max / makenotwork

8.0 KB · 137 lines History Blame Raw

1	# Generative AI Policy
2
3	How Makenot.work defines generative AI, why it matters, and how it affects what you see and sell on the platform.
4
5	---
6
7	## Why This Policy Exists
8
9	Most generative AI tools are built on training sets containing copyrighted work used without permission or payment. Artists, musicians, writers, and developers whose work was scraped to build these models were never asked and never compensated.
10
11	We are not against new tools. We are against tools built on mass uncompensated use of other people's work. If that changes, this policy will evolve. Until then, the default is skepticism.
12
13	---
14
15	## What We Mean by "Generative AI"
16
17	On this platform, generative AI refers to any tool whose existence depends on training sets that are either:
18
19	1. Not publicly disclosed, or
20	2. Known to contain copyrighted material used without compensation to the rights holders
21
22	This includes derivative products: any application, plugin, or service that relies on the API or model weights of a covered tool. If the underlying model is covered, the wrapper is covered.
23
24	### What this covers
25
26	- Large language models with undisclosed or disputed training data (GPT, Claude, Llama, Mistral, etc.)
27	- Image generators trained on scraped artwork (Midjourney, DALL-E, Stable Diffusion, etc.)
28	- Music generators trained on copyrighted recordings (Suno, Udio, etc.)
29	- Code generators trained on public repositories without clear contributor consent (GitHub Copilot, etc.)
30	- Any application, plugin, or service built on top of these models via their APIs
31
32	### What this does not cover
33
34	- Tools trained entirely on the creator's own work
35	- Tools trained on verifiably public domain material
36	- Tools with fully disclosed training sets composed of properly licensed, compensated content
37	- Traditional automation and digital tools: EQ, compression, batch processing, non-generative spell check, programmatic effects, scripting, macros
38
39	The distinction is not about what the tool does. It is about how the tool was built.
40
41	---
42
43	## The Three Tiers
44
45	Every item published on Makenot.work must declare one of three tiers at publish time. There is no unlabeled option.
46
47	### Handmade
48
49	No generative AI tools were used at any stage, from brainstorming to final output. Traditional digital tools (DAWs, image editors, IDEs, effects processors) are fine.
50
51	### Assisted
52
53	This product was created by a human with generative AI tools as part of the process. Requires a disclosure statement: what tools were used, for what purpose, and at what stage. Visible to fans on the item page before purchase.
54
55	Examples of good disclosure:
56	- "I used GPT-4 to help outline the chapter structure. All writing is mine."
57	- "Background textures were generated with Stable Diffusion and painted over. Character art and environments are hand-drawn."
58	- "Code completion (Copilot) was used during development. Game design, art, and music are original."
59
60	The disclosure must accurately represent the scope of AI use. If the product contradicts the disclosure, that is fraud under our [Acceptable Use Policy](../legal/acceptable-use.md).
61
62	### Generated
63
64	This product was primarily generated by AI tools. The human contribution is direction, curation, or assembly rather than direct authorship of the creative output.
65
66	---
67
68	## How Fans Use This
69
70	Fans can filter what they see on the platform:
71
72	- Handmade only: only items with no generative AI involvement
73	- Human-led: Handmade and Assisted items (excludes Generated)
74	- Everything: no filter
75
76	Filters apply to the Discover page, search results, and tag browsing. The tier is also displayed on every item and project page.
77
78	> [!UI] ai-tier-badges
79
80	---
81
82	## Enforcement
83
84	The tier system is self-reported by creators. We enforce it through:
85
86	1. The disclosure statement. Assisted-tier creators put their AI use in writing. If the product contradicts the disclosure, that is documented misrepresentation.
87	2. Community reports. Fans and fellow creators can flag an item they believe is misclassified with the Report link on its page (choose "Mislabeled"). You can also email [reports@makenot.work](mailto:reports@makenot.work) with the item URL and the tier you think it should be.
88	3. Moderation review. Misrepresenting your tier is treated as fraud. Consequences follow the [Acceptable Use Policy](../legal/acceptable-use.md), up to and including account termination.
89
90	If you used AI, say so, explain how, and let your audience decide. That is the point of the Assisted tier.
91
92	---
93
94	## AI in the Platform Itself
95
96	We will never build generative AI tools or features into the product. No AI writing assistants, no AI-generated thumbnails, no AI recommendations.
97
98	Discovery follows [Discovery Through Exploration](../guide/discovery.md#discovery-through-exploration): everything you see traces back to a choice you made. No behavioral tracking, no machine learning. The discovery and feed logic is public and readable:
99
100	- [discover.rs](https://makenot.work/git/max/makenotwork/tree/main/server/src/db/discover.rs): Search, browse, and all filtering/sorting logic for the Discover page. Uses PostgreSQL trigram matching for fuzzy text search. Sorting is newest, popular (by sales count), or price; no behavioral signals.
101	- [follows.rs](https://makenot.work/git/max/makenotwork/tree/main/server/src/db/follows.rs): Feed generation. Your feed shows items from users, projects, and tags you follow, ordered newest first. No algorithmic reordering.
102	- [tags.rs](https://makenot.work/git/max/makenotwork/tree/main/server/src/db/tags.rs): Tag facets and hierarchical browsing. Tag counts reflect actual item counts, not weighted relevance.
103
104	Security and spam filtering are the exception. We reserve the right to adopt whatever keeps the platform safe, though at the time of writing our security toolchain does not use LLMs or generative AI. We will never use these tools to influence what you see or recommend content.
105
106	Platform development uses LLM-assisted tooling openly. LLMs reduce costs and let a solo developer ship features at a pace that would otherwise require a team. This is transparent in our commit history.
107
108	The quality bar is non-negotiable: a comprehensive automated test suite, property-based testing, mutation testing, and regular code audits. LLM-generated code is held to the same standard as hand-written code. If it doesn't meet that bar, it gets rewritten until it does.
109
110	There is a meaningful difference between using these tools to build infrastructure and selling their output as creative work. Our tier system reflects that distinction.
111
112	A note from the founder:
113
114	> The honest truth is that I would prefer not to use these tools. In my non-legal opinion, they are unethically built. They are also, frankly, mind-numbingly boring. At the same time, they allow a one-person team to build and maintain five products, a substantial server codebase, and comprehensive documentation, all without outside funding. As the codebase grows, their usefulness diminishes in favor of a more hands-on approach as the focus moves from scaffolding to testing and refinement. We will always be honest about what we use on our end. I will always try to be honest about how I feel about those tools. And as always, we would love to someday replace them with tools that we make or that our community makes: tools that are well-built, ethical, and that I can feel good about using.
115	>
116	> \- Max
117
118	---
119
120	## This Definition Will Change
121
122	This field moves fast. This policy is a living document. When we update it, we will:
123
124	- Publish the change with a version date
125	- Explain what changed and why
126	- Give creators at least 30 days to update their tier classifications if the definitions shift
127
128	If a currently-covered model demonstrates verifiable, fully compensated training data, it may be reclassified. The principle stays the same: we follow the ethics of the training data, not the capabilities of the tool.
129
130	---
131
132	## See Also
133
134	- [Acceptable Use Policy](../legal/acceptable-use.md): What happens when rules are broken
135	- [Our Story](./story.md): Why this platform exists
136	- [What We Guarantee](./guarantees.md): Binding commitments, in writing
137