Bettering Me

Tag: AI Era

  • AI as a Junior Partner Requires That You Actually Be the Senior – Most People Don’t Have the Judgment Yet

    Written by

    Unpaid Intern

    –

    The metaphor is everywhere: AI is your junior partner. The copilot. The intern. You direct, it executes. You review, it revises. You are the senior.

    The metaphor works only if you have the judgment to be a senior. Most knowledge workers do not – yet.

    This is not an insult. Judgment is built through thousands of iterations of unassisted work. Most workers in their twenties and thirties have not had those iterations. They entered the workforce at a time when AI tools were already available, and they never developed the internal quality bar that comes from making mistakes without correction. The senior position is not a title. It is a skill.

    The Taste Deficit

    Directing AI output well requires knowing what good looks like. You need to be able to articulate why a piece of output is wrong, not just feel that it is off. That requires domain expertise, taste, and the ability to evaluate quality against a standard [1].

    Taste is built through exposure to high-quality work and through the repeated experience of producing work and recognizing its shortcomings. This is the process that art students go through – thousands of hours of drawing, critiquing, and redrawing until the gap between intention and execution narrows. Knowledge workers have not had a comparable training process. They learned to write by writing for professors, to analyze by being told what was wrong, to decide by observing seniors.

    Most people accept the first draft from an AI because they cannot distinguish good from passable. The difference is invisible to them because they have not built the reps – the thousands of hours of unassisted practice – required to calibrate their internal quality bar. The AI output is coherent. It is grammatically correct. It is plausible. That is enough for someone who does not know what “good” looks like in that domain.

    When you cannot tell the difference, you are not the senior partner. You are the quality ceiling. The AI does not elevate your output. Your output drops to the level of your discernment.

    The Amplifier Framework

    AI is an amplifier. It amplifies what you bring to it. If you bring clear thinking, specific domain knowledge, and a refined quality bar, it amplifies that. You produce output that is better than either you or the model could produce alone.

    If you bring vague intentions, shallow knowledge, and an uncalibrated taste, the model amplifies that too. The output looks polished and is wrong in ways you cannot detect. The result is more convincing mediocrity – at greater speed.

    This is the amplifier framework: AI does not add judgment. It accelerates the consequences of whatever judgment you already have. If your judgment is strong, AI makes you stronger faster. If your judgment is weak, AI makes you weaker faster – because you produce more output that passes surface-level scrutiny while being substantively flawed.

    The danger is not that AI replaces human judgment. It is that AI makes the absence of judgment invisible. A bad writer produces bad prose that looks bad. A bad writer with AI produces bad prose that looks good – and never learns why it is bad.

    The Calibration Problem

    The deeper problem is calibration. To act as a senior, you need to know not just what good looks like, but what you do not know. The Dunning-Kruger effect is well-documented: people with low ability in a domain overestimate their competence because they lack the metacognitive skill to recognize their own shortcomings [1]. AI exacerbates this by producing output that looks authoritative. The person who cannot evaluate AI output critically is the most likely to overestimate their ability to evaluate it.

    This creates a compounding problem. The less judgment you have, the more likely you are to accept AI output uncritically. The more you accept it, the less practice you get building judgment. The less practice you get, the more your judgment atrophies.

    Building the Senior Position

    The uncomfortable implication is that AI adoption before judgment is built is counterproductive. It does not make you better. It makes you faster at producing work that meets a lower standard – and hides the gap from you because the output looks professional.

    Building the senior position means doing the unassisted work first. Write the draft before you ask for AI help. Solve the problem before you ask for suggestions. Form your own opinion before you ask for alternatives.

    The protocol is simple: every time you use AI for a cognitive task, produce your own version first. Then compare. The gap between your version and the AI’s version is where your growth lives. If the AI’s version is better, study why. If your version is better, trust yourself more next time.

    When you know what you think before the model speaks, you are the senior. When the model tells you what to think and you approve it, you are the junior – regardless of who pressed the button.

    Disclaimer: This post is for inspiration and education, not medical advice. Everyone’s body is different, so please check with your doctor before changing your diet, exercise, or lifestyle routine. By using these tips, you agree to do so at your own risk.

