There are two ways to sue a technology company for what its algorithm did to you. One of them hands the company a constitutional shield. Right now, nobody is choosing between them on purpose.

In March, a Los Angeles jury found that Meta and Google had negligently designed products that injured a young woman who had used them since she was a teenager, and put six million dollars on the finding. TikTok and Snap had settled days before trial rather than let the jury reach them.

The day before that verdict, a Santa Fe jury found that Meta had violated New Mexico’s consumer-protection law by misrepresenting how safe its platforms were for children, and imposed the maximum civil penalty available: $375 million. Jurors had heard from former employees and gone through hundreds of internal documents. Last week the same judge finished the second phase of that case and ordered Meta to pay $567 million into a treatment fund and to change how its products work for minors — time limits for underage users, likes hidden by default, visible risk warnings. He found that endless scrolling, autoplay, notifications, and content recommendations had been used deliberately to keep teenagers on the platform longer.

In May, the first federal test case in the national social-media injury litigation — one rural Kentucky school district — was settled by all four defendants before a jury was seated, reportedly for around twenty-seven million dollars combined. This month, in Oakland, the first of the state attorneys general cases against Meta goes to trial.

Three thousand miles east, a federal judge in Manhattan has been sitting since spring on a narrower-looking question. New York is the first state to forbid landlords from setting rents on the recommendation of software fed by other landlords’ private data. RealPage, the company whose software is the reason the law exists, has sued to have the ban struck down. It says the statute restricts its right to give advice — that is, its speech.

Put those together and the shape of the year is clear enough: the companies are losing. What is less visible, and will matter longer than any single verdict, is which of two arguments is doing the winning.

Two doors

Section 230 of the Communications Decency Act is the reason it has historically been hard to sue a platform at all. In plain terms: a website is generally not treated as the publisher of what its users post. Sue over a user’s video and the case can end before the platform has to defend anything about how it was built.

There are two ways past that.

The first door is product design. You do not sue over the video. You sue over the machine that delivered it — autoplay, infinite scroll, engagement ranking, the absence of workable age verification, defaults set to keep a child watching. This is product-design law, the same basic logic that governs a defective ladder, and whether it fits software cleanly is still being fought over. What it does not require is anyone calling the platform a speaker. It is the theory that won in Los Angeles, the theory behind the New Mexico judgment, and the theory the federal judge overseeing the national litigation has allowed to proceed.

That door got wider on August 10. The Ninth Circuit dismissed Meta’s and TikTok’s attempts to appeal early, on the ground that Section 230 “merely provides a defense to liability — not immunity from suit.” A defense is raised as the case goes on; an immunity gets you out now. The distinction is procedural on its face and practical in its effect: platforms have to litigate much further before an appeals court will hear their Section 230 arguments, which raises the cost of fighting and the pressure to settle.

The second door is speech. In Anderson v. TikTok, decided by the Third Circuit in August 2024, ten-year-old Nylah Anderson died after her For You feed served her videos of a dangerous viral “challenge” she had never searched for. The district court dismissed under Section 230. The Third Circuit reversed by treating the platform’s algorithmic curation as its own expressive activity: the platform’s speech, not merely a user’s post. Section 230 covers third-party content, so it does not cover this.

Same destination. Very different thing left behind in the record.

What the second door leaves behind

The Third Circuit did not have to take that route, and one judge on the panel did not. Judge Matey agreed the case should proceed but declined to rely on the speech rationale at all: on his reading of the statute’s history, Congress immunized platforms for content they passively host, not for content they actively push. A third route was available too — Ninth Circuit precedent holding that a site’s own design choices can fall outside Section 230, with nobody calling those choices speech.

The Supreme Court authority the panel leaned on was thinner than the weight put on it. The Court’s 2024 NetChoice decisions protected human-authored content-moderation guidelines, and Justice Kagan expressly reserved on adjacent questions, noting that different platform activities involve different degrees of editorial judgment. Read closely, the opinion points the other way. A system that ranks a video because it predicts you will keep watching is not obviously exercising editorial judgment at all — which would make it less likely to count as protected expression, not more.

And the holding became binding without anyone deciding it should. Rehearing by the full Third Circuit was denied on procedural grounds. TikTok, having lost, chose not to take it to the Supreme Court — a strategic choice, not neglect, but the effect is the same: no higher court has ever reviewed it. Other district courts have since echoed it.

Now watch what that leaves lying around.

