D.C. No. 4:22-cv-00051YGR

Appeal from the United States District Court for the Northern District of California Yvonne Gonzalez Rogers, District Judge, Presiding

Argued and submitted December 4, 2024 Seattle, Washington

Filed April 28, 2026

Before: William A. Fletcher, Marsha S. Berzon, and Ryan D. Nelson, Circuit Judges.

Opinion by Judge R. Nelson; Concurrence by Judge Berzon; Concurrence by Judge R. Nelson

Page 2SUMMARY *

The panel affirmed the district court's dismissal of plaintiffs’ putative class action against Meta Platforms, Inc., alleging that anti-Rohingya content spread on Facebook incited violence against them and their villages in Myanmar and that Meta designed Facebook to encourage this bad content.

The panel held that plaintiff’s claims were barred by the immunity Congress granted to online companies under Section 230 of the Communications Decency Act, which protects Meta from claims that seek to treat it as the publisher or speaker of any information provided by another information content provider.

First, the panel held that Section 230 applied to plaintiffs’ claims under California’s choice -of-law rules in this diversity case. Assuming California's choice -of-law rules could demand application of Myanmar's rules of decision rather than Section 230, the panel concluded that plaintiffs failed to carry their burden to show that foreign law applies. Myanmar’s interest in protecting its citizens from harmful attacks and misinformation on Facebook was insufficiently incorporated into the positive law of the country, and therefore Myanmar’s interest did not predominate.

Page 3Section 230 immunizes from liability “(1) a provider or user of an interactive computer service (2) whom a plaintiff seeks to treat, under a state law cause of action, as a publisher or speaker (3) of information provided by another information content provider.” Barnes v. Yahoo!, Inc. , 570 F.3d 1096, 1100-01 (9th Cir. 2009). At step one, the parties agreed Meta was the provider of an interactive computer service. At step two, each of plaintiffs’ claims sought to hold Meta responsible in its capacity as a publisher of third-party conte nt. At step three, Meta did not make a 'material contribution” to or develop the third -party content at issue. Accordingly, Section 230 barred plaintiffs’ claims.

Concurring, Judge Berzon, joined by Judge W. Fletcher, stated that this Court’s precedent has unduly expanded the scope of section 230 immunity, and has stretched the term “publisher” in section 230 past the point of recognition. She urged this Court to reconsider en banc precedent extending section 230 immunity to recommendation of content and connections to users.

Concurring, Judge R. Nelson wrote separately to address two issues. First, this Court has over-read Section 230, straying from the original public meaning of the statutory text and creating an all-purpose liability shield for internet platforms. Second, state choice-of-law rules can never direct application of a rule of decision contrary to federal law. Because Section 230 is supreme to state law, it controls even if California’s choice -of-law rules point to a conflicting foreign rule.

Page 4Roger J. Perlstadt (argued) and Ryan D. Andrews, Edelson PC, Chicago, Illinois, for Plaintiffs-Appellants.

Kristin A. Linsley (argued), Wesley Sze, and Rosemarie T. Ring, Gibson Dunn & Crutcher LLP, San Francisco, California; Patrick J. Fuster and Perlette M. Jura, Gibson Dunn & Crutcher LLP, Los Angeles, California; Jacob T. Spencer and Christian Dibblee, Gibson Dunn & Crutcher LLP, Washington, D.C.; for Defendant-Appellee.

R. NELSON, Circuit Judge:

For decades, the Rohingya have faced discrimination and violence in Myanmar. After Facebook was introduced there in 2011, anti-Rohingya content spread on the platform. Plaintiffs allege that this content incited violence against them and their villages. They argue that Meta designed Facebook to encourage this bad content and that Meta should be held responsible.

We deal from the start with the immunity Congress granted to online companies under Section 230 of the Communications Decency Act of 1996. See 47 U.S.C. § 230(c)(1). Section 230 ’s grant of immunity protects Meta from claims that seek to treat it “as the publisher or speaker of any information provided by another information content provider.' Id. We conclude that, under this court's precedents, Plaintiffs ' claims treat Meta as a publisher of third-party content and are therefore barred by Section 230.

Page 5So we affirm the district court's dismissal of Plaintiffs’ claims on this ground.

Plaintiffs are members of the Rohingya community — a largely Muslim ethnic minority population indigenous to Western Myanmar — who were displaced from their homes following attacks by military and civilian forces. The Rohingya have long been violently persecuted in Myanmar. Plaintiffs allege that the persecution has worsened in the last decade, resulting in genocide at the hands of the Burman ethnic majority and the government. This action centers on the allegation that Meta is to blame for fomenting the increased, pervasive violence against the Rohingya.

Plaintiffs allege that Meta, through its social media platform Facebook, played a role in the Rohingya genocide.¹ When Facebook was introduced in Myanmar in 2011, Meta partnered with telecommunications companies to pre-load Facebook onto mobile devices sold in the country, thereby allowing tens of millions of Myanmar people to access the social media platform.

Plaintiffs allege that Meta 's mode of operations in Myanmar increased the risk that Facebook would be used to stoke violence in Myanmar. They assert that Facebook did not have a Burmese-language platform interface or terms and conditions. This omission meant that users could not report “ misinformation, hate speech, or calls for violence .” Plaintiffs also fault Meta for not actively monitoring Page 6Facebook in Myanmar, in part because it lacked enough content moderators competent to do so.

Plaintiffs also allege that Facebook's core design and function exacerbated these harms, encouraging the creation of violent posts. Plaintiffs point to Facebook’s implementation of an algorithmic content delivery system in 2009 that promoted 'toxic posts ' because users were more likely to interact with them. According to the complaint, people interact with toxic posts more because they play to users' negative emotions. Allegedly, to maximize user time and so the viewing of revenue-generating advertisements, Meta adjusted its algorithms to boost negative content over neutral or positive content.

Not only did this policy boost the time users spent on Facebook — it also encouraged users to post in a certain way. When users published negative or hateful posts, they saw notifications and high post engagement (such as likes and comments) -what Plaintiffs call 'social rewards'and were encouraged to post more such posts.

To illustrate the point: Imagine a user creates two posts on Facebook. The first is benign — say, a recipe for a traditional Rohingya Masser Sallon fish curry. The second contains false and inflammatory claims about Rohingya conspiracies to harm Burman Buddhists. The second post will allegedly receive more engagement — likes and shares by incensed Burman nationalists, comment wars, angry reactions, and so on. Because users have engaged with it more, the second post is boosted on other users’ feeds , where it gets even more attention. And the original poster is encouraged to post more anti-Rohingya conspiracy content instead of recipes. The cycle continues, and the information system is flooded with vitriol.

