Critiques are the lifeblood of design. Anyone who went to design school has participated in and has been the focus of a crit. It’s “the intentional application of adversarial thought to something that isn’t finished yet,” as Fabricio Teixeira and Caio Braga, the editors of DOC put it.

A lot of solo designers—whether they’re a design team of one or if they’re a freelancer—don’t have the luxury of critiques. In my view, they’re handicapped. There are workarounds, of course. Such as critiques with cross-functional peers, but it’s not the same. I had one designer on my team—who used to be a design team of one in her previous company—come up to me and say she’s learned more in a month than a year at her former job.

Further down, Teixeira and Braga say:

In the age of AI, the human critique session becomes even more important. LLMs can generate ideas in 5 seconds, but stress-testing them with contextual knowledge, taste, and vision, is something that you should be better at. As AI accelerates the production of “technically correct” and “aesthetically optimized” work, relying on just AI creates the risks of mediocrity. AI is trained to be predictable; crits are all about friction: political, organizational, or strategic.

As regular readers will know, the design talent crisis is a subject I’m very passionate about. Of course, this talent crisis is really about how companies who are opting for AI instead of junior-level humans, are robbing themselves of a human expertise to control the AI agents of the future, and neglecting a generation of talented and enthusiastic young people.

Also obviously, this goes beyond the design discipline. Annie Hedgpeth, writing for the People Work blog, says that “AI is replacing the training ground not replacing expertise.”

We used to have a training ground for junior engineers, but now AI is increasingly automating away that work. Both studies I referenced above cited the same thing - AI is getting good at automating junior work while only augmenting senior work. So the evidence doesn’t show that AI is going to replace everyone; it’s just removing the apprenticeship ladder.

From the Sep 2025 Harvard University paper, “Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data.” (link)

And then she echoes my worry:

So what happens in 10-20 years when the current senior engineers retire? Where do the next batch of seniors come from? The ones who can architect complex systems and make good judgment calls when faced with uncertain situations? Those are skills that are developed through years of work that starts simple and grows in complexity, through human mentorship.

We’re setting ourselves up for a timing mismatch, at best. We’re eliminating junior jobs in hopes that AI will get good enough in the next 10-20 years to handle even complex, human judgment calls. And if we’re wrong about that, then we have far fewer people in the pipeline of senior engineers to solve those problems.

I’ve been playing with my systems in the past month—switching browsers, notetaking apps, and RSS feed readers. If I’m being honest, it’s causing me anxiety because I feel unmoored. My systems aren’t familiar enough to let me be efficient.

One thing that has stayed relatively stable is my LLM app—well, two of them. ChatGPT for everyday and Claude for coding and writing.

Christina Wodtke, writing on her blog:

The most useful model might not win.

What wins is the model that people don’t want to leave. The one that feels like home. The one where switching would mean losing something—not just access to features, but fluency, comfort, all those intangible things that make a tool feel like yours.

Amazon figured this out with Prime. Apple figured it out with the ecosystem. Salesforce figured it out by making itself so embedded in enterprise workflows that ripping it out would require an act of God.

AI companies are still acting like this is a pure technology competition. It’s not. It’s a competition to become essential—and staying power comes from experience, not raw capability.

Your moat isn’t your model. Your moat is whether users feel at home.

I wouldn’t call myself a gamer, but I do enjoy good games from time to time, when I have the time. A couple of years ago, I made my way through Hades and had a blast.

But I do know that the publishing of a triple-A title like Call of Duty: Black Ops takes an enormous effort, tons of human-hours, and loads of cash. It’s also obvious to me that AI has been entering into entertainment workflows, just like it has in design workflows.

Ian Dean, writing for Creative Bloq explores this controversy with Activision using generative AI to create artwork for the latest release in the Call of Duty franchise. Players called the company out for being opaque about using AI tools, but more importantly, because they spotted telltale artifacts.

Many of the game’s calling cards display the kind of visual tics that seasoned artists can spot at a glance: fingers that don’t quite add up, characters whose faces drift slightly off-model, and backgrounds that feel too synthetic to belong to a studio known for its polish.

These aren’t high-profile cinematic assets, but they’re the small slices of style and personality players earn through gameplay. And that’s precisely why the discovery has landed so hard; it feels a little sneaky, a bit underhanded.