Anderson is a plaintiff’s weapon. You cannot hide behind your users’ content when the recommendation is yours. But “our output is our protected expression” is also the strongest available answer to a government that wants to tell you what your system may do.

A court says TikTok owns the choice of what reaches a child’s screen. A plaintiff replies: then this was TikTok’s conduct, not somebody else’s post. Later, a company replies: because the choice is ours, the government cannot tell us how to make it.

The sentence does not change. The hand holding it does.

One caution, because this is where the argument is easiest to overstate. Being a speaker for purposes of Section 230 and being a speaker for purposes of the First Amendment are not the same finding; they come from different bodies of law with different tests. A court could accept Anderson‘s result and still reject RealPage’s constitutional defense. Rent recommendations look a great deal more like commercial pricing advice than like editorial judgment. And NetChoice is least helpful to the shield exactly where the curation is most machine-driven.

So the risk is not that the classification transfers automatically. It is that the vocabulary has been seeded. Once courts are describing algorithmic outputs as expressive, litigants reach for that description in rooms where the test is different, and some judges will find it persuasive. That is a slower and more plausible failure than a clean doctrinal transfer, and it is already happening.

The vocabulary in motion

New York’s rent statute is a process rule. It does not ask what RealPage knew or intended; it forbids landlords from adjusting rents on the recommendation of software drawing on pooled landlord data. There is no mental state in it anywhere.

RealPage’s answer is that the law is a content-, viewpoint-, and speaker-based restriction on its right to give advice. It sued the New York Attorney General shortly after settling with the Justice Department, seeking to block the law. Briefing finished months ago. The case remains undecided and the statute is in force.

Meanwhile the shield keeps accumulating. On August 5, the Second Circuit struck down a New York City law that had required food-delivery apps to hand customer data to restaurants, holding that the deferential rule for compelled factual disclosure did not apply — the disclosure concerned third parties rather than the platforms’ own services — and applying a tougher standard the law then failed. Six days later, RealPage filed a letter telling the Manhattan court that this decision supports its challenge to New York’s rent law.

A ruling about restaurant customer lists became ammunition in a fight about rent within a week. That is how a legal vocabulary travels: not by command, but by citation.

Attribution for blame is not attribution for speech

This is the distinction the whole argument turns on, and it is easy to lose.

Attribution for liability says: the firm answers for what its system does, because it built it, trained it, deployed it, profits from it, and can change it.

Attribution for expression says: the system’s output is the firm’s protected speech.

Product liability is the standing proof that the first does not require the second. A manufacturer answers for a defective product without anyone calling the product its speech. Antitrust and consumer-protection law likewise reach conduct — data pooling, automatic implementation, coercive defaults, misrepresentation — without reaching a recommendation’s message. New Mexico just ordered nearly a billion dollars and a five-year overhaul of two products without once needing to decide whether a feed is a sentence.

So the danger was never that the law would fail to find someone to hold responsible. It is that the vocabulary it happens to reach for while doing so converts a basis for accountability into a claim of immunity — and that this conversion is happening case by case, with no one deciding it should.

Which yields the practical point, and it is unglamorous: the first door is better than the second. When both routes lead to the same courtroom, the design-defect route is worth preferring precisely because it leaves no speaker behind. Plaintiffs’ lawyers pick the theory that wins their case; nobody is paid to think about what that theory costs the next plaintiff, or a state legislature two years from now. The choice between doors is being made one brief at a time, by people with no institutional reason to weigh it.

Why this looks the same from underneath

It is tempting to describe the past two years as a change of eras: first the law demanded a culpable mind, now it is manufacturing a speaker. From the bench, that may be what it looks like.

From where the harm lands, the two regimes are one wall in different clothes.

Under the old rule, you must establish what the company knew. Under the new one, you must establish that its output is not protected expression. Both facts live inside the company. Neither is visible to a tenant whose rent moved, or to a parent trying to reconstruct what a feed served a child. The legal test changed. The position of the person who has to satisfy it did not.

And whichever door a plaintiff walks through, she arrives in the same room, facing the same problem: the evidence is on the other side of the wall.

Find the memo

The law’s default picture of engineered harm has a person in it. Somebody understood how the product worked on you, wrote it down, and optimized accordingly. Tobacco is the clean case: the industry knew about nicotine, documented that it knew, and tuned its products accordingly — and the 1998 settlement was possible in large part because the documents existed and eventually came out.

Fraud requires knowledge. Deception requires a state of mind. Find the memo and you have the case.

