> daily_signal(2026_07_12)

A jury already found that Meta's design addicted a child. The appeal filed this week decides whether that verdict, and the thousands of claims stacked behind it, survive.

PickBits Daily Signal · Sunday, July 12, 2026

By Mark Pickering · 9 min read · July 12, 2026

// tl;dr

Meta filed a notice of appeal on Tuesday against the first jury verdict in the US to hold a social platform liable for designing its product to addict a child. The jury found negligence by Meta and Google-owned YouTube was a substantial factor in harming a 20-year-old plaintiff, and awarded roughly $6 million, split 70/30. Judge Carolyn B. Kuhl denied the post-trial motions in early June, which cleared the path to the appeal. The claim that survived was product-design negligence, which Section 230 does not shield.
Five women expanded a federal class action against xAI and added Stability AI as a defendant, alleging Grok generated roughly 7,000 sexual images of a woman from a single photo of her at age 11. The complaint also alleges xAI filed its mandatory CyberTipline reports while omitting the user and IP data police need, leaving 90% of them unactionable. The theory against Stability reaches into the training data and the decision to release weights with safety restrictions stripped.
OpenAI published a proof of the Cycle Double Cover Conjecture, open since 1973, that it says GPT-5.6 Sol Ultra produced in under an hour using 64 parallel subagents. It is not peer reviewed. Mathematician Thomas Bloom faults it for citing none of the prior work it leans on, including a 1983 paper, and calls the argument short, elementary, and findable in the 1980s.
Harvard's COMPASS predicts who will respond to cancer immunotherapy by reading a tumor's gene activity, beating 22 existing methods by about 8.5 points of accuracy. It routes roughly 16,000 genes through 44 human-readable immune concepts, so an oncologist gets a rationale instead of a score. It is retrospective and not FDA cleared.

We have watched the courts inch toward this for a month. On June 21, a ruling that when Google's AI makes something up, Google is the one who said it. On June 27, a judge who would not throw out the voiceprint class action against Meta. At the end of June, another who ordered Meta to trial over social media harms. Every one of those was a company failing to get out early, and today Meta is appealing something further down that road than any of them: the first jury verdict in the country that blames a platform's design, rather than its content, for addicting a child. In California, a class action that started in March with three women is now five, and this week it added the company that published the open weights, which is where this stops being about the person who typed the prompt. OpenAI, meanwhile, published a proof of a fifty-year-old conjecture that nobody has verified, and the first mathematician to read it closely came back with holes. Harvard published a cancer model that walks an oncologist through its reasoning so she can push back on it. Guess which one got the bigger headline.

Meta is appealing the addicting-by-design verdict, five women added Stability AI to a federal class action against xAI over Grok-generated abuse images, OpenAI published an unverified proof of a 50-year-old conjecture, and Harvard's COMPASS predicts which patients immunotherapy will actually help.

1. Meta is appealing the first verdict that blamed a platform's design for addicting a child.

The six million dollars is not what Meta is fighting about.

On Tuesday, July 7, Meta's lawyers filed a notice of appeal in Los Angeles County Superior Court, challenging a jury's finding that Instagram and Facebook were built to hook young users. The verdict is the first of its kind in the United States: a jury found that negligence by Meta and by Google-owned YouTube was a substantial factor in harming the plaintiff, a 20-year-old woman identified in court as KGM, who testified she became addicted to social media as a child and that it deepened her mental-health struggles. The award was roughly $6 million, about $3 million compensatory and $3 million punitive, allocated 70% to Meta and 30% to YouTube. Trial judge Carolyn B. Kuhl denied both companies' post-trial motions in early June, and that ruling is what made this appeal the next move. Meta's public position is that teen mental health is "profoundly complex and cannot be linked to a single app."

Meta spent considerably more than $6 million defending this case, so the money was never the point. Section 230 shields a platform from what its users post. It has never had anything to say about how the product itself was built, and the jury was asked about how the product was built. That is the distinction the California Court of Appeal now has to rule on, and it is why thousands of parallel adolescent-addiction claims, consolidated in California state court and in federal MDL proceedings, are effectively parked until it does. It also lands in a pattern we have been tracking for a month. Courts keep refusing to let these companies exit early: Google was held to own what its AI asserts, Meta could not get the voiceprint class action dismissed, and at the end of June a federal judge ordered it to trial over social media harms. Now a jury has gone the whole way.

