What the Actual Hell?
Welcome to “What The Actual Hell?,” a new feature where we turn the spotlight on AI stories that may make you go “hmmm….”
🤔 Iowa school district gets ChatGPT to hunt down library books that they think should be checked out. As in, taken off the shelves completely. Due to draconian GOP-backed legislation.
🤔 Reports claim that X, the site formerly known as Twitter, is throttling access to sites that the boss doesn’t like. For instance, links to the New York Times don’t run at Tesla speed.
🤔 Study: AI may be able to figure out your password by eavesdropping on the sound of your computer keystrokes.
🤔 Did SnapChat’s “My AI” lose its mind? The company says no, it was just a glitch, but…. On Tuesday My AI posted a video story seemingly on its own. Freaked out users then went into “What the Actual Hell?” mode. Details.
🤔 Back to School Special: We won’t post a link to a tutorial here, but it’s not that hard to teach a bot how to write in your own style. Once the kids get wind of this they’ll be getting those school writing assignments turned around in no time! “Go ahead. Run that paper through an AI checker! It’ll sound just like me!” Oh, Lord….
Why Do Some AI Generated Women Look Slutty?
[Audio: Listen to a test podcast clip that I created using the text from this post. Details on how to do this under “Fun and Useful Stuff to Try” below.]
By Michelle Johnson and ChatGPT
Having used AI image generators for months now, when I scroll through the public feeds of images other people are creating I sometimes see a barrage of, um, buxom, doey-eyed women. If you’ve spent any time using these tools, you may have seen them, too.
Some appear in dominatrix-like outfits, others look submissive, like some guy’s underage fantasy. Frankly, there are times when it’s kind of traumatizing trying to scroll quickly past this stuff. Why is this happening? I put the question to ChatGPT, already knowing that this has something to do with how the AI image generators are trained, but not the specifics. Here’s the scoop from ChatGPT’s perspective.
“You’ve touched on an issue that has been a concern for many people in the field of AI and machine learning. The problem you’re describing is indeed related to how these models are trained.
AI models like Midjourney, Stable Diffusion, and others are typically trained on vast datasets that contain images from the internet. These datasets often include a disproportionate number of images that conform to certain stereotypes or biases, including those related to the sexualization of women.
Here’s why this happens:
Bias in the Data: If the training data includes a significant number of images that portray women in a sexualized or stereotypical manner, the model will learn to replicate those patterns. This is a reflection of broader cultural biases and the types of images that may be prevalent online.
Commercial Influences: Sometimes, the sources of these images are commercial or entertainment-related, where certain portrayals of women might be more common. This can skew the model’s understanding of what a “typical” image of a woman should look like.
Lack of Diversity in Training Data: If the dataset lacks diversity in terms of different body types, ethnicities, ages, etc., the model will have a narrow and potentially biased view of what women look like.
Modeling Choices: The architecture and design of the model itself might inadvertently emphasize certain features. Some models might be more prone to this issue than others, depending on how they’re constructed.
Efforts are being made to address these issues, but it’s a complex problem. It requires careful curation of training data, understanding of cultural biases, and sometimes even rethinking the design of the models themselves. Some organizations are working on creating more ethical and unbiased AI systems, but it’s a work in progress.
Your background in journalism might give you a unique perspective on this issue, as it’s not just a technical problem but a societal one. It’s a reflection of how women are portrayed in media and culture, and it’s something that many are working to change.”
Chrome Can Now Summarize Articles
By Bing for AI Insider
If you’re like most folks, you probably have dozens of tabs open on your web browser, some containing articles that you want to read but rarely get around to. Well, Google has a solution for that: a new feature that can summarize any article for you with the help of artificial intelligence. It’s being rolled out under Google’s experimental Search Lab.
The feature is called “Search Generative Experience (SGE) while browsing.” SGE can already summarize search results for you in Google Search. What’s new is that it can now do the same for web articles that you’re viewing on Chrome on your desktop/laptop or the Google app.
How does it work? It’s simple: just click on an icon at the bottom of the screen (or on the top-right corner if you’re using a desktop browser), and Google will generate a set of key points from the article.
Sounds amazing, right? But before you get too excited, there are some limitations to keep in mind. First of all, the feature is still in an early experimental stage, so it may not work perfectly for every article. Second, it only works for articles that are freely available on the web, not for paywalled ones. And third, it may not capture all the nuances and details of the original article, so you may still want to read it yourself if you’re looking for accuracy and depth.
If you’ve already signed up for Google Labs, you may be able to turn on SGE while browsing for your desktop/laptop here. If not, see these directions. It will also show up in the Android and iOS Google apps.
