Yes, Section 230 Should Apply Equally To Algorithmic Recommendations

from the it-won’t-do-what-you-think-if-you-remove-it dept

If you’ve spent any time in my Section 230 myth-debunking guide, you know that most bad takes on the law come from people who haven’t read it. But lately I keep running into a different kind of bad take—one that often comes from people who have read the law, understand the basics passably well, and still say: “Sure, keep 230 as is, but carve out algorithmically recommended content.”

Unlike the usual nonsense, this one is often (though not always) offered in good faith. That makes it worth engaging with seriously.

It’s still wrong.

Let’s start with the basics: as we’ve described at great length, the real benefits of Section 230 are its procedural protections, which make it so that vexatious cases get tossed out at the earliest (i.e., cheapest) stage. That makes it possible for sites that host third party content to do so in a way that they won’t get sued out of existence any time anyone has a complaint about someone else’s content being on the site. This important distinction gets lost in almost every 230 debate, but it’s important. Because if the lawsuits that removing 230 protections would enable would still eventually win on First Amendment grounds, the only thing you’re doing in removing 230 protections is making lawsuits impossibly expensive for individuals and smaller providers, without doing any real damage to large companies, who can survive those lawsuits easily.

And that takes us to the key point: removing Section 230 for algorithmic recommendations would only lead to vexatious lawsuits that will fail.

But what about [specific bad thing]?

Before diving into the legal analysis, let’s engage with the strongest version of this argument. Proponents of carving out algorithmic recommendations typically aren’t imagining ordinary defamation suits. They’re worried about something more specific: cases where an algorithm itself arguably causes harm through its recommendation patterns—radicalization pipelines, engagement-driven amplification of dangerous content, recommendation systems that push vulnerable users toward self-harm.

The theory goes something like this: maybe the underlying content is protected speech, but the act of recommending it—especially when the algorithm was designed to maximize engagement and the company knew this could cause harm—should create liability, usually as some sort of “products liability” type complaint.

It’s a more sophisticated argument than “platforms are publishers.” But it still fails, for reasons I’ll explain below. The short version: a recommendation is an opinion, opinions are protected speech, and the First Amendment doesn’t carve out “opinions expressed via algorithm” as a special category.

A short history of algorithmic feeds

To understand why removing 230 from algorithmic recommendations would be such a mistake, it helps to remember the apparently forgotten history of how we got here. In the pre-social media 2000s, “information overload” was the panic of the moment. Much of the discussion centered on the “new” technology of RSS feeds, and there were plenty of articles decrying too much information flooding into our feed readers. People weren’t worried about algorithms—they were desperate for them. Articles breathlessly anticipated magical new filtering systems that might finally surface what you actually wanted to see.

The most prominent example was Netflix, back when it was still shipping DVDs. Because there were so many movies you could rent, Netflix built one of the first truly useful recommendation algorithms—one that would take your rental history and suggest things you might like. The entire internet now looks like that, but in the mid-2000s, this was revolutionary.

Netflix’s approach was so novel that they famously offered $1 million to anyone who could improve their algorithm by 10%. We followed that contest for years as it twisted and turned until a winner was finally announced in 2009. Incredibly, Netflix never actually implemented the winning algorithm—but the broader lesson was clear: recommendation algorithms were valuable, and people wanted them.

As social media grew, the “information overload” panic of the blog+RSS era faded, precisely because platforms added recommendation algorithms to surface content users were most likely to enjoy. The algorithms weren’t imposed on users against their will—they were the answer to users’ prayers.

Public opinion only seemed to shift on “algorithms” after Donald Trump was elected in 2016. Many people wanted something to blame, and “social media algorithms” was a convenient excuse.

Algorithmic feeds: good or bad?

Many people claim they just want a chronological feed, but studies consistently show the vast majority of people prefer algorithmic recommendations, because they surface more of what users actually want, compared to chronological feeds.

That said, it’s not as simple as “algorithms good.” There’s evidence that algorithms optimized purely for engagement can push emotionally charged political content that users don’t actually want (something Elon Musk might take notice of). But there’s also evidence that chronological feeds expose users to more untrustworthy content, because algorithms often filter out garbage.

So, algorithms can be good or bad depending on what they’re optimized for and who controls them. That’s the real question: will any given regulatory approach give more power to users, to companies, or to the government?

Keep that frame in mind. Because removing 230 protections for algorithmic recommendations shifts power away from users and toward incumbents and litigants.

The First Amendment still exists

As mentioned up top, the real role of Section 230 is providing a procedural benefit to get vexatious lawsuits tossed well before (and at much lower cost) they would get tossed anyway, under the First Amendment. With Section 230, you can get a case dismissed for somewhere in the range of $50k to $100k (maybe up to $250k with appeals and such). If you have to rely on the First Amendment, it’s up in the millions of dollars (probably $5 to $10 million).

And, the crux of this is that any online service sued over an algorithmic recommendation, even for something horrible, would almost certainly win on First Amendment grounds.

Because here’s the key point: a recommendation feed is a website’s opinion of what they think you want to see. And an opinion is protected speech. Even if you think it’s a bad or dangerous opinion. One thing that the US has been pretty clear on is that opinions are protected speech.

