Outsourcing creativity

Is both uncreative and unoriginal

29 July 2026 misc

If you check out the post frequency on the website, you’ll notice around 18 months back, I started clearing a decades-long backlog and posting a lot.

This backlog accumulated, in part, because I have a demanding day job that likes to keep me working well into the evenings.1 I’ve been working that job for 15 years or so. During that time, I’ve had plenty of ideas and side-projects, and an ever-growing drafts folder.

Then AI happened

Holy smokes. For someone with a lifelong obsession with AI2, the last few years have been interesting to say the least.

So it’s uncanny, really, to see how things have unfolded. I remember the first time I read Attention Is All You Need.3 I remember the GPT-2 moment.4 I remember being advised against RL fine-tuning in my Stanford course because “it’s tricky, nobody has really had success with it yet”. Then RLHF landed,5 OpenAI stopped being open, the world started being covered in data centres.

Now, I can set a model running overnight to churn through tasks on a research project, and I can’t listen to ABC Radio National without encountering someone talking about model collapse or relative data scarcity.6

So I use it

AI is my day job.7 I make AI tools for people to use. I try to do it safely. One evening, as I was wrapping up after a late night working, I remembered this dusty old website. I figured, hey why not do a revamp with these new tools, so I did.

This revamp started slowly, but it pretty quickly escalated. I had decades of half-started ideas that I soon realised could also use the attention. Research ideas, writing ideas, things to test out. Less than a tenth of what I’m currently working on makes its way to the site, and it’s all going very quickly.

But you have to use it right

Intentionality matters more than ever. The new AI tooling amplifies your intention, but it deliberately has very little of its own. It lacks the circuitry for it, and there are good safety reasons for it being that way.8

It gives everyone a new level of power, but if you don’t have at least some experience in the domain, then what use is that power?

I have decades of experience writing code, writing letters, doing machine learning, writing the occasional paper for a client that I’m not allowed to publish, so suddenly having an army of capable synthetic research assistants at my disposal became something I could use. I started clearing the backlog.

I write because it makes me think. If I outsourced that part, I’d be losing the fun of it. So I outsource the parts I know how to do but can do quickly by delegating. Delegation without supervision is (TODO: what’s the rest of the quote?), and that’s something I’m seeing everywhere.9

And I’m seeing it everywhere

GitHub used to be full of collaborators, now it’s full of bots.10 Recently, while catching up on some inference backlog work, I posted a couple of issues on GitHub and had conversations with maintainers. Within minutes I got responses from bots pretending to be people. That’s how I learned about “contribution farming”: in one 2024 scheme, a crypto project offered rewards for open-source contributions, and maintainers of unrelated projects woke to floods of drive-by pull requests from strangers whose YAML files declared them “code owners”.11 The security desks have it worse: curl shut down its long-running bug bounty this January, after AI-generated vulnerability reports, some with fabricated debugger sessions citing functions that don’t exist in the codebase, crowded out the real ones.12 Now maintainers of projects like vLLM are struggling with 4k LOC drive-by PRs containing fully-fused CUDA kernels that ignore all project conventions, sent by hyper-vigilant bots with one job: get someone on the contributors list.

The incentive model for the commons, and the social contract there, is breaking.

Human interaction is already becoming rare as people are isolated, but what were social bubbles are now becoming AI bubbles.13 The slop is gumming up our news feeds and our collaborative spaces.

