polish is a dead signal

this is the meme that started it. I've been chewing on why it bothers me for about a year

highest-paid-engineer.png

ok so. most of us have been there. the ask comes in around 4pm - usually on the day you're most excited to just not be fucking working anymore. "can you send me a quick write-up on the issue? I need to get something over to the partner team"

so the technical team stops what it's doing. someone who actually understands the problem spends 45 minutes writing it up clearly: what broke, why, what's being done, what we could've done better, how we make sure it doesn't happen again, what the partner needs to know. packaged, tidied up, sent off.

and then an executive - a Chief Client Email Sender™ - opens it, copies it, pastes it into a fresh email, signs their name, and sends it to the external partner. no attribution. no reframing for the audience. no attempt to understand what they just put their name on. grab it, sign it, send it.

you've seen this. you may have lived both sides of it. and here's the thing: this is not an AI story. the CCES™ has been doing this since long before ChatGPT existed. their technical team has always been their personal chatbot - prompt it with a vague question, take the output verbatim, present it as your own thinking, never read for comprehension. AI didn't invent the move. it just gave everyone else access to it.

polish used to be proof of work

for the past few decades, polish was a rational signal to read. a tight deck meant someone spent hours thinking, because there was no way to produce the polish without doing at least some of the work. the suit, the structured memo, the confident prose... proxies for trust in a world where verifying substance directly was expensive and constant. orgs couldn't personally audit everyone's competence, so they graded the wrapper. honestly not stupid. just economics. a signal carries information when it's costly to fake.

but that equilibrium is gone now, and it broke from both ends at once.

first: polish became free. or at least it started costing tokens instead of effort. fluent-sounding prose, clean-ish structure, a confident executive summary - these now cost nothing to produce and therefore certify nothing about the thinking underneath. and no, you can't "just spot the fakes." stanford/penn state research found people distinguish AI text from human writing about 53% of the time. a coin flip gets you 50. training barely moves the needle. meanwhile by late 2025, AI-generated articles crossed the halfway mark of everything new published on the web (per the imperfect detectors we have to measure with, anyway). the wrapper is now infinite and nobody can reliably tell by looking.

second - and this is the part that gets less attention - substance got cheaper to verify at the same time. the distance between an idea and a working artifact has collapsed. the deck about the tool is decreasingly necessary... and even when you do produce it, someone else is using AI to read it lol. in most cases you can have the actual tool by the next business day. the demo, the shipped thing, the reproducible result is now the cheap verification method, not the expensive one.

so the old rule inverted. an org that still reads "well-written" and "nicely formatted" as "well-thought" isn't being traditional. it's running a fraud-detection system whose method of detection has been removed. and it will now select for veneer people at scale, faster than ever.

none of this is a new disease btw. Orwell warned in 1946 that ready-made phrases will "even think your thoughts for you" - that prefab language exists partly to conceal the absence of thinking, including from the writer. Harry Frankfurt said the essence of bullshit is "this indifference to how things really are." the disease is old. what's new is we built a machine that produces the indifference in unlimited quantities, wrapped in perfect grammar, and half the professional world is still grading the grammar.

the tax lands downstream

here's why this isn't just an aesthetic complaint. when someone ships a beautifully wrapped document with no thinking inside, the thinking still has to happen. it just happens later, downstream, by someone else, under worse conditions. we call it rework and pretend it's a process problem. it's a thinking-transfer problem. and off-loaders gonna offload.

the data on this is newer than you'd expect and worse than you'd hope. in 2025 BetterUp Labs + Stanford's Social Media Lab surveyed 1,150 US desk workers and gave the thing a name: workslop. AI content that "masquerades as good work, but lacks the substance to meaningfully advance a given task." their core finding is where the cost lands - downstream, on whoever has to interpret, correct, or redo it. 40% of workers got workslop in the past month. ~2 hours to resolve each incident. their estimate of the invisible tax: $186/employee/month, which at a 10k-person company runs past $9M a year. (self-reported opt-in survey, so hold the dollar figures loosely. the mechanism is the point, and the mechanism is one you've lived: the polished nothing arrives, someone competent unwraps it, discovers it's empty, and does the actual work on a compressed timeline.)

David Graeber saw the pre-AI version of this years ago. his taxonomy of pointless work included flunky jobs, which "exist only or primarily to make someone else look or feel important," and box tickers, whose output exists so the org can claim something happened that didn't. the CCES™ runs both plays at once - converts their team into flunkies and their team's comprehension into their own appearance.

and the counterfeit always gets caught the same way: the follow-up question. the partner replies asking about the second paragraph, and suddenly there's a four-person thread and a prep meeting to get the exec ready to discuss their own email. the thinking they skipped comes due with interest. and the team pays it. again.

the excuse just expired

the standard defense is some version of "it's technical, I don't have time to get it." and for decades that was at least plausible - understanding genuinely required background and hours nobody had.

that excuse died and most people haven't noticed. the same tool the exec could use to fake understanding will happily give them the real thing. same price, same minute. any technical issue, explained at any level, in their vocabulary, with infinite patience and zero judgment. the cost of comprehension has never been lower in human history. "I don't have time to get it" no longer describes a constraint. it describes a choice.

and people are making that choice constantly. a 2025 WalkMe survey found over half of gen z workers admit to faking understanding of AI in meetings, and more than half of c-suite respondents admit hiding how they actually use it. Pluralsight's survey of tech execs put the number who've pretended to know more about AI than they do at roughly nine in ten. vendor surveys, sure, vendor incentives, sure. but directionally they confirm what you can see from your own desk: the tool that could end the faking is mostly being used to do the faking better.

