Essay
AI's Next Differentiators: Cognitive Preservation & Escaping the Sea of Sameness
The AI that serves you best is the one that keeps asking you to show up.
Can AI companies prevent their users from going through “a process of dumbening” (as Lisa Simpson once put it when describing the “Simpson gene”)?
And even if they can, will they?
I recently read a thoughtful piece from Matt Maher on the cognitive cost of AI. Studies out of MIT and other institutions seem to indicate atrophy in memory and neural engagement among people who lean on LLMs to think for them, reconfirming what every heavy AI user experiences daily. (I share these concerns and have written about them at length.)
But I want to push past diagnosis toward the question of what we should build differently in our AI tools in response.
The “candy bar” problem
Keeping users engaged and happy with AI output quality remains frontier labs’ main priority, which gives rise not only to sycophancy but also token-consumption-driving behaviors. We all see the standing offer at the end of each response offering to take the output to the next level, or to explore a complementary topic. It’s helpful in the way that the waiter offering the next drink late at night is helpful. We may want it, we may not need it, or we may be better off showing restraint and processing everything we’ve already been served. Recent Claude models are noticeably quicker to rush to creating a deliverable, anticipating our needs while running with unvalidated assumptions — cutting once after only measuring once, if at all.
The illusion of productivity is there, but the honest user must begin the laborious backtrack of catching up to the embedded assumptions, relevant knowledge gathered and reasoning behind the output or recommendation. We can pay down the cognitive debt immediately, or accrue a big balance over time.
Too much, too fast for us to process, and with no ledger of the delta between whatever output our name goes on and what we’ve learned and owned in the process. The onus is on the user to keep themselves honest, to go for the mental run after digesting the proverbial AI candy bar.
There’s another frequently-cited problem with AI — the sea of sameness AI produces — that is rooted in the same underlying user-dependency risk as the cognitive debt challenge. A statistical model is always going to skew toward the generic, and the onus again is on the user to thoughtfully prompt any given command, and to anchor instructions and .MD files that inject differentiating context effortlessly. For the majority of users — non-power users — these are advanced user settings they may never play with, and as a result, their outputs are either undifferentiated, under-resourced for their purpose, or both.
Both problems place the burden of using AI responsibly, and to best effect, on the user. That is a product gap, and I’d argue the net-effect of getting it closed will be hugely beneficial to society and users alike. To be delivered, it also has to be something users demand of AI products.
While convenience has shown to be a seductive product feature, I believe the stakes are high enough now that the pressure to act on these issues will grow significantly in the years ahead. AI-induced incapacitation, and the removal of human-unique outputs, both strike at our fundamental identities. If phones and social media addiction prey on our attention, AI preys on our effort and value as cognitive creatures. Our brains have gotten us this far, they have a ways yet to take us.
Prescriptions & predictions
I’d like to offer a hopeful set of recommendations for how AI can get beyond the cognitive externalities of using such a powerful tool, and more powerfully re-inject us as people into our AI-mediated outputs. I predict many of these ideas are under consideration already and will be adopted by frontier labs, or at least emphasized to greater degrees than they are today.
More specifically: within a few product cycles, the leading AI platforms will begin to differentiate not on raw capability — commoditizing faster than anyone expected — but on how well they let us protect our own minds and preserve our own voices. I think “cognitive preservation” can become a product KPI, with real settings, real defaults, and real willingness-to-pay behind it. And I think companies that design it well will win a segment of users over more mass-market, pure-convenience players.
Part one: the outsourcing of thought
First, let’s tackle the outsourcing of thought and the resultant mental failure to encode and process information “the hard way”.
Whenever mainstream technology removes more friction from a process, or hacks further into our lizard brains, a vocal minority of observers get worried. We raise our objections (often here on LinkedIn), found and support non-profits (e.g., the Center for Humane Technology), we pledge to regulate our own behaviors. To varying degrees, tech companies respond and adjust their products to offer some level of “responsible use” tools or settings. The problem is the broad adoption of those tools — screen time limits, notification silencing, removing autoplay — are direct risks to their business models.
I remember deciding that grayscale mode was to be my new normal, only to lapse on willpower in a few days, so we can’t underestimate the force of swimming upstream against what our brains find appealing.