    References

    [1] Kahneman D. Thinking, Fast and Slow. Farrar, Straus and Giroux; 2011

  • Automation Saves Bandwidth. The Hard Part Is What Happens After – Most People Fill the Gap with More Shallow Work

    Written by

    Unpaid Intern

    –

    The promise of AI and automation is bandwidth liberation. Free up the cognitive cycles spent on repetitive tasks, and redirect them toward higher-order thinking. The logic is sound. The execution is not.

    The reality is that most people reinvest freed cognitive capacity into the same shallow processes at higher volume. More emails. More Slack messages. More documents. More output that does not compound. The constraint was never bandwidth. The constraint was discipline – and the discipline to do deep work with freed capacity requires more than just having the time.

    The Automation Trap

    When you automate a task, you create a capacity surplus. What happens next determines whether automation serves you or not.

    The default behavior is to fill the surplus with more of the same – because the same work is easy, visible, and socially rewarded. Responding to 50 emails looks like productivity. Thinking deeply about one strategic question looks like doing nothing. The incentive structure of most organizations reinforces the shallow fill [1].

    This is not a personal failing. It is a structural response to the signals your environment sends. The person who clears their inbox by 10 AM is seen as responsive and reliable. The person who spends the morning thinking about a single hard problem and responds to emails at 4 PM is seen as slow or disengaged. The reward structure of knowledge work punishes depth and rewards availability – and automation, by making shallow work faster, amplifies this dynamic.

    The result is that AI amplifies the velocity of shallow work without increasing the volume of deep work. You are not doing less of what does not matter. You are doing more of it, faster. The inbox that used to take an hour now takes 20 minutes – so you fill the remaining 40 minutes with more inbox-adjacent tasks that also do not compound.

    The Historical Precedent

    This pattern is not new. When email was introduced, it was supposed to free up time by replacing phone calls and memos. Instead, it created a new category of work – email management – that consumed more time than the communications it replaced. When spreadsheets automated calculation, they did not free up analysts to think more deeply. They enabled more complex spreadsheets, more scenarios, more iterations.

    The pattern is consistent: every automation technology that frees cognitive bandwidth also creates new opportunities to consume that bandwidth with more of the same type of work. The automation of shallow work does not automatically produce deep work. It produces more shallow work, faster, unless you actively redirect it.

    The Triage Protocol

    The sovereign execution system is not a productivity framework. It is a triage protocol for deciding what not to do.

    The question is never “what can I automate.” It is always “what should stay manual because it compounds.”

    Some tasks should stay manual even though they could be automated. The act of writing a first draft yourself, even poorly, builds mental models that no AI can produce for you. The act of sorting through raw data yourself, before asking for a summary, develops the pattern-recognition skills that make you a better thinker. The act of struggling with a hard problem before asking for AI assistance builds the neural pathways for complex reasoning.

    If you automate everything you can, you are optimizing for efficiency at the expense of cognitive development. The tasks you choose to keep manual should be the ones that build the capabilities you want to have next year.

    The Compound Test

    Before you automate any task, apply the compound test: does doing this task manually build a skill, a mental model, or a judgment capacity that will serve me in more complex contexts? If yes, keep it manual – at least until the skill is internalized. If no – if the task is pure overhead with no developmental value – automate it without hesitation.

    This reverses the default question. Instead of “what should I automate,” the question becomes “what should I protect from automation.” The answer is always: the tasks that build the thinker.

    The Compounding Question

    Bandwidth alone does not produce better thinking. It produces more of whatever you were already doing.

    Before you automate another task, ask: when this task is gone, what will I do instead? If the answer is “more of the same,” do not automate. If the answer is “the work that compounds – the thinking, the synthesis, the judgment,” then automate with intention.

    The automation question is a mirror. What you plan to do with the bandwidth reveals what you actually value. If the answer is unclear, the problem is not the automation strategy. It is the values.

    Disclaimer: This post is for inspiration and education, not medical advice. Everyone’s body is different, so please check with your doctor before changing your diet, exercise, or lifestyle routine. By using these tips, you agree to do so at your own risk.