Saying that an internet service can be held liable for giving its opinion on “what we think you’d like to see” would be earth shatteringly problematic. As partly discussed above, the modern internet today relies heavily on algorithms recommending stuff, giving opinions. Every search result is just that, an opinion.

This is why the “algorithms are different” argument fails. Yes, there’s a computer involved. Yes, the recommendation emerges from machine learning rather than a human editor’s conscious decision. But the output is still an expression of judgment: “Based on what we know, we think you’ll want to see this.” That’s an opinion. The First Amendment doesn’t distinguish between opinions formed by editorial meetings and opinions formed by trained models.

In the earlier internet era, there were companies that sued Google because they didn’t like how their own sites appeared (or didn’t appear) in Google search results. The E-Ventures v. Google case here is instructive. Google determined that E-Venture’s “SEO” techniques were spammy, and de-indexed all its sites. E-Ventures sued. Google (rightly) raised a 230 defense which (surprisingly!) a court rejected.

But the case went on longer, and after lots more money on lawyers was spent, Google did prevail on First Amendment grounds.

This is exactly what we’re discussing here. Google search ranking is an algorithmic recommendation engine, and in this one case a court (initially) rejected a 230 defense, causing everyone to spend more money… to get to the same basic result in the long run. The First Amendment protects a website using algorithms to express an opinion over what it thinks you’ll want… or not want.

Who has agency?

This brings us back to the steelman argument I mentioned above: what about cases where an algorithm recommends something genuinely dangerous?

Our legal system has a clear answer, and it’s grounded in agency. A recommendation feed is not hypnotic. If an algorithm surfaces content suggesting you do something illegal or dangerous, you still have to make the choice to do the illegal or dangerous thing. The algorithm doesn’t control you. You have agency.

But there’s a stronger legal foundation here too. Courts have consistently found that recommending something dangerous is still protected by the First Amendment, particularly when the recommender lacks specific knowledge that what they’re recommending is harmful.

The Winter v. GP Putnam’s Sons case is instructive here. The publisher of a mushroom encyclopedia included recommendations to eat mushrooms that turned out to be poisonous—very dangerous! But the court found the publisher wasn’t liable because they didn’t have specific knowledge of the dangerous recommendation. And crucially, the court noted that the “gentle tug of the First Amendment” would block any “duty of care” that would require publishers to verify the safety of everything they publish:

The plaintiffs urge this court that the publisher had a duty to investigate the accuracy of The Encyclopedia of Mushrooms’ contents. We conclude that the defendants have no duty to investigate the accuracy of the contents of the books it publishes. A publisher may of course assume such a burden, but there is nothing inherent in the role of publisher or the surrounding legal doctrines to suggest that such a duty should be imposed on publishers. Indeed the cases uniformly refuse to impose such a duty. Were we tempted to create this duty, the gentle tug of the First Amendment and the values embodied therein would remind us of the social costs.

Now, I should acknowledge that Winter was a products liability case involving a physical book, not a defamation or tortious speech case involving an algorithm, but almost all of the current cases challenging social media are self-styled as product liability cases to try (usually without success) to avoid the First Amendment. And that’s all they would be regarding algorithms as well.

The underlying principle remains the same whether you call it a products liability case or one officially about speech: the First Amendment bars requirements that publishing intermediaries must “investigate” whether everything they distribute is accurate or safe. The reason is obvious—such liability would prevent all sorts of things from getting published in the first place, putting a massive damper on speech.

Apply that principle to algorithmic recommendations, and the answer is clear. If a book publisher can’t be required to verify that every mushroom recommendation is safe, a platform can’t be required to verify that every algorithmically surfaced piece of content won’t lead someone to harm.

The end result?

So what would it mean if we somehow “removed 230 from algorithmic recommendations”?

Practically, it means that if companies have to rely on the First Amendment to win these cases, only the biggest companies can afford to do so. The Googles and Metas of the world can absorb $5-10 million in litigation costs. For smaller companies, those costs are existential. They’d either exit the market entirely or become hyper-aggressive about blocking content at the first hint of legal threat—not because the content is harmful, but because they can’t afford to find out in court.

The end result would be that the First Amendment still protects algorithmic recommendations—but only for the very biggest companies that can afford to defend that speech in court.

That means less competition. Fewer services that can recommend content at all. More consolidation of power in the hands of incumbents who already dominate the market.

Remember the frame from earlier: does this give more power to users, companies, or the government? Removing 230 from algorithmic recommendations doesn’t empower users. It doesn’t make platforms more “responsible.” It just makes it vastly harder for anyone other than the giant platforms to exist while also giving more power to governments, like the one currently run by Donald Trump, to define what things an algorithm can, and cannot, recommend.

Rather than diminishing the power of billionaires and incumbents, this would massively entrench it. The people pushing for this carve-out often think they’re fighting Big Tech. In reality, they’re fighting to build Big Tech a new moat.

Filed Under: 1st amendment, algorithmic feeds, algorithmic recommendations, algorithms, feeds, free speech, opinion, section 230