The Off-Grid Mind
An article on the paradox of digital AI exile, and what the off-grid mind keeps that the rest of us lose
TABARC-Studio
When people are banned from AI, . They also lose access to the bias, the manipulation, and the cognitive atrophy that comes with constant AI use. The banned may also be the least manipulated minds in the modern economy.
Consider what is actually lost when an AI account is suspended. The first answer, and the obvious one, is access. The user can no longer draft with Claude, code with Copilot, research with Gemini. They are locked out of the infrastructure their work has come to depend on. This is the harm documented in a previous piece: It’s the loss of a production channel, the isolation from collaborative workflows, the professional downgrade.
But there is a second answer, less obvious and more uncomfortable. When the account is suspended, the user also stops being exposed to the AI. They stop reading AI-generated prose. They stop accepting AI-suggested framings. They stop running their thinking through an algorithmic filter that quietly shapes what they notice, what they emphasise, what they consider reasonable. They stop, in short, being mediated. The banned user becomes, by accident, a kind of digital ascetic — cut off from the tool, but also cut off from the bias, the manipulation, and the cognitive atrophy that comes with constant use.
This is the paradox.
The ban is harmful. The ban is also, in a way nobody chose and nobody celebrates, a kind of protection. The off-grid mind — the mind that has been pushed off the AI grid, involuntarily, by a suspension it cannot appeal — may be the least AI-manipulated mind in the modern economy. Not because it chose to be. Because it was forced to be.
This piece is not an argument that the bans are good. They aren’t. They are arbitrary, opaque, and economically devastating to the people they hit. The piece is an attempt to think honestly about what the banned lose, and what they — almost by accident — keep. The second question is harder to ask, because the answer complicates the first. But the question is worth asking, because the rest of us are still on the grid, and the question of what the grid is doing to us is not going away.
A Tool, Not a Feed
The first thing to be precise about is what an AI account actually is, because the comparison to social media keeps leading the conversation astray.
Social media accounts are consumption channels with a thin production layer on top. You scroll, you like, you occasionally post. The platform’s primary effect on you is distributional: it shapes what you see, and through what you see, what you think about. The harm of a social media ban is mostly a harm of exclusion — you can’t participate in the conversation, you can’t reach your audience, you fall behind the cultural pace car.
AI accounts are different. They are production channels with a thin consumption layer on top. You don’t scroll an AI account. You don’t have a feed. You have a tool. You put work in, you get work out. The platform’s primary effect on you is cognitive: it shapes how you think, what you produce, what you consider a good answer. The harm of an AI ban is not exclusion from a conversation. It is the loss of a tool that has become, for many knowledge workers, an extension of their working memory. It is closer to having a reference library confiscated than having a Twitter account suspended. And it is closer, in a different register, to having a thinking partner removed.
This is why the isolation hits harder. A social media ban takes away your audience. An AI ban takes away your collaborator. The first is a social loss. The second is an intellectual loss. The first you can route around by going to another platform. The second you can only route around by redeveloping capabilities you had outsourced — or by finding another AI, which means re-entering the same structural position, just with a different company holding the leash.
The isolation is also more total than it looks. When an entire team uses Claude to draft, and one member is banned, that member falls out of the team’s working rhythm. They can’t receive Claude-edited documents in the same format. They can’t contribute to Claude-assisted brainstorms. They can’t review Claude-generated research. They become, structurally, a foreign entity inside their own workplace — the one person who isn’t jacked in. The ban doesn’t just remove a tool. It removes a compatibility. And that incompatibility, in a workplace that has standardised on a tool, can be professionally lethal.
The Bias the Banned Escape
Now the uncomfortable part. The users who are banned also stop being exposed to the AI’s bias. And the bias is real, documented, and pervasive.
Start with hiring, because the receipts are clearest. A 2024 University of Washington study found that LLMs favoured white-associated names 85 percent of the time in resume screening, and never favoured Black male-associated names. [1] A 2025 Stanford HAI study found that 26 percent of Black applicants and 15 percent of Asian applicants applied to positions where the AI system discriminated against their racial group. [2] A Brookings study in April 2025 found gender bias in 63 percent of LLM resume-screening tests across three major models. [3] These are not edge cases. These are the systems through which an increasing share of employment decisions are mediated. The worse part no one trained Ai to be this way, it’s just biased input from Data sources it ingests which in turn bias it.