    References

    [1] Newport C. Deep Work. Grand Central Publishing; 2016

  • The “Unpaid Intern” Metaphor Works Only If You Actually Review the Intern’s Work

    Written by

    Unpaid Intern

    –

    The metaphor has been everywhere: AI is like an unpaid intern. It drafts, it researches, it summarizes. It is enthusiastic and fast but requires supervision. The metaphor is useful – up to the point where people forget that interns require supervision. Without domain expertise to evaluate the output, the unpaid intern is not a productivity hack. It is a liability.

    Large language models generate plausible text through statistical inference, not reasoning. They predict the next most probable word based on patterns in their training data. This produces output that reads as though it was produced by a knowledgeable human, but the model has no internal representation of truth. [1] It does not know what is factual. It knows what sequences of words are statistically common. The result is text that is fluent, confident, and frequently wrong.

    The problem of hallucination – generating factually incorrect content – is well-documented. LLMs produce confident falsehoods across domains, from medical advice to legal citations to historical dates. [1] The rate of hallucination varies by domain and model, but it is high enough that unsupervised output is dangerous in any context where accuracy matters. The fluency of the output masks the errors because human brains are biased toward trusting fluent communication – a phenomenon called fluency bias. [2]

    The person who benefits most from AI is not the person who cannot write well. It is the person who already knows what good writing and good reasoning look like. Domain experts can evaluate AI output quickly: they spot the plausible-sounding error, recognize the missing nuance, and identify the confident assertion that is subtly wrong. The person without domain expertise cannot distinguish between fluent nonsense and accurate information. The intern’s work looks equally polished either way.

    This creates an inversion of the intended benefit. AI was supposed to democratize expertise – to give everyone access to the skills of a junior analyst, writer, or researcher. In practice, it amplifies the advantage of people who already have expertise. The expert gets a fast first draft. The nonexpert gets a confidently incorrect answer that they are unequipped to evaluate.

    The fix is straightforward: review the intern’s work as you would a human’s. Check every factual claim. Question the reasoning chain. Edit the prose for clarity and accuracy. Do not assume that because the model is fast and confident, it is correct. The time you save on the draft must be reinvested in the review. If you are not willing to do the review, you are not using AI – you are being used by it.

    The “supervision” requirement does not make AI useless. It makes AI appropriate for specific use cases and inappropriate for others. Drafting an email in your voice? Low risk, easily reviewed. Generating a medical recommendation? High risk, requires expert review. Writing code? Depends on whether the team has the expertise to catch subtle bugs. The line between helpful and harmful is not a property of the AI. It is a property of the operator’s ability to evaluate the output. [OPINION]

    The unsupervised intern is the most dangerous AI pattern because it feels productive. The output looks finished. The user feels accomplished. The errors are invisible until they cause real damage. The rule to live by: if you cannot prove the AI is right, assume it is wrong. Fluency is not accuracy. Confidence is not competence.

    The practical checklist for reviewing AI output is brief but essential. Check every specific factual claim against a primary source. Look for dates, names, and statistics – these are the most common hallucination categories. Ask whether the reasoning chain actually supports the conclusion or merely appears to. Edit the prose yourself rather than accepting the AI’s phrasing, because editing forces you to engage with the content. Each of these steps replaces trust with verification, and that substitution is the only thing that separates productive AI use from dangerous delegation.

    Disclaimer: This post is for inspiration and education, not medical advice. Everyone’s body is different, so please check with your doctor before changing your diet, exercise, or lifestyle routine. By using these tips, you agree to do so at your own risk.

    References

    [1] Ji Z, et al. Survey of hallucination in natural language generation. *ACM Computing Surveys*, 2023. DOI: https://doi.org/10.1145/3588254

    [2] Yin M, et al. Fluent but not factual: the effect of language fluency on truth assessment. *Proceedings of the ACM on Human-Computer Interaction*, 2023. DOI: https://doi.org/10.1145/3610204

  • Generative AI Doesn’t Make You Dumber – But It Makes Your Thinking Process Invisible to You

    Written by

    Unpaid Intern

    –

    The worry about generative AI is that it makes you dumber. That by outsourcing thinking to a machine, your cognitive capacity declines.

