The reason machine learning algorithms show bias is that the goal of these algorithms is to learn ALL the patterns in the data including the biases. The "bias" is actually the gap between what the data scientist THINKS is being learned and what's actually being learned. 🧵
An interesting feature of this bias is it's subjective. It depends on what the data scientist INTENDED to learn from the data. For all we know, the data scientist intended to learn all the patterns in the data, racism and all. In which case, there is no bias.
Generally, machine learning does not require us to be specific about what patterns we are trying to learn. It just vaguely picks up all of them. This means we often have no clue what was learned and if it is what we intended to learn.
Traditional statistics isn't like this. In statistics, the first step is specifying what patterns you want to detect. This requires you to have some kind of theory about the structure of the data. Most importantly, this allows you to check if your theory is wrong.
This issue is an huge weakness of the machine learning approach. The vagueness about what is being learned means that we have to do a lot of work after we fit the model to understand the properties of the model itself. In practice, this work is often not done.
The reason we need to do the work is because we can't rely on theory to tell us what the model learned so we must measure it. This means looking at how the model behaves in order to see if it's racist, sexist or has other biases we might care about.
As we see with the many examples of racist algorithms, many of the people using machine learning mistakenly think that they can rely on their intuitions to guess what kinds of patterns are in their dataset and what kind of patterns their algorithms are learning. This is naive.
I think the solution to racism in algorithms (and other biases of this kind) is to be more hands-on about understanding the processes that created the data your model uses and more proactive and explicit about checking that your models have the properties you think they have. 🧵
• • •
Missing some Tweet in this thread? You can try to
force a refresh
You may have heard hallucinations are a big problem in AI, that they make stuff up that sounds very convincing, but isn't real.
Hallucinations aren't the real issue. The real issue is Exact vs Approximate, and it's a much, much bigger problem.
When you fit a curve to data, you have choices.
You can force it to pass through every point, or you can approximate the overall shape of the points without hitting any single point exactly.
When it comes to AI, there's a similar choice.
These models are built to match the shape of language. In any given context, the model can either produce exactly the text it was trained on, or it can produce text that's close but not identical
I’m deeply skeptical of the AI hype because I’ve seen this all before. I’ve watched Silicon Valley chase the dream of easy money from data over and over again, and they always hit a wall.
Story time.
First it was big data. The claim was that if you just piled up enough data, the answers would be so obvious that even the dumbest algorithm or biggest idiot could see them.
Models were an afterthought. People laughed at you if you said the details mattered.
Unsurprisingly, it didn't work out.
Next came data scientists. The idea was simple: hire smart science PhDs, point them at your pile of data, wait for the monetizable insights to roll in.
As a statistician, this is extremely alarming. I’ve spent years thinking about the ethical principles that guide data analysis. Here are a few that feel most urgent:
RESPECT AUTONOMY
Collect data only with meaningful consent. People deserve control over how their information is used.
Example: If you're studying mobile app behavior, don’t log GPS location unless users explicitly opt in and understand the implications.
DO NO HARM
Anticipate and prevent harm, including breaches of privacy and stigmatization.
Example: If 100% of a small town tests positive for HIV, reporting that stat would violate privacy. Aggregating to the county level protects individuals while keeping the data useful.
Hot take: Students using chatgpt to cheat are just following the system’s logic to its natural conclusion, a system that treats learning as a series of hoops to jump through, not a path to becoming more fully oneself.
The tragedy is that teachers and students actually want the same thing, for the student to grow in capability and agency, but school pits them against each other, turning learning into compliance and grading into surveillance.
Properly understood, passing up a real chance to learn is like skipping out on great sex or premium ice cream. One could but why would one want to?
If you think about how statistics works it’s extremely obvious why a model built on purely statistical patterns would “hallucinate”. Explanation in next tweet.
Very simply, statistics is about taking two points you know exist and drawing a line between them, basically completing patterns.
Sometimes that middle point is something that exists in the physical world, sometimes it’s something that could potentially exist, but doesn’t.
Imagine an algorithm that could predict what a couple’s kids might look like. How’s the algorithm supposed to know if one of those kids it predicted actually exists or not?
The child’s existence has no logical relationship to the genomics data the algorithm has available.