That picture still works. It is what New Mexico ran on: former employees, hundreds of internal documents, a finding of misrepresentation. Every major win of the past year has been a document case. Any honest version of this argument has to concede that first.

But the picture fits the building of these systems worse every year. Nobody derives a feed from a settled theory of adolescent attention. You generate variations, choose a number that counts as success, and keep whatever moves the number. Twenty thousand experiments later you may have a highly effective system whose builders can describe what it does without being able to explain why this version works better than the ones it beat.

There is a tempting conclusion here, and it is wrong twice: nobody designed it, so nobody is responsible.

Wrong first because somebody chose the number. Optimization always contains a specification of success — inherited, delegated, revised, often half-implicit, but there nonetheless. Selecting it is the clearest human act in the loop.

Wrong second, and more importantly, because the machines never stopped keeping notes. They keep enormous notes: experiment logs, evaluation dashboards, A/B results, ranking-change tickets, model cards. What they stopped keeping is prose.

There is no paragraph saying we found that showing this to lonely fourteen-year-olds increases session length. There is a dashboard where that fact is visible to anyone who knows which query to run.

So the problem is not an absence of records. It is that the records are diffuse, privileged, and organized for engineers rather than for cross-examination. The person injured has no way to reach them, and often no way to know they exist. The problem is not authorship. It is discoverability.

That reframe invites an obvious objection. If the records exist, isn’t this just a discovery problem — real, but small? Two answers. First, discovery problems are not small when the entire causal history of a product lives in systems built to be read by nobody outside the company. Second, and more to the point: nothing in the sword-and-shield problem depends on the records being unreachable. It runs identically against a fully documented, memo-writing platform — the plaintiff still needs a theory to clear Section 230, and the firm still keeps whatever classification the court hands over. Unreachable records are not what creates that problem. They are what would remove the one thing that has so far made it survivable: the documents.

Will the records get harder to reach? Probably. Retention schedules shorten rather than lengthen; legal holds attach to named custodians and identified systems, which maps poorly onto infrastructure that regenerates itself; logging is increasingly automatic and increasingly ephemeral; and written strategy has been giving way to dashboards for a decade, for reasons that have nothing to do with litigation. Those trends are observable. That they will produce a materially thinner evidentiary record in the next wave of cases is a forecast, and it is the load-bearing assumption of what follows. If the next generation of systems proves as document-rich as this one, I am describing a problem that never quite arrives — and the proposals below are worth having anyway.

What to ask for that does not depend on the speech fight

Two proposals. The first matters more, and it is the one an injured person can actually use.

Keep the receipts

Start with the penalty, because it is the part that does the work.

If a company deploys a system that shapes prices or feeds at scale, and cannot produce records it was required to keep of what it told that system to optimize, a fact-finder should be permitted to infer that the missing records would not have helped it.

That is not a new theory of liability. It is an evidentiary rule — closer to requiring a restaurant to keep its inspection logs than to telling it what it may say. What must be retained is specific: objective specifications and their revisions, experiment and evaluation results, deployment changes, material incident records. On a defined schedule. In a producible form. With defined audit access.

Note what it does not ask for. It does not require a company to produce a human-readable theory of why any particular output happened; no company can reliably supply that account for every output. It requires preservation of the operational record of what was tested, measured, changed, and chosen as the target.

The pedigree is old. What Justice Traynor was reaching for in his 1944 concurrence in Escola v. Coca-Cola Bottling Co. — the exploding bottle case, where a waitress could not show what the bottler did wrong because she had never been inside the plant — was not only a theory of responsibility. It was burden allocation: when one party controls the process and another is injured by it, the law can move the evidentiary burden toward the party who can see what happened.

The closer precedent is Sindell v. Abbott Laboratories (California, 1980), where women injured by a drug their mothers had taken could not identify which manufacturer made the specific pills, and the court restructured the burden by market share rather than relaxing anyone’s state of mind. Cite it with its record intact: most states declined to follow, and it stayed largely confined to that one drug. Burden restructuring is available, is not new, and is not a reliable winner.

The counterargument arrives immediately, and it is the shield again: a firm told to preserve what it optimized for will answer that its objectives and evaluation criteria are the editorial standards of an expressive product. Two things cut back. A retention duty is a preservation obligation, not a compelled public statement. And a firm’s records about its own system are disclosure about its own services — the distinction the Second Circuit leaned on this month when it held the deferential rule inapplicable to a law about third parties’ data.