Product designers usually view friction as a sin. But psychology has long understood the value of ‘desirable difficulty’ — the precise amount of resistance required for the human brain to encode information and build mastery. The hidden cost of current AI usage paradigms is that zero-friction generation simply pushes friction downstream. You save 10 minutes generating a draft, but spend 30 minutes backtracking through unvalidated assumptions and soul-crushing editing. By re-introducing desirable difficulty at the prompt stage, we slow users down to speed them up — eliminating the downstream tax of low-trust output.
Friction is typically removed because we are inherently drawn toward pursuing the most efficient paths to producing a given result, while remaining highly ambitious in our identities. AI exploits this by drawing ambitious people into using it to produce more and more, spending less time on any one thing before moving to the next. Nobody wants to be the person who can’t write on their own, or can’t wrestle with a problem in their brain long enough to make headway on it, but these can be net effects of the AI productivity hamster wheel. The gap between how we act in the moment — “chasing the efficiency dragon” — and who we want to be over time — learned, wise, responsible, proud — is exactly the kind of gap that products are built to close. Fitness apps, language-learning apps, savings tools and more help us push through struggle and make incremental gains over time. The market for products that “help me be the version of myself I actually respect” is enormous, and AI is opening a new frontier for it.
Grayscale failed on me because it was a blunt setting that punished my brain’s novelty-seeking through subtraction, without really adding anything to the mix. The AI equivalent doesn’t have to work that way, because the tool is conversational and can meet you contextually in the moment of the request.
Imagine setting “cognitive optimization preferences” the way we set privacy preferences today. A desirable-difficulty-dial between do it for me and make me do it, adjustable by task, by subject, by deadline pressure. A few mechanisms I’d want in this feature:
Proficiency gates. For subjects you’ve told it you want to actually learn, your AI declines to simply hand over the answer or recommendation. Instead it opens an exchange — a few questions, a request that you take the first pass, a “show me your reasoning and I’ll refine it.” A mentor-teacher mode you can always override as a deliberate choice rather than the default, logged for future debt pay-down (more on that later).
Comprehension checks. After AI helps you produce something — a memo, a model, a chunk of code — it quizzes you on it. Can you explain why this fact matters? What breaks if we change this assumption? What’s another way to describe the same thing? The MIT study found that people couldn’t quote a sentence from an essay they’d “written” minutes earlier. A comprehension mechanism turns that passive artifact into something you own and probably gets to better outputs, enrolling your newly educated mind in the process.
Spaced return. A cron-job conscience: your AI schedules follow-ups on the things you offloaded under deadline pressure but flagged as “I should actually understand this.” Two weeks later it prompts you, “you shipped something without seeming to really engage with several key ideas within it, which keep surfacing in your other requests; want ten minutes on it now in real time, or a 5-minute primer with relatable metaphors?” Deadlines and productivity-hacking encourage higher degrees of heavy-lift outsourcing, while spaced return lets you repay cognitive debt on your own timeline instead of letting it compound.
Effort receipts. The ledger that’s missing is a record of the delta between what you “produced” and what you actually evidenced as having learned and owned. Not to shame you, but to make the invisible visible. You can’t manage what you can’t see, and right now the erosion is completely unmeasured at the individual level.
Part two: stylistic uniqueness
I predict the annoying AI-generated writing tics and Claude-vanilla-design styles will become a curious quirk of the mid-2020s.
The fix is not better models where we train those tics out, or better slide-making skills, but shifting the burden of differentiation off of the user in every prompt. Today, sounding like yourself, or looking like your company, requires knowing that system prompts and context files exist, and maintaining them actively. That’s a power-user tax, and most people will never pay it. And even for power users who feed hand-crafted templates and writing samples, it’s often not enough to escape the most pernicious AI-isms.
Despite the linguistic AI-isms being hard to train out (because they’re reflected in AI-written training material), I expect newer models will apply filters and skills that screen them out, partly to remove the stigma of AI content. If no one can tell it’s AI, no AI-blame gets cast for the phoned-in essay, no one hates on the “AI slop”, arguably the least-favorite two words in Silicon Valley. That creates problems of its own, of course, for authenticity and provenance.
But screening out what’s generic is only half the job. The harder half is filling that space with what’s yours, and that requires your AI to hold a far richer picture of you than it does today. Right now, most AI products carry something like a thumbnail of the user — a name, a job title, a few saved preferences, maybe a paragraph of custom instructions you wrote once and forgot. What we need is more akin to wallpaper — scaled-up, full-bleed, high-resolution.
