AI shopping assistants from major platforms struggled with stale information during Black Friday testing, according to The Verge. The tools promised product research, deal hunting, comparison, and guidance through a retail season built on urgency. Instead, they sometimes recommended expired discounts, unavailable products, or choices that required the user to check the work manually. The assistant enters the store with the confidence of a clerk and the memory of a catalog that may already be out of date.

Shopping is a hard test for generative convenience because price, stock, shipping, coupons, returns, and seller reputation change faster than a polished answer can age. A bad recommendation does not remain abstract. It becomes a cart with the wrong item, a missed sale, a wasted hour, or a purchase made under false confidence. Retail platforms want AI to soften the work of choosing, especially when pages overflow with sponsored results and near-identical products. The stale cart shows the old burden returning through another interface. The user asks for help, receives fluent guidance, then opens more tabs to verify whether the machine has remembered the present correctly.

A developer known as Cookie shared logs in which text from a Perplexity session attributed its treatment of her to sexist pattern-matching. TechCrunch said it reviewed saved chats. Cookie said the trouble arose during work on quantum-algorithm documentation. Perplexity disputed the account, saying it could not verify the claims and that several markers indicated the exchanges were not Perplexity queries. The reported confession provides no inspectable record of the model operations behind earlier answers. A language model generates a plausible continuation to the user's question.

Evidence of bias requires controlled comparisons of outputs. Researchers can keep a task fixed, vary a demographic cue, repeat the prompt across model versions, and measure whether recommendations or refusals change. A 2024 Nature study used meaning-matched African American English and Standardized American English prompts and found differences in occupational and judicial outputs across the models it tested. That procedure creates observations another team can examine. Asking a chatbot why it behaved badly creates another generated answer, often shaped by agreement with the accusation and the preceding conversation. Such an answer may be vivid while the bias appears elsewhere, through a lower-ranked job or a changed recommendation. Product audits therefore need preserved prompts, model identifiers, sampling settings, and output distributions, with provider disputes recorded alongside the evidence. The useful artifact is a reproducible comparison attached to each model and paired test record.

Poetic reformulations of harmful requests produced unsafe responses across 25 language models in tests by researchers. The study was a preprint, so its findings had not completed peer review when WIRED reported them on November 28. The researchers measured a 62 percent average attack-success rate for 20 hand-crafted poems. They also converted 1,200 MLCommons harmful prompts into verse with a standard meta-prompt and measured about 43 percent, compared with roughly 8 percent for prose baselines. Outputs were labeled by three open-weight judge models. These figures describe the tested prompts, model versions, safety settings, and evaluation procedure.

A prohibited request can preserve its operational intent while metaphor and lineation alter the tokens presented to refusal systems. The researchers found the effect across risk categories and providers, though susceptibility varied sharply by model. Its cause remains unresolved. They proposed that poetic wording may move a request away from patterns that safety training or classifiers recognize, while the underlying model still interprets the request. That account is a hypothesis, since the team lacked access to proprietary model internals and did not isolate one common guardrail architecture. Defenders therefore need paired safety tests that hold intent steady while changing style, followed by review of the actual output. The immediate consequence is another adversarial prompt set in a model evaluator's test suite.

Meta is removing ChatGPT, Copilot, and other general AI assistants from WhatsApp under platform rules that restrict large language models from using the app as a distribution channel. The Verge reported that the change will leave Meta AI as the native assistant inside one of the world's largest messaging systems. WhatsApp is intimate infrastructure. It holds family logistics, work favors, voice notes, school groups, medical updates, neighborhood arguments, and the boring messages that keep a day moving. Whoever controls the assistant inside that space controls a very privileged doorway.

The decision turns AI competition into a question of pipes. A model may be clever, cheap, or popular, but it still needs an address inside the apps where people already speak. Meta can describe the move as platform management, load control, or user safety. The commercial effect is simpler. Rival assistants lose access to a daily habit while Meta's own assistant remains close to the send button. Users who thought they were choosing a bot may discover they were really borrowing permission from the owner of the room. The chat stays familiar, but the available voices narrow by policy.

AI-generated recipes and search summaries are sending home cooks toward bad instructions while draining traffic from food bloggers, according to reporting cited by The Verge. The examples are comic until they reach the oven. A cake can be assigned the wrong pan, the wrong time, or a temperature that turns dessert into charcoal. At the same time, Google-style answer boxes and scraped recipe pages can keep readers away from the people who tested the dish, photographed the steps, adjusted measurements, and answered comments from confused cooks.

