A way to think about responsibility that makes sense, both when working with generative AI and personally.
Responsibility. What is it actually?
I’ve done a fair amount of thinking and writing about it, and I still like the Terry Pratchett version. This passage in Wee Free Men has young witch Tiffany Aching facing the end of the world in the absence of her mentor, Miss Tick, and talking (as one does) to a knowing toad:
“You’d better tell me what you know, toad,” said Tiffany. “Miss Tick isn’t here. I am.”
“Another world is colliding with this one,” said the toad. “There. Happy now? That’s what Miss Tick thinks. But it’s happening faster than she expected. All the monsters are coming back.”
“Why?”
“There’s no one to stop them.”
There was silence for a moment. “There’s me,” said Tiffany.
That’s responsibility in the sense of stepping forward and doing the thing that needs to be done, even if it is dangerous and no one would blame you for looking the other way. But you do it for the common good, because other people (or the world) depend on you, and because no one else is willing to do it.
Let’s take that down a notch
There’s a more mundane kind of responsibility, which is really just the list of things that need doing to keep the everyday world running, along with the person who is assigned to that task: who takes the recycling out or watches the kids or gets supper on the table?
In the workplace, responsibility is that list, but with more at stake. That’s because the level of responsibility one has in an organisation is often tied to the amount of money one gets in a pay packet (or a set of stock options) at the end of the month.
I’ve yet to see a clearer breakdown of how that works than the one given by a writer I’ve been following for a while now.
Roger Martin is a management consultant and strategy advisor to companies worldwide. His series of articles on Substack is called Playing to Win/Practitioner Insights (PTW/PI) which may sound intimidating to people who don’t do a lot of business-related reading. But don’t be intimidated! Some of what he writes doesn’t apply to me or my small business, but a lot of it does, and I’ve benefitted from his ideas enormously.
READ: How a core idea of his – “betterment” – helped me.
A recent post, entitled Sharing Responsibility Productively With Both Humans & Artificial Intelligence, sounded promising. Not only do I have many thoughts about responsibility, I also have many thoughts about AI. I was not disappointed.
Martin’s key insight is that there are two ways of assigning responsibility: “I am in charge & you are not, and you are in charge & I am not”. But these two extreme positions are not helpful. Martin says it’s better to think of responsibility as a ladder, explained in this diagram:
Using the ladder metaphor is (to my mind) a really useful way to think about responsibility. Martin writes at length about how this works in a corporate setting, advising people to pick rungs on the ladder that are appropriate for their skills and then to talk to their boss about that, or do the work and then talk to their boss
As someone who was once a manager of people on an online journalism website, I know the qualitative difference between a team member who stands at my desk and says: “I was working on a story and now it’s gone and I don’t know what to do” (Rung 6 on the ladder) and someone who says: “I was working on a story and then (insert actual observations of what happened here), so I phoned IT and they told me that there’s a general problem affecting everyone. I’ve told everyone in the office and put in a ticket to get it fixed.” (Rung 1 on the ladder).
Guess which employee gets recommended for promotion? Beyond that flippant remark though, I wish I had had this mental model at the time. If I had been able to recognise tasks and their places in the levels of reponsibility in this way, and mentor staff to move through them, I would have been a better manager.
How does this relate to generative AI?
In a hyped-up world where it seems every company is deploying AI agents to do stuff (all the stuff! all the time!), the question of responsibility needs some thinking about.
Martin thinks AI enthusiasts who say AI should be operating at the top rung of the ladder (making decisions and informing the human after the fact) are wrong. That’s because AI generally is not fit for purpose at that level.
Why is that?
First, we need to take a small step back, and look at what large language models do. In another post, Martin uses the concept of the Knowledge Funnel to explain how knowledge progresses. It goes:
- from mysteries (why do objects fall to the ground?) to
- heuristics (the theory of gravity) to
- algorithms (mathematical models that describe gravitational pull).
Large language model-based generative AI (LLM/AI) scours large databases, infers heuristics and then creates algorithms. But as it is doing that, says Martin:
(T)he LLM/AI does not ask: what is unique or outstanding? It asks: what is most? LLM/AI democratises heuristics – but it makes them average. LLM/AI will find the mean/median/modal representation of the heuristic in the domain it is asked to search. It won’t find the right tail of the distribution because it isn’t set up to do so – it won’t even try.” (his emphasis)
That analysis of LLMs and AI dates back to March 2024, and so could be regarded as ancient history in AI-land. (For clarity, “the right-tail of the distribution” is the part of a bell curve, with dreadful on the left, average in the middle, and quality on the right).
