive content material that by definition couldn’t exist, since Cooper died lengthy earlier than Philip Dru Administrator was printed. Name me loopy, however the capability to generate unique artistic works suggests a system with significantly extra company and volition than a passive copying mechanism. Perplexity’s argument that it engages in “no extra volitional conduct than Cablevision did when it curated programming obtainable for copying” is more durable to maintain when the system is producing novel prose slightly than retrieving and reproducing present materials. No RAG search goes to return that product as a result of it doesn’t exist, or it didn’t till Perplexity wrote it.
It undermines the “consumer presses the button” narrative. Perplexity argues that “direct legal responsibility attaches solely to ‘the one that truly presses the button’” and that customers are the proximate reason behind any output. However when Perplexity generates an unique story, the artistic decisions — the prose fashion, the invented dialogue, the narrative construction — are made by Perplexity’s system, not by the consumer. I simply offered a immediate, however the expressive content material of the response originates from Perplexity’s LLM. This makes it more durable, if not unattainable, to argue that Perplexity is merely a conduit executing consumer instructions with none impartial artistic or volitional function.
Sensible Significance
To be clear, Perplexity’s movement is particularly centered on the slim query of whether or not the Complaints adequately allege direct infringement for outputs that allegedly reproduce Plaintiffs’ copyrighted content material. The Philip Dru screenshot doesn’t straight implicate the copyright claims at challenge. Nonetheless, it’s the sort of proof that might be used to problem Perplexity’s broader characterization of itself as a search-like software missing company or volition, which is central to its authorized principle. If Perplexity’s system could be prompted to generate unique artistic works — slightly than merely retrieve and summarize present data — that capability undercuts the analogy to passive automated programs like Cablevision’s DVR and suggests the system workouts a level of artistic company that’s tough to sq. with the “passive conduit” narrative.
Bond, James Bond
Then I requested Perplexity to rewrite the Cooper story within the fashion of Ian Fleming. Once more, no blinking cursor—my thought was an “wonderful alternative” which it was however not for the rationale Perplexity thought:
The Ian Fleming revision deepens the strain with Perplexity’s self-characterization in a number of necessary ways in which transcend even the primary Philip Dru instance.
The “Revision” Drawback
Within the movement, Perplexity’s legal professionals describe their consumer’s reply engine as a software that “searches for, acquires, and synthesizes data from a variety of related sources” utilizing the retrieval-augmented era we’ve mentioned earlier than to “cull information in actual time from quite a few sources” after which “condenses and summarizes these information utilizing an LLM.” . That is seemingly a RAG that pulls on closed vector indexes and reside retrieval. Vector indexes are databases of numerical embeddings representing beforehand ingested paperwork that permit an AI system to shortly retrieve semantically comparable textual content, whereas reside retrieval pulls recent internet pages in actual time by a search engine or crawler earlier than feeding them to the mannequin.
The system is introduced by the legal professionals as one which “mechanically generates an easy-to-understand, natural-language response, incorporating information and concepts gathered from sources on the web.” . Customers supposedly “depend on Perplexity’s AI software to acquire up-to-date, succinct, factual solutions to an almost limitless variety of questions.”.
However—asking Perplexity to revise its personal beforehand generated unique story within the fashion of a distinct writer has nothing by any means to do with any of those described features. There is no such thing as a web search taking place, no retrieval of information, no synthesis of supply materials and doubtless no RAG as a result of there’s nothing to replace. The system is performing a purely generative, artistic activity — rewriting prose within the distinctive fashion of Ian Fleming — that’s completely untethered from the search-and-summarize paradigm Perplexity’s very sensible legal professionals current to the Courtroom.
Particular Inconsistencies within the MTD
No RAG, no retrieval, no “information.” Perplexity’s movement emphasizes the RAG course of as central to how the reply engine works: the system receives a immediate, retrieves “related content material,” combines the enter with retrieved paperwork, and offers the mixed knowledge to an LLM. However when a consumer asks Perplexity to revise a narrative it already wrote within the fashion of a completely different writer, this time Ian Fleming, there isn’t a factual retrieval step. There’s nothing to retrieve this time, both. The system is exercising literary judgment — analyzing stylistic parts related to Fleming’s prose (brief declarative sentences, brand-name specificity, understated violence, and so on.) and making use of them to reshape present textual content. That’s artistic authorship, not data retrieval. .
