With AI, are we automating? Or are we delegating?
We keep trying to automate with LLMs. We’re actually delegating, and pretending otherwise makes us bad at it.
Large language models are confusing everyone. Is it a computer? Is it a person? How do we talk to it? The advice is all over the map — sometimes contradictory within the same article. Here’s a good example from the BBC:
Companies design AIs like ChatGPT or Google’s Gemini to behave like people, so it makes sense they can sometimes seem as if they have moods you can manage or personalities you can steer. Don’t be fooled. AI tools are mimics, not living beings. They’re just simulating human behaviour. If you want better answers, stop treating AI like a person and start treating it like a tool.
They then offer a series of tips for how to talk to an AI effectively, which boil down to:
- Ask for multiple options
- Give examples
- Encourage it to ask questions
- Be careful assigning it roles
- Don’t ask leading questions
This isn’t advice you give about a tool. Instead, it is effective communication. If you do this, you are treating AI more like a person, not less.
Is it possible that we’re actually better off pretending that the AI is human? The BBC says “no,” but their advice says “yes.”
The BBC’s confusion is our own. We expect computers to be deterministic, but now suddenly they’re not. We still want to feel like they’re a tool. We want to be able to say words to them, and have them do exactly what we want. We get mad at them when they don’t.
LLMs can “understand” the natural language you say because they mimic human communication. Just as with humans, if that communication is poor, details are missing, or instructions are contradictory, LLMs react like a human and try to make the best of the situation. They don’t throw a syntax error. They just get “weird.”
People seem to think that once you can “program” a computer using natural language, they can just say whatever and the computer will figure it out. That‘s not how it works. We‘ve just replaced the rigor of working with a programming language with the messiness of collaborating with another intelligence. We still have a responsibility to communicate well, and pretending an LLM is “just a tool” doesn’t make that responsibility go away.
Automating vs Delegating
Tools get automated. Once an automation is successful, it can run unattended for long stretches. The system will tick along until something breaks.
We delegate to people. We expect them to fill in the gaps, extrapolate from incomplete information, come back for help or clarifications on instructions. We expect them to figure shit out so we don’t have to.
LLMs do all that “people” stuff. We think “we’re automating” because it’s a computer. We think we ought to be able to shove it in a box somewhere on AWS and stop paying attention, the same way we do any other bit of software. But we’re actually delegating, and we still have the same responsibilities we would if we were working with a human.
But (says the peanut gallery) what about the mechanical tasks LLMs do so well? Tagging, taking notes, summarizing, translating? Surely that counts, right?
Well, no. These tasks are real work that require judgment, creativity, and initiative. Translation is not just word-substitution. Classifying content is not just picking the right tag. Summarizing without clear purpose is just condensing. If this was mechanical work, we’d be able to automate it with deterministic tools. We can’t, which is why LLMs doing it well enough to seem plausible is impressive. Maybe it looks easy, but quality requires a craftsperson’s touch.
In the quest for automated levels of reliability, we end up with the AI equivalent of micromanagement: over-detailed instructions and prompting, many “check-in” points to validate results, modeling entire software engineering teams. Again: we actually model teams collaborating, delegating, double-checking, reporting back. This works fine — sometimes quite well, in fact — as long as we are still there to guide the process, validate the results, making sure what’s coming back is what we wanted. Try to let it run unattended, though, and even a 5% failure rate can be an insurmountable pile of crap. That’s not automating. That’s the boss taking a two-martini lunch. That’s playing golf during working hours.
An LLM is a tool, but one unlike any of our other tools. It is a tool designed to understand and mimic human communication. This is incredibly powerful in some ways — but in the process, LLMs inherit many of the inefficiencies and errors humans are capable of. What’s interesting about the BBC’s advice about talking to an LLM is that it closely tracks advice about delegating.
What’s missing from the advice is what to expect from the results of delegating. Expect deviation from your instructions. Expect your delegate to approach the problem in a different way than you would. Expect the results to require refinement or correction. And if the task is repeated, expect variation in the process and in the quality of results.
It’s hard for the BBC to make this leap, though, because tools do what they’re told, and we’re not supposed to anthropomorphize LLMs.
That impulse is misguided. It’s also impossible to adhere to. Humans anthropomorphize inanimate objects all the time. They gender their cars for crying out loud. It’s just a thing we do, and asking us not to when it does such a good job sounding human is bonkers. It’s queasiness about admitting that we’re working with a mock human. And that prevents us from engaging constructively and realistically with the results we get from the LLM. We don’t expect to “one shot prompt” a person. (Well, we shouldn’t.) Why do we expect to one-shot an LLM that’s imitating one?
My personal experience using LLMs for programming suggests that when I exercise the same skills (and patience) working with an LLM as I do working with human team members, I get better results. Or, at least, I am more satisfied. When delegating, that amounts to the same thing.