and to be precise here, because this is not an "AI bad" essay: your bullshit use of AI and your bullshit use of your technical team are the same behavior. identical in kind. both take someone else's cognition, strip the attribution, skip the comprehension, and pass the output off as your own. the technology is not the pathology. it's just the pathology's newest, cheapest delivery mechanism.

there's a slide making the rounds again, from an internal IBM training deck in 1979. all caps, no design, two lines: "A computer can never be held accountable" - therefore it must never make a management decision. 47 years old, author unknown, and it names exactly what the CCES™ is laundering: accountability. a signature is a claim of accountability, and accountability means being able to answer the follow-up question. sign what you never metabolized and the management decision got made by nobody... with a confident signature at the bottom.

the real enemy is unexamined defaults

ok now the concession, because the argument doesn't hold without it.

polish is not the crime. there are roles where presentation genuinely is the deliverable. a great account lead's product is trust, translation, knowing how something needs to land in a specific room. that's real work, and done with thought it's not veneer at all - it's craft. relaying your team's findings isn't a crime either, that's literally the job. "here's what my team found" is delegation. copy-paste-sign is counterfeiting.

the crime is presentation with nothing behind it, applied by rote, in every room, without ever once asking whether this is the moment to strip the veneer and be real. someone who can play the game and chooses when to play it is exercising judgment. someone who plays it identically everywhere is following rules nobody wrote, for reasons they can't state. Cal Newport calls the org version of this pseudo-productivity: "the use of visible activity as the primary means of approximating actual productive effort." the deck, the status meeting, the performative reply-all. visible activity, standing in for thought.

and that cuts in every direction, including at the meme this whole essay started from - the senior engineer showing up to the exec meeting in adam sandler attire. I love that meme. I've also had to get honest about it. the gym shorts aren't the absence of a costume. they're a more expensive one, and what they say is "I'm so valuable I'm exempt from your rules." it only works because the rules exist. economists call it countersignaling. it's still signaling.

and the exemption isn't for sale at any level of excellence. duke researchers found black women with natural hairstyles rated less professional and less likely to get interview recommendations than identical candidates with straightened hair. plenty of people wear the veneer as armor, not theater, because the penalty for dropping it lands on them and not on the hoodie guy. if your takeaway here is "everyone should dress down"... you've missed it. the wardrobe was never the point. the thinking was.

what it cost me

I should say what this position has cost, because an argument with no scars is just a victory lap.

my career path has a detour in it: an account role in an otherwise product/tech tree. I was decent at it - could read rooms, manage stakeholders, play the game when it needed playing. and it was quietly eating me. I wasn't building anything. I was a glorified pencil pusher whose entire job description was the presentation layer, and it hurt my soul in a way that's hard to overstate. so when the account director path opened up, I turned it down and left. because I needed to go build.

the exit taught the lesson twice. I took a role at Slack pitched as builder work - come make real solutions for onboarding clients. what it actually was: 4-6 week engagements with enterprise clients whose POVs were already locked and who had zero reason to trust a consultant they'd known for a week. and honestly? they'd earned that skepticism. they were used to consultants who absorb less than the full context and hand over polished recommendations anyway - precisely the grade of work AI now does for free. no trust on a minimum engagement, nothing real built without trust. six months and I had to get out.

the way back was starting a Product discipline from scratch at Team One, at a lower level than the title I'd walked away from. four years to climb back to director. but the role finally pulled the whole path into one job: the self-taught tech and design, the internal teams, the small agency, and yeah - the account years too, because stakeholder management is part of the craft when there's something behind it. all of it pointed at product work on platforms millions of people use every month.

I could tell you the detour was strategy. it's more honest to say I twice declined to invest in a currency I could see being debased, and the market took four years to agree with me. that's the price of not playing the game: you don't lose. you're right on a delay. and the delay is unpaid.

stop grading the wrapper

so what do you actually do with this on monday.

if you review work: stop grading the wrapper. before you react to how something reads, ask what the thought is and whether it survives being unwrapped. a document that can't answer a follow-up question isn't a document. it's a costume.

if you produce work: audit your rituals. can you state the reason each one exists? not the rule - the reason. keep the ones with reasons. the ones where the honest answer is "this is just what's done" are quietly converting your hours into someone else's theater.

and if you're the CCES™, or one of the quieter versions: the invitation is not "stop using AI." it's to notice you're holding the greatest comprehension subsidy ever created and spending it on disguise. point it inward. ask it to explain the thing until you actually understand the thing. then write the email yourself, in your own words, for your actual audience, and credit the team that did the work. that version of you is more valuable, not less. and the cost of becoming them just dropped to a rounding error.

one last thing about that IBM slide: nobody knows who wrote it. branch-office training deck IBM never archived. the photo surfaced from a retiree's papers in 2017, and the original was destroyed in a flood two years later. no author, no design, no institution behind it - and it outlived every polished deck of its era on one load-bearing thought. strip everything away and the thinking either holds or it doesn't.

sources, for anyone who wants receipts: the workslop study is BetterUp Labs + Stanford Social Media Lab in Harvard Business Review (sept 2025). AI text detection is Jakesch, Hancock & Naaman in PNAS (2023). Frankfurt is On Bullshit (2005), Graeber is Bullshit Jobs (2018), Orwell is "Politics and the English Language" (1946), Newport is Slow Productivity (2024). countersignaling is Feltovich, Harbaugh & To in the RAND Journal of Economics (2002). the hair bias study is Koval & Rosette in Social Psychological and Personality Science (2021). the faking-AI numbers are WalkMe/SAP via Fortune and Pluralsight's 2025 AI Skills Report (vendor surveys, cited as such). the IBM slide provenance is documented by Simon Willison (feb 2025).

(and yes - I see it. writing an essay about stripping away polish and then dressing it up as a slack thread is itself a costume. the difference is I can answer the follow-up question lol)