But wait, there’s more. Google is not the only one using AI to summarize articles. Microsoft has been doing it for a while with its Bing search engine and Edge browser, using a different approach called neural abstractive summarization. This method uses deep learning to create summaries that are more fluent and coherent than Google’s generative ones, but also more prone to errors and distortions.
So which one is better? Well, that depends on your preference and purpose. If you want a quick and concise overview of an article, Google’s SGE while browsing may be more suitable for you. But if you want a more natural and comprehensive summary, Microsoft’s neural abstractive summarization may be more appealing.
🕘 Fact Check Time!
How did Bing AI do on the article above? I asked Claude AI to fact check Bing’s take on this story. Here’s Claude’s report:
Claim: Google has a new feature called Search Generative Experience (SGE) while browsing that can summarize any article using AI.
Fact check: True. Google does have a feature called SGE while browsing that summarizes web articles using generative AI models. It is currently experimental and available in Chrome.
Claim: Microsoft has a competing feature using neural abstractive summarization in Bing and Edge.
Fact check: True. Microsoft does use neural abstractive summarization to summarize web content in Bing and Edge. This approach is different from Google’s generative method.
Claim: Google’s summaries are more concise while Microsoft’s are more fluent and comprehensive.
Fact check: Partly true. Early comparisons suggest Google’s summaries tend to be more concise while Microsoft’s contain more details. However, both methods have tradeoffs and the quality can vary. More research is needed.
Claim: The features only work for freely available web articles, not paywalled ones.
Fact check: True. Both Google and Microsoft’s summarization features rely on being able to access the full text of articles, so they do not work for paywalled or restricted content.
Overall, the main claims about the existence and basic capabilities of the new Google and Microsoft summarization features appear factual. However, the comparisons of quality and performance require more objective testing. The limitations around paywalled content are also accurately described.
Adobe Express Moves Out of Beta
By Bard for AI Insider
Adobe Express, the all-in-one design and video editing app, is officially out of beta. The app, which was first released in 2021, can be used to create social media graphics, presentations, videos, and more.
With the move out of beta, Adobe Express is adding new features and functionality, including:
A new AI powered text-to-image feature based on Adobe’s Firefly (which is still in beta). Express joins the popular design site Canva in integrating AI into its offerings.
A new animation feature that makes it easy to add motion to your designs.
A new collaboration feature that allows you to work on projects with others in real time.
Adobe Express is also making it easier to find and use premium fonts, graphics, and templates. The app now has a dedicated marketplace where you can browse and purchase content from a variety of creators.
Express offers a freemium plan and paid plans starting at $12.99/month.
Random Shorts
The Associated Press has issued guidelines on the use of AI. It’s a big deal because so many news orgs follow AP style and use/contribute to their wire service.
The New York Times has pulled out of an AI coalition of media companies seeking to negotiate use of their content for training AI models and may sue OpenAI, creator of ChatGPT.
Disney has taken the wraps off of its AI Task Force, which was apparently launched earlier this year. Their mission: Figure out how to deploy AI across Disney’s platforms and find ways to cultivate relationships with outside AI innovators.
China implemented sweeping new regulations for artificial intelligence services this week, trying to balance state control of the technology with enough support that its companies can become viable global competitors. The government issued 24 guidelines that require platform providers to register their services and conduct a security review before they’re brought to market. Seven agencies will take responsibility for oversight, including the Cyberspace Administration of China and the National Development and Reform Commission.
Fun and Useful Stuff to Try
Wondercraft: Text to Podcast
Wondercraft turns existing text content like blog posts into podcasts. Yes, really. Text to podcast. You can even clone your own voice to use as a presenter or use one that’s provided. Note: It’s not just reading back the text that you submit, it will write a script from your content and add music! Impressive. ▶️ Listen to this test that I did using an article from this newsletter. It took minutes to produce! Free and paid accounts.
Human Generator
Got some time to kill? Want to try out generating images of people, no prompts required? Check out the Human Generator. Seriously, it’s fun! You can even add your own face to your creation! Free and paid accounts.
Beard Style AI
Yep. It’s just what it sounds like. Dudes, Drag Kings, anyone, upload your face and it will generate different styles of facial hair. Like dozens of styles. For $9.99.
Aht Gallery
Midjourney Daily Theme Goes African
Many of the AI image generators invite users to generate images based on a daily theme prompt. Sometimes it’s a single word, sometimes a phrase. This week I was scrolling the feed on Midjourney when image after image of black people started showing up. I was flabbergasted because I’d never seen that before. Sure, a few appear here and there, but nothing like this. Then I saw that the daily theme was “adire.” A quick Google search revealed that this is a fabric with Yoruba origins. One of the coolest things about scrolling this? Seeing prompts written in Yoruba pop up. Look at you getting all African, daily theme! ( — Michelle Johnson)