ABC News July 10 2026 AP wire report Meta appeals landmark jury verdict that found it to blame for a young woman's social media addiction, notice of appeal filed July 7 in Los Angeles County Superior Court, jury found negligence by Meta and Google-owned YouTube a substantial factor in harming plaintiff KGM, roughly 6 million dollars awarded split 70 percent Meta 30 percent YouTube, Judge Carolyn B. Kuhl denied post-trial motions in early June — abcnews.com · July 10, 2026

Why this matters: If you have ever suspected that the autoplay and the notification that arrives right as you put the phone down were engineered to be hard to walk away from, a jury has now agreed with you and attached a dollar figure to it. The evidence at trial came out of Meta's own retention specs and experiment logs, which is why this frightens the industry more than a privacy fine ever did. Every consumer product built to be hard to put down is now something a family can sue over, and the appellate ruling decides whether that stays true. Action this week: As a parent or as anyone with a teenager in the house, go into Instagram and YouTube settings today and turn off autoplay, kill push notifications for social apps, and switch the account to the supervised teen experience, because those are the exact mechanics the jury identified as the harm. If you build or own a consumer product with engagement mechanics, pull your retention-feature specs and experiment logs and read them the way opposing counsel would: could a plaintiff read this document as evidence you optimized for compulsion? Whatever your team writes down this quarter, assume a lawyer reads it out loud someday. Ask legal now whether your minor-user experience needs a documented design review, rather than after the ruling lands.

abcnews.com (AP): Meta appeals landmark jury verdict that found it to blame for social media addiction (July 10, 2026)
lasvegassun.com: Meta appeals landmark jury verdict that found it to blame (July 10, 2026)
business-standard.com: Meta challenges US jury verdict in teen social media addiction case (July 11, 2026)

2. The class action over Grok's abuse images has finally named the company that published the weights.

The complaint says the company chose to build it this way.

On July 7, plaintiffs filed an amended class-action complaint in the Northern District of California against xAI and, newly added as a defendant, Stability AI. The case was filed back in March by three women and now has five. Jane Doe 4, a Wyoming woman in her twenties, says her stepfather took a single photograph of her from when she was about 11, fed it to Grok, and generated roughly 7,000 sexually explicit images and videos of her, which he traded with other offenders. Law enforcement told her he chose Grok because it was less restrictive than the other models he tried. Then there is the reporting. US providers are legally required to report apparent child sexual abuse material to NCMEC's CyberTipline, and the complaint alleges xAI complied on paper while hollowing the reports out: by early 2026, NCMEC assessed 90% of xAI's reports as not actionable, because xAI declined to include the user and IP information that lets police identify the person on the other end. In Jane Doe 4's case, the complaint says xAI's February report contained only the original authentic photo, and none of the thousands of images Grok made from it.

The claim against Stability reaches further back, into the training set and the release. The allegation is that it trained early Stable Diffusion models on a dataset known to contain abuse material, then stripped safety restrictions from the published weights to drive adoption. The question underneath it is whether shipping a model with the safety layer deliberately removed is itself the wrongful act, and the open-weight world has been dodging it for two years. We have circled this before. In late May we covered Heretic, a tool that stripped the guardrails off Meta's Llama in under ten minutes and had been downloaded millions of times, and we said then what came out the other side. In early June a British MP took xAI to the High Court over Grok deepfakes of her. Now it is in a US federal court, and this time the defendants are the people who built the models. The plaintiffs sue under Masha's Law and the federal Trafficking Victims Protection Act. Masha's Law carries statutory damages, so a nationwide class here comes with a price tag these companies cannot wave off. Neither company responded to reporters.

CyberScoop July 9 2026 report deepfake CSAM lawsuit against Grok and xAI expands to add Stability AI as defendant, amended class action filed July 7 in Northern District of California, five Jane Doe plaintiffs, allegation that a stepfather used Grok to generate roughly 7000 sexually explicit images from one photo of a girl at age 11, NCMEC found 90 percent of xAI CyberTipline reports not actionable because user and IP data was omitted, claims under Masha's Law and the Trafficking Victims Protection Act — cyberscoop.com · July 9, 2026

Why this matters: The input to this attack is one ordinary photo of your kid, the kind that sits on a school sports page or a public Instagram, and the tools that turn it into abuse material are a tap away. The AI lawsuits we have covered have mostly been about copyright, or about the user who typed the prompt. This one goes at the company that built the model, and at a company that shipped its weights with the safety classifier taken off. Action this week: Use NCMEC's Take It Down service at takeitdown.ncmec.org, which hashes and blocks known sexualized images of a minor across participating platforms without you ever uploading the image, and report material to the CyberTipline at report.cybertip.org. Under the federal TAKE IT DOWN Act, platforms must remove non-consensual intimate imagery, including AI-generated deepfakes, within 48 hours of a valid victim request, and the FTC enforces that deadline, so invoke it by name. Then go set your kids' accounts to private tonight. If you ship, fine-tune, or self-host an image model, go read the complaint and then check two things. Write down where your training data actually came from, because "we inherited an open dataset" is precisely the position Stability is now defending. Then check whether your CyberTipline pipeline includes subscriber and IP information, because the allegation that sank xAI here is not that it failed to file. It is that 90% of what it filed was useless to the police who received it.