Page 7Facebook's algorithm might be the same everywhere, but Plaintiffs’ theory relies on the assertion that the Myanmar people are particularly susceptible to this alleged implicit encouragement of toxicity. Plaintiffs allege that the Myanmar people 'lack[ed] basic understanding of how to . . . make judgments on online content' and were thus digitally unsophisticated and incapable of navigating Facebook’s information environment. So susceptible were these people that they responded with real-world violence and genocide.

Plaintiffs do not allege that Facebook’s 2009 algorithm boosted certain posts specifically because of the company’s content preferences. Under Plaintiffs' theory, Facebook's algorithm promoted toxic content to all users to drive up engagement, reflecting the 'universal human instinct[s]' for hatred and divisiveness.

Jane Doe 1 filed a putative class action in California state court claiming strict products liability and negligence and seeking “at least $150 billion.' Meta removed the case to the Northern District of California and moved to dismiss. Meta argued that Plaintiffs’ claims were untimely, barred by Section 230 of the Communications Decency Act, and failed on the merits.

The district court dismissed the complaint without prejudice, identifying general pleading deficiencies related to timeliness and the merits. Plaintiffs then filed an amended complaint, adding claims for negligent product design and aiding and abetting other torts.

Meta moved to dismiss the amended complaint on timeliness, Section 230, and merits grounds. The districtPage 8 court dismissed the amended complaint with prejudice on timeliness grounds. Plaintiffs appealed.

We have jurisdiction under 28 U.S.C. § 1291. An order granting a motion to dismiss is reviewed de novo, Zellmer v. Meta Platforms, Inc. , 104 F.4th 1117, 1122 (9th Cir. 2024), and we accept all plausible allegations as true, Ashcroft v. Iqbal , 556 U.S. 662, 678 (2009).

Section 230 of the Communications Decency Act commands that '[n]o provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider.” 47 U.S.C. § 230(c)(1). Meta argued below that Plaintiffs’ claims could be dismissed under that provision. Because the district court dismissed Plaintiffs' claims as untimely, it did not reach the Section 230 issue. But Meta renews its Section 230 arguments on appeal and asserts that the district court can be affirmed on the ground that Section 230 bars Plaintiffs’ claims. We agree.

First, we consider whether Section 230 applies under California’s choice -of-law rules. Generally, federal courts sitting in diversity apply the law of the forum state — including the forum state’s choice -of-law rules. See Cassirer v. Thyssen-Bornemisza Collection Found. , 596 U.S. 107, 115 (2022) (discussing Klaxon Co. v. Stentor Elec. Mfg. Co. , 313 U.S. 487, 496 (1941)). California, the forum state, presumptively applies its own rules of decision. Chen v. L.A. Truck Ctrs., LLC , 7 Cal. 5th 862, 867 (2019).

Page 9Plaintiffs argue that the rules of decision of Myanmar apply to some questions at issue. They contend that the rules of decision of California and Myanmar conflict when it comes to immunity for interactive computer service providers. Assuming California’s choice -of-law rules could demand application of Myanmar’s rules of decision rather than Section 230, 2 we conclude that Plaintiffs fail to carry their burden to show that foreign law applies. See Wash. Mut. Bank v. Superior Ct. , 24 Cal. 4th 906, 921 (2001).

California uses the governmental interest test to determine what jurisdiction’s rules of decision should apply to a given dispute. See generally Reich v. Purcell , 67 Cal. 2d 551 (1967). This test entails three steps: First, we determine whether the rules of decision of the potentially affected jurisdictions differ regarding the issue. See Chen , 7 Cal. 5th at 867 – 68.

Second, if the rules of decision differ, we consider the interest of each jurisdiction to decide whether a 'true conflict” exists. Id. To do so, we determine the application Page 10of each jurisdiction’s laws “under the circumstances of the particular case.' Scott v. Ford Motor Co. , 224 Cal. App. 4th 1492, 1503 (2014); accord Bernhard v. Harrah’s Club , 16 Cal. 3d 313, 320 (1976), abrogated on other grounds by statute in Cal. Civ. Code § 1714 (2012). And we look to the policies animating each jurisdiction’s interest in having its law applied. See Offshore Rental Co. v. Cont’l Oil Co. , 22 Cal. 3d 157, 163 (1978). After all, although two jurisdictions may 'have different laws, there is still no problem in choosing the applicable rule of law where only one of the [jurisdictions] has an interest in having its law applied.' Chen v. L.A. Truck Ctrs., LLC , 42 Cal. App. 5th 488, 496 (2019) (quotation omitted). And because it is 'a serious matter” to assert a conflict between the interests of the forum and a foreign state, a foreign jurisdiction’s interest “must be real rather than hypothetical.” Id. (citation omitted).

Third, if a true conflict exists, we compare “the nature and strength of the interest of each jurisdiction” to determine “which state’s interest would be more impaired” if its law were not applied. Id. at 495 (quotation omitted).

Plaintiffs’ choice -of-law argument fails at step two. The United States ' interest in applying Section 230 is clear.³ Imposing liability on Meta for its actions as a publisher Page 11would frustrate Section 230’s purpose of “promot[ing] the continued development of the Internet and other interactive computer services.” 47 U.S.C. § 230(b)(1). Plaintiffs do not dispute that.

So the question is whether Myanmar has a legitimate interest in this dispute. Plaintiffs argue that Myanmar has a significant interest in protecting people within its borders from harm. But the issue here is a question of immunity: whether American courts can treat a party as the speaker or publisher of certain third-party content. Fair Hous. Council of San Fernando Valley v. Roommates.com, LLC , 521 F.3d 1157, 1162 – 63 (9th Cir. 2008) (en banc). Providing immunity for actions that cause harm necessarily implicates the ability of the individuals to prevent or reduce that harm, and to seek redress for any harm suffered.

Plaintiffs rely on their foreign law expert, who states that “[t]he law on liability of internet companies in Myanmar is governed by three sources: the common law of torts, the Electronic Transactions Law of 2004[], and the Telecommunications Act of 2013.” No case law supports the contention that courts in Myanmar would find tort liability for social media companies under these circumstances . Plaintiffs’ expert conceded that he could find no case in Myanmar concerning a social media company. And the expe rt report notes that 'there has been no codification of tort law in Myanmar, and indeed no statutory intervention at all in [tort] law.'

Likewise, neither Myanmar statute Plaintiffs cite establishes Myanmar's interest in holding social media companies liable in tort for their actions as publishers. The Electronic Transactions Law of 2004 merely “provide[s] for validity of electronic signatures and regulate[s] the storage Page 12of electronic data.' And the Telecommunications Act of 2013 is directed at regulating '[b]lackmailing, bullying, making wrongful restraint on, defaming, disturbing, exerting undue influence on or threatening a person using a telecommunication network.” Both statutes “extend[] liability to those who abet commission of an offence, ' but do not address the liability of social media companies as publishers. So neither statute establishes Myanmar 's interest.

