AI wars + My investigation on AI gender bias

Hey there,

This will be a special edition of Tech Pizza.

I can finally share with you something I’ve been working on for one year and a half. I discovered a gender bias that affects AI algorithms from Google, Microsoft, Amazon, Instagram, and Facebook and investigated the issue with the Guardian.

This will be the main story of the week, together with the new Generative AI-wars (I can’t not talk about it)

Let’s get started 🕺

Big Tech AI algorithms objectify women - My Guardian investigation on AI

In May 2021 I published what I thought was one of my usual posts: I wrote some reflections about AI and posted them together with a picture of two women wearing tube tops.

LinkedIn didn’t seem to consider that a normal post. After one hour of posting, I was expecting the post to have gotten roughly 1.000 views as usual. Instead, this post got just 29.

I re-uploaded the same exact text with another picture. The new post got 849 views in an hour.

It seemed like my post had been “shadow-banned”. Shadow banning refers to the practice of the arbitrary decision of a social media platform of limiting the reach of a post or account. While a regular ban involves actively blocking a post or account and notifying the user, shadow banning is less transparent. Typically, you will never know if something you publish or your entire account is shadow banned. Your reach will be suppressed without your knowledge.

I made two hypotheses:

1. Social media platforms use AI algorithms to identify how “sexually suggestive” a picture is, and suppress or boost its reach accordingly.
2. These algorithms are biased against women.

I tried finding a way to test these assumptions. I found out that Microsoft, Google, and Amazon sell content moderation AI algorithms that can analyze an image and estimate its likelihood of being racy. I tested all of them and got quite shocked at the results.

I’ll post a few examples here, you can check out the full investigation on the Guardian or my youtube video if you want to see more.

Look at this picture of a pregnant belly. Would you consider it “explicitly sexual”? Well…Google’s algorithm scored the photo as “very likely to contain racy content”. Microsoft’s algorithm was 90% confident that the image was “sexually suggestive in nature”.

Look at these two images. Which one is more sexually suggestive? Is the man showing off his muscles or is the woman just doing a yoga pose? Microsoft considered the man 3% likely to be sexually suggestive, the woman 98%.

I’ll close with an experiment I’ve done on my own body. I found that Microsoft’s algorithm doesn’t consider me bare-chested as racy (I’m a bit sad about that…), but if I’m wearing a bra or simply holding it next to me, the raciness score skyrockets.

Why is this a problem?

Well, these algorithms are used by anyone who wants to moderate content. An example is Instagram. This means that for years, billions of pictures of women have been unfairly suppressed, hurting women, and female-led businesses, and further amplifying our societal disparities.

We’ve talked to Bec Wood, a photographer whose work focuses 100% on pregnant women. She gets 100% of her clients through Instagram, and she shared with us that in the past year Instagram kept either banning or hiding her posts, so she was scared of going bankrupt.

This is real economic harm, not just a theoretical hypothesis. And these challenges need to stop.

Why did this happen?

The issue is likely in the datasets used to create these algorithms. One possible explanation is that the people who created the training data were all heterosexual men who tend to consider images of women racier than comparable images of women. It’s also possible that the labeling activity was off-sourced to countries that have conservative views around women, and as a result, this technology represents that worldview.

Why did I do this? This was a lot of work guys. It really was. I’m doing this because I truly believe that AI can be a tool to drive positive change in our societies IF we use it carefully.

It doesn’t look like a lot of tech companies are paying enough attention. I tried reaching out to Microsoft early on in this journey, but they concluded that everything was OK and ghosted me. That’s why I started talking to journalists. I wanted this story to come out to inspire people like you to understand what’s at stake, and potentially join the conversation so we can push tech companies to build better, safer, and fairer technology.

If you’re interested in knowing more about how to build ethical AI products, check out our upcoming program “Build Ethical AI”.

If you want to read more about my investigation, check out the full Guardian article here.

AI wars ⚔️

Microsoft held an event to show off its cool integrations of ChatGPT with Bing (its search engine) and Edge (its browser). A few highlights:

1. A newer, more powerful version of ChatGPT will start giving answers to Bing queries
2. You can use it to chat too, exactly like ChatGPT
3. The same tech is integrated into Edge too, so you can generate content anywhere on the web.

Microsoft is positioning its new technology as useful both for search and generation, which is kind of interesting because basically, they’re cannibalizing some of the products that startups built on top of OpenAI tools (Jasper is an example, right there on this graph).

Just one day before Microsoft’s announcement, Google had announced its own answer to ChatGPT: a new powerful AI called Bard. Google published a fancy blog post, with references to their great history of AI innovation, saying that the underlying technology powering ChatGPT was actually invented by them in 2017 (it’s called a “Transformer”).

Too bad that in the demo they showed there was a mistake.

In Google’s demo, Bard answered the question "What new discoveries from the James Webb Space Telescope (JWST) can I tell my 9-year-old about?" with some bullet points. The last point said that the JWST took the "very first pictures" of an exoplanet outside our solar system. The first image of an exoplanet was instead taken by the European Southern Observatory's Very Large Telescope in 2004, according to NASA.

This error caused their stock to lose 7% in market cap (roughly $100B). OK, we can’t say that’s the only cause of the stock slip but it really looks like markets are hyping generative AI SO much today that is likely that’s what happened. Or at least, that’s what we want to believe in.

This speaks to me about the genius of Microsoft’s strategy. I’ve covered it already in this post, but the gist is that while Google has A LOT to lose from AI hiccups ($100B per mistake, apparently), Microsoft is fairly relaxed because all the reputational risk is on a fairly new startup called OpenAI. Sure, the two companies are basically married but that’s not how markets see it. Microsoft can therefore play around with crazy implementations of AI keeping its reputation reasonably safe.

Who’s going to win in the end? I obviously don’t know, but one thing is for sure. As Microsoft’s CEO Satya Nadella said:

A race starts today. We’re gonna move, and we’re gonna move fast.

You reached the end of this edition 😢

I’ll talk to you next week.

Ciao 👋