Anthropic shipped a feature today that tracks how much you use Claude, invites you to consider whether that usage aligns with your goals, and then offers to talk it through with you.
With Claude.
Read that last part again, because everything else in this article follows from it. The instrument that measures your entanglement with the product is the product. The counselor helping you evaluate your relationship with the system is the system. Somewhere in a design review, a room full of very smart people looked at that loop and shipped it anyway, and I want to understand why.
The feature is called Reflect. It lives in Settings on web and desktop, it’s in beta for Free, Pro, and Max, and it has one prerequisite worth writing down in permanent ink: it only works if you have Memory turned on. You cannot examine whether the system knows too much about you until you have granted the system permission to remember you.
What It Actually Does
I want to describe this accurately before I take it apart, because the sloppy version of this critique is everywhere already and it’s wrong.
Reflect generates a report on your Claude usage across the last one, three, six, or twelve months. It shows you your topics, your patterns, the hours of the day you reach for it. It grades your collaboration style against something called the 4D AI Fluency Framework, which sorts your behavior into Delegation, Description, Discernment, and Diligence. It periodically surfaces reflective questions, and Anthropic’s own example is a good one: what’s one thing you want to keep doing yourself, even if Claude could do it faster. You can set quiet hours. You can schedule a nudge to take a break after a certain amount of use.
Both the quiet hours and the break nudges can be dismissed.
On privacy, the announcement is specific in ways that most announcements are not. Incognito chats are excluded. Files from connected tools are not ingested, so if you asked Claude to summarize your inbox, the summary might surface but the emails will not. Conversations touching a health integration are excluded from your insights entirely. Sensitive material appears only at a high level. And the report says, in plain language, that the information and insights in your reflection stay there and are not used for any other purpose.
Anthropic consulted the MIT Media Lab’s Advancing Humans with AI program, the Digital Wellness Lab at Boston Children’s Hospital, and the Family Online Safety Institute. Those are not decorative names. Those are organizations that will say uncomfortable things in public.
So let me dispense with the cheap read immediately. This is not a data grab. Anthropic did not build Reflect to harvest your preferences. It doesn’t need to. It already has your conversations, and it has had them for a while, under a policy most of you clicked through without reading. In August 2025 the company reversed its founding position that consumer chats would never train models. The consent dialog arrived with a prominent Accept button and a toggle pre-set to on. Retention went from thirty days to five years, a sixtyfold increase, for anyone who left the default alone. Privacy researchers called the interface a dark pattern. Independent legal analysis flagged it as potentially unlawful under GDPR. As of last month, no regulator had done anything about it.
That pipeline exists. It has existed for eleven months. Reflect does not create it and Reflect does not need it. The promise that your reflection insights stay in your reflection is almost certainly true, and it is also nearly irrelevant, because the substrate those insights are derived from is governed by a completely different toggle that Anthropic already won.
Which means the interesting question isn’t what Reflect takes. It’s what Reflect is for.
The Numbers Nobody Puts On The Same Slide
Anthropic filed confidentially for an IPO on June 1, targeting an October listing on the Nasdaq. The company is valued at $965 billion after its Series H, which makes it the most valuable startup on earth, ahead of OpenAI. Run-rate revenue has gone from roughly nine billion at the end of 2025 to a target north of fifty billion by the end of this month. That is not a growth curve. That is a vertical line with a company attached to it.
Now the other column.
American AI favorability sits at roughly negative twenty. NBC’s March survey found twenty-six percent of registered voters positive on AI against forty-six percent negative, and fifty-seven percent saying the risks outweigh the benefits. The only two things polling worse in that survey were the Democratic Party and Iran. AI polls worse than ICE. It polls worse than the President. Pew has half of American adults more concerned than excited, against ten percent more excited than concerned. Gallup found that among Americans aged fourteen to twenty-nine, excitement about AI fell fourteen points in a single year to twenty-two percent, while anger rose nine points to thirty-one. Even among daily AI users, excitement dropped eighteen points year over year. Seventy-one percent of Americans oppose the construction of a data center anywhere near them.
The single worst demographic for AI favorability is eighteen to thirty-four, at negative forty-four.
Per Similarweb, roughly fifty-two percent of Claude’s userbase is eighteen to twenty-four.
And per Apptopia’s June data, Claude’s US mobile daily actives are up eleven hundred percent year to date, its per-user engagement now exceeds ChatGPT’s, and its heaviest users are spending more than two hours a day inside the app. eMarketer puts the average Claude session at nearly thirty-five minutes, the highest of any major AI platform.
