Problematic Smart Algorithms
December 12, 2023
This essay is the work of a dumb dinobaby. No smart software required.
We already know that AI is fundamentally biased if it is trained with bad or polluted data models. Most of these biases are unintentional due ignorance on the part of the developers, I.e. lack diversity or vetted information. In order to improve the quality of AI, developers are relying on educated humans to help shape the data models. Not all of the AI projects are looking to fix their polluted data and ZD Net says it’s going to be a huge problem: “Algorithms Soon Will Run Your Life-And Ruin It, If Trained Incorrectly.”
Our lives are saturated with technology that has incorporated AI. Everything from an application used on a smartphone to a digital assistant like Alexa or Siri uses AI. The article tells us about another type of biased data and it’s due to an ironic problem. The science team of Aparna Balagopalan, David Madras, David H. Yang, Dylan Hadfield-Menell, Gillian Hadfield, and Marzyeh Ghassemi worked worked on an AI project that studied how AI algorithms justified their predictions. The data model contained information from human respondents who provided different responses when asked to give descriptive or normative labels for data.
Normative data concentrates on hard facts while descriptive data focuses on value judgements. The team noticed the pattern so they conducted another experiment with four data sets to test different policies. The study asked the respondents to judge an apartment complex’s policy about aggressive dogs against images of canines with normative or descriptive tags. The results were astounding and scary:
"The descriptive labelers were asked to decide whether certain factual features were present or not – such as whether the dog was aggressive or unkempt. If the answer was "yes," then the rule was essentially violated — but the participants had no idea that this rule existed when weighing in and therefore weren’t aware that their answer would eject a hapless canine from the apartment.
Meanwhile, another group of normative labelers were told about the policy prohibiting aggressive dogs, and then asked to stand judgment on each image.
It turns out that humans are far less likely to label an object as a violation when aware of a rule and much more likely to register a dog as aggressive (albeit unknowingly ) when asked to label things descriptively.
The difference wasn’t by a small margin either. Descriptive labelers (those who didn’t know the apartment rule but were asked to weigh in on aggressiveness) had unwittingly condemned 20% more dogs to doggy jail than those who were asked if the same image of the pooch broke the apartment rule or not.”
The conclusion is that AI developers need to spread the word about this problem and find solutions. This could be another fear mongering tactic like the Y2K implosion. What happened with that? Nothing. Yes, this is a problem but it will probably be solved before society meets its end.
Whitney Grace, December 12, 2023