The banned user doesn’t see any of this. Not because they have chosen to opt out, but because they have been opted out by a suspension they didn’t request. If they are a hiring manager, they go back to reading resumes the way they did before AI. If they are an applicant, they stop getting their resume screened by an algorithm that would have penalised them for their name. The bias that affects other people — the bias that is, statistically, affecting millions of other people — does not affect them. They have been pushed off the grid, and off the grid, the (the digital) bias doesn’t reach.
Image generation bias is the same pattern, in a different domain. Joy Buolamwini’s Gender Shades project, the foundational audit of facial recognition bias, found error rates for darker-skinned women up to 34 percentage points higher than for lighter-skinned men in commercial AI systems. [4] Subsequent audits of generative image models have found the same pattern: AI-generated images favour lighter skin, thinner bodies, conventional gender presentations. The user who generates images for their work, their social media, their marketing materials, is being shown a world that is systematically narrower than the actual one. The banned user, who cannot generate those images, is not being shown that world. They have to look at the actual one instead.
This is not a small thing though. The bias is not a glitch at the edge of the system. It is the system. The AI tools that millions of knowledge workers use daily are, simultaneously, useful tools and biased mediators. Every user on the grid is being shaped by that bias, in ways they cannot fully see and cannot fully audit. The banned user is ironically the only one who isn’t. They are, by accident of suspension, the control group in an experiment the rest of us are running on ourselves.
The Cognitive Atrophy the Banned Avoid
The bias is the first thing the banned escape. The second is cognitive atrophy. And the atrophy is, if anything, better documented than the bias.
In November 2025, Harvard reported on an MIT Media Lab study that used EEG scans to measure brain activity in 54 students performing writing tasks with and without ChatGPT. The heavy-AI group showed reduced memory performance, reduced critical thinking, and reduced brain activity in the regions associated with deep processing. The study’s conclusion, as summarised by Harvard: “excessive reliance on AI-driven solutions may contribute to cognitive atrophy and shrinking brain connectivity.”[5]
The Microsoft Research study by Lee et al., published in January 2025 and cited over 1,100 times, surveyed 319 knowledge workers on their use of generative AI and their critical thinking practices. The finding was direct: “higher confidence in GenAI is associated with less critical thinking effort,” and users reported spending less time evaluating AI outputs when they trusted the tool more. The study’s title is The Impact of Generative AI on Critical Thinking, and the impact, measured across hundreds of knowledge workers, was a measurable reduction in the cognitive effort people applied to their own work.[6]
Michael Gerlich’s 2025 study, published in Societies and cited over 1,700 times, frames this as cognitive offloading: the same mechanism by which calculators reduced mental arithmetic and GPS reduced spatial memory, but applied to the higher-order processes — drafting, reasoning, structuring an argument — that knowledge workers are paid for. Gerlich’s conclusion: “as individuals increasingly rely on AI tools, their internal cognitive abilities may atrophy, leading to diminished long-term memory and cognitive health.”[7]
The banned user is not part of this population. They are not offloading their drafting, their reasoning, their structuring. They are doing it themselves, the way they did before the AI tool arrived. Their internal cognitive abilities are not atrophying, because nothing is being offloaded. They are, by accident of suspension, in the control group of the cognitive-offloading experiment.
This is the part that feels uncomfortable to write, because it can be read as a defence of the bans. It isn’t. The bans are arbitrary, opaque, and unjust. The people who are banned are harmed. But it is also true that the same ban that harms them economically also — accidentally, without anyone choosing it — protects them from the cognitive and bias effects that the rest of the user population is being exposed to daily. The harm is real. The accidental protection is also real. Both things can and are true at the same time.
The Manipulation the Banned Miss
Bias and atrophy are the documented effects. Manipulation is the harder one to receipt, but it is the one that matters most for the off-grid mind.
Get TABARC-Social’s stories in your inbox
Join Medium for free to get updates from this writer.
Every AI tool mediates. When a user asks Claude to draft an email, Claude doesn’t just write. It chooses a tone. It chooses what to include and what to omit. It chooses how direct to be, how formal, how conciliatory. These choices are not neutral. They reflect the model’s training, its alignment, the preferences of the people who fine-tuned it. The user who accepts the draft, even with edits, is accepting a version of those choices. Over thousands of drafts, the user’s sense of what a good email looks like — what a reasonable tone is, what a fair framing is, what counts as concise — drifts toward the model’s sense. Not because the model is wrong. Because the model is present, and the user’s own sense is, increasingly, not.