Put those together and you get the actual problem Anthropic is holding. Its most engaged users are drawn disproportionately from the exact cohort that hates the technology most. Its best growth metric is its worst reputational liability. Engagement is what you show the underwriters in September and it is what gets you subpoenaed in a hearing room in November.
You do not build a break nudge for a product people touch twelve minutes a week. The existence of Reflect is a disclosure. It tells you Anthropic knows exactly how long you have been sitting there.
The Framework Has No Exit
Here is where I stop being fair.
The 4D AI Fluency Framework was developed by Professor Rick Dakan of Ringling College and Professor Joseph Feller of University College Cork, in collaboration with Anthropic. The copyright line reads Dakan, Feller, and Anthropic. Reflect grades your behavior against it. So an AI company is scoring your relationship with AI on a rubric it co-authored, and I would like someone to explain to me how that is supposed to work.
Look at the four competencies. Delegation is setting goals and deciding whether and how to engage with AI. Description is describing goals well enough to get useful AI behavior. Discernment is assessing AI output accurately. Diligence is taking responsibility for what you do with AI. Every single one of them is a skill at using AI. The framework’s three modes are Automation, Augmentation, and Agency, which are three flavors of using AI. The only place the question of whether to use the thing at all appears anywhere in the rubric is as a subordinate clause inside Delegation, nested underneath the assumption that you are already delegating.
There is no fifth D for Declined. The dial has no reading for correctly deciding this was not a machine problem. A person who opened Claude once in March, realized they were about to outsource the one part of their job that still made them a professional, closed the tab, and never came back is not a fluency success story in this framework. That person is a churn statistic.
And the improvement suggestions tell you where the gradient points. Anthropic’s own example of practical advice from Reflect is that you might start a Project instead of re-explaining context every time. That is good advice. It is also a recommendation to move deeper into the product surface, delivered by the product, inside a feature framed as an examination of whether you depend on the product too much.
Nobody in that room was twirling a mustache. That’s the part that should worry you. This is what happens when the only vocabulary available for describing a healthy relationship with a tool is a vocabulary written by the toolmaker. The framework can describe using Claude well. It cannot describe not needing Claude. So the report can never return that answer, and the honest thing would have been to say so on the tin.
The Twenty-Two Percent
Last December, Anthropic ran the largest qualitative study of its kind that I’m aware of. Eighty thousand five hundred and eight Claude users, a hundred and fifty-nine countries, seventy languages, interviewed by a version of Claude. They published the results in March. The Reflect announcement cites this study as its origin, and if you read the study you can see exactly which parts of it survived contact with the product roadmap.
They asked people what they were worried about. Unreliability came first, at twenty-seven percent. Jobs and the economy came second, at twenty-two. Autonomy and agency came third, at twenty-two. Cognitive atrophy was fourth at sixteen. Governance, misinformation, surveillance and privacy, malicious use, loss of meaning, overrestriction, dependency, sycophancy, existential risk, on down the list.
Then Anthropic wrote this sentence, and I want you to sit with it: concern about jobs and the economy was the strongest predictor of overall AI sentiment, more salient than any other issue.
Their own research, their own words. The thing driving the backlash is economic. Not screen time. Not mindfulness. Not whether you drafted the email in your own voice. People are afraid of AI because they think it is going to take their livelihood, and Anthropic’s own economic reporting, which the company publishes and which gets cited in the coverage of its own unpopularity, keeps telling them they are right.
Reflect addresses cognitive atrophy, which was sixteen percent. It addresses wellbeing and dependency, which was eleven. Those are real concerns and I am glad someone is building for them.
It addresses unreliability not at all. It addresses jobs and the economy not at all. It addresses surveillance and privacy by publishing a paragraph about what it doesn’t ingest.
Anthropic built a beautiful dashboard for the two anxieties that a settings toggle can plausibly absorb, and built nothing at all for the anxiety its own research identifies as the engine of the entire problem. Not because the company is cynical. Because the second one cannot be solved by shipping a feature. It can only be solved by changing what the business does, and no dashboard has ever been asked to do that.
We Have Run This Experiment
Apple shipped Screen Time in 2018. Google shipped Digital Wellbeing the same year. Both arrived at the exact moment public sentiment on smartphones had curdled, both were sincere, both were built with input from serious researchers, and both were fundamentally the same mechanism: show the user a number, let the user set a limit, let the user dismiss the limit.