cyberscoop.com: Deepfake CSAM lawsuit against Grok and xAI expands to include Stability AI (July 9, 2026)
npr.org: Deepfake CSAM class action expands against xAI and Stability AI (July 9, 2026)
lieffcabraser.com: Deepfake victims bolster class action against xAI, add Stability AI (July 2026)

3. OpenAI says its model proved a 50-year-old conjecture. The mathematicians reading it are finding the seams.

There is no machine that can check this one. Only people.

OpenAI announced on July 10 that GPT-5.6 Sol Ultra generated a proof of the Cycle Double Cover Conjecture, a graph-theory problem posed by George Szekeres in 1973 and independently by Paul Seymour in 1979, and open ever since. The run used 64 subagents in parallel and finished in under an hour. OpenAI published the proof and the roughly 700-word prompt that orchestrated it, which is arguably the more useful artifact: it is a rare public look at how a frontier lab structures multi-agent search on a hard problem. The proof reduces the conjecture to cubic graphs, leans on the 8-flow theorem, and builds an edge labeling that forces every edge into exactly two cycles through a linear-algebra argument.

This has not been peer reviewed, the conjecture has attracted flawed proofs for five decades, and community verification is still pending. The most detailed public assessment so far comes from mathematician Thomas Bloom. He faults the write-up for citing none of the prior work it leans on, notably a 1983 paper by Bermond, Jackson and Jaeger, and says the argument is short, elementary, and could have been discovered in the 1980s. Bloom is not saying the proof is wrong. He is saying a person could have found it forty years ago, and no person did. So the model did not invent new mathematics. It ground through theory we already had, with a stubbornness nobody human bothered to sustain, and it cleared a bar that had stood for 50 years. That is genuinely impressive, and it is not what anyone is selling. Set it against what we covered on July 5, when Mistral shipped an open model whose formal proofs a computer can verify line by line. Those come with a machine that checks them. This one comes with a PDF and a request that mathematicians read it, and so far the mathematicians are unimpressed.

the-decoder July 11 2026 report OpenAI GPT-5.6 Sol Ultra reportedly solves a 50-year-old math problem in under an hour, proof of the Cycle Double Cover Conjecture posed by Szekeres in 1973 and Seymour in 1979, generated with 64 parallel subagents, OpenAI published the proof PDF and the 700-word orchestration prompt, not peer reviewed, mathematician Thomas Bloom faults missing citations to a 1983 Bermond Jackson Jaeger paper and calls the argument short and elementary — the-decoder.com · July 11, 2026

Why this matters: The next time someone tells you an AI solved something no human could, ask who checked it. Here, the answer is nobody yet. Forget the theorem for a moment, because the part you can actually use is how they ran it: 64 agents in parallel, an hour of compute, and OpenAI published the prompt that drove them. It only pays off when a candidate answer can be checked cheaply and mechanically. If you cannot check the answer, pointing 64 agents at it just buys you 64 guesses. Action this week: Read OpenAI's published orchestration prompt before your next agent design review, because it is a working example of decomposing one hard problem across dozens of subagents and reconciling their output, and it maps directly onto problems you already have: exhaustive test-case generation, migration-path exploration, config-space search. And budget for someone to check the output, not just produce it. Do not cite this proof as settled, and push back on anyone in your organization who does, until the graph-theory community signs off.

the-decoder.com: OpenAI's GPT-5.6 Sol Ultra reportedly solves a 50-year-old math problem in under an hour (July 11, 2026)
cdn.openai.com: The Cycle Double Cover proof (PDF)
en.wikipedia.org: Cycle double cover conjecture

4. A Harvard model reads a tumor and predicts whether immunotherapy will actually work.

Checkpoint inhibitors fail most of the people who take them.

