Authors Challenge Publisher Claims as Anthropic Copyright Settlement Payouts Begin

01Anthropic copyright settlement enters payout phase as authors challenge claims by publishers and literary agencies

Authors expecting payments from Anthropic’s $1.5 billion copyright settlement are challenging claims filed by publishers and literary agencies that they say are seeking money allocated to their books. Some writers received emails this week notifying them that another party had claimed part or all of their payments.

Anthropic settled the class-action lawsuit last year after a judge ruled that training AI models on copyrighted material qualified as fair use, but obtaining that material through piracy did not. The settlement received final approval in July and covers nearly 500,000 titles, with $3,000 available for each pirated work.

The allocation depends on the book’s publishing status. For a traditionally published book still in print, the author and publisher generally split the payment equally. An author should receive the full amount for a self-published book or one whose rights were returned after the publisher allowed it to go out of print.

Writers Beware, a publishing-industry watchdog blog, said the complaints it received repeatedly described two problems: publishers claiming works whose rights had already reverted to the author, and publishers claiming 100% of payments that should be divided 50-50.

Mystery and thriller writer April Henry, for example, said HarperCollins claimed a book whose rights had reverted at least 17 years earlier. Her account illustrates how old publishing records can collide with the settlement’s allocation process, although the source does not establish whether the claim was ultimately accepted.

Some publishers have described their claims as mistakes and asked Anthropic to correct them. Writers Beware’s Victoria Strauss said poor recordkeeping could plausibly explain the disputes. Authors Guild CEO Mary Rasenberger likewise attributed the problems to inadequate records and a confusing settlement process rather than a deliberate attempt by publishers to take authors’ money.

Literary agencies have also appeared in claim notices. Strauss said Writers Beware received complaints involving several agencies, even though agents are not rights holders in the books they sell. The source does not explain the contractual or administrative basis of those agency claims, or establish whether any will be upheld.

Authors can dispute their payment allocations, but the timing of a rights reversion matters. To claim the entire payment for a previously published book, an author must show that the rights reverted before August 10, 2022, the settlement’s designated “download date.” The total number of affected authors and works—and the ultimate validity of the contested claims—remain unknown.

02Kalanick’s Atoms reportedly prepares hiring and acquisitions, explores robotaxi collaboration with Uber

Atoms, the startup founded by Uber co-founder Travis Kalanick, is beginning to reveal a possible route into the autonomous-vehicle industry after saying little about its broader ambitions. Earlier this summer, the company announced a $1.7 billion funding round led by Andreessen Horowitz, but Kalanick did not spell out exactly what the capital would support.

Atoms is now preparing a hiring spree and acquisitions that could turn it into a major autonomous-vehicle player, TechCrunch reported, citing the Financial Times. The report did not specify how many people Atoms plans to hire, which companies it may acquire, or when it expects to commercialize any products.

One possible application is robotaxis—self-driving vehicles used for paid passenger trips. According to the report, Atoms has discussed with Uber how the ride-hailing company could use its robotaxi technology. Those talks have not been described as a formal partnership or customer order, and no technical details or deployment timetable were disclosed.

Uber already has a financial interest in Atoms. The company invested $100 million in the startup, an amount previously confirmed by TechCrunch. It has also established partnerships with a long list of autonomous-vehicle companies, giving it experience working with outside developers of self-driving systems. For now, however, the reported discussions with Atoms remain exploratory.

Atoms has taken a more concrete step through its acquisition of Pronto, an autonomous-mining startup led by Anthony Levandowski, Uber’s former self-driving chief. The deal brings an existing autonomous-vehicle business and an experienced industry executive into Atoms, making the acquisition one of the clearest signs of the direction Kalanick may pursue.

Robotaxis are not the whole plan. Sources cited in the report emphasized that they represent only part of Atoms’ intended business, although no fuller description was provided. That leaves the startup’s ultimate product range unresolved even as its financing, planned recruitment, acquisition strategy, Pronto deal, and talks with Uber establish an emerging autonomous-vehicle path.

03Why Restaurant AI Menus Look Smooth, Symmetrical, and Unsettling: Diffusion Models’ Structural and Texture Weaknesses

Restaurants, cafés, and brands are increasingly using generative AI to create promotional food images. The results range from obviously malformed dishes—wormlike noodles, hole-riddled burritos, and stringy chicken—to subtler menus in which every sandwich is unnaturally symmetrical, rounded, and smooth.

These distortions begin with how many image generators work. Diffusion models start with visual noise and progressively turn it into an image, recovering broad structures before adding fine details. According to Oxford computer-vision professor Chris Russell, the basic shape may already be wrong when the model overlays convincing colors and textures, allowing a malformed ingredient to look polished without becoming physically coherent.

Thin, continuous structures are particularly difficult. Behavioral scientist Giovanbattista Califano said diffusion models struggle to generate forms that must remain connected and end at a logical point. Noodles and tendrils can therefore spread into unrelated parts of a dish, while repeating details such as bubbles and seeds may cross sensible boundaries. The models statistically reproduce what food looks like without understanding ingredients, objects, or their physical relationships, so textures appropriate for concrete or other materials can appear on something meant to be edible. The Verge

A separate process helps explain why menus from different restaurants can look alike even when they contain no glaring structural errors. TechCrunch’s interviewees said training data and model optimization favor pleasing, inoffensive imagery, stripping away differences and pushing outputs toward glossy lighting, rounded forms, smooth surfaces, and familiar chain-restaurant aesthetics. The precise composition of individual models’ training data remains unknown.

Repeated editing can intensify that sameness. TechCrunch reproduced an experiment in which an AI-generated menu was edited many times and found that the food became progressively rounder and smoother. Restaurants changing prices, names, or small design elements may subject an image to this process repeatedly, though the source does not quantify the degradation caused by each edit. Interviewees characterized the effect as convergence or homogenization, not necessarily the more severe “model collapse” associated with excessive training on AI-generated output.

Consumers’ discomfort is strongest when the result nearly passes as real. Researchers at Germany’s University of Duisburg-Essen found an “uncanny valley” effect: almost-real AI food images prompted more disgust and unease than clearly artificial ones. TechCrunch