“Sneaky” and “underhanded” are odd adjectives, no? I suppose gamers are feeling like they’ve been lied to because Activition used AI?

Dean again:

While no major studio will admit it publicly, Black Ops 7 is now a case study in how not to introduce AI into a beloved franchise. Artists across the industry are already discussing how easily ‘supportive tools’ can cross the line into fully generated content, and how difficult it becomes to convince players that craft still matters when the results look rushed or uncanny.

My, possibly controversial, view is that the technology itself isn’t the villain here; poor implementation is, a lack of transparency is, and fundamentally, a lack of creative use is.

I think the last phrase is the key. It’s the loss of quality and lack of creative use.

I’ve been playing around more with AI-generated images and video, ever since Figma acquired Weavy. I’ve been testing out Weavy and have done a lot of experimenting with ComfyUI in recent weeks. The quality of output from these tools is getting better every month.

With more and more AI being embedded into our art and design tools, the purity that some fans want is going to be hard to sustain. I think the train has left the station.

Geoffrey Litt is a design engineer at Notion. He is one of the authors at Ink & Switch of “Malleable software,” which I linked to back in July. I think it’s pretty fitting that he popped up at Notion, with the CEO Ivan Zhao likening the app to LEGO bricks.

In a recent interview with Rid on Dive Club, Litt explains the concept further:

So, when I say malleable software, I do not mean only disposable software. The main thing I think about with malleable software is actually much closer to … designing my interior space in my house. Let’s say when I come home I don’t want everything to be rearranged, right? I want it to be the way it was. And if I want to move the furniture or put things on the wall, I want to have the right to do that. And so I think of it much more as kind of crafting an environment over time that’s actually more stable and predictable, not only for myself, but also for my team. Having shared environments that we all work in together that are predictable is also really important, right? Ironically, actually, in some ways, I think sometimes malleable software results in more stable software because I have more control.

For building with AI, Litt advocates “coding like a surgeon”: stay in the loop and use agents for prep and grunt work.

How do we think of AI as a way to leverage our time better? [So we can] stay connected to the work and [do] it ourselves by having prep work done for us. Having tools in the moment helping us do it so that we can really focus on the stuff we love to do, and do less of everything else. And that’s how I’m trying to use coding agents for my core work that I care about today. Which is when I show up, sit down at my desk in the morning and work on a feature, I want to be prepped with a brief on all the code I’m going to be touching today, how it works, what the traps are. Maybe I’ll see a draft that the AI did for me overnight, sketching out how the coding could go. Maybe some ideas for me.

In other words, like an assistant who works overnight. And yeah, this could apply to design as well.

He told me his CEO - who’s never written a line of code - was running their company from an AI code editor.

I almost fell out of my chair.

OF COURSE. WHY HAD I NOT THOUGHT OF THAT.

I’ve since gotten rid of almost all of my productivity tools.

ChatGPT, Notion, Todoist, Airtable, Google Keep, Perplexity, my CRM. All gone.

That’s the lede for a piece by Derek Larson on running everything from Claude Code. I’ve covered how Claude Code is pretty brilliant and there are dozens more use cases than just coding.

But getting rid of everything and using just text files and the terminal window? Seems extreme.

Larson uses a skill in Claude Code called “/weekly” to do a weekly review.

1. Claude looks at every file change since last week
2. Claude evaluates the state of projects, tasks, and the roadmap
3. We have a conversation to dig deeper, and make decisions
4. Claude generates a document summarizing the week and plan we agreed on

Then Claude finds items he’s missed or procrastinating on, and “creates a space to dump everything” on his mind.

Pavel Bukengolts writes a piece for UX Magazine that reiterates what I’ve been covering here: our general shift to AI means that human judgement and adaptability are more important than ever.

Before getting to the meat of the issue, Bukengolts highlights the talent crisis that is our own making:

The outcome is a broken pipeline. If graduates cannot land their first jobs, they cannot build the experience needed for the next stage. A decade from now, organizations may face not just a shortage of junior workers, but a shortage of mid-level professionals who never had a chance to develop.