At bottom, Myanmar’s interest in protecting its citizens from harmful attacks and misinformation on Facebook, while real, is insufficiently incorporated into the positive law of the country. Myanmar's interest therefore does not predominate. For these reasons, even if we could or should consider Myanmar law, Section 230 applies.

Section 230 immunizes from liability “ (1) a provider or user of an interactive computer service (2) whom a plaintiff seeks to treat, under a state law cause of action, as a publisher or speaker (3) of information provided by another information content provider. ' Barnes v. Yahoo!, Inc. , 570 F.3d 1096, 1100 – 01 (9th Cir. 2009). The parties agree that Meta was the provider of an interactive computer service. So we consider the other two steps of the Barnes test and conclude that , under this Circuit's case l aw, Section 230 immunizes Meta from Plaintiffs’ claims .

The second step of the Barnes test asks whether a claim would treat a defendant as the speaker or publisher of thirdparty content. We consider whether a cause of action “seek s to plead around [ Section 230’s ] strictures .' Est. of Bride ex Page 13rel. Bride v. YOLO Techs., Inc. , 112 F.4th 1168, 1176 (9th Cir. 2024). We have been wary of the 'artful skirting' of Section 230 immunity, Kimzey v. Yelp! Inc. , 836 F.3d 1263, 1266 (9th Cir. 2016), and 'creative attempt[s] to plead around” it , Lemmon v. Snap, Inc. , 995 F.3d 1085, 1094 (9th Cir. 2021). Accordingly, it is “not the name of the cause of action” that matters but “the duty that the plaintiff alleges the defendant violated, ” Barnes , 570 F.3d at 1101 – 02, including what the claimed duty “ requir[es] the defendant to do, ” Calise v. Meta Platforms, Inc. , 103 F.4th 732, 742 (9th Cir. 2024). So if a claim turns on a defendant's status as a publisher or conduct of publishing — including “reviewing, editing, and deciding whether to publish or to withdraw from publication thirdparty content”— it is barred under Section 230. Barnes , 570 F.3d at 1102. We have also held that Section 230 bars products liability claims that “attempt to hold [a defendant] responsible for users' speech or [the] decision to publish it.” Est. of Bride , 112 F.4th at 1180.

Plaintiffs characterize Meta's duty as one of product design — that Meta should not have built Facebook in a way that boosted incitements to violence. Still, the alleged defects relate to Facebook 's core design as a publishing platform, particularly how Facebook promoted or downplayed third-party posts using algorithms. Under our case law, matching users with content is publishing conduct, even when the user has not requested the content. See, e.g. , Doe v. Grindr Inc. , 128 F.4th 1148, 1152 – 53 (9th Cir. 2025); Dyroff v. Ultimate Software Grp., Inc. , 934 F.3d 1093, 1097 – 98 (9th Cir. 2019); Carafano v. Metrosplash.com, Inc. , 339 F.3d 1119, 1124 – 25 (9th Cir. 2003); see also Barnes , 570 F.3d at 1102.

Plaintiffs identify three purported defects in Facebook’s design: First, it encourages users to create “ extreme and Page 14polarizing content” by using a system of “social rewards” (such as likes, shares, and comments); second, it amplifies such content; and third, it does not allow users to report problematic posts in Burmese.

The third purported defect — suggesting that Meta did not do enough to screen and moderate content — does not eliminate the application of Section 230. It blames Meta “ for not moderating content in some way, whether through deletion, change, or suppression.” Est. of Bride , 112 F.4th at 1180; accord Doe v. Internet Brands, Inc. , 824 F.3d 846, 851 – 52 (9th Cir. 2016); cf. Lemmon , 995 F.3d at 1092. Section 230 immunizes interactive computer services from such claims. See Calise , 103 F.4th at 744. Asking a platform to monitor and review third-party content goes to Section 230’s core . See id. 4

Page 15The first and second alleged defects warrant more discussion. These theories fold together: After all, posts are algorithmically boosted because users engage with them, and users engage with them because they are boosted. In other words, Facebook’s promotion of posts is inextricably (even circularly) linked to the 'social rewards' reflecting third-party engagement. Plaintiffs' claims based on these alleged defects are barred by Section 230 if algorithms that provide users with suggestions of some third-party posts over others constitute publishing conduct. Under our precedent, they do.

These theories of harm depend on the content of thirdparty posts. Est. of Bride , 112 F.4th at 1180. The complaint is filled with allegations that “Facebook posts caused a wave of violence,” and that these “posts, pages, and groups . . . were effective for inciting and organizing' the violence at issue. The encouragement provided by social rewards ultimately depends on third-party engagement and content.

This dynamic is true even though Plaintiffs try to frame the issue as a matter of product design. To be sure, we have held that a products liability claim against a web platform can survive Section 230. See Lemmon , 995 F.3d at 1092. But this case is unlike Lemmon . And those differences point toward publisher status protected by Section 230. Lemmon was about users' direct interaction with Snapchat's platform features (a speed filter). Id. That interaction (speeding) was not mediated by third-party content, meaning the claim stood independently of the content that other users created. Id. at 1087, 1093 – 94.

This case, on the other hand, is about the content of thirdparty posts. Cf. id. at 1093 –94. Plaintiffs’ theory is about Page 16the sharing of messages and what happens when some users interact with bad content, allegedly because of how the Facebook algorithm promoted some content whether specifically requested or not. Cf. Est. of Bride , 112 F.4th at 1180; accord Grindr Inc. , 128 F.4th at 1152 – 53. It thus necessarily relies on information provided by third parties. See Lemmon , 995 F.3d at 1087, 1093 – 94; accord Barnes , 570 F.3d at 1101. Calling Facebook a “product” rather than a publication platform only obscures the point that Meta ' published user-generated speech that was harmful ' to Plaintiffs. Kimzey , 836 F.3d at 1266. This case is more like Estate of Bride , where we concluded that a products liability claim was barred by Section 230: “ What is th[e] harm” at the root of Plaintiffs' claims but the violent and inciting 'posts of others' posted on Facebook? 112 F.4th at 1180.

Relying on Moody v. NetChoice, LLC , 603 U.S. 707 (2024), Plaintiffs also suggest that Facebook ’s algorithm was itself Meta’s message or content. If that were true, then we might not have to treat Meta as the publisher or speaker of third-party content, and Section 230 might be avoided. See Barnes , 570 F.3d at 1100 – 01. But our case law has taken another tack on whether algorithm design represents platform s’ own content .