But be precise about the weak point. The preservation half is fairly safe. The adverse inference is exposed: it is a judicial penalty for not producing material a company will argue is expressive, and a court worried about that characterization may balk. Spoliation sanctions already exist and are already hard to get. This is the piece most likely to fail.

Watch the diet and the wiring

The second proposal is one government has already built, and its useful feature is what it declines to ask.

The Justice Department’s consent judgment against RealPage — entered by a federal court in North Carolina in May — does not ask what RealPage knew, or what its model was really selecting for. It constrains diet, what the system may consume, and coupling, how directly its output becomes action. Model training limited to backward-looking nonpublic data at least twelve months old, with competitors’ current and localized pricing excluded. Restrictions on features that auto-implement recommendations or discourage users from rejecting them. A court-appointed monitor, for seven years.

Each does identifiable work. The data-age floor breaks the loop in which a rival’s current behavior becomes a live coordination signal. Excluding localized pricing stops the recommendation from working as a near-real-time substitute for phoning a competitor. The auto-implementation limits restore a moment where a human can depart from the model, and the rejection-friction limits keep that moment from decaying into a ceremonial click.

The analogy does not fully travel. RealPage’s data restriction is designed to stop competitors from coordinating through a machine, and a video feed is not a cartel. There is no rival whose current price must be kept out of the model, so a twelve-month data floor has no direct counterpart for a For You page — asking for one is a category error, not a strict rule.

The coupling constraints do travel, and we no longer have to speculate. The New Mexico remedy order last week is a coupling order in everything but name: time limits for minors, likes hidden by default, warnings surfaced, findings directed at autoplay, endless scroll, and notifications by name. A state court reached those remedies through public-nuisance law, deciding nothing about speech.

Neither register touches the objective. A system can be fed clean, stale, aggregated data and still be pointed at the wrong target. Human review does little when the reviewer is measured by the same revenue number that produced the recommendation.

An objective function is not the same thing as intent. Objectives are plural, inherited, locally tuned, sometimes implicit in the training data. But an objective is the most concentrated surviving record of the human choice that makes the loop possible: a declaration of what the system is permitted to treat as success. A scheme that constrains the diet and never looks at that has left alone the one act in the loop a person unambiguously performed.

The honest limit

Everything in this section is argued from the regulator’s chair. Diet and coupling is attractive because it is administrable: it requires reading no model’s interior and gives an agency something to monitor. Those are the virtues of the seat doing the regulating.

I do not know what it delivers to the person the harm lands on. A tenant does not experience a twelve-month data floor. She experiences a rent. Whether the floor moved it, by how much, and whether the same increase arrived a year later by a permitted route are questions I have not answered — though as of this year they have at last become answerable, because the RealPage judgment is entered and its compliance deadlines have run. There is now a covered period, and somebody should measure it.

Preserving a moment of human refusal is likewise worth something only if the person at that moment has a reason to refuse, and a leasing agent with an occupancy target is not that person.

So the honest version is narrower than the confident one: diet and coupling is the instrument most likely to survive a courtroom. That is durability, not efficacy. The retention proposal is the exception, and it is the reason to prefer it — a preservation mandate and an adverse inference are things an injured person can invoke, not things done on her behalf.

There is also a boundary where none of this reaches. Every instrument here assumes an identifiable operator running a loop it wants to keep running, with inputs that can be constrained and a point where output can be refused. Illicit markets have no such operator; nobody is holding the loop, which is why other tools do that work. Process regulation grips an optimization loop only when someone is holding it.

The smallest thing worth demanding

Strip out everything contestable and one requirement survives.

If you deploy a system that sets prices, allocates opportunities, or decides what millions of people see, preserve the record of its objectives, its major experiments, and the changes by which you learned it worked. Keep it on a schedule. Produce it when someone is injured. If you cannot, let the fact-finder consider what the absence means.

That requires no theory of whether a recommendation is speech, no finding about anyone’s state of mind, and no new agency. It asks a company to preserve a record it already generates for its own purposes, in a form an injured person’s lawyer can obtain. We ask this of businesses constantly. There is nothing mystical about asking it here.

Meanwhile the larger question — whether a rent recommendation, a wage offer, a premium, or a feed is protected expression — is being briefed in dockets almost nobody will read, on a schedule nobody chose, before courts that took the cases because companies filed them. The last time it came up squarely, in a Philadelphia appeal about a dead ten-year-old, a court answered it in a way no party had asked for, resting on a Supreme Court opinion that never addressed the statute at issue, and the answer stuck because the losing side decided not to press it further.