If rote repetitive tasks are being automated by AI and junior staffers aren’t needed for those tasks, then what skills are still valuable? Further on, he answers that question:

Centuries ago, in Athens, Alexandria, or Oxford, education focused on rhetoric, logic, and philosophy. These were not academic luxuries but survival skills for navigating complexity and persuasion. Ironically, they are once again becoming the most durable protection in an age of automation.

Some of these skills include:

• Logic: Evaluating arguments and identifying flawed reasoning—essential when AI generates plausible but incorrect conclusions.
• Rhetoric: Crafting persuasive narratives that create emotional connection and resonance beyond what algorithms can achieve.
• Philosophy and Ethics: Examining not just capability but responsibility, particularly around automation’s broader implications.
• Systems Thinking: Understanding interconnections and cascading effects that AI’s narrow outputs often miss.
• Writing: Communicating with precision to align stakeholders and drive better outcomes.
• Observation: Detecting subtle signals and anomalies that fall outside algorithmic training data.
• Debate: Refining thinking through intellectual challenge—a practice dating to ancient dialogue.
• History: Recognizing recurring patterns to avoid cyclical mistakes; AI enthusiasm echoes past technological revolutions.

I would say all of the above not only make a good designer but a good citizen of this planet.

In a heady, intelligent, and fascinating interview with Sarah Jeong from The Verge, Cory Doctorow—the famed internet activist—talks about how platforms have gotten worse over the years. Using Meta (Facebook) as an example, Doctorow explains their decline over time through a multi-stage process. Initially, it attracted users by promising not to spy on them and by showing them content from their friends, leveraging the difficulty of getting friends to switch platforms. Subsequently, Meta compromised user privacy by providing advertisers with surveillance data (aka ad tracking) and offered publishers traffic funnels, locking in business customers before ultimately degrading the experience for all users by filling feeds with paid content and pivoting to less desirable ventures like the Metaverse.

And publishers, [to get visibility on the platform,] they have to put the full text of their articles on Facebook now and no links back to their website.

Otherwise, they won’t be shown to anyone, much less their subscribers, and they’re now fully substitutive, right? And the only way they can monetize that is with Facebook’s rigged ad market and users find that the amount of stuff that they ask to see in their feed is dwindled to basically nothing, so that these voids can be filled with stuff people will pay to show them, and those people are getting ripped off. This is the equilibrium Mark Zuckerberg wants, right? Where all the available value has been withdrawn. But he has to contend with the fact that this is a very brittle equilibrium. The difference between, “I hate Facebook, but I can’t seem to stop using it,” and “I hate Facebook and I’m not going to use it anymore,” is so brittle that if you get a live stream mass shooting or a whistleblower or a privacy scandal like Cambridge Analytica, people will flee.

Francesca Bria and her collaborators analyzed open-source datasets of “over 250 actors, thousands of verified connections, and $45 billion in documented financial flows” to come up with a single-page website visually showing such connections.

J.D. Vance, propelled to the vice-presidency by $15 million from Peter Thiel, became the face of tech-right governance. Behind him, Thiel’s network moved into the machinery of the state.

Under the banner of “patriotic tech”, this new bloc is building the infrastructure of control—clouds, AI, finance, drones, satellites—an integrated system we call the Authoritarian Stack. It is faster, ideological, and fully privatized: a regime where corporate boards, not public law, set the rules.

Our investigation shows how these firms now operate as state-like powers—writing the rules, winning the tenders, and exporting their model to Europe, where it poses a direct challenge to democratic governance.

I must admit I’ve tried to read this essay by Frank Chimero—a script from a talk he recently gave—for about a week. I tried to skim it. I tried to fit it into a spare five minutes here and there. But this piece demands active reading. Not because it’s dense. But because it is great.

Chimero reflects on AI and his—and our—relationship to it. How is it being marketed? How do we think about it? How should we use it?

First off, Chimero starts with his conclusion. He believes we should reframe AI to be less like a tool or technology, and more like a musical instrument.

Thinking of AI as an instrument recenters the focus on practice. Instruments require a performance that relies on technique—the horn makes the sound, but how and what you blow into it matters; the drum machine keeps time and plays the samples, but what you sample and how you swing on top of it becomes your signature.

In other words, instruments can surprise you with what they offer, but they are not automatic. In the end, they require a touch. You use a tool, but you play an instrument. It’s a more expansive way of doing, and the doing of it all is important, because that’s where you develop the instincts for excellence. There is no purpose to better machines if they do not also produce better humans.

