My Mid-2026 AI Policy

This is not actually a policy, sorry, more of a once-every-six-months “How am I using/feeling about AI”-style post, but everyone is out here posting “policies” now so I thought it’d be fun. In fact, you can tell it’s not a policy because almost all of these are “I do” or “I have”, not “I will” or “I won’t”! As always, it’s a free substack!

• I just aesthetically do not like most AI products, or AI features incorporated into products, especially when branded as such. This has gotten me to generally shift to more AI-free options for many common tools, such as changing my default search engines and browsers to ones that offer either explicit opt-in or explicit opt-out of AI summaries or AI tools. If there is an option to disable an AI feature or avoid an upgrade that adds it, I will generally take that option.

• That said, elephant in the room, I still use a bunch of Google products (Youtube, Search, Sheets, and Photos) so even if I’m not interacting with AI tools, I’m still exposed to them, and certainly end up at least being distracted by them.

  • Specifically, despite changing all of my default search engines, I probably fall back to Google somewhere between 1/3 and 1/2 of the time, as the actual search results, as in the links, are often better, even if they’re buried under AI summaries. I try strenuously to avoid the summaries.

  • The exception to the exception is Google Photos: I have no problems using AI search (presumably powered by AI image tagging) in Google Photos, and do!

• I use AI translation, especially when traveling1. In theory I’m interested in always-on translation via like AirPods etc, but haven’t actually done that.

• If I ever do want to affirmatively ask an AI question, I use Claude via the web app.

  • AI as a chatbot tool is most useful to me in cases where the input is complicated, followed by cases where the output is complicated. So I primarily use Claude for cases where there is no feasible way to get my question down to less than a sentence or some keywords (including if the input is a photo), or where the output involves collation between a lot of disparate sources (and I don’t care to do that collation myself.)

  • I explicitly don’t use any desktop or mobile apps because I like the additional friction of forcing it to be a couple clicks and an occasional re-login.

  • I use the free plan and always leave it on the “Extra effort” setting, both because: I do want the best answer; I like the idea that I’m costing Anthropic a bunch of money for my bullshit; and I like that it takes >1 minute to answer and will presumably hit a daily usage limit at some point, which again helps keep it intentional rather than habitual.

  • I probably do this ~0-5x a week, averaging maybe 2.5x/week? By way of data, I asked 8 questions in June, 12 questions in July.

  • If I do lookup something in Claude, I always admit to that as the source, and try to say it sheepishly or with a some level of eye-rolling. If you say “Claude told me” earnestly and with no deprecation, then know I will gladly fill in the gap and will (internally) be doing the eye-rolling on your behalf.

• I generally don’t use AI photo editing or photo generation, but I’m not opposed to it in principle; it’s more downstream of just not really needing it. That said, I would never privately or publicly share an AI-generated or significantly AI-edited photo without noting it as such, and would probably not share it at all unless it was on either pole of 1) truly indistinguishable as AI, or 2) weird and inhuman in an aesthetically or philosophically interesting way.

• I am sure there are lots of things I use in my life that use predictive algorithms etc.

• I did use AI (ChatGPT) to generate the first iterations of the icon for my dumb little cocktail app (though that actual icon was a vector image I made in Pixelmator effectively as a modified trace of the AI icon.) I think I made all the alternate icons by hand, but I don’t remember?

  • Regardless, I have no qualms with doing this again in the future, either the literal same “AI for the brainstorm, make the real icon by hand” or just using the AI icon itself if it was good enough.

• Obviously, the app itself uses Apple Intelligence for its image and text parsing.

• I use AI coding tools; I’ll talk more about that below, and have talked about it in the context of this app previously.

• I do not use AI for any of the writing or editing of this substack2. I don’t have a stance against using it for proofreading, but this is a free substack so I just don’t care to!

• Part of this is philosophical, where the entire point of this is to do the act of writing, so why would I automate that? Insofar as this substack has a purpose it’s because 1) I find it fun and 2) I want to build the muscle of actually writing, so AI for any sort of drafting or revision would be fully self-defeating.

