One of the advantages of generative AI tech is its natural language capabilities. That means you don’t need to be a programmer, engineer or scientist to “talk” to a gen AI chatbot and prompt it to create text, illustrations and other images, video, audio, photographs, and even programming code in seconds.
But the “magic” here has a dark side, including biases, hallucinations and other problems with how the tools themselves work. There’s also a growing problem with people leaning into these easy-to-use and powerful gen AI engines to create fake photos or deepfake videos with an eye toward misleading, confusing or just flat-out lying to an intended audience.
This week, we have examples of both.
First up: Just a week after Google’s Gemini text-to-image generator had to hit the pause button because it was delivering offensive, embarrassing and biased images – Google CEO Sundar Pichai sent the tool back to testing after saying the results were “completely unacceptable” – Microsoft is now reckoning with issues in its Copilot Designer AI generator. That reckoning comes after a company engineer wrote to the Federal Trade Commission expressing concerns about disturbing and violent images created by the tool.
Microsoft engineer Shane Jones said he was “actively testing the product for vulnerabilities, a practice known as red-teaming,” CNBC reported. The product, originally called Bing Image Creator, is powered by OpenAI’s technology. (OpenAI is the maker of ChatGPT and text-to-image converter Dall-E.) Jones said the AI service produced images of “demons and monsters alongside terminology related to abortion rights, teenagers with assault rifles, sexualized images of women in violent tableaus, and underage drinking and drug use.”
All of those images, CNBC added after re-creating Jones’ tests, run “far afoul of Microsoft’s oft-cited responsible AI principles.” Jones said Microsoft ignored his findings despite repeated efforts to get the company to address the issues.
“Internally the company is well aware of systemic issues where the product is creating harmful images that could be offensive and inappropriate for consumers,” Jones states in the FTC letter, which he also sent to Microsoft’s board and published on LinkedIn, according to The Guardian.
Microsoft told CNBC and The Guardian that it’s “committed to addressing any and all concerns employees have in accordance with our company policies and appreciate employee efforts in studying and testing our latest technology to further enhance its safety.”
The second example has to do with people creating fake images with AI. This week’s entry has to do with supporters of ex-President Donald Trump, who created photos that depict the now presidential candidate surrounded by fake Black voters as part of misinformation campaigns to “encourage African Americans to vote Republican,” The BBC reported after investigating the sources of the fabricated images.
One of the creators of the fake photos, a Florida radio show host named Mark Kaye, told the BBC that “he’s not out there taking pictures of what’s really happening” and that if US voters are swayed into thinking that they should vote for Trump based on photos that depict him with supposed Black voters, it’s up to them if they’re fooled.
“If anybody’s voting one way or another because of one photo they see on a Facebook page, that’s a problem with that person, not with the post itself,” Kaye told The BBC.
The images have started to appear as Trump “seeks to win over Black voters who polls show remain loyal to President Joe Biden, the Los Angeles Times reported. “The fabricated images … provide further evidence to support warnings that the use of AI-generated imagery will only increase as the November general election approaches.”
Sound far-fetched? The Center for Countering Digital Hate issued a report called “Fake Image Factories” that found popular AI image tools “create election disinformation in 41% of cases, including images that could support false claims about candidates or election fraud.” The 29-page report has many examples, if you’re one of those people who cares about truth and/or doing your own research.
Here are the other doings in AI worth your attention.
Biden asks Congress to ban AI voice impersonations, but…
In his State of the Union address, President Biden asked Congress to “ban voice impersonation using AI” and enact legislation that allows us to “harness the promise of AI and protect us from its peril,” the Hill reported.
Biden’s call comes after scammers created AI-generated robocalls that copied his voice and encouraged Democratic voters not to cast a ballot in the New Hampshire presidential primary. That led the Federal Communications Commission in February to ban robocalls using AI-generated voices.
The New Hampshire example definitely shows the dangers of AI-generated voice impersonations. But do we have to ban them all? There are potential use cases that aren’t that bad, like the Calm app having an AI-generated version of Jimmy Stewart narrate a bedtime story.
In December, Stewart, the beloved actor who died in 1977 after giving legendary performances in Rear Window, It’s a Wonderful Life and Harvey, was sort of brought back to life to read a piece called It’s a Wonderful Sleep Story for listeners of the meditation app, Variety reported. (It also shared a clip if you want to hear it.)
Of course, what makes this different from the NH example is that Stewart’s family and his estate gave their permission for the voice clone. And Calm clearly labeled the 45-minute story as being brought to listeners by “the wonders of technology.”
Anthropic gives Claude a boost to take on ChatGPT, Gemini
Anthropic, the San Francisco-based AI rival to OpenAI’s ChatGPT and Google’s Gemini, released Claude 3 and boasted that its family of AI models now exhibit ‘human-like understanding.”
Saying it was a “bold though not entirely unprecedented assertion by a maker of gen AI chatbots,” CNET’s Lisa Lacy summarizes the new features, noting that “the Claude 3 family can handle more complicated queries with higher accuracy and enhanced contextual understanding.” She also reminds us that a gen AI chatbot isn’t an artificial general intelligence (AGI) and that Claude, like its competitors, doesn’t truly understand the meaning of words as we humans do.
Still, there’s enthusiasm for Claude’s update, which Anthropic said is also better at analysis and forecasting; content creation; code generation; and conversing in languages like Spanish, Japanese and French. Claude 3 Opus is available to subscribers for $20 per month. Claude 3 Sonnet, a less powerful version, is free.