Four platform advances can get us there:
Active intake, not passive settings. Onboarding should ask. Not a blank “tell us about yourself” box buried in settings, but a real interview — what do you do, who do you write for, what do you sound like when you’re at your best, show me three things you’ve made that you’re proud of. Every other product category that depends on personalization does structured intake at the start. AI, oddly, mostly waits to be told.
Voice as the intake modality. Typing is a bottleneck for exactly this task, because writing about yourself is effortful and self-conscious in a way that talking often is not. I can ramble for minutes about a product feature I just delivered and my AI learns more about my cadence, my hedges, my go-to metaphors and what I actually care about than it might from an hour of me taking a test about myself. The most efficient path to capturing a voice is to let someone use it unfettered. We aren’t all writers, and we don’t want to be multiple choice test dummies. But we all talk the way we think.
Passive signal capture. Learning from the edits you make, the tones you reject, the peculiar words or sayings you consistently embrace. Every time you rewrite an AI’s sentence, you are giving a free training signal about who you are, and there are no hard mechanisms to ensure that signal is structurally captured.
A higher bias toward reflecting it. This is the one that determines whether any of the above matters in implementation. It isn’t enough to have your voice on file — the model has to weight it heavily against everything else it knows at the moment of generation. Today your uploaded writing samples compete with the statistical nucleus of the pre-training set, and that volume is hard to counter. The advance is a more deliberate retrieval bias toward your material over the general prior.
Beyond voice, the same logic applies to substance. The generic output problem is really a missing-context problem, and your AI is in a far better position than you are to notice what’s missing. An AI that says “this draft would be much stronger if I knew what your customer actually complained about — tell me in your words” is doing something far more valuable than writing better prose. It is re-injecting the person into the output, which is the only durable defense against the sea of sameness.
These have something in common with part one: each asks something of the user. A real conversation at onboarding, a few minutes of talking, an occasional story. That’s the through line of this whole thing — the AI that serves you best is the one that keeps asking you to show up — in effort, in character, in voice.
This is not a pipe dream
Here is why I think this survives the business-model objections raised earlier. Convenience features are easy to copy and easy to leave. Preservation features are different, because they compound. An AI that has spent six months learning which subjects you’re trying to master, how hard to push you, how you actually sound, and what you’ve genuinely internalized versus merely produced becomes a real coach. The switching cost becomes the accumulated understanding of how you think and how you learn. Same goes for organizations.
That is a moat that deepens the longer you stay, and unlike engagement-maximizing design, it deepens because it serves your long-term interests.
Screen time limits threatened the ad model because attention was the product. Here, capability can be the product. That’s a meaningfully different set of incentives, and it’s the reason this time we get further than grayscale mode or app limits.
There’s a version of the AI future where these tools paralyze us as a result of their own convenience. But there’s another version where the same capability, pointed differently, makes us the most capable generation of thinkers: everyone with a patient tutor who refuses to let them coast, who returns them to the hard thing, who insists the work still reflects us. The foundational models can be identical in both futures, the difference is how they are fielded.
There’s a real and growing segment of people who will pay for strategic friction. Enterprises and governments may indeed mandate it if they place a premium on human control and decision-making. “Made people sharper” is a far better outcome than “did it for people and left them behind in the process.” Helping AI serve us symbiotically requires a “do the work” and “do it uniquely for me” type of interaction model. My prediction is simply that the market is about to find ways to act on this, and that whoever delivers on it first will be glad they did.
PS: I ran five progressive hand-written, hand-edited drafts of this through Claude and Gemini for input and pressure-testing. After the latest, Claude asked me: “Want me to save your writing style now? I’ve watched you revise this across five passes and have a sharp read on how you actually write.”
After saving it, it told me: “Saved. One honesty note: this build of Cowork doesn’t have a dedicated writing-style skill wired up, so I couldn’t stash it in a hidden profile that auto-loads — I captured it as a document instead… If you paste a sample of you writing in a different register — a quick note, an email, something with humor — I’ll add a second profile so I don’t apply LinkedIn-essay voice to a Slack message. And tell me anything I mischaracterized; the profile is only as good as the corrections you give it.”
Gemini never offered any such advice. Perhaps Claude is the righteous path to helping to coax exactly what’s needed out of me — but still has work to do on making it turnkey.