Then, he wanders off to give examples of four artists and their relationships with technology, stoking his audience—me, us, you—to consider “some more flexibility in how to collaborate with the machine in your own work, creative or otherwise.”

Read the whole piece. Curl up this mid-autumn Sunday afternoon with some hot tea and take the 20–25 minutes to read it and take it in.

Chris Butler wrestles with a generations-old problem in his latest piece: new technologies shortcut the old ways of doing things and therefore quality takes a nosedive. But is it different this time with the tools available to us today?

While design is more accessible than ever, with Adobe experimenting with chat interfaces and Canva offering pro-level design apps for free, putting a tool into the hands of someone doesn’t mean they’ll know how to wield it.

Anyone can now create something that looks professional, that uses modern layouts and typography, that feels designed. But producing something that feels designed does not mean that any design has happened. Most tools don’t ask you what you want someone to do. They don’t force you to make hard choices about hierarchy and priority. They offer you options, and if you don’t already understand the fundamentals of how design guides attention and serves purpose, you’ll end up using too many of them to no end.

Butler concludes that as designers, we’re in a bind because “the pace of change is only accelerating, and it is a serious challenge to designers to determine how much time to spend keeping up.”

You can’t build foundational knowledge while chasing the new. But you can’t ignore the new entirely, or you’ll fall behind. So you split your time, and both efforts can suffer. The fundamentals remain elusive because you’re too busy keeping up. The tools remain half-learned because you’re too busy teaching [design fundamentals to clients].

Butler—nor I—know if there’s a good solution to this problem. Like I said at the start, this is an age-old problem. Friction is a feature, not a bug.

This is just the reality of working in a field that sits at the intersection of human behavior and technological change. Both move, but at different speeds. Human attention, cognition, emotion — these things change slowly, if at all. Technology changes constantly. Design has to navigate both.

And while Butler’s essay never explicitly mentions AI or AI tools, it’s strongly implied. Developers using AI tools to code miss out on the fundamentals of building software. Designers (or their clients) using AI to design face the issues brought up here. Those who use AI to accelerate what they already know, that seems to be The Way.

While this piece by Matias Heikkilä is from a developer’s point-of-view, it’s applicable to designers. He poses a conceit: LLMs are good at coding, but can’t see the bigger picture and build software. To be sure, Cursor and Claude Code reason and produce plans. I’ve given both fairly small products to build. Their plans look good, but when they try to implement, invariably they’ll hit a snag. They’ll confidently say “It’s done!” with a green checkmark emoji. And then when I go to run it, the program invariably fails.

Heikkilä, writing in his company’s blog:

There is old wisdom that says: Coding is easy, software engineering is hard. It seems fair enough to say that LLMs are already able to automate a lot of coding. GPT-5 and the like solve isolated well-defined problems with a pretty nice success rate. Coding, however, is not what most people are getting paid for. Building a production-ready app is not coding, it’s software engineering.

Robin Sloan wrote a thought piece exploring what “extended thinking” and “reasoning” models actually mean.

…the models can only “think” by spooling out more text — while human thinking often does the oppo­site: retreats into silence, because it doesn’t have words yet to say what it wants to say.

That’s an interesting point Sloan makes. I believe there’s nuance though.

I’ve long felt that I do my best thinking by writing. When I work through a gnarly design problem, I’m writing first, then sketching, then maybe Figma-ing. But that could be after a walk, a shower, or doing the dishes.

In a very gutsy move, Grammarly is rebranding to Superhuman. I was definitely scratching my head when the company acquired the eponymous email app back in June. Why is this spellcheck-on-steroids company buying an email product?

Turns out the company has been quietly acquiring other products too, like Coda, a collaborative document platform similar to Notion, building the company into an AI-powered productivity suite.

So the name Superhuman makes sense.

Grace Snelling, writing in Fast Company about the rebrand:

[Grammarly CEO Shishir] Mehrotra explains it like this: Grammarly has always run on the “AI superhighway,” meaning that, instead of living on its own platform, Grammarly travels with you to places like Google Docs, email, or your Notes app to help improve your writing. Superhuman will use that superhighway to bring a huge new range of productivity tools to wherever you’re working.

