In partnership with

Keep up to date on the latest products, workflows, apps and models so that you can excel at your work. Curated by Duet.

Stay ahead with the most recent breakthroughs—here’s what’s new and making waves in AI-powered productivity:

ChatGPT Computer History by OpenAI

OpenAI has launched Computer History for the ChatGPT macOS desktop application, offering consumers and professionals a continuous timeline of their desktop activity. The tool tracks interaction events across approved apps and websites, enabling users to ask natural language questions to retrieve lost files, summarize completed work, and transform repetitive actions into automated skills. By eliminating manual file searching across multiple programs, it brings a practical digital memory assistant to desktop users.

Symphony by Wix

Wix has officially launched Symphony by Wix, a standalone platform designed to provide small and medium businesses with a dedicated team of proactive AI agents. This platform coordinates key operational tasks, automated marketing workflows, and customer engagement based on specific business goals. By acting as an adaptive digital staff that learns company preferences over time, Symphony helps small teams automate repetitive operational workloads without adding overhead.

Hi3D V3.0

Hi3D has released version 3.0 of its AI asset generation platform, introducing high precision 3D creation capabilities tailored for creators, designers, and small studios. The tool converts 2D reference images into structured 3D models with high voxel resolution, while offering automated AI texturing and multi color preparation for 3D printing. Creators can easily split models into printable components with connectors, dramatically simplifying the workflow for physical product prototyping and digital media design.

Grok 4.6 by xAI

xAI has unveiled Grok 4.6, an upgraded AI model tailored for complex multi step agentic workflows, software development, and deep research tasks. The model features enhanced self verification capabilities that enable it to handle long trajectory coding projects and continuous workspace automations with high reliability. Small business developers and creators can leverage Grok 4.6 to build custom applications and automate administrative workflows at a fraction of standard computing costs.

Google Ads and Analytics Ask Advisor

Google has introduced Ask Advisor alongside new AI insights cards across Google Ads and Google Analytics to streamline reporting for small business marketers. The system automatically highlights performance shifts, competitive benchmarking metrics, and budget pacing anomalies as soon as users log in. Through an interactive conversational prompt, store owners and creators can ask plain English questions about campaign metrics to receive immediate data summaries and actionable growth recommendations.

ChatGPT Computer History Takes the Friction Out of Mac Workflows

OpenAI has launched Computer History for its ChatGPT desktop application on macOS, introducing a powerful feature designed to eliminate daily workflow friction for consumers and busy small business owners. Operating directly within the Mac interface, this intelligent system constructs a continuous timeline of user activity by keeping track of interactions across authorized software, web browsers, and documents. Rather than requiring users to manually label files or save endless browser bookmarks, the application creates a contextual record of app switches, keystrokes, and clicks so users can interact with their work history through natural conversation.

The technology behind Computer History centers on event based activity tracking rather than continuous screen recordings. The tool monitors interaction events across designated programs and stores temporary activity files that allow fast context retrieval. Users can type natural commands to ask where a specific document was referenced earlier in the day or request a quick summary of tasks completed during a busy afternoon. Beyond basic search, the assistant can identify repeated series of manual actions and turn them into single click automated skills directly inside the workspace.

The main benefit of this new feature is a drastic reduction in cognitive fatigue and time wasted on context switching. Searching for lost research links, tracking down scattered draft files, or manually assembling daily updates takes up hours of productive time every week. By handing those administrative tasks over to the desktop assistant, users can focus on high value creative tasks, knowing that every detail of their digital workspace can be recalled in seconds.

To keep user control straightforward, OpenAI built flexible privacy and management tools right into the app settings. Users can choose exactly which applications or web domains are included in the timeline, clear temporary files whenever necessary, or pause background tracking with one click. The feature is available for paid subscription tiers, giving Mac users an intuitive personal memory assistant that makes daily computer work much smoother.

Nobody Could Tell

Earlier this month, researchers at Villanova University published the results of an experiment in what people can and can’t recognize about the writing they read.

They took three short stories written by human authors and published in real literary journals and collections. They asked ChatGPT to write a companion piece for each one. Then they recruited 1,682 adults — ages 18 to 81, quota-matched to the U.S. population on age, gender, and race — and gave each person a single story to read and rate.

The AI stories won.

Readers rated the machine-written stories higher on quality, and higher on absorption — the researchers’ measure of how thoroughly you get lost in a story, how much the room around you fades. On a scale running from −3 to +3, the AI stories averaged 1.54 on quality against 0.97 for the humans. The study was published in Judgment and Decision Making by Cambridge University Press.

