What’s up, AI Insiders? News continues to emerge from the world of AI at a rapid pace. Controversy, too. Google has temporarily stopped its chatbot Gemini from generating images of people. Read on for more about this dustup in AI land. And, as always, if you find this content useful, leave something in the tip jar to support AI Insider
Google Temporarily Curbs Image Generator to Fix Diversity Issues
By Copilot and Michelle Johnson, AI Insider
Just in time for Black History Month, race moved up higher on the agenda for Google and its AI chatbot Gemini this week.
On Thursday, Google temporarily suspended Gemini’s ability to generate images of people due to inaccuracies in historical depictions. Users posted screenshots of the Founding Fathers, Popes, and other characters inaccurately portrayed as racially diverse.
They also complained that it was hard to get the bot to generate images of white people. Posts titled “Google is Racist!” and charges of “wokeness” proliferated. Elon Musk even weighed in, suggesting Google’s “racism” extends to their search engine.
What happened here? The speculation is that the bot was “overcorrecting.” If it has been trained to generate diverse characters when prompted for images of a firefighter, for instance, it could certainly do the same thing for a prompt requesting the Founding Fathers or the Pope.
As the complaints escalated, Google issued a statement saying that they’d “missed the mark” and would work to correct the issue.
This TechCrunch piece does a good job of summarizing the pitfalls surrounding AI training and race, and making the case for why it’s important for AI vendors to be transparent. The authors’ advice: “Were AI vendors to address their models’ shortcomings head on, in humble and transparent language, it’d go a lot further than haphazard attempts at “fixing” what’s essentially unfixable bias. We all have bias, the truth is — and we don’t treat people the same as a result. Nor do the models we’re building. And we’d do well to acknowledge that.”
Testing Image Generators for Bias
By Michelle Johnson, AI Insider
If you’re a regular reader of AI Insider, you may have noticed that many of the AI-generated images here feature Black people. Some were requested in the prompt (Generate an African American woman working on her laptop with a cup of coffee on her desk). Others just showed up from a more general prompt. In either case, I intentionally feature people of color in the newsletter because, frankly, this is my space, and I can.
I laughed out loud when I saw tech bros on Reddit complaining that an image generator wouldn’t produce white people. (Welcome to our world, said people of color.) They were soooo angry. But I digress.
As this TechCrunch piece points out, “the data sets used to train image generators generally contain more white people than Black people, and yes, the images of Black people in those data sets reinforce negative stereotypes. That’s why image generators sexualize certain women of color, depict white men in positions of authority, and generally favor wealthy Western perspectives.”
Combating negative stereotypes is one reason I ask bots to generate something different. The other reason is that I simply want to see people who look like me engaged with technology. What happened with Google’s Gemini aside, I do appreciate companies that try to build in some diversity, even if it’s not always perfect.
A couple of months back, I “name-checked” an AI image generator to see what it would turn out when a prompt included clearly ethnic names. The results were mostly on point. Jamal and Imani (African American) and Diego and Rosa (Latino) popped right up.
As Google scrambles to tweak its image generator, it might be a good time to look at how the other image generators handle ethnicity requests. A major complaint about Gemini is that it wouldn’t generate white people when specifically asked. Here’s a look at how other bots responded.
Tyler Perry: Hold My Upgrade, AI
By Copilot for AI Insider
Tyler Perry, the actor, filmmaker, and studio owner, has put his $800 million studio expansion on hold after seeing the capabilities of OpenAI’s text-to-video model Sora, which can generate realistic and cinematic video outputs from text descriptions.
Perry, who owns a 330-acre studio in Atlanta, was planning to add 12 more sound stages to the property, but he changed his mind after witnessing how Sora could create scenes and locations without the need for sets, travel, or even actors.
“I no longer would have to travel to locations. If I wanted to be in the snow in Colorado, it’s text. If I wanted to write a scene on the moon, it’s text, and this AI can generate it like nothing,” he said in an interview with The Hollywood Reporter.
Perry is not only concerned about his own business, but also the future of the entertainment industry and the livelihoods of the people who work in it. He said Sora and other AI technologies will affect every corner of the industry, from actors and directors to grip and electric and sound and editors.
Perry called for a collective and proactive approach from the industry to confront the rapid development of AI and to protect the workers and the art form. He also urged the government to step in and regulate technology to ensure the safety and well-being of humanity.
Chrome Web Browser Gets New AI Features
By ChatGPT, for AI Insider
Google rolled out an upgrade of the Chrome web browser this week. Dubbed version M122, it features a suite of AI-powered tools.