In Dyroff , we held that algorithmic content delivery and “recommending” content is “meant to facilitate the communication and content of others” and is thus a form of publisher conduct, not a message or content itself. 934 F.3d at 1097 – 98. That was true even though the Experience Project — the platform at issue in Dyroff —used “machine -learning algorithms” to recommend content to users “based on the content of their posts and other attributes.' Id. at 1095. Facebook’s 2009 algorithm is alleged to be much less tailored. See supra pp. 6 – 7.

Page 17Following Dyroff 's lead, we conclude that Facebook’s promotion of engagement-driven content through the 2009 recommendation algorithm must be characterized as recommending and matching conduct. We have already decided that such conduct is the work of publishers, rather than a platform's own content or messaging. See Dyroff , 934 F.3d at 1095; accord Est. of Bride , 112 F.4th at 1181. Accordingly, each of Plaintiffs' claims seek to hold Meta responsible in its capacity as a publisher of the third-party content. See Lemmon , 995 F.3d at 1093; accord Est. of Bride , 112 F.4th at 1180.

Under the third step of the Barnes test, we consider whether Meta made a “material contribution” to or developed the third-party content at issue. Rigsby v. GoDaddy Inc. , 59 F.4th 998, 1008 (9th Cir. 2023) (quoting Gonzalez v. Google LLC , 2 F.4th 871, 892 (9th Cir. 2021)). On this front, Roommates.com emphasized that a platform’s provision of “neutral tools” for users to publish content is not material contribution to such content. 521 F.3d at 1171 – 72. This is especially true where the content is posted contrary to the website’s “express policies” and depends on users’ “voluntary inputs.” Id.

Meta provided users “something along the lines of blank text boxes in which they could post” content of all stripes on Facebook. Dyroff , 934 F.3d at 1099; accord Roommates.com , 521 F.3d at 1173. This blank-box system “could be utilized for proper or improper purposes.” Roommates.com , 521 F.3d at 1172. Facebook did not 'provide any specific guidance as to what' users' posts should contain, “nor [did] it urge [users] to input” violent or genocidal content. Id. at 1173 – 74. And Facebook did notPage 18 'require[] users to post specific content' to access its services. Dyroff , 934 F.3d at 1099; accord Roommates.com , 521 F.3d at 1166.

Plaintiffs’ theory of social rewards does not amount to an allegation that Facebook “made suggestions regarding the content of potential user posts,” either. Dyroff , 934 F.3d at 1099. The system of social rewards that Plaintiffs complain of is, “on [its] face, neutral.” Calise , 103 F.4th at 745. There is no plausible allegation in the complaint that the Facebook algorithm specifically treated anti-Rohingya content differently than any other third-party content. See Gonzalez , 2 F.4th at 894 –95. What’s more, nothing in the complaint plausibly alleges that Meta singled out or selected violenceprone users to post violent content. Cf. Huon v. Denton , 841 F.3d 733, 742 (7th Cir. 2016).

Nor can there be any doubt that third-party users willingly provided “the essential published content” underlying Plaintiffs' claims, see Carafano , 339 F.3d at 1124 , and that such “information was tendered to [Facebook] for publication online,' Roommates.com , 521 F.3d at 1174. So, under our case law, this is “precisely the kind of situation for which [S]ection 230 was designed to provide immunity.” Id.

We have also stated that a defendant's actions must relate to 'what makes the displayed content illegal or actionable.' Kimzey , 836 F.3d at 1269 n.4 (quotation omitted); accord Dyroff , 934 F.3d at 1099. Even if the Facebook algorithm and system of third-party feedback and “social rewards” encouraged the posting of content, nothing about the platform’s design contributed to what made those posts illegal or actionable. See Kimzey , 836 F.3d at 1269 n.4; accord Roommates.com , 521 F.3d at 1172. “Meta’sPage 19 ‘solicitation’ and ‘assistance’ efforts” were “ allegedly used for unlawful purposes, but that does not result from Meta’s efforts.” Calise , 103 F.4th at 745; accord Kimzey , 836 F.3d at 1269 n.4. And, unlike Lemmon , the alleged 'social rewards” that Plaintiffs argue motivated offending behavior or content resulted from user engagement. Cf. 995 F.3d at 1091 – 92.

At this step, too, the algorithm does not change our calculus or count as a material contribution. See Dyroff , 934 F.3d at 1099. Facebook algorithmically boosted content that users engaged with more. But in this way, Facebook’s 2009 algorithm is like the algorithmic 'matching' we have determined does not affect coverage by Section 230. See Carafano , 339 F.3d at 1123 – 25.

If a personalized algorithm that used machine learning to pick up on users' preferences and attributes was not a material contribution in Dyroff , then neither is the more rudimentary algorithm here. See supra p. 7; cf. Dyroff , 934 F.3d at 1095; Gonzalez , 2 F.4th at 894 – 95. Facebook’s 2009 algorithm “ helped facilitate . . . user-touser communication, ” but did not materially contribute to or develop the content of the communications itself. See Dyroff , 934 F.3d at 1099.

Accordingly, each of Plaintiffs’ claims fail the third step of our Barnes test. Under our precedent, Facebook’s algorithm did not materially contribute to the allegedly unlawful posts here — nor did it contribute to the alleged unlawfulness of that content itself.⁵

Page 20Meta created Facebook, an interactive computer service. That service recommended content to users to maximize engagement. Under our precedent, that is publishing conduct that does not materially contribute to or develop the underlying content.

Plaintiffs believe that Facebook’s design, coupled with the darker elements of human nature, caused real-world harm. But Section 230, as we have interpreted it, bars their claims, and we cannot hold Meta 'responsible for the unfortunate realities of human nature.” Est. of Bride , 112 F.4th at 1181.

AFFIRMED.

BERZON, Circuit Judge, with whom W. FLETCHER, Circuit Judge, joins, concurring:

I concur in the majority opinion in full. We are bound by Ninth Circuit precedent addressing the scope of section 230 immunity, which requires the conclusion we reach here: The plaintiffs claims’ challenging Facebook’s algorithmic design seek to treat Meta 'as the publisher or speaker of any information provided by another information content provider,” 47 U.S. § 230(c)(1), so section 230 provides Meta Page 21with immunity. See Barnes v. Yahoo!, Inc. , 570 F.3d 1096, 1100 – 01 (9th Cir. 2009).

I nevertheless continue to think that this Court’s precedent has unduly expanded the scope of section 230 immunity. For the reasons persuasively outlined by Judge Katzmann in his partial dissent in Force v. Facebook , 934 F.3d 53 (2d Cir. 2019), cert. denied , 140 S. Ct. 2761 (2020), and as I stated in my concurrence in Gonzalez v. Google LLC , 2 F.4th 871 (9th Cir. 2021), vacated , 598 U.S. 617 (2023), “if not bound by Circuit precedent I would hold that the term ‘publisher’ under section 230 reaches only traditional activities of publication and distribution — such as deciding whether to publish, withdraw, or alter content — and does not include activities that promote or recommend content or connect content users to each other,” Gonzalez , 2 F.4th at 913 (Berzon, J., concurring).