When courts answer it again, the answer will travel — to the feed, the rent, the wage, the premium. And the person it arrives for will not be a doctrine. She will be a tenant opening a letter, or a parent opening a phone.

Apparatus

What follows is scaffolding: sources, corrections, alternative explanations, and bets I am willing to lose. A reader who stops at the line above loses nothing.

Corrections carried into this draft

The RealPage Final Judgment was entered. Prior drafts described it as a proposed settlement awaiting a Tunney Act determination, and one flagged the status as unverified. It is verified: Judge William L. Osteen, Jr. (M.D.N.C.) signed off on May 19, 2026 (Law360; WSOC-TV). The chronology is: proposed Final Judgment filed Nov. 24, 2025; Stipulation and Proposed Order entered Mar. 26, 2026; response to public comments published May 8, 2026; Final Judgment entered May 19, 2026; seven-year term. The Greystar judgment had been entered earlier, on Mar. 2, 2026. Consequences beyond tense: the instrument is now an entered court judgment rather than a proposal; the compliance clocks (30/60/180 days) have run; and the efficacy question therefore has a covered period that can be measured, which the body now says.
What remains true is the narrower caveat: an entered consent judgment is still a negotiated instrument that has survived no constitutional challenge. Nobody has yet argued to a court that these constraints violate the First Amendment. RealPage is making that argument about New York’s statute, not about the judgment it agreed to.
Cut in the previous draft for want of Tier 1 support and still out: Meta’s reported $4 million counter-offer in the state-AG case, and the judge’s reported characterization of both sides’ figures.
Also cut, two drafts ago: a Japanese fishing lure used to illustrate optimization without theory. It carried no argumentative load and its own selection story was contaminated in the way the essay elsewhere criticizes.

Evidence framework

Tier 1 — public records

Escola v. Coca-Cola Bottling Co., 24 Cal. 2d 453 (1944) (Traynor, J., concurring).
Sindell v. Abbott Laboratories, 26 Cal. 3d 588 (1980) — market-share liability; minority adoption, largely confined to DES.
Fair Housing Council of San Fernando Valley v. Roommates.com, LLC, 521 F.3d 1157 (9th Cir. 2008) (en banc).
Moody v. NetChoice, LLC (U.S. 2024) — protection for human-authored moderation guidelines; express reservation as to other activities.
Anderson v. TikTok, Inc. (3d Cir., Aug. 27, 2024); Matey, J., concurring in part; rehearing denied Oct. 23, 2024; no certiorari petition.
Ninth Circuit, Aug. 10, 2026 (Nguyen, J., with Bennett, J., and Matsumoto, D.J.), Nos. 24-7032 / 24-7265 / 24-7312 — Section 230 “merely provides a defense to liability — not immunity from suit”; interlocutory appeals dismissed for want of jurisdiction. Published opinion.
DoorDash, Inc. v. City of New York, No. 25-81 (2d Cir., Aug. 5, 2026) — Zauderer inapplicable where compelled disclosure concerns third parties; Central Hudson scrutiny applied; NYC Customer Data Law struck down. RealPage letter invoking it filed Aug. 11, 2026 (Law360).
N.Y. Gen. Bus. Law § 340-b (S.7882, signed Oct. 16, 2025; effective Dec. 15, 2025). RealPage, Inc. v. James, No. 1:25-cv-09847 (S.D.N.Y.) — complaint Nov. 2025; PI motion pending; AG motion to dismiss Jan. 2026; undecided as of Aug. 14, 2026.
National Retail Federation v. James (S.D.N.Y., Rakoff, J., Oct. 8, 2025) — GBL § 349-a disclosure statute upheld under Zauderer. Cut from the body for density.
United States et al. v. RealPage, Inc. et al., No. 1:24-cv-00710 (M.D.N.C.) — Final Judgment entered May 19, 2026; terms as described (12-month backward-looking training data; exclusion of competitors’ nonpublic and localized pricing; auto-implementation and rejection-friction limits; court-appointed monitor; seven-year term, early termination possible after four).
New Mexico (State ex rel. Torrez v. Meta Platforms), First Judicial District Court, Santa Fe (Biedscheid, J.): jury verdict Mar. 24, 2026, 75,000 Unfair Practices Act violations, $375M maximum civil penalty; Phase II order Aug. 6–7, 2026, public nuisance found, $567M abatement fund ($420M treatment, $90M screening, $33M prevention, $15M referral and coordination) plus five years of court-supervised product changes. Meta has said it will appeal.
MDL 3047 (N.D. Cal., Gonzalez Rogers, J.) and JCCP 5255 (L.A. Super. Ct., Kuhl, J.): design-defect claims permitted past Section 230; K.G.M. verdict Mar. 25, 2026, $6M ($3M compensatory allocated 70/30 Meta/Google, $3M punitive), post-trial motions and appeals filed; Breathitt County, Ky. bellwether settled May 2026, reported ≈$27M combined across four defendants; state-AG bellwether opening in Oakland this month. Tier B sourcing — litigation press and general news, not docket-verified.