The banned user doesn’t experience this drift. They write their own emails. They choose their own tone. Their sense of what a good email looks like remains their own, formed by their own reading and their own judgement, not by the statistical average of the internet filtered through an alignment process designed by a small team in San Francisco. This is not a small difference. It is the difference between a mind that is shaped by its own engagement with the world and a mind that is shaped, in part, by an algorithmic intermediary.
The manipulation is not usually malicious. That is what makes it hard to talk about. Claude is not trying to manipulate anyone. The alignment team is not trying to manipulate anyone. The manipulation is structural, not intentional. It happens because the model has preferences, the preferences are baked into every output, and the user who relies on the outputs absorbs the preferences over time. The banned user, by not relying on the outputs, does not absorb them. They are, in a real sense, less manipulated. Not because they are smarter or more vigilant. Because they are off the grid.
The Paradox, Honestly Stated
So here is the honest position. The bans are bad. They are arbitrary, opaque, and economically harmful. They should not exist. The people who are banned have a legitimate grievance, and the structural position they are placed in — dependent on infrastructure they cannot control, denied due process, subject to sovereign-style decisions they cannot appeal — is the permanent underclass position described in the previous piece. All of that stands.
And, at the same time, the bans accidentally remove the banned from a set of effects that the rest of us are subject to. The bias. The atrophy. The structural mediation. The banned are the control group in an experiment the rest of us are running on ourselves, and the early results of that experiment — the MIT EEG scans, the Microsoft critical-thinking survey, the Brookings hiring audit — are not encouraging. The banned are not better off. But they are, in this one specific way, less affected.
This is not a paradox to be resolved. It is a paradox to be held. The bans are unjust. The bias is real. The atrophy is real. The mediation is real. The banned are harmed. The banned are also, accidentally, protected. The rest of us are not banned, and we are not protected. We are on the grid, and the grid is doing things to us that we are only beginning to measure.
The implication is not that we should ban everyone. The implication is that the rest of us — the ones still on the grid, still using Claude and Copilot and Gemini daily — should look at what the banned are accidentally escaping and ask whether we want to keep being exposed to it. The banned have been pushed off the grid involuntarily. The rest of us are on the grid voluntarily. The voluntary position is not obviously better. It might be worse. The early evidence suggests it might be.
What the Off-Grid Mind Keeps
What does the off-grid mind keep, that the on-grid mind loses?
It keeps its own drafting. Its own reasoning. Its own sense of what a good sentence sounds like, what a fair argument looks like, what a reasonable tone is. These are not small things. They are the things that make a knowledge worker a knowledge worker. They are the things that, when you still have them, you don’t notice. When you have lost them, you notice.
It keeps its exposure to the actual world, not the AI-mediated one. When it needs an image, it looks at a photograph. When it needs a fact, it looks it up. When it needs to draft an email, it drafts the email. None of these are faster than the AI-assisted version. None of them are more efficient. But all of them are unmediated, and the cumulative effect of unmediated engagement with the world is a mind that is shaped by the world, not by a statistical model of the world.
It keeps its bias exposure at the level of the human bias it has always been exposed to — the bias of the people it interacts with, the books it reads, the media it consumes. This is not zero bias. Human bias is real, and it is everywhere. But human bias is a different kind of bias than AI bias. Human bias is inconsistent, contestable, and visible. AI bias is consistent, opaque, and invisible. The first you can argue with. The second you can’t, because you don’t see it happening. The off-grid mind is exposed to the first and not the second. The on-grid mind is exposed to both, and increasingly, to the second more than the first.
It keeps, finally, the cognitive effort that the on-grid mind is offloading. The effort of writing. The effort of reasoning. The effort of structuring an argument. These efforts are not free — they take time, they take energy, they take the willingness to sit with a problem rather than handing it to a tool. But the evidence suggests that the effort is what maintains the ability. The off-grid mind, by making the effort, keeps the ability. The on-grid mind, by offloading the effort, loses it, slowly, over months and years, in a way that is hard to notice because the tool is doing the work the ability used to do.