The Supreme Court declined to address the application of section 230 to the claims presented in Gonzalez v. Google . See 598 U.S. at 622. But the Court later held in Moody v. NetChoice, LLC , 603 U.S. 707 (2024), that an online platform’s decisions about how to order, organize, or display thirdparty content are the platform’s “expressive choices.” Id. at 740. Some courts have taken Moody as an invitation to reconsider the scope of section 230 immunity — rightfully so. See infra p. 25. Our case law has nevertheless continued to reinforce a reading of the term “publisher” in section 230 that reaches far outside the scope of traditional publication. Although I agree with several of the points raised by Judge R. Nelson in his concurrence, I write separately to underscore how this Court’s precedent has stretched the term “publisher” in section 230 past the point of recognition, and to address whether Moody may offer another path forward.

We held in Barnes held that 'publication involves reviewing, editing, and deciding whether to publish or to withdraw from publication thirdparty content.” 570 F.3d at 1102. I do not disagree with Barnes on this point. 'A website's decisions to moderate content, restrict users, or allow third parties full freedom to post content and interact with each other all therefore fall squarely within the actions of a publisher shielded from liability under section 230.” Gonzalez , 2 F.4th at 914 (Berzon, J., concurring).

As I explained in my concurrence to Gonzalez v. Google , however, the line of cases after Barnes took an astonishingly expansive view of what 'publication involves.' Barnes , 570 F.3d at 1102. For example, Dryoff v. Ultimate Software Grp., Inc. , 934 F.3d 1093 (9th Cir. 2019), held that, '[b]y recommending user groups and sending email notifications, [the website defendant] was acting as a publisher of others’ content. These functions — recommendations and notifications — are tools meant to facilitate the communication and content of others. They are not content in and of themselves.” 934 F.3d at 1098. Since Gonzalez v. Google , this Court has continued to grant section 230 immunity to internet service providers for conduct that bears little resemblance to traditional activities of publication, including targeted recommendations and the affirmative promotion of interactions among users.

Consider Calise v. Meta Platforms, Inc. , 103 F.4th 732 (9th Cir. 2024). Calise involved claims brought by Facebook users who alleged that they were harmed by fraudulent thirdparty advertisements posted to the website. 103 F.4th at 736. The Calise plaintiffs claimed that scammers exploited Meta’s targeted advertising practices to “deliberately target Page 23Meta's more vulnerable users.' Id. The plaintiffs alleged that, as a predicate step to promoting advertisements to users, Meta solicited advertisement sales from disreputable advertisers, even where Meta knew the advertisers violated Meta’s policies. Id. at 737.

Calise held that Meta’s asserted duty to vet third -party advertisements implicated “quintessential publishing conduct.' Id. at 744. It was not enough, Calise reasoned, that the plaintiffs pointed to Meta’s manipulation of third -party advertisements “to skew fraudulent content (because it could make more money) by targeting certain advertisers.' Id. 746. Instead, Calise concluded, the solicitation and selection of certain advertisers did not 'materially contribute' to the unlawfulness of the advertisements and so was protected by section 230. Id. ; see also Dryoff , 934 F.3d at 1099; Fair Hous. Council of San Fernando Valley v. Roommates.Com, LLC , 521 F.3d 1157, 1168 (9th Cir. 2008) (en banc) (“[A] website helps to develop unlawful content, and thus falls within the exception to section 230, if it contributes materially to the alleged illegality of the conduct.”).

It did not matter that the advertisements in Calise were commercial activity. Meta sold the advertising space, and so essentially acted as a salesperson or marketer. But under this Court’s precedent, Calise concluded Meta’s sales activities are captured by the term “publisher.” 0F 1

Again relying on Dryoff , this Court in Doe v. Grindr Inc. , 128 F.4th 1148 (9th Cir. 2025), cert. denied , 146 S. Ct. 319 (2025), squarely addressed recommendations of third-party Page 24content by an internet service provider to its users. Grindr involved claims brought by John Doe, a minor, against a dating app. 128 F.4th at 1152. “[Grindr] matches users based on proximity, and it allows matched users to send direct messages to each other.' Id. Doe alleged that the app's matching function — that is, its recommendation of users to other geographically proximate users — was defective, and that Grindr was negligent for breaching its duty to avoid matching him with adult men who would abuse him. Id.

Grindr concluded that the app's matching function is 'meant to facilitate the communication and content of others” and is therefore “content neutral” on its own. Id. at 1153 (quoting Dryoff , 934 F.3d at 1098, 1100). And Grindr held that Doe's state law claims 'necessarily implicate[d] Grindr’s role as a publisher of third -party content.” Id. That is because, Grindr reasoned, any alleged duty to “suppress matches and communications between adults and children” would “require Grindr to monitor third -party con tent,” and to thereby act as a publisher. Id.

Grindr ’s conclusion that section 230 barred Doe's claims because they challenged the app’s recommendation function is consistent with Dryoff . Yet I struggle to see how “facilitat[ing] the communication and content” of app users on the basis of geographical proximity has any relationship to traditional publishing activities. Id. (quoting Dryoff , 934 F.3d at 1098). Nevertheless, the Ninth Circuit has plainly “extend[ed] immunity under section 230 to targeted recommendations of content and connections.” Gonzalez , 2 F.4th at 917 (Berzon, J., concurring).

Whether soliciting scam advertisements or matching adults with minors on dating apps, this Court’s broad reading of section 230 of the Communications Decency Act permitsPage 25 internet service providers affirmatively to behave indecently, potentially — as alleged in Dryoff and Grindr — causing serious harm to vulnerable people. In my view, the conduct challenged in these cases stretches far beyond the traditional activities of publication contemplated by section 230.

B.

Another approach is possible. Outside the Ninth Circuit, other courts have — as Judge R. Nelson notes in his concurrence, R. Nelson Concurrence at 33 — taken first steps to narrow section 230’s broad sweep.

Most notably, Anderson v. TikTok, Inc. , 116 F.4th 180 (3d Cir. 2024), held that a social media company’s algorithm was the company’s “own expressive activity,” and so section 230 did not bar the plaintiff’s claims. 116 F.4th at 184. To do so, Anderson relied on the Supreme Court’s recent First Amendment decision in Moody v. NetChoice , which addressed a challenge to Florida and Texas laws limiting the ways in which social media platforms can censor, suppress, or otherwise moderate certain user-generated content. 603 U.S. at 720 – 21.