• But also: any time I have experimented with using any tool for any edits or revisions beyond proofreading the results have been abysmal.

• I on occasion use Claude for researching, as in finding sources to verify externally. I have used it (or ChatGPT) I think once for brainstorming or trying to help figure out what was going on, specifically in the Trump Worldtimer post. I have once or twice used it as, I guess pseudo “fact-checking” for a claim I was making that I was unsure was true.

• The real answer is: I don’t do much of this, so unsure to what degree I would/will use it.

• Like with Dottxt, I think I’m generally fine with it as a research, clarification, and gut-checking tool, and less interested in it the more it touches the actual writing. In particular, like with Dottxt, any fiction writing would be for fun and for ego, so why would I want to automate or outsource a thing I’m doing for fun??

• In the off-chance that I ever were to write a thing for someone else or pitch a thing or whatnot, I would gladly follow whatever the AI policies are of the relevant publisher! This includes “No AI whatsoever” if that’s the vibe! Not fiction writing, but that is what I did for eg: my final essay in my Shakespeare class last year!

• For playwriting specifically, I am potentially interested in AI text-to-speech for the purposes of virtual “table reads”, in the absence of being plugged into an acting or playwriting scene in my city 😢

• I have for almost a decade been using AI (aka “machine learning”) tools for denoising and audio repair, specifically iZotope.

• I use Descript for part of my editing workflow, which uses AI transcription and allows you to edit audio (in part) by editing the transcript

• I also use some of Descript’s additional AI tools like highlighting or attempting auto-edits of filler words or retakes. I do not use any of it for larger content or structure edits.

• And have I think three times used Descript’s text-to-speech to generate myself saying about half a sentence that would otherwise be a retake.

  • I have no problem using this at that scale and for myself, and would have no problem using it at that scale for my co-host if they were okay with it.

  • Because The A-Dub Club is low-listenership entertainment I would be comfortable doing larger retakes or edits or new additions with it.

  • I would not consider this acceptable for source audio for journalism or narrative nonfiction. I think I would be fine with it for narration.

  • I would consider it acceptable, with caveats, for audio fiction.

  • Maybe the actual policy version is: for source audio in journalism/narrative nonfiction, always unacceptable. For entertainment and fiction, acceptable provided 1) there’s consent of the person being audio cloned, and 2) its acceptability is in inverse proportion to how much you’re “replacing” someone who would otherwise be paid.

  • If, in a fiction context, it was a fully AI-generated performance I think I would find that acceptable in direct proportion to how much are you using that to create something that wouldn’t have been possible before (either because you’re using some element of the inhumanity in interesting ways, or the content is sufficiently non-commercial such that the choice is “use the robots or don’t make it at all.”)

• I don’t think I have, but I would have no problem using Claude for focused research or fact checking (again, to identify sources or do replicable calculations, rather than as a source in and of itself.)

• Like with any writing, in a hypothetical where I was paid for audio work, I would gladly follow whatever the AI policies are of the relevant commissioner, including “don’t use it at all”!

• I use, and am strongly encouraged to use3, AI coding tools. Specifically I’ve settled on:

  • Copilot for AI autocomplete in VS Code4

  • Claude Code as both a chatbot for questions/Google alternative, and for agentic feature development and bug fixes. I run it in a sidebar in VS Code and almost exclusively in manual mode with whatever is the most powerful model available to me at the time.

  • I take advantage of Claude.mds and Skills and whatnot that other team members have built. I have only contributed to those in a limited capacity.

  • I am experimenting with using local coding models as replacements for Claude Code for personal coding, because I will never ever ever in a million years personally pay money for an AI tool. I will sooner never touch a coding side project again than pay for AI in a personal capacity.

  • I reserve the right to maybe change the above stance if I were to become a freelance or solo dev (but I suppose that would still not be paying in a “personal” capacity if it was a business expense.)

• For coding agents in particular, I have a somewhat bimodal usage pattern:

  • If I don’t care about what I’m doing and there’s nothing interesting about it, I start by giving it to the agents and make (usually significant) tweaks and fixes to their output as needed.

