AI usage policy

Dynomight wrote a disclosure document about how they use AI. I already had one, but the race dynamics of blog social signalling suggest I might want to hoist that content from where it hid amongst the other technical details of this site to some specialized page that is All About AI. I… don’t find it particularly interesting to be an AI straight-edge. I use AI for all kinds of things, and am constantly auditioning new ways that it can be useful, as well as constantly testing whether it is actually useful. That is more-or-less literally my job.

I use AI to process the articles on this site in various ways.

1. Copy-editing. A lightweight language model reviews my sloppy, typo-riddled text, and removes the typos. It’s basically a mildly aggressive spellchecker. I manually accept or reject all such updates, albeit with varying levels of scrutiny.

2. Summarising. Those little “Wherein …” summaries at the top of the post are generated by a model coaxed into imitating Victorian novels.

3. Indexing. An embedding model digests each post and produces a vector representation that is stored in a vector database. This generates the “suspiciously similar” recommendations for posts and the 3d blog map on the home page.

4. Scoring. An LLM rates each post for quality (0–10), assessing how interesting, useful and/or coherent it is. These scores are not displayed directly to you, but influence the site’s front page ordering, since they feed into the “frecency” (freshness-weighted quality) score that I use for default page ordering.:

   \[ \text{frecency} = Q2^{-t/H} \] where Q is quality (0–1), t is age in days, and H is a half-life. This ensures the front page favours recent writing, but a really spicy good post can sneak ahead of newer, scruffy ones.

5. Literature reviews.* The main place I will get AI to draft actual text. I use AI to read and summarise papers outside my discipline, and to generate a first draft of a literature review. I label these with a 🤖🤖🤖 in the header to flag that they are AI-generated. This is especially useful when I’m trying to get up to speed on a new topic, and the existing work is 💩💩💩. Where I am especially keen to make sure no one mistakes it for my own writing, I additionally label it with a callout block.

   For example, the ageing literature review has substantial content generated by AI. This is an area where the default online advice is politicised and compromised by snake-oil merchants, so a careful AI-generated review is better than the default, which is whatever health influencers are getting kickbacks from this month.

   Lest I be misunderstood, the AI-generated content is not at all instant. I do not one-shot a literature review and then serve it up to you, for that would be a waste of time at the current state of the technology. An AI-assisted literature review takes me ~two days to assemble, and I spend substantial time fact-checking, revising and so on. Nonetheless, it is still not something I would call my own because the prose sucks and I have not internalised the area well enough to make really good calls about omissions and fuckups.

   An alternative would be to spend a month on each lit review I create to scope out a new research area. That would be higher quality, but it would also restrict the number of areas I could explore and probably be too hi-fi a process for the quick and dirty search I am performing. I could alternatively not share such lit reviews, but since they are still somewhat effortful, expensive, useful artefacts, this seems like a waste.

6. Automation of tedious RA tasks: I use AI to automate various maintenance tasks on this site, including stuff like renaming and re-organising blog posts, dividing them into sub-pages, and searching the internet for missing references.

7. Self tutoring: When I want to learn a new topic area and it is important that I understand it well, I use AI to do socratic tutoring to rapidly get up to speed on new concepts, and to generate exercises to test my understanding. You probably observe this only indirectly, by the reduction in the number of stupid things I say.

8. Feedback: I regularly use AI to provide feedback on my arguments, mathematics, and code. Once again, you only see this indirectly.

9. All Kinds of Zany Experiments. Code generation, exploratory theorem proving, general-purpose assistants, and agent harness design, style transfer, sovereign compute are all experiments in using AI to do things that I do not have enough hours in my day to do without augmentation. Some of those end up touching my blog text in incidental ways. I trust there will be more of these and better ones as time goes on.

My rule-of-thumb is that if AI might help me provide more value to the world than would the alternative, it’s worth trying it out. You are welcome to disagree with me about what “value” means and where the line is. You are also welcome to point out any mistakes I make in my use of AI, and to suggest better ways of using it.

Oh, I publish the code for my more general experiments online.

You might enjoy

• danmackinlay/Filibotster: live detection of AI-generated speeches
• danmackinlay/stylebot: prose style transfer for Dans
• danmackinlay/mathx: miniature math proof assistant
• danmackinlay/pdf-ingest-skill: Agent skill: route PDFs to the right local converter (markitdown / Docling / marker / MinerU) with token- and memory-budget discipline
• danmackinlay/gemini\_deep\_research\_client: : Get the google gemini deep research pay-as-you-go rather than by unaffordable subscription, and preen the results to ha ve better provenance tracking and iteration.
• I’m kinda hopeful that I will find some good angles in Civic tech also, but this is WIP

I think it might be worthwhile mentioning where I don’t use AI, and why.

1. I aim not to use AI deceptively, without labelling the content as AI-assisted.
2. I aim not to use it in lieu of ever exercising those of my own skills that I need to keep sharp.
3. I try to avoid it becoming an addiction.