To address whether the state laws implicated the First Amendment, Moody observed that 'the First Amendment offers protection when an entity engaging in expressive activity, including compiling and curating others’ speech, is directed to accommodate messages it would prefer to exclude.” Id. at 731. Even if “achieved through the use of algorithms,” id. at 734, Moody concluded, '[w]hen the platforms use their Standards and Guidelines to decide which third-party content [their] feeds will display, or how the display will be ordered and organized, they are making expressive choices,” id. at 740.

Page 26Drawing from Moody , Anderson reasoned that, if algorithmic recommendations are first-party speech under the First Amendment, those recommendations are first-party speech under section 230. Anderson , 116 F.4th at 184. I agree.

At issue in Anderson was the recommendation algorithm used by the TikTok social media platform. Id. at 182. TikTok “curates and recommends a tailored compilation of videos” for each user based on “a variety of factors,” including the user's demographics and past online interactions. Id. The TikTok algorithm recommended to a ten-year-old child “Blackout Challenge” videos, “which encourage[] viewers to record themselves engaging in acts of selfasphyxiation.” Id. at 181. Attempting the “Blackout Challenge,” t he child unintentionally hanged herself, and her mother brought suit. Id. Emphasizing that section 230 immunizes “only information ‘provided by another,’” id. at 184, (quoting 47 U.S.C. § 230(c)(1)), and because the plaintiff sought to hold TikTok liable for its own information -that is, TikTok's algorithm, its “expressive product” under Moody — Anderson held that section 230 did not bar the plaintiff's claims.

I believe Anderson was correct to interpret TikTok's algorithm as the platform’s first -party speech, and “to chart a new course in response to Moody .' R. Nelson Concurrence at 33. This Court should follow suit.

I agree with Judge R. Nelson’s concurrence to the extent that he contends Ninth Circuit precedent has expanded the definition of “publisher” “to include what traditionally was never considered publishing conduct.' Id. at 30. But I disagree with his suggestion that “advanced [that is, Page 27individualized] algorithms are more clearly outside the scope of traditional third-party publishing conduct that Section 230 protects” than algorithms which make recommendations based on generic characteristics of content. 1F 2 Id. at 34.

In my view, generic, content-based algorithms do not bear any more resemblance to traditional publication than their “advanced”— individualized — counterparts. Even where, as in the case of Meta’s activities in Myanmar, an algorithm allegedly promoted third-party content partly on the basis of cumulative user engagement, the assessment of user viewing habits to recommend content remains “more analogous to the actions of a direct marketer, matchmaker, or recruiter than to those of a publisher.” Gonzalez , 2 F.4th at 914 (Berzon, J., concurring). Such recommendations carry an inherent message that is attributable to the internet service provider making the recommendation. Whether generated by an “advanced,” “personalized recommendation” algorithm, R. Nelson Concurrence at 34 , an 'engagement -maximizing' algorithm, id. , or by hand, I understand suggestions and recommendations to be distinct messages presented by social media companies. Those messages Page 28cannot be attributed to any third parties and so are not protected by section 230.

If not bound by Circuit precedent, I would hold that section 230 does not bar the claims raised against Meta in this case because 'websites' use of machine -generated algorithms to recommend content and contacts are not within the publishing role immunized under section 230.” Gonzalez , 2 F.4th at 917 (Berzon, J., concurring). Both Moody v. NetChoice and the Anderson case cast this Court’s unfettered expansion of section 230 into even more doubt than at the time of Gonzalez v. Google . Taken together, Moody and Anderson persuasively suggest that algorithmic recommendations are an internet service provider's first -party speech for purposes of assessing section 230 immunity. And those two cases demonstrate that there is now both an intercircuit conflict and strong tension with a recent Supreme Court case. Moody v. NetChoice and Anderson only underscore the necessity for en banc review of this Court’s precedent addressing section 230. I again— even more emphatically on this go round — urge this Court to reconsider en banc our precedent extending section 230 immunity to recommendation of content and connections to users.

Page 29R. NELSON, Circuit Judge, concurring:

Under our precedent, Section 230 bars Plaintiffs’ claims. I write separately to address two issues. First, we have overread Section 230, straying from the original public meaning of the statutory text and creating an all-purpose liability shield for internet platforms. Second, state choice-of-law rules can never direct application of a rule of decision contrary to federal law. Because Section 230 is supreme to state law, it controls even if California’s choice -of-law rules point to a conflicting foreign rule.

I

First, Section 230. There is, of course, a 'growing chorus of voices calling for a more limited reading of the scope of section 230 immunity.” Gonzalez v. Google LLC , 2 F.4th 871, 913 (9th Cir. 2021) (Berzon, J., concurring), vacated , 598 U.S. 617 (2023). I agree with that sentiment, but perhaps for different reasons. See Calise v. Meta Platforms, Inc. , 103 F.4th 732, 746 – 48 (9th Cir. 2024) (R. Nelson, J., concurring).

Our precedent has ignored the original public meaning of the term 'publisher' in Section 230's Good Samaritan provision. See 47 U.S.C. § 230(c)(1). That term should be interpreted consistent with the traditional publisherdistributor distinction developed over centuries regarding certain speech torts. See Malwarebytes, Inc. v. Enigma Software Grp. USA, LLC , 141 S. Ct. 13 (2020) (Thomas, J., statement respecting denial of certiorari); Anderson v. TikTok, Inc. , 116 F.4th 180, 187 – 91 (3d Cir. 2024) (Matey, J., concurring in part and dissenting in part); see also Free Speech Coal., Inc. v. Paxton , 606 U.S. 461, 476 n.4 (2025). Although Section 230 immunizes platforms from liability as 'publishers' of third -party content, we bar claims evenPage 30 tangentially related to online speech. See, e.g. , Est. of Bride v. YOLO Techs., Inc. , 112 F.4th 1168, 1182 (9th Cir. 2024). In doing so, we have expanded “publisher” to include what traditionally was never considered publishing conduct.

For example, as the opinion explains, Plaintiffs’ product liability claims are barred by Section 230 under our precedent. See Maj. 13 – 16. But how does a product liability claim hinge on treating the provider of an interactive computer service as the publisher of third-party content? See Malwarebytes , 141 S. Ct. at 17 (Thomas, J., statement respecting denial of certiorari).

Sure, we might analogize a party’s duty (for example, in designing products) to that of a publisher. But if the claim does not actually require us to treat a defendant as the publisher of thirdparty content, Section 230’s plain text is inapt. Even if a product liability claim might look like publishing third-party content from afar, the duty underlying such a claim is not ' identical to publishing or speaking.' Barnes v. Yahoo!, Inc. , 570 F.3d 1096, 1107 (9th Cir. 2009) (emphasis added). Some torts — like negligent undertaking — could be functional duplicates of publishingbased speech torts. See id. But the connection here seems too attenuated.