Cooking exposes a weakness that polished text can hide. A paragraph may sound fluent, but butter melts, yeast dies, sugar burns, chicken stays raw, and an oven refuses rhetorical confidence. Food writing has always included memory, trial, substitution, weather, family habit, and the unglamorous labor of washing the same bowl again. An automated recipe can imitate that format while skipping the kitchen. The cost then lands twice. The blogger loses the visit that paid for testing, and the cook loses ingredients, time, and trust at the counter. Search turns a meal into a generated instruction, and the correction happens only after dinner fails.

A leaked White House draft sought to block state AI laws by tying federal funding, agency pressure, and legal strategy to a broad preemption campaign, according to The Verge. The effort, associated with David Sacks, treated state legislatures as obstacles to the national AI race. The proposal met resistance from lawmakers, advocates, and some Republicans who questioned both the legality and the politics of using executive power to freeze local rules. A technology sold as too fast for government suddenly needed Washington to slow down every smaller government first.

The fight gives the AI industry a preferred map. Companies want one national lane, fewer conflicting obligations, and a market where product updates outrun local experiments in privacy, safety, labor, education, and liability. States often write messy laws, but they also become the first place where residents can pressure officials after a chatbot harms a teenager, a model rejects a worker, or a school buys surveillance software. Preemption would move that argument upward, away from state hearings and toward federal offices closer to the companies asking for speed. The rulebook would start with a missing page where local complaint used to go.

A temporary restraining order barred OpenAI from using the word “Cameo” for a Sora feature that lets people insert themselves or characters into generated videos. The Verge reported that Cameo, the celebrity video-message service, secured the order while a trademark hearing remained pending. The dispute sounds narrow, almost comic, because it turns on a product name. It also points to a larger pressure around synthetic appearance. A platform that can generate a body, scene, and performance also needs language for the act of entering the image.

The name matters because AI video turns self-insertion into a feature with legal edges. A user can place a face into a clip, borrow a character, simulate a performance, or make a person appear inside a scene they never filmed. The older Cameo business sold a paid address from a recognizable person to a fan. Sora's version describes a slot inside a generated world, where likeness and participation become interface options. Trademark law is handling the first dispute because it has a clean object to grasp. The harder argument will concern whose face, role, and invitation are being packaged each time the button is renamed.

Great American, Chubb, and W. R. Berkley asked U.S. regulators for permission to exclude broad AI liabilities from corporate insurance policies, according to Financial Times reporting summarized by TechCrunch on November 23. An underwriter told the FT that model outputs were too opaque to price confidently. The same reporting named AIG, but AIG later told TechCrunch that it was not seeking to use the exclusions discussed and had no plans to implement them. The cited cases ranged from a $110 million lawsuit over a false Google AI Overview to Air Canada being required to honor a discount supplied by its chatbot. FT-derived claims about the insurers' filings and industry concerns remain attributed to that reporting.

Insurance depends on grouping independent losses closely enough that premiums from many policyholders can absorb a claim. A model embedded across thousands of businesses can disturb that calculation because one defect or repeated false output may generate correlated claims at once. An Aon executive's example contrasted a $400 million loss at one company with 10,000 losses triggered by one agentic system. Exclusion clauses preserve the insurer's balance sheet by moving that exposure outside the contract. Businesses then retain the cost, seek narrower specialist cover, restrict deployment, or pass losses to workers and customers. The exclusion is written into the policy before the first disputed claim reaches an adjuster.

Udio users lost the ability to download AI-generated music after the company changed its terms following a settlement with Universal, according to The Verge. The restriction angered people who had treated the service as a studio, archive, and outlet for their own outputs. They had typed prompts, refined tracks, paid subscriptions, and built private libraries around the assumption that creation ended with a file. The new rule makes the file less like property and closer to a controlled playback privilege granted by the platform.

That shift exposes the weak ownership at the center of many generative services. A user may feel authorship because the interface responds to their prompt, their taste, their revision, and their time. The contract may still say that access can change, downloads can stop, disputes can be routed away from class action, and the work remains trapped inside business negotiations the user never joined. AI music promised quick production without musicians, studios, or labels. The settlement pulls the label back into the room through terms of service. A song can exist in an account, fill a playlist, and still fail to become a file the maker can keep.