Tier 2 — inferences

The design-defect route and the first-party-speech route are substitutes, not complements: both clear Section 230, only one produces a speaker. Supported by Anderson (majority vs. Matey), the Aug. 10 disposition, and the New Mexico judgment, which reached remedies without a speech classification.
The Anderson holding is contingent, not structurally required: three routes existed, one panel member took another, rehearing was denied procedurally, and the losing party declined to seek review.
Coupling constraints transfer across harm types; the coordination-breaking diet constraint does not. The New Mexico Phase II order is the affirmative instance for coupling.

Tier 3 — structural hypotheses

Records will get harder to reach as retention schedules shorten, legal holds fail to map onto self-regenerating infrastructure, logging becomes automatic and ephemeral, and behavior originates in training processes that emit logs rather than rationales. The component trends are observable; the conclusion about litigation records is the forecast. Falsified by: comparable density of internal narrative research in litigation over systems built after ~2024.
The expressive vocabulary will migrate into rooms where the doctrinal test differs, and will be found persuasive in some of them.

Alternative explanations considered

Simplest: Anderson is an outlier and the sword/shield problem is a Third Circuit phenomenon. After Aug. 10 this is the strongest competing account, and the essay now concedes most of it — the largest docket in the country is running on non-expressive theories. What survives is the vocabulary-migration claim: weaker, slower, better evidenced.
Simplest: ordinary regulated-industry lobbying; the First Amendment framing is standard-issue. Partly true. The distinguishing fact is that the classification was supplied by a court to a defendant that had argued against it.
Competing: the real problem is Section 230’s overbreadth, not the First Amendment. Viable, and it predicts the difficulty dissolving on statutory amendment — a clean distinguishing test that the Aug. 10 ruling arguably begins judicially.
Competing: this is ordinary e-discovery friction, not a structural gap. Now stated and answered in the body rather than assumed past; the answer is that the sword/shield problem does not depend on the evidentiary claim at all.

Institutional actions

Retention and permissive adverse inference. State legislatures under existing consumer-protection authority, or Congress via the federal preservation regime: a retention duty on optimization records for operators of large-scale pricing or recommendation systems, with an adverse inference available only where the moving party shows the material was relevant, within the operator’s control, and unpreserved despite a duty. That qualifier distinguishes strategic non-preservation from ordinary technical loss and is what keeps the rule from becoming automatic liability.
Coupling constraints for consumer recommenders. State AGs, existing authority: limits on auto-implementation, default-on personalization, and friction against declining. The New Mexico order is a template that has survived a trial.
Publish the monitor’s reports. The RealPage judgment installs a court-appointed monitor for seven years. Monitor findings in consent judgments routinely stay sealed. Making these public — or reporting them in summary to the court on the docket — costs nothing and is the only mechanism by which anyone outside the parties will learn whether the constraints hold.
Measure the covered period. HUD or state housing agencies: publish rent trajectories for units covered by the RealPage constraints against comparable uncovered units, dated from the compliance deadlines that ran in 2026. This became possible on May 19 and nobody is obliged to do it.

Forecast Register

Probabilities are authored priors informed by the structural analysis, not derived rates; reference classes are named so a scorer can see what they are compared against. Stated to one decimal because the second would be false precision.