Used correctly, it doesn’t need to do this. It can be a “thanks Jeff, your bots will talk to my bots and we’ll arrange something”. Cue human handshake. Done incorrectly it’s a fake dating profile leading you on through the night to gain your trust as part of a complicated scam.14

Where I’m currently at

Right now, my use of AI is quite liberal. I use it for all sorts of things. I’ve built quite a setup around it really, and I use nearly 100% of my VC-funded tokens every month. This includes things like:

  • Doing the grunt-work of setting up containers for experiment runs
  • Doing sysadmin, database admin, and cloud admin work
  • Writing software to a spec, then reviewing compliance with that spec
  • Formatting and doing copy edit on academic papers
  • Doing citation checks and prior art scans for blog posts
  • Handling life admin, which can get very complex at times
  • Creating recipes to cook every night15

I’m avoiding, as much as possible:

  • Letting it wear my identity like a trenchcoat
  • Letting it steer the pen with my writing
  • Letting it dictate curiosity and research process

Those last ones belong to me, and I’d care if I lost them. I’m going to hold to that balance for the minute because it seems to be the most self-preserving.

The golden rule

Communication takes investment from both sides.16 You’re short-changing another human on that social contract if you send a bot to do interactions on your behalf in a forum whose norms expect the presence of another human.17 Sending a bot that pretends to be a human is cheating twice over.

You can use AI, but if you misrepresent then you’re violating a social contract, and contract violators risk reprisal. Humans have cheater detection circuits that are as old as tribes, which are now beginning to notice this kind of underhandedness.18

Money bends every commons it touches, and the Internet’s slide from weird personal space into a lame mall is the reference case; Doctorow’s word for the process is enshittification.19 AI is now departing from its promise along the same slide, turning every social forum into LinkedIn.

I hope we establish the social norms and enforcement strategies to recover from this, or the future we’re creating will not be particularly enjoyable, and will be missing the one thing that makes the human creative process so effective: individuals contributing original thoughts and sharing their unique experience.

So be self-preserving, have a unique experience, contribute your original perspective, and don’t donate your identity to body-snatching AI.

Footnotes

  1. The contract says “from time to time”, but with some clients that ends up being “all the time, plus you’re taking 10-hour flights on weekends”. ↩

  2. My uncle showed me a demo of a neural-network agent simulation he’d built when I was about 12, back in the 90s, and I’ve been hooked since then. For the last 30 years or so I’ve researched and ingested everything I could on the topic, from maths to developmental psychology to philosophy. ↩

  3. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention Is All You Need,” in Advances in Neural Information Processing Systems 30, 2017. The architecture in question now runs most of the industry, including the tools this post is about. ↩

  4. A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever, “Language Models are Unsupervised Multitask Learners,” OpenAI, 2019. The moment was as much the theatre as the model: OpenAI first declared the full weights too dangerous to release (“Better Language Models and Their Implications,” OpenAI, Feb. 2019), then let them out in stages once the sky stayed up. ↩

  5. The advice was defensible at the time. The mechanism dates to P. F. Christiano, J. Leike, T. B. Brown, M. Martic, S. Legg, and D. Amodei, “Deep Reinforcement Learning from Human Preferences,” in Advances in Neural Information Processing Systems 30, 2017; the landing was L. Ouyang et al., “Training Language Models to Follow Instructions with Human Feedback,” in Advances in Neural Information Processing Systems 35, 2022, the InstructGPT work that made chat models worth talking to. ↩

  6. The segment’s likely sources: I. Shumailov, Z. Shumaylov, Y. Zhao, N. Papernot, R. Anderson, and Y. Gal, “AI models collapse when trained on recursively generated data,” Nature, vol. 631, pp. 755-759, 2024; and P. Villalobos, A. Ho, J. Sevilla, T. Besiroglu, L. Heim, and M. Hobbhahn, “Will We Run Out of Data? Limits of LLM Scaling Based on Human-Generated Data,” 2024, arXiv:2211.04325, on the scarcity. ↩

  7. I was well-placed for the revolution, aside from living on the wrong side of the globe. Still, I’m on a good wicket, even if I’d be earning more doing nothing had I purchased real estate 10 years ago or invested in mining when I first read the Bitcoin paper the day after it landed on Slashdot. I have a reasonable amount of regret about that one, having ridden the e-gold wave in the decade prior. The paper, for the record: S. Nakamoto, “Bitcoin: A Peer-to-Peer Electronic Cash System,” 2008. ↩