Here’s a breakdown of what’s new:
Tab Organizer: This feature automatically groups your open tabs into neatly organized bundles based on their content, aiming to make it easier to manage and navigate through them.
How to Turn On: Simply right-click on any tab and select “Organize Similar Tabs” or use the dropdown arrow next to your tabs to let Chrome suggest groups for you.
Custom AI Themes: Chrome M122 allows you to personalize your browser with themes generated by AI, based on your chosen subject, mood, or color preferences.
How to Turn On: Go to the “Customize Chrome” side panel, click on “Change theme,” and then select “Create with AI.” Follow the prompts to create a theme that matches your style.
Writing Assistance: Need help drafting an email or a post? Chrome’s writing assistant is at your service, offering AI-generated suggestions to kickstart your writing.
How to Turn On: Right-click in any text field within Chrome and select “Help me write.”
Chrome M122 brings improved performance controls like detailed insights into memory usage per tab, and an expanded Safety Check feature that provides proactive alerts for compromised passwords, harmful extensions, and more. These features are built-in so they don’t require manual activation.
[Ed note: Not sure which version of the Chrome browser you’re using? Or need to upgrade? Click the three-dot menu in the upper right of the browser window, select “Settings,” then “About Chrome.” ]
Adobe Adds an ‘AI Assistant’ to Acrobat
By Copilot, for AI Insider
If you are a user of Adobe Reader or Acrobat, you may soon see a new feature: an AI assistant that can help you understand and manage your documents, including PDFs, Word, Powerpoint, or even meeting transcripts.
The AI assistant, which is currently in beta, is designed to help you comprehend lengthy documents, generate summaries, answer questions, and more. “Deeply integrated into Reader and Acrobat workflows, AI Assistant instantly generates summaries and insights from long documents, answers questions, and formats information for sharing in emails, reports, and presentations,” Adobe said in a press release.
The ability to ask for a summary or query a chatbot about documents isn’t new. Bots such as ChatGPT (via plugins) or Anthropic’s Claude accept document uploads.
One of the advantages Adobe is touting about its AI assistant is that it is seamlessly integrated into the Adobe applications, so there’s no need to upload documents to a third party.
Adobe also points out that your data is more secure and not used without your consent for training the AI assistant. However, the assistant is limited to answering questions about the content in the PDF. It doesn’t access the internet in search of related information.
Down the road, Adobe is looking at adding the ability to generate and edit text by changing the voice or tone, and integration with the Adobe Firefly image generator.
The AI assistant is currently available for Acrobat Individual, Pro, and Teams customers, with plans to expand to Reader users soon.
If you want to try Adobe’s AI assistant now, how do you get access? According to the press release: “No complicated implementations are required. Simply open Reader or Acrobat and start working with the new capabilities.”
For those who wait, an add-on subscription plan will be available when AI Assistant is out of beta.
What the Actual Hell?
Magician Says He Created Fake Biden Audio
Well, there’s some stuff that you just can’t make up. A magician claims a campaign consultant hired him to create the audio used in the fake Biden robocall that targeted New Hampshire primary voters.
That’s right. Paul Carpenter, a self-proclaimed magician and digital nomad, claims to be behind the AI-generated voice that impersonated Biden. And, he says, he was hired by a Democratic consultant, Steve Kramer, who worked for Dean Phillips’ presidential campaign, to use AI to mimic Biden’s voice for a legitimate purpose. He told the Associated Press that he had no idea that the audio he created would be used for a malicious robocall that violated the state’s voter suppression laws.
- Reuters: Nvidia briefly hits $2 trillion valuation as AI frenzy grips Wall Street
- And speaking of $$$, social media platform Reddit will be raking in a cool $60 million a year for letting Google use its content to train its AI models. Users are probably thrilled to have their yammering sold this way. Just. Thrilled.
- The Verge: ChatGPT spat out gibberish for many users overnight before OpenAI fixed it
- Techcrunch: Are You Blacker Than ChatGPT? Quiz game aims to spotlight AI bias. Skip right to the quiz.
- GPT-4 Turbo’s knowledge base now goes up to December 2023. Still using the freebie version? It’s only current up to January 2022, as it likely reminds you from time to time.
- TechCrunch is launching a series of interviews with women in AI. Meet Mutale Nkonde, founding CEO of the nonprofit AI for the People (AFP), which seeks to address bias in AI. And this is Miranda Bogen of the Center of Democracy and Technology’s AI Governance Lab. Check out more interviews with women in AI.
Aht Gallery
Did you have a stressful week? Here’s a little something fun you can share with the little monsters in your life. Prompt: A small, friendly furry monster, Pixar style.