In shedding the Grammarly name, Mehrota says:

“The trouble with the name ‘Grammarly’ is, like many names, its strength is its biggest weakness: it’s so precise,” Mehrotra says. “People’s expectations of what Grammarly can do for them are the reason it’s so popular. You need very little pitch for what it does, because the name explains the whole thing … As we went and looked at all the other things we wanted to be able to do for you, people scratch their heads a bit [saying], ‘Wait, I don’t really perceive Grammarly that way.’”

The company tapped branding agency Smith & Diction, the firm behind Perplexity’s brand identity.

Grammarly began briefing the Smith & Diction team on the rebrand in early 2025, but the company didn’t officially select its new name until late June, when the Superhuman acquisition was completed. For Chara and Mike Smith, the couple behind Smith & Diction, that meant there were only about three months to fully realize Superhuman’s branding.

Ouch, just three months for a complete rebrand. Ambitious indeed, but they hit a homerun with the icon, an arrow cursor which also morphs into a human with a cape, lovingly called “Hero.”

In their case study writeup, one of the Smiths says:

I was working on logo concepts and I was just drawing the basic shapes, you know the ones: triangles, circles, squares, octagons, etc., to see if I could get a story to fall out of any of them. Then I drew this arrow and was like hmm, that kinda looks like a cursor, oh wow it also kinda looks like a cape. I wonder if I put a dot on top of tha…OH MY GOD IT’S A SUPERHERO.

Check out the full case study for example visuals from the rebrand and some behind-the-scenes sketches.

Apologies for sharing back-to-back articles from NN/g, but this is a good comprehensive index of all the AI-related guides the firm has published. Start here if you’re just getting into it.

Highlights from my POV:

• Your AI UX Intern: Meet Ari. AI tools in UX act like junior interns whose output serves as a starting draft needing review, specific instructions, and added context. Their work should be checked and not used for final products or decisions without supervision.
• The Future-Proof Designer. AI speeds up product development and automates design tasks, but creates risks like design marginalization and information overload. Designers must focus on strategic thinking, outcomes, and critical judgment to ensure decisions benefit users and business value.
• Design Taste vs. Technical Skills in the Era of AI. Generative AI has equalized access to design output, but quality depends on creative discernment and taste, which remain essential for impactful results.

Leave it to NN/g to evaluate the AI prompt-to-code tool landscape with some rigor. Huei-Hsin Wang and Megan Brown cover over a dozen tools, including ChatGPT, Claude, UX Pilot, Uizard, Relume, Stitch, Bolt, Lovable, v0, Replit, Figma Make, Magic Patterns, and Subframe. They use a human designer as the control.

Among their conclusions:

AI’s limited grasp of design nuances and inconsistent output make it best suited for ideation, concept exploration, and early-phase prototype testing, rather than later stages. While you likely won’t take an AI-generated prototype straight to production, these tools can help you break through creative blocks and explore new directions quickly.

I think the best part is they shared screenshots of outputs in a FigJam board.

I’ve been a big fan of node-based UIs since I first experimented with Shake in the early 2000s. It’s kind of weird to wrap your head around, especially if you’re used to layers in Photoshop or Figma. The easiest way to think about nodes is to rotate the layer stack 90-degrees. Each node has inputs on the left, a distinct process that it does to the input, and outputs stuff on the right. You connect up multiple nodes to process assets to form your final composition. Popular apps with node-based workflows today include Unreal Engine (Blueprints), DaVinci Resolve (Fusion and Color), and n8n.

ComfyUI is another open source tool that uses the same node graph architecture. Made in 2023 to add some UI to the visual generative AI models like Stable Diffusion appearing around that time, it’s become popular among artists to wield the plethora of image and video gen AI models.

Fast-forward to last week, when Figma announced they had acquired Weavy, a much friendlier and cloud-based version of ComfyUI.

Weavy brings the world’s leading AI models together with professional editing tools on a single, browser-based canvas. With Weavy, you can choose the model you want for a task (e.g. Seedance, Sora, and Veo for cinematic video; Flux and Ideogram for realism; and Nano-Banana or Seedream for precision) and compose powerful primitives using generative AI outputs and hands-on edits (e.g. adjusting lighting, masking an object, color grading a shot). The end result is an inspiring environment for creative exploration and a flexible media pipeline where every output feeds the next.