  • If I do care or find it interesting, I start manually and escape hatch with an agent for either uninteresting parts or places where I’m stuck (especially if googling or searching for other examples internally has already failed me.)

  • The one exception is if I’m doing front-end work for our internal tools platform, I will delegate most of that to AI. Not that it’s not interesting or I don’t care, but because I am not skilled in this domain. This is because there is a single universal truth of AI tools, which is that they are always best at the thing that you are least familiar with. Every back-end engineer thinks AI is better at front-end than back-end, and every front-end engineer thinks the opposite. I have gotten into this literal argument several times before5.

  • The correct way to use agents is to write tests by hand and have agents do the implementation. I almost always do the opposite, because this is a deficiency I have in my craft.

  • I never run multiple agents or work on more than one coding problem at the same time.

  • I may multithread by having an agent work on a coding problem in the background while I work on non-code stuff (research, architecture, PR reviews, meetings, other problem solving, administrivia, etc.) Though I still try to avoid this.

  • This overall pattern is inspired both by my continued subjective experience of the relative quality of the state of the art tools (albeit using the state of the art models with a minimum of additional scaffolding); my continued belief that multitasking is never worth it; my desire to maintain skills; and by the fact that there is almost nothing more boring in this world for me than asking a robot to do something for you. Once the robot is involved I emotionally check out and it takes considered effort to get myself back into caring.

  • Speculating and editorializing, I imagine I could be more gung-ho on AI tools if they came with an explicit or an implicit trade of “any time you save with them you get to spend on whatever you want,” rather than the implicit trade of “you get to spend either more time prompting more AI, or more time fixing the results of having prompted AI.” If this was my mechanism for the famed Google 20% time, that’s more interesting.

  • As it stands now, any time savings I do get I try to reorient towards thinking, researching, or being a better teammate (answering questions, helping teammates, writing docs, doing PR reviews, etc.)

  • Also, I continue to be cynical about the fact that I do see some net productivity improvements to using these tools in both myself and colleagues6….. but it still pales in comparison to the productivity improvements I have seen by following known best practices like having a well-designed org, having a good product manager, having a clear corporate vision and clear ownership of how that translates into strategy and tactics, etc etc. Most places do not follow known best practices!

  • By way of example, I guess I would rank AI coding tools as about the same productivity boost as adding Scrum to a low-performing team, and an order of magnitude less of a boost than removing Scrum from a high-performing team7.

• I use, in the sense that I am subjected to, AI code review at various points from both Github Copilot and Claude. This is not by my choice and would not be my choice, with the one exception I don’t mind it for external contributions into repos I’m less familiar with. That can be useful for catching obvious misses.

• I occasionally use either Grafana’s built in AI analyzer or Claude wired up to a Grafana MCP for analysis of logs and observability for things that that are either particularly urgent and tricky, or particularly boring.

• I never use AI for any professional writing or presentations.

• I very rarely use AI summarization, usually only Slack and generally only when being tagged into massively long and convoluted threads.

• I am subjected to AI search on internal platforms. I try to find my ways around it.

• I don’t generally use AI for any sort of business intelligence or data purposes directly, but I have used Claude to help me craft SQL queries that I can run manually.

• I am sure there are many tools I use that visibly or invisibly use recommendation algorithms and other forms of machine learning and AI.

• Like with everything in the artistic/hobby section, I would gladly follow an AI policy that limited AI usage! I suppose though the difference is at this moment in the tech industry, I feel I’m more likely to find policies that mandate it vs policies that forbid it 🙁

So that’s AI! I use it, and as evidenced by the fact I use it I see at least perceived benefit in some circumstances8. I am still in a zone where if there was a Butlerian Jihad tomorrow I would be able to go back to being normal without much incident (provided Google went back to being good as well.) I would like to stay in that zone as long as possible.

Compared to the last check-in, I definitely use it more in a professional capacity, but that’s balanced by using it significantly less in a personal capacity. And I have politically hardened against the AI industry to a frankly shocking degree! I’m not yet a single-issue voter on data center moratoriums…… but it’s certainly a tie-breaker! I would not have guessed this a year ago!