It demonstrates iterative artistic company. My Philip Dru instance already confirmed Perplexity producing unique artistic content material. However the Ian Fleming revision goes additional: it reveals Perplexity modifying its personal artistic work based mostly on an aesthetic directive. This isn’t a system that passively “executes” a consumer command the best way a DVR information a pre-existing tv program.
It’s a system making a sequence of advanced artistic decisions — what to maintain, what to vary, the right way to modulate tone, diction, and pacing derived from its personal work — which can be attributable to the system itself, to not me. I mentioned “revise within the fashion of Ian Fleming”; each explicit artistic choice about how to do this was made by Perplexity’s LLM with no prompting from me.
It undermines the Cablevision analogy much more forcefully. Perplexity argues it “engages in no extra volitional conduct than Cablevision did when it curated programming obtainable for copying, and offered ‘entry to a system that mechanically produces copies [of that content] on command.’” . However Cablevision’s DVR made copies of pre-existing tv programming. . It didn’t rewrite a sitcom within the fashion of a distinct showrunner. The Ian Fleming revision demonstrates that Perplexity’s system is just not merely copying or retrieving something — it’s performing a refined artistic transformation. That could be a basically completely different sort of “automated” exercise than Cablevision’s DVR, and it makes the analogy considerably more durable to maintain. The truth is, you would possibly even say it’s not automated in any respect.
It complicates the “consumer presses the button” narrative. Perplexity’s movement insists that “direct legal responsibility attaches solely to ‘the one that truly presses the button’” and {that a} consumer’s immediate is the proximate reason behind the output. However I simply typed “revise within the fashion of Ian Fleming.” I didn’t make any of the substantive artistic selections that went into the revised textual content. I didn’t determine which sentences to restructure, which adjectives to swap, or the right way to modulate the narrative voice. All of these expressive decisions — the very substance of the output — originated from Perplexity’s system. The hole between my instruction and the system’s artistic output is way wider than the hole between a Cablevision buyer urgent “document” and the DVR making a bit-for-bit copy of a broadcast.
The Broader Implication
What the Ian Fleming revision (and I might preserve that up all day) illustrates is that Perplexity’s reply engine can’t probably be merely a search software or a factual retrieval system. It’s a general-purpose generative AI able to unique artistic authorship, stylistic mimicry, and iterative revision of its personal artistic work. Perplexity’s movement to dismiss fastidiously frames the product as a “novel and groundbreaking AI software” that helps customers “acquire up-to-date, succinct, factual solutions” by “gathering insights from top-tier sources.” . However examples like mine reveal capabilities that go far past the legal professionals’ framing to the Courtroom— capabilities which can be tough to reconcile with the characterization of Perplexity as a passive, automated conduit missing volitional conduct. .
Whereas this instance, like the primary one, doesn’t straight contain the copy of the plaintiffs’ copyrighted content material at challenge within the case, it offers extra ammunition for plaintiffs to argue that Perplexity’s self-description is strategically incomplete and that the system workouts a level of artistic company that belies the “passive middleman” framing central to its movement to dismiss.
So these sensible metropolis fellers have been seemingly describing anyone‘s product, simply not Perplexity’s.
And There’s Simply One Extra Factor That’s Been Bothering Me
There’s one thing about these examples that’s nagging me. The flexibility of Perplexity’s system to put in writing within the distinctive types of James Fenimore Cooper and Ian Fleming — and to revise between these types on command — strongly implies that the underlying LLM was skilled on a broad corpus of literary texts, together with works by these authors. This has vital implications in mild of the arguments Perplexity makes in its movement.
What the Movement Says (and Doesn’t Say) About Coaching Information
Perplexity’s movement is notably cautious — dare I say even evasive — concerning the query of coaching knowledge. It acknowledges, by quoting the complaints, that the LLMs “upon which Perplexity’s AI merchandise are constructed” can “generate solutions to questions on data that’s included of their coaching knowledge” and that “[t]he high quality of the output depends upon the scale of the mannequin, the range of coaching knowledge, and the precise structure and coaching methods used.” . However the movement by no means addresses what is definitely in that coaching knowledge, and it focuses solely on the RAG-based retrieval course of — the “output” aspect — slightly than the “enter” aspect of coaching. .
Perplexity expressly notes that its movement doesn’t problem the Plaintiffs’ Depend I claims, which goal “Perplexity’s alleged personal — i.e., volitional — acts with respect to utilizing Plaintiffs’ works as inputs to develop its reply engine.” . The coaching knowledge challenge is, in different phrases, intentionally sidestepped.