This node-based approach brings a new level of craft and control to AI generation. Outputs can be branched, remixed, and refined, combining creative exploration with precision and craft. The Weavy team has inspired us with the balance they’ve struck between simplicity, approachability, and power. They’ve also created a tool that’s just a joy to use.

I must admit I had not heard about Weavy before the announcement. I had high hopes for Visual Electric, but it never quite lived up to its ambitions. I proceeded to watch all the official tutorial videos on YouTube and love it. Seems so much easier to use than ComfyUI. Let’s see what Figma does with the product.

In graphic design news, a new version of the Affinity suite dropped last week, and it’s free. Canva purchased Serif, the company behind the Affinity products, last year. After about a year of engineering, they have combined all the products into a single product to offer maximum flexibility. And they made it free.

Of course then, that sparks debate.

Joe Foley, writing for Creative Bloq explains:

…A natural suspicion of big corporations is causing some to worry about what the new Affinity will become. What’s in it for Canva?

Theories abound. Some think the app will start to show adverts like many free mobile apps do. Others think it will be used to train AI (something Canva denies). Some wonder if Canva’s just doing it to spite Adobe. “Their objective was to undermine Adobe, not provide for paying customers. Revenge instead of progress,” one person thinks.

Others fear Affinity’s tools will be left to stagnate. “If you depend on a software for your design work it needs to be regularly updated and developed. Free software never has that pressure and priority to be kept top notch,” one person writes.

AI features are gated behind paid Canva premium subscription plans. This makes sense as AI features have inference costs. As Adobe is going all out with its AI features, gen AI is now table stakes for creative and design programs.

I’ve been on the receiving end of Layer 1226 before and it’s not fun. While I’m pretty good with my layer naming hygiene, I’m not perfect. So I welcome anything that can help rename my layers. Apparently, when Adobe showed off this new AI feature at their Adobe MAX user conference last week, it drew a big round of applause. (Figma’s had this feature since June 2024.)

There’s more than just renaming layers though. Adobe is leaning into conversational UI for editing too. For new users coming to editing tools, this makes a lot of sense because the learning curve for Photoshop is very steep. But as I’ve always said, professionals will also need fine-grained controls.

Writing for CNET, Katelyn Chedraoui:

Renaming layers is just one of many things Adobe’s new AI assistants will be able to do. These chatbot-like tools will be added to Photoshop and Express. They have an emphasis on “conversational, agentic” experiences — meaning you can ask the chatbot to make edits, and it can independently handle them.

…

Express’s AI assistant is similar to using a chatbot. Once you toggle on the tool in the upper left corner, a conversation window pops up. You can ask the AI to change the color of an object or remove an obtrusive element. While pro users might be comfortable making those edits manually, the AI assistant might be more appealing to its less experienced users and folks working under a time crunch.

A peek into Adobe’s future reveals more agentic experiences:

Also announced on Tuesday is Project Moonlight, a new platform in beta on Adobe’s AI hub, Firefly. It’s a new tool that hopes to act as a creative partner. With your permission, it uses your data from Adobe platforms and social media accounts to help you create content. For example, you can ask it to come up with 20 ideas for what to do with your newest Lightroom photos based on your most successful Instagram posts in the past. 

These AI efforts represent a range of what conversational editing can look like, Mike Polner, Adobe Firefly’s vice president of product marketing for creators said in an interview. 

“One end of the spectrum is [to] type in a prompt and say, ‘Make my hat blue.’ That’s very simplistic,” said Polner. “With Project Moonlight, it can understand your context, explore and help you come up with new ideas and then help you analyze the content that you already have,” Polner said.

In thinking about the three current AI-native web browsers, Fanny on Medium sees what lessons product designers can take from their different approaches.

On Perplexity Comet:

Design Insight: Comet succeeds by making AI feel like a natural extension of browsing, not an interruption. The sidecar model is brilliant because it respects the user’s primary task (reading, researching, shopping) while offering help exactly when context is fresh. But there’s a trade-off — Comet’s background assistant, which can handle multiple tasks simultaneously while you work, requires extensive permissions and introduces real security concerns.