Anthropic’s investments highlight the rapid pace of updates in the gen AI space, noted The New York Times. Google recently released Gemini (formerly called Bard) and OpenAI just updated ChatGPT with a new feature called Read Aloud that can read responses from your prompts to you in 37 languages using five different voice options.
But the paper also noted that the leading AI companies have been “distracted by one controversy after another. They say the computer chips needed to build AI are in short supply. And they face countless lawsuits over the way they gather digital data, another ingredient essential to the creation of AI. (The New York Times has sued Microsoft and OpenAI over use of copyrighted work.)”
Not to mention that the models aren’t exactly delivering stellar results all the time. That just reinforces what I said above: All this AI “magic” has its dark sides.
OpenAI, being sued by Elon Musk, promises to act “responsibly”
OpenAI signed on to an open letter saying that it will work with rivals including Meta, Google, Salesforce, ElevenLabs, Microsoft and Mistral and with other tech companies and startups to “build, broadly deploy, and use AI to improve people’s lives and unlock a better future.”
The March 4 letter, written by Silicon Valley investor Ron Conway and his firm SV Angel, was posted on X days after X CEO Elon Musk, an early investor in OpenAI who’s now working on a rival chatbot, sued OpenAI and its CEO, Sam Altman. Musk argued that they were putting profit above the future of humanity and thereby violating the founding principles of the company. (The lawsuit, quite the read, can be found here. OpenAI’s response is here.)
Altman, in an X post responding to Conway, said he was “excited for the spirit of this letter and ron’s leadership in rallying the industry!”
Altman also got the endorsement of Silicon Valley entrepreneur Reid Hoffman, who worked with Musk as part of the PayPay Mafia. In a video posted on X, Hoffman said that while Musk cares “very deeply about humanity,” he endorses Altman’s “collaborative approach to AI” as the one that will get the “best outcomes for humanity.”
“Musk is a solo-entrepreneur,” Hoffman said. “His gut tends to be, AI is only gonna be safe if I make it … I am the person that can make it happen versus we should bring in a collaborative group. And I’m more with the collaborative group.”
Meanwhile, OpenAI announced that Altman rejoined the board of directors (after he was temporarily ousted as CEO by the prior board in November) and said it added three new directors on March 8: Sue Desmond-Hellmann, former CEO of the Bill and Melinda Gates Foundation; Nicole Seligman, former general counsel at Sony; and Fidji Simo, CEO and chair of Instacart.
Stay tuned. This soap opera is far from over.
AI researchers call for more transparency in evaluating LLMs
Speaking of open letters, a group of over 100 AI researchers have signed on to one arguing that gen AI companies need to open up their systems to investigators as part of required safety checks before their tools are released to hundreds of millions of people.
“The researchers say strict protocols designed to keep bad actors from abusing AI systems are instead having a chilling effect on independent research,” The Washington Post reports. “Such auditors fear having their accounts banned or being sued if they try to safety-test AI models without a company’s blessing.”
The open letter, titled A Safe Harbor for Independent AI Evaluation, focuses on three things. First, the researchers argue that “independent evaluation is necessary for public awareness, transparency and accountability.” Second, they claim that AI companies’ current policies “chill” independent evaluation. And third, they say AI companies “should avoid repeating the mistakes of social media platforms, many of which have effectively banned types of research aimed at holding them accountable.”
What are they talking about?
“The effort lands as AI companies are growing aggressive at shutting outside auditors out of their systems,” the Post adds. “OpenAI claimed in recent court documents that New York Times’s efforts to find potential copyright violations was hacking its ChatGPT chatbot. Meta’s new terms says it will revoke the license to LLaMA 2, its latest large language model, if a user alleges the system infringes on intellectual property rights. Movie studio artist Reid Southen, another signatory, had multiple accounts banned while testing whether the image generator Midjourney could be used to create copyrighted images of movie characters. After he highlighted his findings, the company amended threatening language in its terms of service.”
As a reminder, President Biden’s October 2023 executive order on AI has called for AI companies to put in place safeguards, including testing and other evaluations, to show that their systems are safe before releasing them to the public. But as we’ve seen, deciding the best way to achieve that goal remains contentious.
Microsoft, OpenAI aim to dismiss parts of NYT copyright suit
One of the ongoing issues around gen AI is what data is being used to train the LLMs powering popular chatbots like ChatGPT, Microsoft Bing and Google Gemini. That’s why copyright holders have been filing suits against these for-profit companies, saying their works – books, stories and other content – have been co-opted without the authors’ permission or compensation.
One of the most notable suits was filed by The New York Times in December against OpenAI and Microsoft (which uses OpenAI’s tech for Bing). The newspaper claims that the duo have essentially stolen millions of its stories to train its LLM and that the chatbots now compete with media companies as a source of reliable information. The tech companies so far have argued that they are able to scrape content under the fair use doctrine because they’re not reproducing the entire copyrighted material.
Now both Microsoft and OpenAI have spelled out their arguments against the paper. On March 4, Microsoft filed a 31-page motion in the district court in New York, arguing that LLMs are like videocassette recorders and that chatbots don’t undercut the market for news articles and other materials they were trained on. That motion came a week after OpenAI asked the New York court to also dismiss parts of the NYT’s lawsuit, saying in its 35-page motion on Feb. 26 that ChatGPT is “not in any way a substitute for a subscription to The New York Times.”
This is an important issue for anyone creating content – what content (words, pictures, audio, code) can be used to train these powerful AI systems – and for anyone building an AI engine, since training data is the lifeblood of gen AI.
We’ll see what the courts decide.
Editors’ note: CNET is using an AI engine to help create some stories. For more, see this post.
