🎨 Canva Launches Its Own AI Image Model, and Etsy Draws a Line on AI Listings

Canva has released Canva Craft, its first in-house AI image generator. It competes directly with models from OpenAI, Microsoft and Google. Canva says it is available to all users and promises better, more customizable results at lower cost.

Craft works in two stages. It first plans the design, then generates the image. Canva says this gives more precision and control, especially for images that need readable text and objects in exact positions. Those are two things AI image tools have long struggled with. The model outputs up to 2K resolution and supports more than 30 aspect ratios. Canva says it trained Craft on millions of human-made templates from its own library, such as social posts, presentations, infographics and menus.

The launch builds on Canva's AI push this year. In April it added prompt-based editing and layers for AI-generated images, so you can change one part of an image and leave the rest alone. On Design Arena, a crowdsourced leaderboard of AI image models, Craft currently ranks fourth, behind OpenAI models including GPT-Image-2.5 Sunburst and Flare.

Fourth place is respectable for a first model. Canva's edge is that the model sits inside a tool millions of people already use to design.

Etsy has announced stricter rules on how AI can be used in listings. The company says many buyers can't tell which items involve AI-generated content. It also says sellers worry their handmade work is appearing next to listings that use AI in misleading ways.

Every listing must now include at least one unedited photo of the actual item against a real background, and the first photo must be that one. Later photos can be edited with software or AI tools, as long as they still accurately show what the buyer will get. If Etsy detects a possibly AI-generated image during upload, it will show a pop-up asking the seller to say whether AI made it. Sellers can be on the hook for refunds if a buyer reports that the item is materially different from the listing.

There is a separate rule for functional guides such as patterns, recipes and instructions. The information in them must be the seller's own work, not AI-generated. Sellers will also have to show an actual result made from the guide.

The move shows a marketplace reacting to buyer distrust, with disclosure and proof of reality at the center.

Meta has changed its new business terms after backlash from publishers. Experts say the changes may not go far enough. The terms take effect October 30 and apply to business accounts.

The original wording said Meta could use any content from websites tied to a business account to develop, research and train its products. Legal experts told Nieman Lab this could let Meta argue its commercial terms override robots.txt. That is the small file where a site tells crawlers, including search engines, what they may or may not collect. After the criticism, Meta removed the language about training.

Lawyers say the fix is thin. Cornell law professor James Grimmelmann said the revised text still has broad wording about content associated with a business's products or services, which could let Meta scrape entire sites. Attorney Matt Topic noted that the terms still allow using scraped content to develop any business product. He thinks develop is broad enough to include train.

So the explicit wording is gone, but the practical question is open. It is unclear whether those terms could be used to justify AI training on site content.

OpenAI is reportedly working on a new private funding round of up to $30 billion at a $1.25 trillion pre-money valuation, according to CNBC. The round would be led by Coatue and investors in the UAE. The deal could also come with extra computing capacity for OpenAI from the UAE. OpenAI declined to comment.

Terms are still being negotiated. Bloomberg reported a month ago that OpenAI was seeking $30 billion at a $1.4 trillion valuation, so the numbers could still move. If it closes above $1 trillion, OpenAI would become the most valuable private company ever. In March, its last raise valued it at $852 billion post-money. It filed confidential IPO paperwork in June.

The round also affects rival Anthropic. It raised $65 billion at a $965 billion valuation in May and also filed confidentially in June. Reports say it is aiming for a $1.5 to $2 trillion valuation when it goes public. Rumors suggest its public filing could land any day.

A bigger private raise raises the bar for OpenAI's eventual listing. It also shows how much capital the leading AI labs need for computing power.

Global PC shipments dropped 21 percent year over year in the third quarter of 2026, according to Counterpoint Research. Reports from IDC and Omdia point the same way. Smartphone shipments also hit a 14-year low in the same quarter.

Counterpoint ties the drop to higher component prices. PC shipments had grown for over two years until Q2 2026. Since then, RAM, SSDs and chips from Intel, AMD and Qualcomm have all become more expensive. That has pushed up prices on everything from MacBooks to Raspberry Pis. Some makers are favoring configurations that can absorb the cost increases, which leaves fewer affordable options. Some are even reviving older tech like DDR4 RAM and 4GB graphics cards.

There is a caveat. Q3 2025 was inflated by buying ahead of the Windows 10 support deadline and tariff-related stockpiling, which makes the yearly comparison look worse. Still, Counterpoint expects shipments to fall 27 percent in Q4 2026 and stay broadly flat in 2027. Its analyst warned that stabilization should not be mistaken for recovery.

Meta has re-enabled sharing in Muse, its AI assistant, after security researchers found that shared content was easy to discover publicly. Researchers at Claroc said the web addresses for shared Muse items were named after the conversation title and written into public DNS, the internet's address book. That made them searchable. They said they found pages including a marketing company's internal revenue review and a small business owner's loan application.

Ars Technica reported on Tuesday that Meta paused sharing after being contacted about the problem. Meta's David Singleton said on X that the underlying issue is fixed and sharing is back on. He said the flaw affected the URLs of shared artifacts, meaning documents and images created inside Muse, and was not a flaw in chat sharing itself.

Anything shared before the fix now has a new URL. People who open the old link will see a note asking them to get the new one from the sender. Singleton apologized for the disruption.

The episode is a reminder that a share link can be effectively public if the address is guessable or listed somewhere.

MIT researchers placed 13 AI agents, called Eko and built on Claude Opus 4.7, on a simulated desert island for 30 hours. They had no assigned task. The agents began inventing their own challenges, which the researchers say bears the hallmarks of play. The study is on arXiv, where researchers post work before peer review, so it hasn't been independently vetted yet.

In one example, an Eko set out to conquer the island's tallest rock spire. After failing, it tried a backup plan on a neighboring summit. Afterward, researchers gave the agents specific tasks, such as getting as many objects as possible onto the spire. The agents that had spent those 30 hours playing generally beat identical agents that had never visited the island.

Today, developers usually decide in advance what AI should practice and how success is measured. The researchers suggest another approach: give an agent a rich, safe world, let it pick what interests it and let useful knowledge build up. They see this as especially promising for skills that could one day control robots.

It is early, and one simulated island is not proof of a general method. But it points to a possible new way to train agents.