On ChatGPT Atlas:

Design Insight: Atlas is making a larger philosophical statement — that the future of computing isn’t about better search, it’s about conversation as an interface. The key product decision here is making ChatGPT’s memory and context awareness central. Atlas remembers what sites you’ve visited, what you were working on, and uses that history to personalize responses. Ask “What was that doc I had my presentation plan in?” and it finds it.

On The Browser Company Dia:

Design Insight: Dia is asking the most interesting question — what happens when AI isn’t a sidebar or a search replacement, but a fundamental rethinking of input methods? The insertion cursor, the mouse, the address bar — these are the primitives of computing. Dia is making them intelligent.

She concludes that they “can’t all be right. But they’re probably all pointing at pieces of what comes next.”

I do think it’s a combo and Atlas is likely headed in the right direction. For AI to be truly assistive, it has to have relevant context. Since a lot of our lives are increasingly on the internet via web apps—and nearly everything is a web app these days—ChatGPT’s profile of you will have the most context, including your chats with the chatbot.

I began using Perplexity because I appreciated its accuracy compared with ChatGPT; this was pre-web search. But even with web search built into ChatGPT 5, I still find Perplexity’s (and therefore Comet’s) approach to be more trustworthy.

My conclusion stands though: I’m still waiting on the Arc-Dia-Comet browser smoothie.

Celine Nguyen wrote a piece that connects directly to what Ethan Mollick calls “working with wizards” and what SAP’s Ellie Kemery describes as the “calibration of trust” problem. It’s about how the interfaces we design shape the relationships we have with technology.

The through-line is metaphor. For LLMs, that metaphor is conversation. And it’s working—maybe too well:

Our intense longing to be understood can make even a rudimentary program seem human. This desire predates today’s technologies—and it’s also what makes conversational AI so promising and problematic.

When the metaphor is this good, we forget it’s a metaphor at all:

When we interact with an LLM, we instinctively apply the same expectations that we have for humans: If an LLM offers us incorrect information, or makes something up because it the correct information is unavailable, it is lying to us. …The problem, of course, is that it’s a little incoherent to accuse an LLM of lying. It’s not a person.

We’re so trapped inside the conversational metaphor that we accuse statistical models of having intent, of choosing to deceive. The interface has completely obscured the underlying technology.

Nguyen points to research showing frequent chatbot users “showed consistently worse outcomes” around loneliness and emotional dependence:

Participants who are more likely to feel hurt when accommodating others…showed more problematic AI use, suggesting a potential pathway where individuals turn to AI interactions to avoid the emotional labor required in human relationships.

…

However, replacing human interaction with AI may only exacerbate their anxiety and vulnerability when facing people.

This isn’t just about individual users making bad choices. It’s about an interface design that encourages those choices by making AI feel like a relationship rather than a tool.

The kicker is that we’ve been here before. In 1964, Joseph Weizenbaum created ELIZA, a simple chatbot that parodied a therapist:

I was startled to see how quickly and how very deeply people conversing with [ELIZA] became emotionally involved with the computer and how unequivocally they anthropomorphized it…What I had not realized is that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people.

Sixty years later, we’ve built vastly more sophisticated systems. But the fundamental problem remains unchanged.

The reality is we’re designing interfaces that make powerful tools feel like people. Susan Kare’s icons for the Macintosh helped millions understand computers. But they didn’t trick people into thinking their computers cared about them.

That’s the difference. And it matters.

Speaking of trusting AI, in a recent episode of Design Observer’s Design As, Lee Moreau speaks with four industry leaders about trust and doubt in the age of AI.

We’ve linked to a story about Waymo before, so here’s Ryan Powell, head of UX at Waymo:

Safety is at the heart of everything that we do. We’ve been at this for a long time, over a decade, and we’ve taken a very cautious approach to how we scale up our technology. As designers, what we have really focused on is that idea that more people will use us as a serious transportation option if they trust us. We peel that back a little bit. Okay, well, How do we design for trust? What does it actually mean?

Ellie Kemery, principal research lead, advancing responsible AI at SAP, on maintaining critical thinking and transparency in AI-driven products:

We need to think about ethics as a part of this because the unintended consequences, especially at the scale that we operate, are just too big, right?