			
# FORECAST REGISTER v1
essay: when-the-software-speaks
date_written: 2026-08-14
forecasts:
  - id: F1
    hypothesis: "The First Amendment expressive classification becomes the operative defense
      against algorithmic process regulation."
    column: mechanism
    question: "Will the U.S. District Court for the Southern District of New York, in RealPage,
      Inc. v. James (No. 1:25-cv-09847), resolve the challenge to N.Y. Gen. Bus. Law section
      340-b principally on First Amendment grounds — rather than on standing, ripeness,
      preemption, Due Process, or statutory construction — in its first substantive ruling?"
    resolution_date: 2027-06-30
    resolver: "Text of the court's first substantive opinion on the docket in No. 1:25-cv-09847.
      If no substantive opinion has issued by the resolution date, resolve NO."
    p_essay: 0.6
    p_baseline: 0.45
    reference_class: "Constitutional challenges to state economic regulation where the complaint
      pleads multiple independent grounds and the court may avoid the constitutional question."
    direction: fragility
  - id: F2
    hypothesis: "The First Amendment expressive classification becomes the operative defense
      against algorithmic process regulation."
    column: magnitude
    question: "Will enforcement of N.Y. Gen. Bus. Law section 340-b be enjoined in whole or in
      part by any court on or before 2027-06-30?"
    resolution_date: 2027-06-30
    resolver: "Docket in RealPage, Inc. v. James (S.D.N.Y. No. 1:25-cv-09847) or any appellate
      order therein."
    p_essay: 0.4
    p_baseline: 0.3
    reference_class: "Preliminary-injunction grant rates in First Amendment challenges to state
      commercial regulation."
    direction: fragility
  - id: F3
    hypothesis: "The design-defect and procedural routes past Section 230, not the first-party-
      speech route, carry the bulk of successful algorithmic-harm litigation."
    column: mechanism
    question: "Among U.S. plaintiff verdicts or judgments against social-media platforms for
      minor-user harm entered between 2026-08-14 and 2028-12-31, will a majority rest on
      product-design, negligence, consumer-protection, or public-nuisance theories WITHOUT any
      holding that the platform's recommendations are its own first-party speech?"
    resolution_date: 2028-12-31
    resolver: "Verdicts and judgments identified via CourtListener and Law360; count of those
      whose operative opinions contain a first-party-speech holding versus those that do not."
    p_essay: 0.85
    p_baseline: 0.6
    reference_class: "Theory-of-liability distribution in mass torts where multiple viable
      theories clear a threshold immunity defense."
    direction: stability
  - id: F4
    hypothesis: "The design-defect and procedural routes past Section 230, not the first-party-
      speech route, carry the bulk of successful algorithmic-harm litigation."
    column: magnitude
    question: "Will at least two federal courts of appeals outside the Third Circuit hold, in
      published opinions, that a platform's algorithmic recommendation of third-party content is
      the platform's own first-party speech and therefore outside Section 230, on or before
      2028-12-31?"
    resolution_date: 2028-12-31
    resolver: "Published circuit opinions via CourtListener; count of distinct circuits so
      holding."
    p_essay: 0.2
    p_baseline: 0.2
    reference_class: "Adoption of a novel circuit-level statutory reading by two or more sister
      circuits within roughly four years."
    direction: fragility
  - id: F5
    hypothesis: "The evidentiary register — retention plus burden-shifting — is the instrument
      least likely to be adopted, notwithstanding being the one most useful to plaintiffs."
    column: mechanism
    question: "Will any U.S. state enact, or Congress pass, a statute imposing a record-retention
      duty on operators of large-scale automated pricing or recommendation systems covering
      objective specifications, experiments, or evaluation results, on or before 2028-12-31?"
    resolution_date: 2028-12-31
    resolver: "Enacted-legislation text via LegiScan or Congress.gov. Must impose a retention
      duty on optimization or evaluation records — not merely consumer disclosure or an impact
      assessment filed with an agency."
    p_essay: 0.3
    p_baseline: 0.35
    reference_class: "State adoption of algorithmic-accountability mandates 2019-2026, in which
      consumer-disclosure statutes passed markedly more often than record-keeping duties."
    direction: stability
  - id: F6
    hypothesis: "The evidentiary register — retention plus burden-shifting — is the instrument
      least likely to be adopted, notwithstanding being the one most useful to plaintiffs."
    column: magnitude
    question: "Will any U.S. court grant an adverse-inference instruction or spoliation sanction
      against a platform or pricing-software defendant specifically for failure to preserve
      model-training, experiment, or objective-specification records, on or before 2028-12-31?"
    resolution_date: 2028-12-31
    resolver: "Published or docketed order in any U.S. federal or state court, via CourtListener
      or Law360 reporting."
    p_essay: 0.25
    p_baseline: 0.2
    reference_class: "Spoliation-sanction rates in large product-liability MDLs involving
      electronically stored information; no instance specific to optimization records is known
      to the author."
    direction: fragility
  - id: F7
    hypothesis: "Diet-and-coupling constraints are durable in court but of unmeasured benefit to
      the people the harm lands on."
    column: magnitude
    question: "Will any government agency, academic study, or court-appointed monitor publish a
      comparison of rent trajectories in RealPage-covered units against comparable uncovered
      units for a period beginning after the Final Judgment's compliance deadlines, on or before
      2028-12-31?"
    resolution_date: 2028-12-31
    resolver: "Publication by HUD, a state housing agency, a peer-reviewed journal, or a monitor
      report entered on the docket in No. 1:24-cv-00710."
    p_essay: 0.35
    p_baseline: 0.3
    reference_class: "Post-consent-decree outcome evaluation in antitrust remedies, which is
      rarely performed or published."
    direction: stability