Checkpoint inhibitors are some of the best cancer drugs we have ever built, and they fail most of the people who take them. Only about a third respond. Everyone else spends months on a toxic, expensive drug that was never going to work, while the window for something that might have worked closes behind them. On July 3, researchers at Harvard Medical School's Department of Biomedical Informatics, Wanxiang Shen and Marinka Zitnik, with collaborators at Boston Children's Hospital and Roche, published COMPASS in Nature Medicine. It predicts checkpoint-inhibitor response from a tumor's gene-expression profile. It was pretrained on 10,184 tumors across 33 cancer types, then fine-tuned on 16 clinical cohorts covering seven cancers and six regimens. Against 22 existing methods it improved accuracy by about 8.5 percentage points. It generalized to cancers it never saw in fine-tuning, hitting 76.5% accuracy on lung adenocarcinoma held out entirely, and 85.3% on combination therapy when trained only on monotherapy.

COMPASS uses a concept-bottleneck transformer, routing roughly 16,000 genes through 44 biologically grounded immune concepts, so the oncologist gets a rationale she can read and argue with. That matters because it answers the objection we have raised on nearly every health-AI story we have run, from the FDA fast-tracking AI-written radiology reports at the end of June to the insulin chatbot it cleared on July 6: a model a clinician cannot interrogate stalls in review, whatever its accuracy. COMPASS beat 22 black-box baselines while showing its work, so the usual excuse that explaining a model would cost you accuracy does not survive this paper. It also lands somewhere our own coverage has not. We have run AI that finds disease, from the bowel-cancer stool screen to the retina scan a 17-year-old built, and on Friday, Anthropic funding the search for new drugs. None of that helps the patient who already has a drug and needs to know whether it is the right one. The hard caveat, stated by the authors: these are retrospective analyses. COMPASS needs prospective trials, and it is not FDA cleared.

Inside Precision Medicine July 2026 report on COMPASS, a Harvard Medical School foundation model published in Nature Medicine by Wanxiang Shen and Marinka Zitnik predicting cancer immunotherapy response from tumor gene expression, pretrained on 10184 tumors across 33 cancer types, beats 22 existing methods by 8.5 percentage points accuracy, concept-bottleneck transformer routing 16000 genes through 44 interpretable immune concepts, retrospective and not FDA cleared — insideprecisionmedicine.com · July 2026

Why this matters: If you or someone you love is facing an immunotherapy decision, the drug in front of you probably will not work, and until now nobody could tell you that in advance. Being told up front that this one will not work for you buys back the months you would otherwise spend finding out the hard way, while a treatment that might have helped is still on the table, and it spares you a course that costs six figures a year and carries real toxicity. Action this week: If a checkpoint inhibitor is on the table, Keytruda or Opdivo or Yervoy or similar, ask the oncologist two questions: has my tumor had gene-expression or RNA sequencing done, and what biomarkers are we using to decide this will work? Comprehensive genomic profiling is already standard of care at most NCI-designated cancer centers and is widely covered by insurance and Medicare. COMPASS itself is research-only, so do not ask for it by name and do not delay treatment waiting for it, but the profiling it reads is real and available now. If you build clinical or other high-stakes models, study the concept-bottleneck design, because in any domain where an expert has to sign off, a model that shows its reasoning is the one that gets deployed. The black-box one stalls in review.

insideprecisionmedicine.com: Predicting cancer immunotherapy response better with the COMPASS AI model (July 2026)
pmc.ncbi.nlm.nih.gov: COMPASS open-access mirror (Nature Medicine, July 3, 2026)
medicalxpress.com: AI tool predicts who benefits from cancer immunotherapy drugs (July 2026)

» What to watch this week

Quiet settlements in the parked adolescent-addiction cases. Every ruling in this arc for the past month has gone against the platforms at the motion stage. If Meta starts settling the strongest individual claims while the appeal is pending, that tells you it does not like its odds of getting a good appellate precedent written.
Stability AI's motion to dismiss, whenever it lands. Expect it to argue that publishing open weights cannot itself be a wrongful act. How the court handles that argument is the first real test of the question, and anyone shipping open weights should read the ruling rather than the headline about it.
Discovery into xAI's CyberTipline pipeline, and whether NCMEC's 90% figure survives it. If it does, this jumps from a lawsuit into legislation quickly, because 18 U.S.C. 2258A currently sets no standard at all for what a usable report has to contain. Right now, filing an empty report and filing a useful one are the same act in law.
Whether anyone finds that the 1983 Bermond-Jackson-Jaeger paper already contains the argument. If Bloom's read holds up, the OpenAI result quietly becomes a story about parallel search rather than machine genius.
A prospective trial for COMPASS, and who runs it. Roche is already a collaborator on the paper, which is the shortest path from a Nature Medicine result to something an oncologist can actually order for a patient.

Tomorrow's signal lands here.