For this article we spoke with the team behind World, including Tiago Sada and Lily Gordon at Tools for Humanity, on how they try to solve this problem.
Great framework idea. Where proof of human is tough today is in agentic contexts. If an agent inherits a valid device credential then they have proven they’re human without being human.
Although you did touch on this. It doesn’t really matter if an agent is human. Are they authorized and are they using a system appropriately. I should be able to delegate work to a personal agent and not have that shut down all the time because I’m not viewing ads.
When I write software today, I am adding agentic interfaces so they don’t have to fake being human. They can have appropriate access through and appropriate interface.
It will be interesting to see where this topic goes for sure.
PS. I’ve worked with “big shoe companies” that have this problem. They have a couple of other detection vectors nobody has figured out yet. I’m not able to say what they are but they’re pretty simple and have a lot of bot makers fooled.
Doesn't the human uniqueness check using iris example require trusting the hardware or client to some extent, and rely on an adversary not being able to reverse engineer it and/or steal keys.
If you can generate the same fake iris random value in a reproducible way and present it to the server, then your fake client or fake hardware can pretend to be one of many different fake people on demand.
that is exactly what World has solved. The orb doesn't store your iris code, the manner in which it is computed and verified is regionally multi party and all on-device in the orb. you can read more at world.org
doesn't this require additional things like special devices, physical control of the devices and monitoring the devices constantly so that you can revoke the keys if one goes missing, is tampered with or is swapped with a decoy. None of these appear to be in the system design. A limited number of devices that a small group of trusted people can track is useful too.
If it was something like a mobile app or hardware that the user had possession of, then there are potentially billions of devices owned by many people that you don't know if you can trust and some you definitely can't. They would have as much time as they wanted to look at as many instances of the software and hardware they wanted, on hardware they control and in their own facilities...
Alex, is writing so hard now? Your opening section is riddled with AI speak.
"The reason they all fail in the same way is worth pausing on.", "The defenses websites have relied on for years are failing in predictable ways…", "Each of them helps for a while and then stops working.", "None of them answers the question that actually matters here". -> These are overused styles by LLMs. Then the perfect parallels, X is a proxy for Y, etc. And then the "Phone numbers can be bought in bulk. Fingerprints can be randomized. IPs can be rotated through residential proxy networks." -> Overused patterns by LLMs.
Have been a great fan of YOUR WRITING. Please don't lose your voice!
Hi Adam, thanks so much for the feedback! We use AI for research and editing, but we never use it to write the first draft. We’ll definitely be more thoughtful going forward and make sure we stay true to our writing style. I don’t like AI generated writing either, so I completely get where you are coming from.
I worked on federated identity and security at Nokia for 10 years. I have a patent for agent identity from that work. This work from world.org looks great and deeply thought out. It will need additional tools to handle the Enterprise needs. Thank you for this write up. Very informative.
The delegation section is the sleeper here. Capping agent activity per verified human quietly flips the whole bot-defense model... instead of asking "is this a human?" you ask "how many humans does this traffic represent?" That seems like it may be right question for the next decade for a lot of different use cases.
But bots-accounts and human-or-robot validation are two different things? Humans creating account and letting bots manage would be able to manipulate the system. The gap in retrospect was to allow for the account creation to take place. One workaround is to do a lottery ( during a sale ). But if bot accounts out number the legitimate ones, it’s pretty much the same situation.
Great framework idea. Where proof of human is tough today is in agentic contexts. If an agent inherits a valid device credential then they have proven they’re human without being human.
Although you did touch on this. It doesn’t really matter if an agent is human. Are they authorized and are they using a system appropriately. I should be able to delegate work to a personal agent and not have that shut down all the time because I’m not viewing ads.
When I write software today, I am adding agentic interfaces so they don’t have to fake being human. They can have appropriate access through and appropriate interface.
It will be interesting to see where this topic goes for sure.
PS. I’ve worked with “big shoe companies” that have this problem. They have a couple of other detection vectors nobody has figured out yet. I’m not able to say what they are but they’re pretty simple and have a lot of bot makers fooled.
Doesn't the human uniqueness check using iris example require trusting the hardware or client to some extent, and rely on an adversary not being able to reverse engineer it and/or steal keys.
If you can generate the same fake iris random value in a reproducible way and present it to the server, then your fake client or fake hardware can pretend to be one of many different fake people on demand.
that is exactly what World has solved. The orb doesn't store your iris code, the manner in which it is computed and verified is regionally multi party and all on-device in the orb. you can read more at world.org
doesn't this require additional things like special devices, physical control of the devices and monitoring the devices constantly so that you can revoke the keys if one goes missing, is tampered with or is swapped with a decoy. None of these appear to be in the system design. A limited number of devices that a small group of trusted people can track is useful too.
If it was something like a mobile app or hardware that the user had possession of, then there are potentially billions of devices owned by many people that you don't know if you can trust and some you definitely can't. They would have as much time as they wanted to look at as many instances of the software and hardware they wanted, on hardware they control and in their own facilities...
Alex, is writing so hard now? Your opening section is riddled with AI speak.
"The reason they all fail in the same way is worth pausing on.", "The defenses websites have relied on for years are failing in predictable ways…", "Each of them helps for a while and then stops working.", "None of them answers the question that actually matters here". -> These are overused styles by LLMs. Then the perfect parallels, X is a proxy for Y, etc. And then the "Phone numbers can be bought in bulk. Fingerprints can be randomized. IPs can be rotated through residential proxy networks." -> Overused patterns by LLMs.
Have been a great fan of YOUR WRITING. Please don't lose your voice!
Hi Adam, thanks so much for the feedback! We use AI for research and editing, but we never use it to write the first draft. We’ll definitely be more thoughtful going forward and make sure we stay true to our writing style. I don’t like AI generated writing either, so I completely get where you are coming from.
I didn't even read the rest of the article I'm afraid it's also written by an AI.
I worked on federated identity and security at Nokia for 10 years. I have a patent for agent identity from that work. This work from world.org looks great and deeply thought out. It will need additional tools to handle the Enterprise needs. Thank you for this write up. Very informative.
Nice article alex 👍
The delegation section is the sleeper here. Capping agent activity per verified human quietly flips the whole bot-defense model... instead of asking "is this a human?" you ask "how many humans does this traffic represent?" That seems like it may be right question for the next decade for a lot of different use cases.
For global leaders and capital thinkers, this phenomenon is less about geology and more about systems thinking:
Value does not equal accessibility
Resources without infrastructure remain theoretical
Nature produces abundance. but not always in usable form
Mount Erebus becomes a metaphor for emerging economies and technologies:
There are assets data, energy, rare materials being generated continuously.
But without the systems to capture, refine, and deploy them, they remain unrealized wealth.
Implication
In the next decade, competitive advantage will not come from discovering new resources.
It will come from engineering the capability to capture what already exists. but is currently unusable.
Because the world is full of “gold dust”. but only a few will know how to turn it into gold.
But bots-accounts and human-or-robot validation are two different things? Humans creating account and letting bots manage would be able to manipulate the system. The gap in retrospect was to allow for the account creation to take place. One workaround is to do a lottery ( during a sale ). But if bot accounts out number the legitimate ones, it’s pretty much the same situation.
How about assigning a UUID (Universal Unique ID)?
Blockchain matters because ownership becomes easier to verify.A clear record saves many arguments and several expensive suits.