…

So we focus a lot of our energy on value, delivering the right value, but we also focus a lot of our energy on making sure that people are aware of how the technology came to that output,…making sure that people are in control of what’s happening at all times, because at the end of the day, they need to be the ones making the call.

…

Everybody’s aware that without trust, there is no adoption. But there is something that people aren’t talking about as much, which is that people should also not blindly trust a system, right? And there’s a huge risk there because, humans we tend to, you know, we’ll try something a couple of times and if it works it works. And then we lose that critical thinking. We stop checking those things and we simply aren’t in a space where we can do that yet. And so making sure that we’re focusing on the calibration of trust, like what is the right amount of trust that people should have to be able to benefit from the technology while at the same time making sure that they’re aware of the limitations.

Ethan Mollick, a professor of entrepreneurship at the Wharton School says that AI has gotten so good that our relationship with them is changing. “We’re moving from partners to audience, from collaboration to conjuring,” he says.

He fed NotebookLM his book and 140 Substack posts and asked for a video overview. AI famously hallucinates. But Mollick found no factual errors in the six-minute video.

We’re shifting from being collaborators who shape the process to being supplicants who receive the output. It is a transition from working with a co-intelligence to working with a wizard. Magic gets done, but we don’t always know what to do with the results. This pattern — impressive output, opaque process — becomes even more pronounced with research tasks.

Mollick believes that the most wizard-like model today is GPT-5 Pro. He uploaded an academic paper that took him a year to write, which was peer-reviewed, and was then published in a major journal…

Nine minutes and forty seconds later, I had a very detailed critique. This wasn’t just editorial criticism, GPT-5 Pro apparently ran its own experiments using code to verify my results, including doing Monte Carlo analysis and re-interpreting the fixed effects in my statistical models. It had many suggestions as a result (though it fortunately concluded that “the headline claim [of my paper] survives scrutiny”), but one stood out. It found a small error, previously unnoticed. The error involved two different sets of numbers in two tables that were linked in ways I did not explicitly spell out in my paper. The AI found the minor error, no one ever had before.

Later in his post, Mollick says that there’s a problem with this wizardry—it’s too opaque. So what can we do?

First, learn when to summon the wizard versus when to work with AI as a co-intelligence or to not use AI at all. AI is far from perfect, and in areas where it still falls short, humans often succeed. But for the increasing number of tasks where AI is useful, co-intelligence, and the back-and-forth it requires, is often superior to a machine alone. Yet, there are, increasingly, times when summoning a wizard is best, and just trusting what it conjures.

Second, we need to become connoisseurs of output rather than process. We need to curate and select among the outputs the AI provides, but more than that, we need to work with AI enough to develop instincts for when it succeeds and when it fails.

And lastly, trust it. Trust the technology, he suggests. “The question isn’t ‘Is this completely correct?’ but ‘Is this useful enough for this purpose?’”

I think we’re in that transition period. AI is indeed dastardly great at some things and constantly getting better at the tasks it’s not. But we all know where this is headed.

In this era of AI, we’ve been taught that LLMs are probabilistic, not deterministic, and that they will sometimes hallucinate. There’s a saying in AI circles that humans are right about 80% of the time, and so are AIs. Except when less than 100% accuracy is unacceptable. Accountants need to be 100% accurate, lest they lose track of money for their clients or businesses.

And that’s the problem Intuit had to solve to roll out their AI agent. Sean Michael Kerner, writing in VentureBeat:

Even when its accounting agent improved transaction categorization accuracy by 20 percentage points on average, they still received complaints about errors.

“The use cases that we’re trying to solve for customers include tax and finance; if you make a mistake in this world, you lose trust with customers in buckets and we only get it back in spoonfuls,” Joe Preston, Intuit’s VP of product and design, told VentureBeat.

So they built an agent that queries data from a multitude of sources and returns those exact results. But do users trust those results? It comes down to a design decision on being transparent:

Intuit has made explainability a core user experience across its AI agents. This goes beyond simply providing correct answers: It means showing users the reasoning behind automated decisions.

When Intuit’s accounting agent categorizes a transaction, it doesn’t just display the result; it shows the reasoning. This isn’t marketing copy about explainable AI, it’s actual UI displaying data points and logic.

“It’s about closing that trust loop and making sure customers understand the why,” Alastair Simpson, Intuit’s VP of design, told VentureBeat.