4 ms·

> One ambitious goal of companies developing AI agents is to have them interact with other kinds of software as humans do, by making sense of a visual interface

> One ambitious goal of companies developing AI agents is to have them interact with other kinds of software as humans do, by making sense of a visual interface and then clicking buttons or typing to complete a task. Can it avoid dark patterns? If it can't, then it's not a solution to anything. It's more like an additional problem. It also opens the door to the concept of dark agents (hey, this bot _always_ installs the Altavista search bar).

ethbr1 2y ago
Are dark patterns harder than captchas? I'd say no.
- latexr 2y ago
  > Are dark patterns harder than captchas? Yes, yes they are. A captcha is a simple system which returns a binary pass/fail. A dark pattern plays with your understanding of reality to make you choose what is beneficial to the company at your expense. AI systems will be gamed to follow dark patterns too.
brookst 2y ago
Why would it not avoid dark patterns? Assuming the reinforcement learning takes desired task and correct outcome into account, it would take work to train them to fall for dark patterns.
- latexr 2y ago
  > Why would it not avoid dark patterns? What do you think is going to happen if an AI system which avoids current dark patterns starts being widespread? Bad actors aren’t going to shrug their shoulders and stop, they’re just going to make systems which trick those AIs. Adapting a dark pattern is significantly faster and more cost effective than training a new AI to bypass those new approaches.
  - brookst 2y ago
    Maybe, but if there is a diversity of AI implementations plus real humans using the system, it may not be practical to fool everyone. Dark patterns work because human psychology is fairly homogeneous. It’s possible that AI will be similar, but I suspect that won’t be the case.
    - latexr 2y ago
      Of course it won’t fool everyone, just like not everyone is fooled now. That’s why we’re able to put a name to the practice, some people saw through it. That just goes to show human psychology is not as homogenous as you make it out to be. There is a vastly greater diversity of humans than AI implementations. It is delusional to think the current crop AI, trained on human behaviour and interactions, won’t be able to be tricked just as easily or even more so, when we’ve been doing it since the start.
      - brookst 2y ago
        “Delusional” is pretty strong for a handwave assertion. And you really think human neuro- and cognitive architecture is more diverse than the many many many different ML architectures? That’s a bold claim. We have wide ML models, and deep, and LORA, a million flavors of transformers and then a million non-transformer architcthres. I’m very suspicious of your absolute certainty and complete dismissal of contrary opinions.
- alganet 2y ago
  I presume Google Mariner is using the digitized behavior of googlers searching for products, analytics from shopping websites (I would use it), even ad interactions (why not?). If it works, boom, lots of unproductive stale data turned into a potential product (if someone builds this, I want 5%). The problem is: that data has a lot of users under the influence of dark patterns. People shopping in stupid ways, falling for tricks. They probably know this and improved search against some of these patterns in the past. But it's not perfect, right? Search isn't. Large models fall for tricks all the time, you have to assume the other way around. They will fail, and you can reduce that failure rate, but never zero it.
philipov 2y ago
A new meaning for the Principal-Agent Problem... or maybe just the same meaning in a new context.
abrichr 2y ago
We designed https://github.com/OpenAdaptAI/OpenAdapt https://github.com/OpenAdaptAI/OpenAdapt to ground the agent to human demonstrations, so that if you as a human don't engage in the dark pattern, neither will the agent.

ethbr1 2y ago

Are dark patterns harder than captchas? I'd say no.

latexr 2y ago
> Are dark patterns harder than captchas? Yes, yes they are. A captcha is a simple system which returns a binary pass/fail. A dark pattern plays with your understanding of reality to make you choose what is beneficial to the company at your expense. AI systems will be gamed to follow dark patterns too.

brookst 2y ago

Why would it not avoid dark patterns? Assuming the reinforcement learning takes desired task and correct outcome into account, it would take work to train them to fall for dark patterns.

latexr 2y ago
> Why would it not avoid dark patterns? What do you think is going to happen if an AI system which avoids current dark patterns starts being widespread? Bad actors aren’t going to shrug their shoulders and stop, they’re just going to make systems which trick those AIs. Adapting a dark pattern is significantly faster and more cost effective than training a new AI to bypass those new approaches.
- brookst 2y ago
  Maybe, but if there is a diversity of AI implementations plus real humans using the system, it may not be practical to fool everyone. Dark patterns work because human psychology is fairly homogeneous. It’s possible that AI will be similar, but I suspect that won’t be the case.
  - latexr 2y ago
    Of course it won’t fool everyone, just like not everyone is fooled now. That’s why we’re able to put a name to the practice, some people saw through it. That just goes to show human psychology is not as homogenous as you make it out to be. There is a vastly greater diversity of humans than AI implementations. It is delusional to think the current crop AI, trained on human behaviour and interactions, won’t be able to be tricked just as easily or even more so, when we’ve been doing it since the start.
    - brookst 2y ago
      “Delusional” is pretty strong for a handwave assertion. And you really think human neuro- and cognitive architecture is more diverse than the many many many different ML architectures? That’s a bold claim. We have wide ML models, and deep, and LORA, a million flavors of transformers and then a million non-transformer architcthres. I’m very suspicious of your absolute certainty and complete dismissal of contrary opinions.
alganet 2y ago
I presume Google Mariner is using the digitized behavior of googlers searching for products, analytics from shopping websites (I would use it), even ad interactions (why not?). If it works, boom, lots of unproductive stale data turned into a potential product (if someone builds this, I want 5%). The problem is: that data has a lot of users under the influence of dark patterns. People shopping in stupid ways, falling for tricks. They probably know this and improved search against some of these patterns in the past. But it's not perfect, right? Search isn't. Large models fall for tricks all the time, you have to assume the other way around. They will fail, and you can reduce that failure rate, but never zero it.

philipov 2y ago

A new meaning for the Principal-Agent Problem... or maybe just the same meaning in a new context.

abrichr 2y ago

We designed https://github.com/OpenAdaptAI/OpenAdapt https://github.com/OpenAdaptAI/OpenAdapt to ground the agent to human demonstrations, so that if you as a human don't engage in the dark pattern, neither will the agent.