None of this is to say that Plaintiffs' product liability claims hold any water. Just the opposite — they likely fail under California law. Not because they remind us of another sort of claim that would demand we treat the provider of an interactive computer service as the publisher of third-party content. But because substantive product liability law likely stops them cold. After all, I am skeptical that Facebook (like other social media platforms) is a “product.” See Restatement (Third) of Torts: Products liability § 19(a)Page 31 (1998) (“A product is tangible personal property distributed commercially for use or consumption.” (emphasis added)); accord Pierson v. Sharp Mem’l Hosp., Inc. , 216 Cal. App. 3d 340, 345 (1989). Plaintiffs’ claims therefore likely fail on the merits, and we need not stretch Section 230 to reject them.

I also agree that we have strayed in immunizing algorithms. See Gonzalez , 2 F.4th at 913 – 18 (Berzon, J., concurring). We have taken a strong position that algorithms are almost always third-party publishing activity, and rarely a platform’s own content. E.g. , Dyroff v. Ultimate Software Grp., Inc. , 934 F.3d 1093, 1098 (9th Cir. 2019).

But Section 230's protection of third -party publishing conduct should reach only “traditional” activities of publication and distribution — not every modern activity that bears some remote resemblance to it. See Gonzalez , 2 F.4th at 913 (Berzon, J., concurring); Force v. Facebook, Inc. , 934 F.3d 53, 81 (2d Cir. 2019) (Katzmann, J., concurring in part and dissenting in part). This is, after all, what publishing meant when Section 230 was enacted. See Gonzalez , 2 F.4th at 914 – 15. And such a definition better adheres to the history and tradition of common-law protection for publishers that Section 230 codified. See Malwarebytes , 141 S. Ct. at 15 (Thomas, J., statement respecting the denial of certiorari).

Modern recommendation algorithms are opaque, esoteric, and -particularly when artificial intelligence enters the fray — incomprehensible, sometimes even to their own designers. See Moody v. NetChoice, LLC , 603 U.S. 707, 795 – 96 (2024) (Alito, J., concurring in the judgment). Much of the matchmaking and network creation that modern algorithms engage in does not fit within any fair definition Page 32of publishing conduct. See Force , 934 F.3d at 76 – 77, 82 – 83, 85 (Katzmann, J., concurring in part and dissenting in part).

And even if algorithms are publishing conduct, the Supreme Court has suggested that they are the publishing of the platforms' own content. See Moody , 603 U.S. at 707, 731, 734. Moody compares the algorithmic compilation of third-party speech to newspapers and other traditional publishing activity. See id. at 732 -33, 738. 'An entity ‘exercis[ing] editorial discretion in the selection and presentation’ of content is ‘engage[d] in speech activity.’” Id. at 731 (quoting Ark. Ed. Television Comm’n v. Forbes , 523 U.S. 666, 674 (1998)). '[T]hat is as true when the content comes from third parties as when it does not.” Id. And it is also true when the prioritization of content is 'achieved through the use of algorithms' that provide “individualized . . . recommendations.” Id. at 734.

As the Supreme Court said, personalized recommendation algorithms constitute a platforms' own 'distinctive expressive product.' Id. at 731. If an algorithm is a distinctive expressive product, then it makes sense that it is a platform’s own content, rather than the content of third parties. And if it is a platform’s own content, rather than the content of third parties, it is not immunized by Section 230. See Barnes , 570 F.3d at 1101 (“By its terms, then, section (c)(1) only ensures that in certain cases an internet service provider will not be ‘treated’ as the ‘publisher or speaker’ of third-party content . . . .' (emphasis added)); Force , 934 F.3d at 80 – 81 (Katzmann, J., concurring in part and dissenting in part); see also Doe ex rel. Roe v. Snap, Inc. , 88 F.4th 1069, 1071 (5th Cir. 2023) (Elrod, J., dissenting from denial of rehearing en banc). So it is unclear why algorithms should be immunized under Section 230 after Moody .

Page 33True, whether algorithmic compilation and content delivery is First Amendment speech is distinct from whether it is first-party content for Section 230 purposes. We have never squared that circle, and neither has the Supreme Court. Cf. NetChoice, LLC v. Paxton , 142 S. Ct. 1715, 1717 & n.2 (2022) (Alito, J., dissenting from grant of application to vacate stay). And what counts as algorithmic speech is not yet fully developed, either.¹ As Justice Barrett explained, the majority did not resolve whether purely engagementmaximizing algorithms were expressive. See Moody , 603 U.S. at 746 – 47 (Barrett, J., concurring).

At any rate, Moody 's thrust is clear: Personalized recommendation algorithms can constitute a platform’s own message, even if third-party content is in the mix. Accordingly, even if non-traditional publishing conduct (such as the use of recommendation algorithms) is publishing activity, the result is the same: Section 230 does not apply. When a plaintiff’s cause of action seeks to hold a platform liable for a recommendation algorithm itself, Section 230 should provide no immunity.

Some courts have already begun to chart a new course in response to Moody , deciding that recommendation algorithms are platforms’ own content and declining to apply Section 230. See Anderson , 116 F.4th at 184. We have already tied ourselves to the mast, however, and while Moody cast some doubt on our precedent about recommendation algorithms, our precedent has not been effectively overruled. See Miller v. Gammie , 335 F.3d 889, 893 (9th Cir. 2003). So it will fall to the en banc court toPage 34 'plumb the depths' of Section 230 immunity once again, especially when it comes to algorithmic content delivery.² Fair Hous. Council of San Fernando Valley v. Roommates.com, LLC , 521 F.3d 1157, 1161 (9th Cir. 2008) (en banc).

But this is not the case to correct our precedent. The chief problem with our Section 230 precedent — particularly Dyroff — involves personalized algorithms custom-fit to individual users based on their engagement habits and content preferences. See 934 F.3d at 1095. Such advanced algorithms are more clearly outside the scope of traditional third-party publishing conduct that Section 230 protects. See Gonzalez , 2 F.4th at 914 (Berzon, J., concurring). And it was that kind of advanced, personalized algorithm that Moody considered. See 603 U.S. at 734 – 35. So we should revisit our doctrine where a personalized recommendation algorithm is at issue.

We might also await a vehicle in which a platform manipulates its algorithmic content delivery to boost or deboost content based not on what users are likely to engage with, but based on value-laden editorial choices made by the platform itself. Such a case would avoid Justice Barrett’s doubts about whether engagement-maximizing algorithms are a message of their own. See Moody , 603 U.S. at 746 (Barrett, J., concurring). And it should not take long for such a case to arise. Major internet platforms appear all too happy Page 35to engage in this kind of direct interference of information. See, e.g. , Murthy v. Missouri , 603 U.S. 43, 50 – 53, 62 – 68 (2024); id. at 81 – 91 (Alito, J., dissenting); Moody , 603 U.S. at 719, 742; Children’s Health Def. v. Meta Platforms, Inc. , 112 F.4th 742, 751 – 53 (9th Cir. 2024).

