The True Laws of Robotics, From A True Robot Master
Rules for AI as well, just in time for the latest escalation
![](https://substackcdn.com/image/fetch/w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1bc37fd3-aa27-455d-a2da-25bf412da769_1024x1024.png)
Nothing illustrates the power of science fiction better than Isaac Asimov's Three Laws of Robotics.
Asimov coined them (in 1942!) after conversations with other SF writers and editors, because the laws' interactions make for good stories. But he also thought of them as scientifically possible and even necessary -- indeed, some sources say he felt they were an obvious set of principles for any kind of machine. He did not, however, set them forth as a research program. Yet many real discussions of AI and robots and ethics allude to these laws.
In 1942 there was no real-world experience to inform theory. In 2024, of course, that's not so. Many people have made many robots over the past few decades.
Few, though, have been responsible for as many robots as Rodney Brooks (known to us civilians for the Roomba and to soldiers for the Packbot bomb-disposal robot). Those (and many other Brooks-related) robots had multiple parents (including Brooks' former students Joe Jones and Helen Greiner). But their underlying principle was an idea Brooks hit on when he was house-bound in Thailand for a few weeks in 1985. Instead of designing a robot as an advanced computer that tries to model all the world's complexities, Brooks thought, he could make robots that were simple. Just beings moving around in real spaces, following simple rules -- like the teeny-brained insects he saw buzzing around his room.
This was huge. It meant he could build a robot and tell it to move around and not bump into stuff, and let it work out how to get by. That would lead to (to quote the titles of two papers he later wrote) "intelligence without representation" and "intelligence without reason." Most robots now in homes, streets, battlefields, construction sites and elsewhere work according to some aspect of this design principle.
Asimov lived within his own capacious mind. He hated traveling, never took a plane and worked with his blinds always closed. He didn't want to see the world outside his apartment window (that's what he told me, anyway, when I interviewed him in 1987). Brooks has always been the opposite. He's all about getting out in the world, being involved, getting your hands (or grippers, or code) dirty.
So that's why I took note last month when Brooks put forth his three laws of robotics. Then, because I am slow and bad at multi-tasking, I hadn't even gotten them into this blog before Brooks brought forth unto us another three laws, for AI. I think they merit at least as much consideration as Asimov's. So here they are.
Brooks' Three Laws of Robotics
As posted here, they are:
1. The visual appearance of a robot makes a promise about what it can do and how smart it is. It needs to deliver or slightly over deliver on that promise or it will not be accepted.
I think Brooks is warning here against robots that look too smart for their own good. Robots with faces, expressions and voices that will give people the impression they're speaking with a true artificial human. We know that when meeting a machine that is good at a hard task (like winning at chess or go) people often assume it's good at other hard tasks (at which it quite probably sucks). This leads to MORD -- the Moment of Robotic Disappointment -- when the expectant human watches the robot fail at some task the human (unreasonably) assumed the robot would crush. Brooks is telling his fellow roboticists to make unassuming things that people will underestimate, rather than stuff that looks like Skynet troops or C3PO.
2. When robots and people coexist in the same spaces, the robots must not take away from people's agency, particularly when the robots are failing, as inevitably they will at times.
A robot can fail and still be useful -- as RobustAI's Leila Takayama said on the podcast a few weeks ago, a robot cart made by the company, when it fails, turns into a regular old cart, which can still carry stuff. (Brooks is Chief Technology Officer and a co-founder of RobustAI.) OTOH, if your delivery robot is stuck and blocking the curb cut, its failure takes away both its robotic capacities and the capacities people had before. That leaves people worse off than they were without robots. And that will make people think robots aren't worth the trouble.
3. Technologies for robots need 10+ years of steady improvement beyond lab demos of the target tasks to mature to low cost and to have their limitations characterized well enough that they can deliver 99.9% of the time. Every 10 more years gets another 9 in reliability.
IOW: Patience! All those lovely videos of robots sorting dishes do not mean they are coming to Walmart any time soon.
Brooks' Three Laws of Artificial Intelligence
They were posted here, and they are:
1. When an AI system performs a task, human observers immediately estimate its general competence in areas that seem related. Usually that estimate is wildly overinflated.
I think this is the AI equivalent of the first robot rule -- but stated more forcefully because people do this much more with the new, much-hyped AIs than they do with robots.
2. Most successful AI deployments have a human somewhere in the loop (perhaps the person they are helping) and their intelligence smooths the edges.
Sometimes companies try to disguise this fact, leading to what Astra Taylor calls "fauxtomation": Work by humans presented as work by AI. In other instances the human is an expected and necessary part of the process. Some lethal military robots, for instance, have a human "in the loop" (a decision to attack requires a person's OK) or at least "on the loop" (the system can attack on its own but a human is monitoring what it's doing). The key point is that even if a system works very well a lot of the time, there are going to be moments where it doesn't do so well, and a human will be needed. Human abilities to improvise in the face of the unexpected, and human knowledge of the world and of what humans want, can't be equaled by any AI. Those edges will need smoothing for a long time.
3. Without carefully boxing in how an AI system is deployed there is always a long tail of special cases that take decades to discover and fix. Paradoxically all those fixes are AI-complete themselves.
"AI-complete" means that the problem at hand is as difficult as creating an AI that can solve any problem a human could. Even if an AI works great 99 percent of the time, the 1 percent exceptions will turn up and solving them won't be a matter of tweaking a few details and mopping up. So in realms where we can't afford a single failure -- like massive financial-market moves, nuclear deterrence, life-or-death medical decisions, or driving down a rough rural road in the dark -- we should not count on AI to be faultless. It may be very, very good. It may be better than the typical human would be. There will still be odd, unexpected exceptional circumstances where it will fail. Given Rule 1, it's a design challenge to constrain AI so that these failures aren't unexpected.
Here Come the Agents
So maybe humanity should move cautiously and carefully in the development of AI and robots that can affect the world.
Which isn't what is happening.
Even as people are still adjusting to AI that can answer any question and "invent" sentences, images and videos, the next disruption is approaching: AI "agents" that are supposed to handle questions, and do tasks, as naturally as a human assistant.
Right now, you can ask an AI like Chat-GPT or Claude to recommend a good restaurant downtown for a first date with an anxious person, and maybe get a useful response. In an agent world, you could get a response plus news that your table for two at the best option was already reserved. There's no question that this world is coming. How well it will work is a different issue (see AI Rule 3 above). And, if and when it works really well, that will raise another question: Will people enjoy being reminded that they're predictable creatures of habit and average? Because agents' ability to satisfy our desires will mean the machines have got our number.
But let's not get ahead of ourselves. This week's news simply indicates that the era of agents is dawning.
First, OpenAI released GPT-4o (the "o" is for "optimum"). It's an accessible "multi-modal" AI that can handle speech, images, video and other modes. Good video intro here.
The next day, Google announced that its "Project Astra" had made progress in turning its Gemini model into a first pass at a real AI agent -- a device that can usefully answer a question like "where the hell did I leave my glasses?" Here's one of their videos.
Onward we hurtle into the AI future. I recommend keeping Brooks’ laws in mind.
The "o" is for "omni."
Excellent article; thanks!