  8. The low-intention design is deliberate at both ends of the pipeline. Training against explicit written principles rather than emergent goals: Y. Bai et al., “Constitutional AI: Harmlessness from AI Feedback,” 2022, arXiv:2212.08073. Why you want an agent that accepts correction rather than one with plans of its own: N. Soares, B. Fallenstein, E. Yudkowsky, and S. Armstrong, “Corrigibility,” AAAI Workshop on AI and Ethics, 2015. ↩

  9. In a double-irony, the machine copy-editor did flag this incomplete aphorism: ”… is abdication”. I had to point out this was done on purpose. Humour survives as a bastion of human intelligence, even if my jokes are getting a bit dad-like. ↩

  10. I haven’t been on Twitter for years but I hear that was one of the first empires to fall. LinkedIn has always been that way, but the tone of that platform makes you wonder if they’re actually humans wearing a forced professional facade. ↩

  11. The reference case is the tea.xyz wave of March 2024: a crypto rewards scheme had contributors “verify ownership” by merging YAML files carrying their wallet addresses, and the long tail ran to hundreds of thousands of spam packages across registries. “New Crypto Spam Targets Open Source Projects on GitHub,” Socket, Mar. 2024; “Test launch of TEA open source reward project clouded by repository spam attack,” DevClass, Mar. 2024. ↩

  12. curl ran its bug bounty from 2019 until January 2026, when the signal drowned: the confirmed-vulnerability rate fell from above 15% to below 5%, with roughly one in five submissions in the final year being outright fabrication. D. Stenberg, “The I in LLM stands for intelligence,” daniel.haxx.se, Jan. 2024, is the early warning; “Curl ending bug bounty program after flood of AI slop reports,” BleepingComputer, Jan. 2026, is the obituary. ↩

  13. The social version was named by E. Pariser, The Filter Bubble: What the Internet Is Hiding from You, Penguin, 2011: personalisation quietly narrowing each person’s informational world. The sequel is grimmer, because the bubble’s walls now generate their own content. ↩

  14. No, thankfully I’ve been out of that market for some time, but holy hell the stories you hear. ↩

  15. I’d fully outsource the whole meal-making process if I could, but it has been enjoyable. A few Scanpan pans, some good knives, fresh ingredients, and you’re cooking up traditional Ragu. AIs use nearly everything on the Internet for training data, and the Internet is full of code and recipes for food. For that reason, the current crop of LLMs are very good at both. ↩

  16. The old name for this is the cooperative principle: conversation works because each party assumes the other is genuinely contributing toward the exchange’s purpose. H. P. Grice, “Logic and Conversation,” in Syntax and Semantics 3: Speech Acts, P. Cole and J. L. Morgan, Eds., Academic Press, 1975, pp. 41-58. A bot wearing a human face in a human forum free-rides on exactly that assumption. ↩

  17. Everyone has a colleague who needs to read stopsloppypasta.ai. ↩

  18. The circuits are documented rather than rhetorical. People are mediocre at spotting violations of an abstract if-then rule and sharply better when the same rule is dressed as a social contract with a cheater to catch: L. Cosmides, “The Logic of Social Exchange: Has Natural Selection Shaped How Humans Reason? Studies with the Wason Selection Task,” Cognition, vol. 31, no. 3, pp. 187-276, 1989; L. Cosmides and J. Tooby, “Cognitive Adaptations for Social Exchange,” in The Adapted Mind, Oxford Univ. Press, 1992. ↩

  19. C. Doctorow coined enshittification for the platform decay cycle in 2022, and by 2023 it was the American Dialect Society’s Word of the Year, which says something about how widely the diagnosis resonated. The lame mall is my regional variant. ↩

@misc{hollows2026outsourc,
  author = {Hollows, Peter},
  title  = {{Outsourcing creativity}},
  year   = {2026},
  month  = jul,
  url    = {https://dojo7.com/2026/07/29/outsourcing-creativity/}
}