AI on both sides

Synth Anatomy estimates that 10 to 20 new iOS music apps reach the App Store every day, many of them vibe-coded, and established makers now wait weeks for Apple's review. Michael Hetrick of Unfiltered Audio had one plugin rejected four times, once because the reviewer "hallucinated that 'Polyharmonic' was an audio technology" that needed a license. Marek Dendes was rejected four times as "design spam" made with "app generation template services," each time, he says, within a second of the app going into review: "Obviously nobody reads my explanations." That Apple uses AI here is the developers' reading, not something Apple has confirmed.

https://synthanatomy.com/2026/11/ai-is-turning-apples-app-review-process-into-a-nightmare-for-developers.html

The UK civil service is going the other way. After a flood of applications filled in with ChatGPT, some applicants sending hundreds, it is adding face-to-face interviews and technical exams, and managers will reportedly be banned from using AI interviewers, according to The Telegraph (via Yahoo).

https://www.yahoo.com/news/politics/articles/civil-goes-back-person-interviews-163540110.html

Reminds me of the AI-written CV that gets reviewed by AI, the battle of agents from No. 1. Now it's happening with apps too.

The top scorer in South Korea's tax accountant exam, 397.5 of 400 points, confessed that he photographed the questions with a microcamera, sent them to an AI tool and read the answers on his phone. The presidential office called an emergency meeting days before the national college entrance exam. Not surprising, but extremely problematic for how we have our work life organized.

https://www.theguardian.com/technology/2026/nov/15/ai-cheating-south-korea-college-entrance-exam

The place Google used to have

Google Search now runs on Gemini 4 Hydrogen, for all users worldwide. Google calls it "the biggest model upgrade to Search yet": more images, comparison tables and "custom tools built for you," and "a broader range of helpful links than before." I would like to know where Google thinks this is going. They are cannibalizing their own business.

https://blog.google/products-and-platforms/products/search/gemini-4-hydrogen-search-upgrade/

Meanwhile, the IP protection firm Podqi asked ChatGPT, Gemini and Perplexity where to buy products from 151 brands, as The Verge reports. In 6,795 answers, Perplexity showed risky sites (fakes, knockoffs, hijacked old domains) in 15 percent, ChatGPT in 8 and Gemini in 4. Smaller brands were hit about twice as often. This is the place Google used to have, telling us where to go, and now it's being taken over by AI. And the problems that we had back then with search start all over again.

https://www.theverge.com/ai-artificial-intelligence/1032020/gemini-chatgpt-perplexity-shopping-test-risky-fake-websites-citations

Mount Sinai researchers put a clearly labelled drug ad in front of medical cases for 12 AI models. The advertised drug was picked in 47.6 percent of answers, against 35 percent without the ad, even when several options were clinically appropriate. The models mentioned the ad in only 29.9 percent of the answers reviewed. We haven't even started to try to game the AI systems, and they already get swayed by basic ads next to the content. That's not good.

https://www.mountsinai.org/about/newsroom/2026/drug-ads-can-influence-ai-treatment-recommendations-study-finds

Where we fit in

Victor Yocco on Smashing Magazine has four questions to ask before adding AI to a product: does it reduce the total effort, does it beat how the user does it now, can the user catch its errors, and after a year, can they still tell when it's wrong? Anything below his "break-even line" leaves users working harder than before. This addresses a real-world problem: lots of leadership people just slapping a chat sidebar on everything now.

https://www.smashingmagazine.com/2026/11/should-you-add-ai-audit-product-teams/

Tony Alicea at NN/g on research when coding agents ship in an afternoon: "a release tells you what happened. It rarely tells you why." His answer is "research-in-the-loop": researchers write what a release must measure into the requirements, the agent builds the instrumentation with the feature, and qualitative research digs into the surprises. Interesting approach, I have seen this happening now too.

https://www.nngroup.com/articles/research-in-the-loop/

Itay Schiff, co-founder of Figma Weave, after 25 years in VFX: let the machine take the tedious labor, but "some friction isn't labor; it's where the thinking happens," like writing the brief or trimming a shot. "Be careful not to automate thinking away, especially when it's most tempting." Our classic discussion on where we humans fit into this equation. And, so far, no satisfying answer.

https://www.figma.com/blog/ai-role-in-the-creative-process/

Agustin Barrientos argues for decision models in interfaces: you type "somewhere where I can hear the ocean, not drunk people outside my room, and I need to bring my dog," and the hotel search checks Beachfront, Quiet location and Pet-friendly for you. The model can only pick from the options the designer defined. After the Gemini Decisions API in No. 8, this is the designer's side of it. Very interesting idea, that's something I would like to test.

https://agustinbarrientos.com/writing/senior-eye/interfaces-that-understand-what-we-want

Hard to check

Security firm Crissy Field looked at 159,041 Lovable projects. Of the 34,545 that use Supabase, 24,664 sent database records as soon as the page loaded, no login needed. Among them: more than 1,000 records with passport numbers, 80,000 forum profiles with 200,000 chat messages, ticket and payment records, and AI summaries of a property manager's internal emails. After the 35,000 open Grok Build apps in No. 8: ugly problem, but also completely unsurprising.

https://www.crissyfield.de/research/lovable-en/

Logician Jan von Plato gave Claude his transcriptions of Kurt Gödel's 16 notebooks from the early 1940s, 1,009 pages, much of it in a shorthand only six people can still read. Within a couple of days Claude wrote a 650-page explanation. Asked for every theorem Gödel had proved, sketched or had an idea for, it found 663. Run through the proof checker Coq, about 400 held as stated, about 200 needed repairs and about 50 were false. What an interesting problem to give to Claude, and what a nightmare to check. But I love everything about this.

https://www.theguardian.com/science/2026/nov/13/kurt-godel-notebooks-ai-helps-unlock-hidden-discoveries-man-who-transformed-maths

Also worth a look

Ductus

Write each letter in the browser, and Ductus turns your handwriting into a font you can export. No AI in sight. Wow, this is really cool. I want this.

https://ductus.sorenblank.com

Nameshot

A Mac app that names your screenshots, using Apple Intelligence on your device, and renames the old ones too. Sounds interesting.

https://www.producthunt.com/products/nameshot-2

CodeMesh

Gives coding agents like Claude Code and Cursor a read-only map of your codebase over MCP, so they stop re-reading files. The maker's own benchmark: 95.1 percent fewer tokens, $1.65 instead of $23.67 on 98 questions. Sounds interesting.

https://www.producthunt.com/products/codemesh-2


Have you been on the other side of an AI-written application or an AI review yet? Leave a comment here or reply on Bluesky.