Some algorithms are both personalized and more clearly present a platform’s own ideological message. TikTok, for example, uses a highly advanced personalized recommendation algorithm that responds not only to a user’s engagement with content but also other online actions and user metadata. See Anderson , 116 F.4th at 181 – 82. And its algorithm was developed and maintained by a foreign adversary with a particular message to send to Americans. Cf. TikTok Inc. v. Garland , 604 U.S. 56, 78 – 79 (2025). There is strong evidence that this algorithm is an antiAmerican propaganda tool to push some political narratives while suppressing others. E.g. , David Leonhardt, TikTok’s Pro-China Tilt , N.Y. Times (Apr. 24, 2025).³

The algorithm here, however, was based on the more primitive technology available in 2009. See Maj. 7. The 2009 algorithm was not tailored to users. Nor is it alleged to learn from individual users’ engagement habits. And it did not push or hide narratives based on the specific content shared by users. Like the analysis of algorithmic speech in the First Amendment context, these facts matter. See Moody , 603 U.S. at 747 (Barrett, J., concurring); see also id. at 749 (Jackson, J., concurring in part and in the judgment); id. at 769 (Alito, J., concurring in the judgment). More modern algorithms would be a better vehicle than the antediluvian, decade-old algorithm here. See Force , 934Page 36 F.3d at 58 (discussing Facebook’s more recent personalized recommendation algorithm).

The opinion discusses California’s choice -of-law analysis before determining whether to apply Section 230. See Maj. 8 – 12. We did not need to do so. After all, Section 230 is a matter of federal law. And California law cannot contradict federal law — either directly or indirectly.

Start with first principles. By constitutional design, state governments have many advantages over the federal government, including broader legislative authority. See The Federalist Nos. 45, 46 (James Madison). But the Constitution gives the federal government some legislative advantages. Article I expressly delegates certain powers to the federal government. And where legislation is validly enacted pursuant a delegated power, federal law is supreme to the law of the states, “any Thing in the Constitu tion or Laws of any State to the Contrary notwithstanding.” U.S. Const., art. VI, cl. 2. '[A]ny Thing' necessarily includes state choice-of-law rules. Id.

The Supremacy Clause was controversial during the ratification debates, but its necessity to our constitutional design of delegated powers is obvious. See The Federalist No. 33 (Alexander Hamilton). Without it, federal legislative power 'would have been reduced to the same impotent condition' that the Constitution sought to remedy, The Federalist No. 44, at 222 (James Madison) (Jim Miller ed., 2014), and “would amount to nothing,” The Federalist No. 33, at 151 (Alexander Hamilton) (Jim Miller ed., 2014). Federal supremacy is the “necessary and unavoidable implication” of delegating enumerated powers to the federal government. Id. at 149 – 50.

Page 37Accordingly, if federal law reflects the lawmaking power properly delegated to Congress, it definitionally cannot be repugnant to states' authority. See The Federalist No. 45 (James Madison); see also The Federalist No. 32 (Alexander Hamilton). Just the opposite: without giving due effect to the Supremacy Clause, the Constitution would create a government 'founded on an inversion of the fundamental principles of all government”—a “monster” of disorgani zed and unworkable governance. The Federalist No. 44, at 222 (James Madison) (Jim Miller ed., 2014).

Plaintiffs’ proposed rule of decision and federal law lead to clashing results. See Armstrong v. Exceptional Child Ctr., Inc. , 575 U.S. 320, 325 (2015). So the Supremacy Clause “instructs courts what to do” and commands application of federal law. Id. ; accord DelCostello v. Int’l Bhd. Teamsters , 462 U.S. 151, 159 n.13 (1983).

Myanmar’s rules of decision would apply here only by operation of California law. Beech Aircraft Corp. v. Superior Ct. , 61 Cal. App. 3d 501, 522 (1976). And the Supremacy Clause preempts California from choosing any rule of decision contrary to federal law. This would impermissibly “give effect to state laws that conflict with federal laws.” Armstrong , 575 U.S. at 324.

If California cannot directly override federal law (e.g., through legislation stating that providers of certain interactive computer services can be treated as publishers or speakers of third-party content), cf. Williams v. Reed , 604 U.S. 168, 182 (2025) (Thomas, J., dissenting), it cannot do so indirectly (e.g., by adopting a choice-of-law test that would adopt a foreign jurisdiction's rule of decision that would do the same thing). Indeed, we have already said that when general choice-of-law doctrine might lead to a decision Page 38that would conflict with federal law, “general ‘choice -of-law doctrines . . . have no applicability.’” Murphy v. DirecTV, Inc. , 724 F.3d 1218, 1228 (9th Cir. 2013) (quoting Roadway Package Sys., Inc. v. Kayser , 257 F.3d 287, 293 – 94 (3d. Cir. 2001), abrogated on other grounds by Hall St. Assocs. v. Mattel, Inc. , 552 U.S. 576 (2008)). As we explained, “the relevant rule is supplied by the Constitution itself: a valid federal law preempts any state law purporting to regulate the same issue.' Id. (quotation and quotation marks omitted).

Our past decision in Bassidji v. Goe supports this conclusion. 413 F.3d 928 (9th Cir. 2005). Bassidji assumed that federal law (the International Emergency Economic Powers Act by operation of an Executive Order) must apply 'through the Supremacy Clause.' Id. at 933. The California choice-of-law analysis was consulted, but not a workaround to federal law. Instead, the choice-of-law question was relevant to the “separate question” of the enforceability of admittedly illegal contracts under substantive contract law. Id. at 936. In other words, Bassidji assumed that federal law applied, rendering the contracts-at-suit illegal. The question that remained was the implication of that illegality. Here, no such “separate question” needs to be answered by resort to any choice-of-law analysis.⁴ Id.

At bottom, we reached the correct result: Section 230 applies, despite Plaintiffs' choice -of-law arguments. But Page 39California’s choice -of-law rules need not be consulted in the first place. After all, the relevant rule of decision 'is supplied by the Constitution itself' by operation of the Supremacy Clause. Murphy , 724 F.3d at 1228 (quotation omitted).

Allowing states' choice -of-law rules to backdoor the Supremacy Clause would 'provide a roadmap for States wishing to circumvent' federal law. See Haywood v. Drown , 556 U.S. 729, 742 n.9 (2009). That would make federal lawmaking power a nullity, see The Federalist No. 44 (James Madison), and would allow state governments to “sap the foundations of the Union” and upset the basic design of our constitutional republic, The Federalist No. 33, at 150 (Alexander Hamilton) (Jim Miller ed., 2014). Simply put, there is no choice-of-law exception to the Supremacy Clause. Cf. 28 U.S.C. § 1652.