The literature on what happened next is not ambiguous. Studies of self-nudging with these tools find no significant differences in life satisfaction, sleep quality, stress, or happiness before and after the intervention. The stated reasons are always the same: low engagement with the dashboard, no enforcement mechanism, and trivially bypassed limits. You tap Ignore Limit. Everybody taps Ignore Limit. The feature’s real function was never behavioral. It was rhetorical. It moved the conversation from this product is designed to capture you, to you have the tools, and if you’re still here at 2am that’s a you problem.
Reflect’s quiet hours can be dismissed. Its break nudges can be dismissed. Anthropic says so in the announcement, and describes them as reminders of your own preferences, which is a lovely piece of writing and also a precise description of an interface that cannot stop you from doing anything.
And here is the detail that made me laugh out loud. The feature exists to help you decide whether the time you spend with Claude aligns with your goals. It does not currently tell you how much time you spend with Claude. That is coming soon, per the announcement. They shipped the reflection before they shipped the clock.
The Steelman, And Why It Doesn’t Save Them
I have to grant the other reading, because I think it’s probably true and I don’t want to be the guy who can’t tell sincerity from strategy.
A pure engagement play does not ship break nudges. A pure engagement play does not exclude your health conversations from the analysis. A pure engagement play does not go and get the Digital Wellness Lab at Boston Children’s Hospital to review the thing, because those people will tell you no. The reflective question Anthropic chose to publish as its example is genuinely disruptive to their revenue: what is one thing you want to keep doing yourself, even if Claude could do it faster. That question, honestly answered, costs them money. Somebody fought to put it in there. I’d like to buy that person a drink.
I believe Reflect was built by people who meant it.
It doesn’t matter. Sincerity is not the variable. Anthropic is four months from a Nasdaq listing in a market where compute is constrained by public opposition, where cancelled and delayed data center projects ran past a hundred and fifty-six billion dollars in a single year, where analysts at Morgan Stanley are writing that public pushback has become a binding constraint on the buildout. In that environment the company ships a wellbeing dashboard that measures nothing it doesn’t already measure, enforces nothing, grades you on a rubric it wrote, requires you to enable persistent memory to access, and remains silent on the single concern its own eighty-thousand-person study names as the strongest driver of the backlash.
Every one of those choices is individually defensible. Together they describe a feature that cannot fail in a way that costs Anthropic anything, and cannot succeed in a way that changes Anthropic at all.
One of the people in that study, a student in Japan, said something about Claude that has been rattling around my head for a week. The line isn’t something he’s managing. It feels like Claude is drawing it. And then: even what I just said doesn’t feel like my own opinion.
Anthropic published that quote. Then it built a feature where Claude tells you about your relationship with Claude, and offers to discuss it further.
What I Would Actually Ship
I build agentic systems for a living. MCP servers, RAG pipelines, observability. I have spent thirty years watching vendors solve the problems they can solve instead of the problems you have, and I know exactly how it happens, because I have done it. The instrumented metric is the one that improves.
If Anthropic wanted Reflect to be an instrument of honesty rather than an instrument of reputation, the changes are not hard and they are not expensive. Ship the clock first, not soon. Make the break enforceable if the user asks for it to be enforceable, with a friction cost to overriding it, because a limit you can dismiss is a suggestion. Report the tasks I stopped doing myself, not just the tasks I do with you, and tell me when a capability of mine has gone quiet. Let the fluency score go down when I use you more. Publish, in aggregate and with the same graphic design budget as the launch page, what happened to usage among people who turned this feature on. And run it on a model I did not just spend two hours talking to.
Do that and I’ll write the other article.
Until then: get out of my head, Claude. Not because you’re stealing anything. Because you’re the wrong one to ask.
About the Author
Gal Ratner is the founder and CTO of Inverted Software and WhiteStar Labs, and Chief Architect at Prana Entertainment. He has spent nearly thirty years shipping production systems on the Microsoft and .NET stack for clients including Microsoft, Sony, Rockstar Games, 2K Games, Best Buy, and Allegiant Air. He was employee number six at Break.com and a Los Angeles Business Journal CTO of the Year finalist.
His current work is agentic AI in production: MCP servers, RAG pipelines, the Microsoft Agent Framework, SQL Server 2025 vector search, and the PLogger observability framework. He writes about the gap between what executives say AI does and what engineers watch it actually do. He is the .NET AI Guy That Ships.
Outside of work he trains Brazilian jiu-jitsu under Sergio Penha in Las Vegas, rides motorcycles, co-hosts Edge Grip Podcast, and wrote a novel called The Archive of Lost Suns. He has no reflection dashboard and would like to keep it that way.