When it sounds like a plan, but isn't one
Large language models give polish even to poor ideas.
“It’ll never be enough
It’ll never measure up
Turning the depth of the ocean to the size of a cup
But aren’t we good at turning beauty into clichés?”
— Madison Cunningham, from “Beauty Into Clichés”
In a series of videos, @FatherPhi asks various GenAI services (ChatGPT, Gemini, and Claude, among them) whether he should walk or drive to a carwash that’s only 100 meters away. Each provides a coherent and confident response with supporting arguments. They all advise him to walk.
This is one of many such errors generated by large language models – competent-sounding but wrong. They also generate extraordinary and transformative results. The trick is in knowing the difference.
For the carwash example, the point here isn’t the glaring error. The point is the polish. The responses from these systems are clear and clean. The flow and structure are convincing. Every answer actually “sounds like a plan.” But some of those answers can lead you to a carwash without a car.
Generative AI raises the floor on clean and competent communication. That polish is both a feature and a flaw.
The underlying models are developed and trained on massive amounts of (stolen) content to produce strings of next-most-likely words. As a result, when you ask GenAI for a plan, you genuinely get a response that “sounds like a plan” because the model is pattern-matching what plans sound like. Quite often, that’s entirely fine for a wide array of computer code, cogent coaching, crisp emails, and clean reports.
But the polish of the answer is no longer a proxy for its value. The insights are dressed in the same clothing as the dreck.
So, how do you determine whether something that sounds like a plan is an aligned and actionable plan? You look beyond the polish and the prose to the assumptions, connections, and conclusions behind them. You bring your uniquely human intelligence to the task – your embodied, situated, and social self:
Embodied – is the answer aligned with genuine lived experience, with sensory, emotional, physical reality, not generically but specifically for each person it touches?
Situated – does the answer make sense in and for its particular place, using not just most-likely next words but most-resonant and most-relevant for the geographies and ecologies involved?
Social – does the answer consider and connect with the human relationships, shared assumptions, conventions, and cultures in play?
All of this requires deep discernment, domain expertise, and lived experience. Even if you and your team don’t use AI, you’ll need to develop and defend this discernment in response to others who do, and against the inevitable flattening of even human-generated text in an AI-generated world.
Thanks to the pattern-prodigy of generative AI, almost everything can sound like a plan. It’s our job now to sort and sift the signal from the noise.
From the ArtsManaged Field Guide
Function of the Week: Program & Production
Program & Production involves developing, assembling, presenting, and preserving coherent services or experiences.
Framework of the Week: Calibrating Uncertainty
Calibrating Uncertainty is a framework for decision-making that involves assessing the chance and cost of being wrong. It helps prioritize actions by determining whether to invest in thorough information gathering or to proceed with small, experimental steps based on the potential risks and consequences.
Photo by Randy Jacob on Unsplash
Totally. AI can help generate ideas, but it can't replace the lived experience, cultural context, and human judgment that turn an idea into meaningful work.