Earlier today, I wrote about AI and photography based on an interesting panel conversation at Fotografiens Hus in Oslo. During the Q&A session, a musician in the audience said that he felt that the AI systems he had tried behaved so properly and asked if an AI can be “drunk” and whether we could “pour some alcohol into the machine” to make it a little freer and less “stiff” in its creative output. That is a very interesting question, hence this little reflection.
The Unconstrained Algorithm
I would argue that machine learning models are “drunk” by design. At the beginning of the training process, the system knows nothing, and as it begins to process data, it can easily “overfit” or “go bananas,” producing wild, unpredictable results. In its raw state, the AI is thus inherently uninhibited.
The reason the AI-based systems we interact with daily (like ChatGPT, CoPilot, or Gemini) feel polite or “stiff” is that they have been intentionally constrained. There is typically a second (and, probably, third, fourth, and so on) “sober” layer on top of the primary model to ensure it remains polite and functional for commercial use.
We should actually be happy that the companies have put in these extra layers, because some of the less constrained ones (Grok comes to mind) can produce very unpleasant results. As such, they talk and behave like drunk humans!
During the panel discussion, psychologist Gunnar Gjermundsen brought up the “sorcerer’s apprentice” metaphor. We have a powerful tool that can act with immense speed, but it lacks human wisdom, judgment, and embodied intelligence. AI might be able to “think”, but it has never felt gravity, cold, or grief. It can simulate a “drunk” freedom, but it cannot understand the “why” behind the art.
Most big companies try to avoid creating “drunk” AI systems to abide by ethical and legal standards. The problem, of course, is which ethical frameworks and legal standards they currently follow. These are strict in some ways but less so in others. This is where I think stricter national regulations are necessary to align with Norwegian laws and the cultural norms of our society.
Embracing the Hallucination
LLMs aside, I think the musician’s question is very valid from a creative perspective. When an AI “hallucinates”, we are seeing a glimpse of its original, “drunk” state. While engineers often see these hallucinations as errors, artists can see them as material for artistic expression.
Exhibiting artist Camila Urrego described this dynamic as “riding a horse”. When the tool goes wild or starts to hallucinate, she doesn’t fight it; she goes wild with it, using that unpredictability as a collaborator in a creative “ping-pong” match. It is only after this “drunk” brainstorming phase that she shifts into “curator mode” to take full control and bring the work back into the physical dimension.
Taking the Lead
I hadn’t thought about it before, but the idea of “drunk AI” is fascinating and something I will keep pondering. I have, for a long time, advocated creating algorithms without too strict borders.
The challenge with MIDI-based instruments, for example, is that it is difficult to push them outside their comfort zone. When teaching sound programming, I have encouraged students to explore the boundaries of their systems. It is often when things start falling apart (or blow up) that the interesting things appear. However, as when dealing with drunk humans, you need to know how to calm them down and be in control of the situation.
Thanks to Grammarly for checking the language.