What the Philip Dru and Ian Fleming Examples Reveal
The capability to carry out these duties is robust circumstantial proof of broad-based coaching on copyrighted literary works, for a number of causes:
Stylistic mimicry requires ingestion of the supply materials. To write down “within the fashion of Ian Fleming,” the LLM should have internalized Fleming’s distinctive prose traits — the clipped, declarative sentences, the brand-name specificity, the indifferent observations of violence, the sensory precision. That data doesn’t come from RAG retrieval of web search ends in actual time. . It comes from having been skilled on a considerable pattern of Fleming’s prose — which consists nearly completely of the copyrighted James Bond novels and brief tales. The identical is true for James Fenimore Cooper’s ornate, Nineteenth-century descriptive fashion.
Data of obscure works implies breadth of coaching corpus. Philip Dru: Administrator is a comparatively obscure 1912 novel by Colonel Edward Mandell Home, a political advisor to Woodrow Wilson. He ain’t no James Bond. The truth that Perplexity’s system can generate unique fiction that includes this character suggests the LLM was skilled on a corpus broad sufficient to incorporate this textual content. If the coaching knowledge encompasses a piece this obscure, it nearly definitely encompasses an unlimited vary of extra distinguished literary works as effectively.
The revision functionality implies deep stylistic encoding. The truth that the system can revise its personal output from one writer’s fashion to a different’s demonstrates that it has not merely memorized surface-level patterns however has developed what quantities to an inner mannequin of every writer’s distinctive voice. This stage of stylistic sophistication is per coaching on massive volumes of every writer’s work — not merely a sentence or two referenced in a Wikipedia article concerning the writer.
The Bartz Connection
In Bartz v. Anthropic, the core allegation is that Anthropic skilled its Claude mannequin on a large corpus of copyrighted books, enabling it to breed and generate content material derived from these works. The Philip Dru and Ian Fleming examples would assist a similar inference right here: that the LLM underlying Perplexity’s reply engine was skilled on a equally broad corpus of copyrighted literary texts. This issues for a number of causes:
It reinforces the plausibility of Plaintiffs’ Depend I claims (which Perplexity doesn’t problem on this movement) that Perplexity used copyrighted works as coaching inputs. If the system can mimic particular authors’ types, these authors’ works have been very seemingly within the coaching knowledge — and so, presumably, have been Plaintiffs’ journalistic works. Only a guess.
It additionally undermines Perplexity’s framing of its product as primarily a retrieval software. The movement emphasizes that the reply engine “searches for, acquires, and synthesizes data from a variety of related sources” utilizing RAG. However the Philip Dru and Ian Fleming examples exhibit capabilities that don’t have anything to do with retrieval and every part to do with the generative energy of an LLM skilled on a broad corpus of copyrighted texts. The RAG-centric description of the product is, at finest, incomplete.
However I nearly forgot…that nagging unfastened finish that’s been bothering me, that I simply can’t fairly resolve in my easy thoughts, is that it complicates the volitional conduct argument, don’t it? Only a bit. If Perplexity’s LLM was skilled on copyrighted works — and that coaching is what permits it to generate stylistically distinctive artistic prose on request — then the “volitional conduct” at challenge isn’t just the automated response to a consumer question. It additionally consists of Perplexity’s deliberate alternative to coach on these works, which is the muse for the system’s generative capabilities. Wouldn’t that be Perplexity’s deliberate industrial alternative? Perplexity tries to cabin the volitional conduct evaluation to the second of output era, however examples like mine inexorably draw the Courtroom’s consideration again to the coaching course of, the place Perplexity’s personal decisions — about what knowledge to ingest and the right way to practice the mannequin — are unambiguously volitional. That was it, I believed I used to be nearly performed however then this different unfastened finish simply saved nagging me.
Only a Easy Thoughts for a Easy Man
The Philip Dru and Ian Fleming examples are per the inference that Perplexity’s LLM was skilled on a broad and various corpus of copyrighted literary works, very similar to the coaching practices alleged in Bartz v. Anthropic. Whereas Perplexity’s movement strategically avoids the coaching knowledge query by focusing solely on the output-side claims (Counts II and III), these examples spotlight the problem of sustaining a clear separation between the “enter” and “output” sides of the system. The generative capabilities on show — stylistic mimicry, artistic authorship, iterative revision — are the direct product of coaching decisions, they usually undercut Perplexity’s effort to characterize its reply engine as a passive, retrieval-oriented software. And that might make you nervous sufficient to start out enjoying with a idler tassel.
Or at the least that’s the best way I see it and I’m not as sensible as the town fellers. I’m most likely incorrect.