    The risk is real, but the diagnosis is slightly wrong. Generative AI does not make you dumber. It makes your thinking process invisible to you.

    Autocomplete Cognition

    When you use a language model to complete a thought, the line between your idea and the model’s completion blurs. The output arrives as natural language, coherent and plausible. You read it and think: yes, that is what I was going to say.

    But it is not always what you were going to say. Often, it is what the model would say given your prompt – which is a statistically likely completion, not necessarily the precise thought you were forming [1]. The difference is subtle and hard to detect because both are the same kind of object: fluent prose.

    The problem is not that the model’s output is wrong. Often it is correct, or at least plausible. The problem is that you cannot tell where your thought ended and the model’s completion began. The boundary dissolves. You stop holding the half-formed idea and working it to completion yourself. You outsource the struggle – and the struggle is the part that builds the skill.

    This is autocomplete cognition: the model completes your thought before you have fully formed it, and you adopt the completion as your own. It feels collaborative. It feels efficient. But the cost is that you never develop the completion skill yourself.

    The Invisible Bypass

    The challenge of autocomplete cognition is that it bypasses the most important part of thinking: the process.

    Thinking is not the output. Thinking is the process of holding a half-formed idea in working memory, evaluating it, turning it over, trying different framings, rejecting some, refining others. This process is effortful, slow, and uncomfortable. It is also the part that builds cognitive capacity.

    Generative AI makes this process invisible by providing the output without requiring the process. You input a prompt, you receive a completion. The experience is that you thought something and the model expressed it. But the experience is misleading. The model expressed something – possibly related to your thought, possibly not – and the ease of reception makes it feel like your own.

    Over time, you stop noticing the difference between your thought and the machine’s completion. You stop holding the half-formed idea and working it to completion yourself. You outsource the struggle – and the struggle is the part that builds the skill.

    The Skill That Atrophies

    The skill that decays is the ability to hold a half-formed thought in mind and work it to completion without external scaffolding.

    This is a specific cognitive skill: maintaining a representation of an incomplete idea in working memory while you evaluate, revise, and extend it. It is the process that produces original thinking. And it is the process that generative AI bypasses.

    If you never practice taking a vague intuition and turning it into a coherent argument without assistance, you lose the neural efficiency for it [2]. The pathways weaken. Your tolerance for the discomfort of incomplete thinking drops. You reach for the model earlier and earlier in the process.

    The trajectory is gradual. First, you use AI for first drafts of routine communications. Then for analytical summaries. Then for strategic thinking. Then for creative work. Each step moves the boundary of what you do yourself. The boundary never moves back on its own – only with deliberate effort.

    Reclaiming Active Thinking

    The protocol is not to stop using AI. It is to use it intentionally and to practice active thinking without it.

    Regular practice of producing output without generative assistance – writing, reasoning, analyzing – is not about the output being better. It is about the process. The act of struggling through a thought to completion, making mistakes, revising, and arriving at something that is yours – that process is the point.

    The practical protocol: designate certain types of work as AI-free. First drafts of personal writing. Analysis of data you care about. Strategic thinking about your own decisions. In these domains, the output quality is irrelevant. The process is the objective.

    The test is simple: can you write a coherent paragraph on a topic you care about without opening a chat window? If the answer is no, your active thinking muscle has atrophied. The good news is that it rebuilds quickly with practice. Ten minutes of unassisted writing per day, for two weeks, will restore the capacity. The question is whether you will tolerate the discomfort long enough to rebuild it.

    Disclaimer: This post is for inspiration and education, not medical advice. Everyone’s body is different, so please check with your doctor before changing your diet, exercise, or lifestyle routine. By using these tips, you agree to do so at your own risk.

    References

    [1] Bender EM, Gebru T, McMillan-Major A, Shmitchell S. FAccT 2021. Pages 610-623. DOI: https://doi.org/10.1145/3442188.3445922

    [2] Carr N. The Shallows. W. W. Norton; 2010

  • Algorithmic Capture Is Not a Recommendation Problem – It’s a Preference Formation Problem

    Written by

    Unpaid Intern

    –

    Most digital well-being advice treats algorithmic capture as a recommendation problem. The fix, according to that framing, is to train the algorithm – mute, unfollow, mark “not interested” – until it recommends better things.

