Analysis | Chatbots' flaws aren't stopping tech giants from putting them everywhere – The Washington Post
A newsletter briefing on the intersection of technology and politics.
A newsletter briefing on the intersection of technology and politics.
Happy Tuesday! Hope you didn’t invest in Truth Social. Send news tips to: will.oremus@washpost.com.
Below: Social media companies’ political ad-tracking tools are falling short. First:
When tech entrepreneur Vivek Wadhwa opened WhatsApp on Friday to talk with colleagues in India, the app popped up a message suggesting he try out its new AI chatbot, Meta AI. Wadhwa, a former Washington Post opinion contributor whose latest venture is a diagnostics start-up called Vionix Biosciences, thought he’d give it a spin.
Within minutes, the bot had infuriated him. Wadhwa first asked it who he was, then what he’s famous for, and finally what he’s “infamous” for, he told The Tech 202 by phone Monday. In response to the last question, the bot described a 2014 blog post accusing Wadhwa of plagiarism, complete with a byline, date and formal citation. As you may have guessed at this point, the blog post doesn’t exist.
By now, such “hallucinations” are a familiar problem for AI chatbots. My colleague Pranshu Verma and I reported last year on a similar episode in which OpenAI’s ChatGPT fabricated a sexual harassment scandal about law professor Jonathan Turley. Since ChatGPT’s release in late 2022, there have been countless reports of it and other “generative AI” tools from industry stalwarts such as Microsoft and Google spouting misinformation, showing bias and giving inappropriate responses.
But the risk of chatbots going awry doesn’t seem to be slowing tech giants’ march to integrate them into the world’s largest online platforms.
Meta has recently begun rolling out its Meta AI chatbot more broadly, including in WhatsApp, Instagram and Messenger across India and parts of Africa, bringing the tool to markets where it serves hundreds of millions of people. And on Monday, the New York Times reported that the company is approaching Instagram influencers in the United States on a project called “Creator AI” that would let an AI avatar trained on their voice interact with their fans on their behalf.
For a company that has prided itself on AI research but fallen behind rivals Microsoft, Google and OpenAI on commercialization of the latest language and image models, the moves represent an ambitious leap toward integrating AI into people’s everyday interactions. Meta envisions people asking its chatbots questions, using it to create images or just chatting with it, the company said when it announced Meta AI in September.
Given what we know about chatbots so far, it’s a leap of faith.
AI chatbots can dazzle with their ability to converse on a wide range of topics and generate images of whatever people dream up (with some limits). But what the industry calls “hallucinations” — responses that aren’t grounded in reality — have come to seem more like a feature than a bug. Lately, companies including Microsoft, Google, OpenAI and now Meta have tried to patch the problem by connecting AI models to the web and letting them search for information and cite their sources, with mixed results. Meanwhile, attempts to train chatbots to shed offensive biases picked up from their training data often fall short or go too far.
Meta isn’t the only company forging ahead with chatbot integrations despite the risks. Last year, its rival Snapchat launched My AI, billed as a friendly conversation partner, and my colleague Geoffrey A. Fowler quickly found it gave teens inappropriate advice. Google and Microsoft are increasingly embedding AI language tools into their search engines and productivity tools, even as Microsoft’s have been used to generate pro-Nazi memes and Google’s showed a penchant for overcorrecting to avoid racial stereotypes. Elon Musk’s X recently put its Grok AI tool to work writing trending headlines — several of which quickly turned out to be false. Even Amazon has released an AI chatbot for shopping, which my colleague Shira Ovide reviewed as “dumb as rocks.” (Amazon founder Jeff Bezos owns The Washington Post.)
At this point, Meta AI apparently fabricating a controversy about Wadhwa hardly comes as a surprise.
Like others before it, Meta’s response is that these tools are a work in progress.
“Meta AI and our suite of generative AI features help people on our apps be more creative and productive,” spokesman Kevin McAlister said in an emailed statement. “As we said when we launched these new features in September, this is new technology and it won’t always be perfect, which is the same for all generative AI systems. Since we launched, we’ve constantly released updates and improvements to our models and we’re continuing to work on making them better.”
Wadwha said he doesn’t buy that defense. “Already in India, WhatsApp is being used to influence the elections, being used to stir up hatred, misinformation,” he said. “Now imagine an AI chatbot that people trust, which produces very credible-sounding references. … This is dangerous.”
Despite the downsides, tech giants see AI chatbots as strategically important, said Matt Navarra, a U.K.-based social media consultant and industry analyst.
“What looks set to happen is that most of the platforms, whether it be TikTok or X, Instagram or Facebook, will all have their own versions of an AI system chatbot where you can engage with it to get answers to problems or find information or perform searches,” Navarra said. While people might find the features “gimmicky” for now, they give the tech companies another way to glean users’ personal data, keep them on the platform, and use those interactions to further train their AI systems for future applications, which they hope will someday be “more meaningful.”
In the meantime, he said, we can expect “many more headlines” about biases and inappropriate responses, as those problems aren’t likely to be solved anytime soon.
Correction: This story originally mischaracterized Vivek Wadhwa’s startup. It is a diagnostics startup, not a robotics startup.
The tools social media companies have developed to help people track political advertising on their platforms are all “plagued by missing data, bugs, shoddy features, and unacceptable shortfalls,” a new report has found. My colleague Gerrit De Vynck reports on the findings from Mozilla, the maker of the Firefox web browser, and CheckFirst, a Finnish internet transparency organization.
Over the past several years, companies like Meta, Snapchat, TikTok and Google have built “ad repositories” meant to allow researchers and journalists to see who is paying for political ads on those platforms. Regulation coming into effect later this year in the European Union also mandates companies be transparent about political ad funding.
But the investigation from Mozilla and FirstCheck found that ad transparency tools from Snapchat, X (formerly Twitter) and Microsoft Bing were all practically unusable. Tools from Google had “bare minimum data and functionality,” while those from Apple’s app store, LinkedIn, Meta and TikTok had “big gaps in data and functionality.” No platform was “ready for action,” the report said.
The tools have been an important data source for activists and researchers as billions of people head to the polls this year around the world, including in the United States and India.
“Our research shows most of the world’s largest platforms are not offering up functionally useful ad repositories,” said Claire Pershan, Mozilla’s E.U. advocacy lead. “The current batch of tools exist, yes — but in some cases, that’s about all that can be said about them.”
Battle lines drawn as US states take on big tech with online child safety bills (The Guardian)
X reverses course, tells Brazil court it will comply with rulings (Reuters)
Judge halts Texas probe into Media Matters’ reporting on X (Ars Technica)
Adobe explores OpenAI partnership as it adds AI video tools (Reuters)
Microsoft to invest $1.5bn in Abu Dhabi AI group G42 (Financial Times)
SpaceX could finally face competition. It may be too late. (By Christian Davenport)
Meta is temporarily shutting Threads down in Turkey on April 29th. (The Verge)
Broadcom questioned by EU over VMware licensing changes (Reuters)
Scammers are targeting teenage boys on social media — and driving some to suicide. (Bloomberg)
They criticized Israel. This Twitter account upended their lives. (By Pranshu Verma)
Best AI x-risk coverage so far. https://t.co/2fWKKma1WQ
That’s all for today — thank you so much for joining us! Make sure to tell others to subscribe to The Technology 202 here. Get in touch with Cristiano (via email or social media) and Will (via email or social media) for tips, feedback or greetings!
