This page describes Draftlab’s on GenAI, the ethical constraints that guide us, and the practical decisions around AI use we commit to in everyday work.
- Revision: October 2024.
If I didn’t care to write it, you shouldn’t bother reading it
We take pride in our ability to express complex information to other humans. We believe that it must be harder for us to create the content, than for you to consume it. Generative AI over-explains; its essays lack complex structures and lose themselves in circular logic; it says little using too many words. We expect the opposite of ourselves.
Human first, human last
Generative Artificial intelligence is an assistive tool in service of human creativity. Humans define projects, goals, strategies and implementation pathways. GenAI can assist in tasks that it is particularly well suited for, such as parsing large quantities of information or executing rote tasks.
Any content that our organisation produces will always have a human responsible for it, verified, edited, and re-written with humans in mind.
When it comes to computer code, we only publish and use code that we can comfortably and confidently explain, line by line, to a peer.
Generative AI always hallucinates and can never be trusted
GenAI has no understanding of things, actions, intentions, shapes or relations. It can generate content that can be plausible, correct, or insightful to humans. This is due to the ability that humans have to understand things; differentiate between truth and hallucination: genAI cannot tell the difference between the two.
For this reason we will never trust GenAI output and will always verify and review any output it creates.
Aggressive de-biasing of model output
Generative AI is built on top of massive corpora of information that statistically over-index the global minority, perspectives of people in power, and historical inequalities in access to content creation.
Whenever we engage with Generative AI, we question and de-bias its outputs from an inclusive, rights-based perspective.
Only enterprise models or local private LLMs
We do not let AI train their models with our input. To achieve that, we either use enterprise versions of common AI tools (if legal protection is deemed enough for the threat model of our partners and beneficiaries), or private, locally hosted LLMs.
Our enterprise AI of choice is Claude. We have enterprise access to Gemini and use it sporadically, when relevant. We explicitly do not use ChatGPT or Grok.
No PII except in local, private LLMs
We do not share Personally Identifiable Information (PII) with any GenAI except locally hosted private LLMs that have no access to the internet.
No generated multimedia content
We believe GenAI multimedia outputs (sound, images, animations) hold negative value and do not use them. If we can hire a human artist or content creator, we will.
We are not opposed to supporting basic scaffolding of designs, diagramming and information architectures with AI: we draw the line at creative expression. We appreciate that it is a blurry line.