But my experience, and from a lot of what I read, is that this is still the case. Even when being deployed for advanced use cases, there’s not a lot of evidence to think that language-based AI is doing much more than synthesising and parroting, or that coding agents can be trusted to work fully on their own. (See Ed Zitron’s latest post, which says that at real estate tech firm Zillow, LLM outputs requiring human review have increased by nearly 50%; code deployments and pull requests have increased by 39%; and software reviewer load has increased by 29,000 hours each month, creating a massive burden on the 1,500 or so engineers working at the company.).
Back to responsibility
Martin’s core point is this: “(R)esponsibility sharing is not a one-size-fits-all proposition. It is context-specific, and each context – i.e. each task – requires a customized division.”
What that means is that context matters when deciding where and how to use generative AI. If a task needs an “average” answer, then by all means let the AI do the task and tell you about it later. But if you need something more complex, think again. Says Martin:
I would never assign AI to the top rung – ever – because AI doesn’t even know how to try to find such a solution. In fact, I wouldn’t ask it to take second-rung responsibility either… I might ask it to generate options (third rung down) as food for thought for me. But more likely, I would take the lead and ask it to provide critiques of my thinking along the way. (my emphasis)
For me, that’s the key thing in working with AI: the human needs to take the lead, to make the decisions, in advance, about what is an appropriate sharing of responsibilities.
How then to use generative AI?
When I read about the complexities that the use of AI brings to workplaces, I wonder if it is all worth it. I attended a session of the AI in Publishing Worflows conference (organised by PurePub.ai) called From Tools to Teammates. I was struck by how much meta-level work seemed to be going on to orchestrate agents to do things like marketing, research and manuscript screening. Would it not be quicker and easier to stick to human-mode, I wondered. What’s the benefit of handing tasks to technology if just as much work goes into supervising the work of agents as might be needed to have humans do the tasks?
I know that carrying a burdensome workload can make people feel like Atlas in the Greek myth, carrying the sky on his shoulders (it’s often thought that the sphere hs s carrying is the world, but it ain’s so). And that the promise of a shiny new toy that could help carry that load is enticing. But it’s possible to simply be engaging in making more work – without any prospect of long-term benefit – while playing with these toys.
Do the thinking work
Enter Martin’s Ladder of Responsibility, which can serve as a lens to look at that workload and see how it might be divided among team-mates or, indeed, AI agents.
I, for instance, am happy to hand the task of taking meeting notes to a tool called Granola, which operates at a level in the ladder where it does the job and then hands the output to me, pretty much as a done deal. But I am not happy to design and build a tool to schedule my work days, because I think I need to take responsibility for that – I am the only one who can know what any given day looks and feels like.
My one-person business is of course a lot less complex than (say) a publishing house. But in both cases, it makes sense to get AI tools to do the work they are fit to do, based on a task-by-task analysis of who carries responsiblity and accountability for the work. There’s work and thinking involved there, too, but using generative and agentic AI because it seems like the latest time- and money-saving tech could be causing more problems than it solves.
Atlas no doubt would have loved to lay down his burden, and get on with his mythical life. But it wasn’t that simple; he had been condemned to carry the sky for eternity. We modern-day humans are no different – the temptation to find the way to the free lunch is always there. But we have a responsibility to ourselves to think things through, and to do the work that is ours to do.
Main picture: A view of a bronze statue of Atlas at the Rockefeller Center in New York. The statue was created by the sculptor Lee Lawrie with the help of Rene Paul Chambellan and was installed in 1937, says Wikipedia. Photo by Afif Ramdhasuma, Pexels
Previous Sensible Guide articles
Yes, we did make a mess. Taking responsibility has nothing (and everything) to do with the personal – What does responsibility mean? Where does it start, and where does it end? Tomatoes are a good place to begin.
Leadership in the time of plague – A post from 2020, when the Covid-19 pandemic concentrated everyone’s minds on questions of leadership. But perhaps we would be better served (then and now) looking to ourselves and our responsibilities?
Being a creator in the AI swamp – a way forward – If you work in the creative industries (writing, editing), artificial intelligence looms like a tsunami over you. My thoughts on a way forward… (a post from March 2024).