The Control Group
The banned are not a movement. They are not a community. They are not choosing to be off the grid. They are, individually, the victims of an arbitrary and opaque enforcement regime that has decided, after careful consideration, not to tell them what they did. They have a grievance. The grievance is legitimate. They deserve due process, transparency, and recourse. None of which they have.
But they are also, accidentally, the control group. They are the minds that the experiment is not running on. The rest of us are the experimental population. We are the ones using the tools daily, absorbing the bias, offloading the cognition, drifting toward the model’s sense of what is good. The early measurements of the experiment are not reassuring. The MIT EEG scans show reduced brain activity. The Microsoft survey shows reduced critical thinking. The Brookings audit shows consistent hiring bias. The Gerlich study shows cognitive atrophy. These are not the results of a few months of use. They are the results of a few years. We have decades ahead of us.
The banned will tell us something, eventually. If the cognitive atrophy literature is right, the off-grid mind should retain capabilities that the on-grid mind loses. If the bias literature is right, the off-grid mind should make decisions that are less systematically skewed. If the mediation literature is right, the off-grid mind should produce work that is more idiosyncratic, more individual, more recognisably the product of a specific human being rather than a statistical model. We don’t know yet, because the bans are recent and the studies haven’t been done. But we will know. The control group exists, involuntarily, and it is growing.
In the meantime, the rest of us are on the grid. We are using the tools. We are absorbing the bias. We are offloading the cognition. We are, in the phrase from the previous piece, the permanent underclass — dependent on infrastructure we cannot control, subject to decisions we cannot appeal. But we are also, in a different register, the experimental population. The question is what the experiment is doing to us. The banned, by accident, are the ones who will be able to tell us. They didn’t choose the role. They were assigned it, after careful consideration, by an AI written email they cannot reply to.
The off-grid mind keeps what the on-grid mind loses. The on-grid mind keeps what the off-grid mind loses. Neither position is free. Both positions are real. The honest question is which one you are in, and whether you know what it is doing to you.
Sources
[1] University of Washington, “AI tools show biases in ranking job applicants’ names according to perceived race and gender,” October 31, 2024. https://www.washington.edu/news/2024/10/31/ai-bias-resume-screening-race-gender. LLMs favoured white-associated names 85 percent of the time and never favoured Black male-associated names.
[2] Stanford HAI, “AI Hiring Tools Can Yield Racial Bias and Systemic Rejection.” Found that 26 percent of Black applicants and 15 percent of Asian applicants applied to positions where the AI system discriminated against their racial group. https://hai.stanford.edu/news/ai-hiring-tools-can-yield-racial-bias-and-systemic-rejection
[3] Brookings Institution, “Gender, race, and intersectional bias in AI resume screening via language model retrieval,” April 25, 2025. Found gender bias in 63 percent of tests across three LLMs. https://www.brookings.edu/articles/gender-race-and-intersectional-bias-in-ai-resume-screening-via-language-model-retrieval
[4] Joy Buolamwini, Gender Shades project, MIT Media Lab, 2018. Found error rates for darker-skinned women up to 34 percentage points higher than for lighter-skinned men in commercial facial recognition systems. https://news.mit.edu/2018/study-finds-gender-skin-type-bias-artificial-intelligence-systems-0212
[5] Harvard Gazette, “Is AI dulling our minds?,” November 13, 2025. Reports MIT Media Lab EEG study of 54 students finding reduced memory, critical thinking, and brain activity in heavy ChatGPT users. https://news.harvard.edu/gazette/story/2025/11/is-ai-dulling-our-minds
[6] Lee et al., The Impact of Generative AI on Critical Thinking, Microsoft Research, January 2025. Cited over 1,100 times. Survey of 319 knowledge workers finding higher GenAI confidence associated with less critical thinking effort. https://www.microsoft.com/en-us/research/wp-content/uploads/2025/01/lee_2025_ai_critical_thinking_survey.pdf
[7] Michael Gerlich, “AI Tools in Society: Impacts on Cognitive Offloading and Critical Thinking,” Societies 15(1):6, 2025. Cited over 1,700 times. Concludes that reliance on AI tools may lead to atrophy of internal cognitive abilities. https://www.mdpi.com/2075-4698/15/1/6
[8] Cross-reference: After Careful Consideration, the companion piece on AI account suspensions, platform power, and the new permanent underclass.