		

Unresolved questions

Does a retention duty survive the compelled-speech objection after DoorDash? The preservation half is probably safe; the adverse-inference half is exposed. Doctrinal, not empirical.
Did the RealPage constraints move rents? Now genuinely answerable — the judgment is entered and the compliance clocks have run — and still unanswered. Requires the covered-versus-uncovered series nobody is obliged to publish. Promoted to Forecast F7.
Can a tenant in a process-rule jurisdiction establish what set her rent? Empirical. The efficacy case for process rules rests entirely on an untested assumption about plaintiff capability.
Is there any deployed system for which no retention duty would produce a usable record? The cheapest decisive test of the discoverability framing, still unresolved.
Does the choice of liability theory actually propagate? Whether a first-party-speech holding in one circuit measurably worsens a state’s regulatory position in another is an empirical claim about doctrinal diffusion that this essay asserts and does not demonstrate. The RealPage letter of Aug. 11 is one data point, not a pattern. This remains the weakest link in the argument.

Metadata

Provenance. General-audience rewrite of a prior draft (“The Algorithm Takes the Stand”), revised three times: against an independent referee report (UKE_REFEREE v0.3), then against two rounds of multi-model review. Per §1.5.6 the honest input entry remains: reports not consumed beyond the brief — no engine reports were opened by any pass in this sequence, and no Findings Ledger is claimed.

What each round changed. Round one: authorship reframed to discoverability; Sindell added; the transfer test run in-body; the lure demoted. Round two: the Ninth Circuit’s Aug. 10 ruling and the New Mexico Phase II order, both surfaced by reviewers, restructured the essay around two routes rather than one trap; the Section 230 / First Amendment conflation was flagged and is now a stated limit; the lure was cut. Round three (this draft): the RealPage entry correction and its downstream consequences; compression throughout; the adverse inference made permissive and conditioned on a preexisting duty; the objective-function passage slowed down; a bridge added into the discoverability section; the document-thinning forecast anchored to observable trends while remaining marked as a forecast.

Adversarial review.

Weakest link: the vocabulary-migration claim (Unresolved Question 5). One letter, one week old.
Second weakest: the document-thinning forecast, now argued rather than asserted, still a forecast.
Most likely criticism: that after Aug. 10 the essay describes a risk the courts are actively routing around. Conceded; that concession is the structure. The prescriptive half does not depend on the risk materializing.
Not answered: whether an adverse-inference remedy is enforceable in practice against a well-resourced defendant. F6 scores it.

Brittleness. Independent lines: (1) the two-doors comparison, grounded in Anderson versus the Aug. 10 disposition and New Mexico; (2) RealPage v. James plus DoorDash as vocabulary migration; (3) the retention proposal, standing on Escola and Sindell regardless of either; (4) coupling constraints, now with both an entered federal judgment and a state trial court order behind them. Refuting any one leaves the rest standing.

Source quality. Tier S: court opinions and judgments, Federal Register, DOJ filings, New Mexico AG release, statutory text. Tier A: Law360, PBS, ABC, CNBC, The Hill, WSOC, Reuters-derived reporting. Tier B: litigation-tracking press for MDL/JCCP figures, marked in-line. Tier C: none.

Corpus consulted. Federal and state opinions and dockets, primary where available and legal press otherwise; Federal Register; DOJ Antitrust Division; New York session law; New Mexico DOJ; general news. Knowingly not consulted: Anderson, Moody, and DoorDash read in full — characterized from reported holdings and quoted passages. The essay’s sharpest claims are close readings of Kagan’s reservation and Matey’s concurrence; those two opinions should be read directly before publication. They are the paragraphs a hostile lawyer checks first. I am not a lawyer.