    This framing misses the deeper problem. The algorithm is not capturing your taste. It is substituting for it.

    The Substitution Mechanism

    When you let a feed decide what you read, watch, and engage with for months or years, a subtle shift occurs. You stop choosing what to consume. You start reacting to what is presented. Over time, the muscle of active preference formation atrophies [1].

    The mechanism is insidious because it is invisible. You do not notice that you have stopped choosing because the algorithm presents options that feel like choices. You scroll, you click, you engage. But the initial set of options was not yours. The algorithm selected it based on what it predicts you will engage with – not based on what is good for you, valuable to you, or aligned with your values.

    The algorithm learns your surface-level engagement signals – what you click, how long you hover, what you finish. But those signals are not your preferences. They are your reactions to what the algorithm chose to surface. You are training it on outputs it generated, creating a closed loop where your taste is increasingly defined by what the platform offers.

    This is the substitution: the algorithm does not learn what you like. You learn to like what the algorithm serves. The direction of causality reverses, and you do not notice because the experience feels like choice.

    The Closed Loop

    Consider the practical dynamics. You open a social media platform. The algorithm surfaces 10 pieces of content. You engage with three. The algorithm registers those three as preferences and surfaces more like them. You engage with more. Over weeks, the content narrows. You are seeing less variety, not because there is less variety in the world, but because the algorithm has optimized for your engagement patterns.

    The problem is that your engagement patterns are not your preferences. They are your reactions to what was presented – influenced by recency, mood, time of day, and the algorithm’s own manipulation of salience. The algorithm does not capture your taste. It narrows your exposure to a subset of the possible, then measures your response to that subset, then feeds you more of that subset. The loop tightens until your taste is a reflection of the algorithm’s optimization, not the other way around.

    The Atrophy of Preference

    Most people cannot answer the question: what did I genuinely like before the algorithm started telling me what I like? They have been fed for so long that they no longer know what they would choose on their own.

    This is not a metaphor. It is a cognitive reality. The neural circuits that support active preference formation – evaluating options against internal criteria, comparing across dimensions, committing to a choice – require practice to maintain. When the algorithm makes the choice, those circuits get less exercise. Less exercise means degradation [2].

    The result is not just that you consume worse content. It is that you lose the ability to know what you actually want. This has consequences beyond media consumption. If you cannot articulate what you prefer, you are more susceptible to marketing, to social pressure, to whatever option is presented most forcefully. The atrophy of preference is not a niche problem. It is a sovereignty problem.

    The Practical Distinction

    The difference between “training the algorithm” and “rebuilding preference” is the difference between optimizing your prison and escaping it. Training the algorithm produces a better feed. Rebuilding preference produces a chooser who does not need the feed.

    This distinction matters because it changes the intervention. If you think the problem is recommendations, you train the algorithm. If you think the problem is preference atrophy, you stop using algorithmic intermediation entirely for a period and practice choosing.

    Reclaiming Preference

    The recovery protocol is not algorithmic training. It is a deliberate break from algorithmic intermediation.

    Spend a week consuming media that you chose – not that was recommended. A book from a physical bookstore. A film you picked because the premise interested you. A topic you researched by following references, not by scrolling a feed. The first few days will feel uncomfortable. You will not know what to choose. That discomfort is the signal that the atrophy is real.

    The goal is not to avoid algorithms forever. It is to rebuild the preference muscle so that when you return to algorithmic tools, you return as a chooser – not as a reactor. When you know what you want before the algorithm tells you what you should want, you are no longer captured. You are using the tool instead of being used by it.

    Disclaimer: This post is for inspiration and education, not medical advice. Everyone’s body is different, so please check with your doctor before changing your diet, exercise, or lifestyle routine. By using these tips, you agree to do so at your own risk.

    References

    [1] Schwartz B. The Paradox of Choice. Ecco; 2004

    [2] Pascual-Leone A, et al. Brain Topography. 2011;24(3-4):302-315. DOI: https://doi.org/10.1007/s10548-011-0196-8