Prompting is still annoyingly important

Nearly two years into the ChatGPT era, a pattern has become clear across the charity teams we work with. The people who get useful results from AI tools are the ones who've learned how to talk to them well. The people who try it once, get something mediocre, and conclude "AI isn't that good" are usually the ones who typed a vague request into a chat box and expected magic.

This isn't their fault. The entire pitch for generative AI has been that it's as easy as having a conversation. Just type what you want, in plain English, and the machine will figure it out. That pitch is misleading. Not because AI tools are bad, but because expressing what you actually want is surprisingly difficult.

Language requires effort

Daniel Kahneman's distinction between System 1 (fast, automatic) and System 2 (slow, deliberate) thinking is useful here. Clicking a button is System 1. Selecting from a menu is System 1. Constructing a sentence that clearly expresses your intent, with enough context for a machine that has no idea who you are or what you're trying to achieve, is System 2. It requires active thought.

That matters because most people using AI at work are trying to save time. They're fitting it between meetings, emails, and the other seventeen things on their list. They don't want to carefully construct a paragraph of context before getting a draft email. They want to type "write me a fundraising email" and move on. What they get back is generic, because the request was generic. The machine isn't reading between the lines. It can't. You have to be explicit about everything: who you are, who you're writing to, what tone you want, what you're trying to achieve, what constraints matter.

Anthropologist Edward T. Hall described cultures as high-context (where meaning is implied) and low-context (where everything must be spelled out). AI tools are radically low-context. Every assumption you don't state is an assumption the machine will fill in with something average. Most people communicate in high-context mode by default, especially at work where shared understanding is taken for granted. The mismatch is where mediocre AI outputs come from.

We've seen this before

This isn't the first time the tech industry has promised natural language as the universal interface. Smart speakers were going to change how we interacted with technology. In practice, most people learned three commands, found the experience frustrating, and let the devices collect dust. Voice interaction research at consumer electronics companies consistently showed high failure rates. Not because the speech recognition was bad, but because people struggled to express what they wanted through language alone when the system couldn't see what they were looking at, didn't share their context, and couldn't ask clarifying questions naturally.

The text box in ChatGPT is a more capable version of "Hey Google," but the underlying interaction problem is the same. Language is cognitively expensive, and we're asking people to use it as their primary interface with a tool that has no shared context.

What this means for charities

There's a practical consequence for any charity rolling out AI tools. The gap between "we gave everyone access to ChatGPT" and "people are getting real value from ChatGPT" is largely a prompting gap. Bridging it takes practice, and practice takes time that most charity staff don't have.

This is a hidden cost of AI adoption that rarely appears in the business case. The tool is cheap or free. Learning to use it well is neither.

It also points to something about how AI tools should be designed for specific workflows. When prompting is built into the product, when someone has already figured out the right context, constraints, and structure for a particular task, the user doesn't need to do that cognitive work themselves. A purpose-built tool for analysing beneficiary feedback doesn't require the user to write a paragraph explaining what beneficiary feedback is and how they'd like it analysed. That thinking has been done in advance.

The charities getting the most from AI right now tend to be the ones that have moved beyond general-purpose chat interfaces for their important workflows. Not because ChatGPT isn't capable, but because relying on every staff member to prompt it well, every time, for every task, is a strategy that depends on humans consistently doing something that's cognitively hard. That's a risky bet.