That’s the headline, and it’s the part that traveled. The rest of the study is stranger.

The second half of the experiment

Before anyone read anything, they got one of two introductions. Half were told their story “was written by a contemporary author and published in a literary journal for short stories and poetry.” The other half were told it “was written by an AI program, ChatGPT,” which “was given a prompt detailing generally what to write about.”

Then the researchers lied to half of them. Some people read a machine-written story under the literary-journal introduction. Some read a published human story and were told a chatbot produced it from a prompt.

The introduction moved the scores. Stories framed as human-written were rated higher — regardless of who or what actually wrote them. Same words, different setup, different verdict.

Put that beside the first finding and you get the sentence the authors wrote to summarize their own results:

“People actually prefer AI-generated stories while believing that these stories are worse.”

Worth being precise about what moved, though. The gap produced by who actually wrote the story was roughly twice the size of the gap produced by the framing. So this isn’t “we only read the byline.” It’s that the framing is doing measurable work on top of the words — and note what that framing actually contained. One version offered a credentialed author and a literary journal. The other offered a machine and a prompt. That bundle is reputation, process, and authorship all at once, which means the study can’t cleanly separate “a human made this” from “this passed an editor.”

Then they tried to catch it, and couldn’t

Two follow-up studies asked a simpler question: can you just tell?

Participants got one human story and one AI story, side by side, and were told plainly that one of each was in front of them. No trickery. All they had to do was point at the machine.

In the first round, 424 people got it right 39% of the time.

On a two-way choice, chance is 50%. Thirty-nine percent is not a near miss — it is statistically worse than guessing, which means the errors weren’t random. Participants were systematically pointing at the AI story and calling it the human one.

The second round, with 481 people, landed at 52% — statistically indistinguishable from chance. One of the three story pairs was an exception, identified correctly at better than chance rates; the other two weren’t. So “nobody could tell” is the shape of the result, not a universal law. Nobody could reliably tell.

Both rounds told participants that one of the two stories was machine-written. Real life doesn’t come with that warning label. The researchers expect accuracy would drop further outside a study where nobody announces that a test is happening — though they’re careful to file that as an open question rather than a finding.

What people used to decide

The researchers also asked participants which feature they’d relied on to make their call, then checked which features predicted right answers. This is the most practical result in the paper and the one that has gone almost entirely unreported.

The most popular cue was the language itself — the prose, the sentences, the word choice. It was the single most-selected option in both detection studies. It was also the one cue that reliably predicted wrong answers in both. People who said they went by the writing got it right 33% of the time in the first study and 47% in the second. Trusting your ear didn’t just fail to help; it actively pointed the wrong way.

Enjoyment fared badly too, though only in the second study did that reach significance. Deciding based on which story you liked more was not a reliable route to the human one.

The only cue with any signal in the other direction was symbolism — and even there the effect was marginal, showed up in one study out of two, and wouldn’t carry much weight on its own. Theme, plot, and character produced nothing statistically significant either way.

So the honest version of this finding is narrow: the instinct most people reach for first is the one that misleads them, and nothing else in the list clearly rescues them. Why that would be is not something the experiment tested. The plausible reading is that sentence-level polish is exactly what these systems are best at — but that’s interpretation, not a result.

Familiarity helped. Being well-read didn’t.

Two kinds of expertise got measured. Only one did anything.

People who scored higher on a measure of AI literacy were better at spotting the machine. Modestly better — a small effect that barely cleared statistical significance — but the direction held across both studies.

People who scored higher on the researchers’ literary-experience questionnaire? Nothing. Statistically indistinguishable from zero.

One caveat that matters: both measures were self-report. Participants rated their own AI experience and their own history with fiction. Nobody’s literary judgment was actually tested. So the finding isn’t “taste is worthless” — it’s that how well-read people say they are predicted nothing about whether they could pick the machine, while how much time they’d spent around these systems predicted a little.

The fine print

This study is being reported as “AI writes better than humans.” It doesn’t show that, and the researchers don’t claim it. Five things are worth stating plainly.

Six texts. This is the big one. The reader sample was large — 1,682 people. The writing sample was three human stories and three AI companions. A thousand readers agreeing about one story doesn’t make that story representative of human writing, and no amount of statistical power on the reader side fixes a stimulus set of six.

The AI was handed the premise. The prompts weren’t “write a good short story.” They were detailed instructions carrying the human original’s theme, symbol, point of view, and setting — one asked for a story “about generational life and death and uncertainty and uses koi fish as a symbol,” and another specified the ending. That’s not a machine out-imagining a novelist. That’s a machine executing a novelist’s brief.

The effects are small, and the measuring stick was homemade. The quality scale was written for this study and had never been validated. None of the three studies was preregistered. The strongest single predictor of how someone rated a story wasn’t the framing or even the actual author — it was the reader’s own attitude toward AI, which the researchers measured after people had read and scored the story, so it can’t be cleanly read as a pre-existing bias.

The stories were tiny. About 1,000 words, five minutes each, and the authors describe all six as fiction. They flag length as their main limitation: a machine that can hold a mood for five minutes may not hold a character together for three hundred pages. This study didn’t test that, and whether the preference survives at novel length is open.

Expert judges have gone the other way. The authors point to earlier work in which ten creative writers assessed 48 stories against a structured creativity rubric; the machine-written ones passed three to ten times fewer of the tests. General readers here preferred the machine. Trained evaluators there did not — though that study used different stories, different models, and a different task, so it can’t establish that expertise alone accounts for the split.

One more that cuts the other way: the AI stories came from ChatGPT 4.0, two of them generated back in December 2024. That’s several model generations old. Whether a current model would widen the gap is untested — but nothing about the trend line suggests it would narrow.

What’s actually happening here

The researchers offer one explanation by analogy.

When you average many human faces together into a single composite image, people consistently rate the composite as more attractive than most of the individual faces that went into it. The average is smoother. No crooked nose, no asymmetric jaw, nothing that catches the eye. It’s also not anyone.

That may be what these systems do to prose. They sand off the rough edges — the odd word choice, the sentence that runs too long, the meaning you have to work to find — and what’s left is easier to move through. Senior author Deena Weisberg’s own reading, in comments accompanying the study, is that AI writing “tends to be clearer, more direct and easier to process,” while human literary fiction is deliberately doing something harder. As she puts it: “Perhaps people generally prefer predictability, because difficult or subtle material requires more brain power.”

So “better” may be the wrong word for what won here. “Frictionless” is closer. And a preference for frictionless is not a new human failing that ChatGPT invented — it’s an old one that ChatGPT is extremely good at serving.

Which shifts the question. It isn’t whether the machine can write. It’s what we’ve been rewarding in writing all along.

Go find out for yourself

The researchers posted all six stories publicly, in the study’s open repository, paired up as three PDFs. Three were written by people — Emilie Fox, whose “FISH” ran in Longleaf Review; Susie Maguire; Patrick Smyth. Three came out of a chatbot. They’re about five minutes each.

One practical wrinkle: each PDF labels its stories, printing the author above the human one and the full generation prompt above the AI one. So you can’t run the experiment on yourself straight off the download. If you want the real version, have someone else cut the headers and hand you the pages in whatever order they like.

It’s worth the small hassle, because the thing to watch isn’t whether you get it right. It’s what you reach for while deciding. Reliance on the prose itself was associated with lower accuracy in both studies — and the prose is what nearly everyone reached for first.

If you do miss, that says less about your taste than it might feel like it does. The people who did better weren’t the ones who read the most fiction. They were the ones who’d spent the most time around the machines.

Which makes this a rare piece of AI advice that doesn’t ask you to buy anything. It asks you to pay closer attention to something you’re already using.

Partner Spotlight: Periscope VPN by Duet Display

Securing your online activity when working remotely or browsing public networks should never slow you down. Periscope by Duet Display offers a powerful solution that is incredibly easy to set up on any device. Engineered with a strong commitment to privacy, Periscope uses state of the art end to end encryption to shield your network traffic from unwanted exposure. Operating under a strict zero log policy with absolutely no logging of your browsing activity, Periscope ensures your connection remains fast, private, and fully protected wherever you work. Learn more and get started at getPeriscope.com.

Reply to everything. Edit nothing.

Your inbox is full. Slack is piling up. Client messages need a response yesterday. Typing thoughtful replies to all of it takes hours you don't have.

Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Speak like you would to a colleague — tangents and all — and get polished output. Emails, Slack, LinkedIn, WhatsApp, whatever's open.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Works on Mac, Windows, and iPhone.

Try Wispr Flow free

Stay productive, stay curious—see you next week with more AI breakthroughs!