I’ll admit that the title of their June 8 research paper — ChatGPT Is Bullshit — in the journal Ethics and Information Technology, is what got me to download and read the 10-page report by three lecturers at the University of Glasgow: Joe Slater, James Humphries and Michael Townsen Hicks.
And though I thought they were kidding at first — there’s a section on bullshit distinctions (general bullshit, hard bullshit and soft bullshit) — the more I read, the more I agreed with their pitch that the term “AI hallucinations” to describe the falsehoods AI chatbots produce is an inaccurate way to describe the output, because of how it may mislead us in how we think about generative AI today and going forward.
They start by making a notable point. “Because these programs cannot themselves be concerned with truth, and because they are designed to produce text that looks truth-apt without any actual concern for truth, it seems appropriate to call their outputs bullshit. We think this is worth paying attention to.”
The trio explained why they wrote their paper in a July 17 essay in Scientific American (which includes an amazing reference to “Shakespeare’s paradigmatic hallucination in which Macbeth sees a dagger floating toward him.”)
Here’s their pitch: “Among philosophers, ‘bullshit’ has a specialist meaning, one popularized by the late American philosopher Harry Frankfurt. When someone bullshits, they’re not telling the truth, but they’re also not really lying. What characterizes the bullshitter, Frankfurt said, is that they just don’t care whether what they say is true. ChatGPT and its peers cannot care, and they are instead, in a technical sense, bullshit machines.”
The word we use to describe the output from these AI systems matters, they argue, for three reasons.
First, terminology “affects public understanding of technology. … If we use misleading terms, people are more likely to misconstrue how the technology works.”
Second, “How we describe technology affects our relationship with that technology and how we think about it.” Take, for instance, people who’ve been lulled into a false sense of security “by ‘self-driving’ cars. We worry that talking of AI ‘hallucinating’ — a term usually used for human psychology — risks anthropomorphizing the chatbots.” That is, AI users attributing human features to computer programs.
That leads us to the third reason why it matters, and which I think is the most important point. “If we attribute agency to the programs, this may shift blame away from those using ChatGPT, or its programmers, when things go wrong. … It is crucial that we know who is responsible when things go wrong.”
Overall, the researchers conclude that characterizing chatbot “inaccuracies” as bullshit rather than hallucinations might bring some much needed perspective to the hype and drama surrounding gen AI today.
“Calling chatbot inaccuracies ‘hallucinations’ feeds in to overblown hype about their abilities among technology cheerleaders, and could lead to unnecessary consternation among the general public. It also suggests solutions to the inaccuracy problems which might not work, and could lead to misguided efforts at AI alignment amongst specialists,” they write in the closing of their paper.
“It can also lead to the wrong attitude towards the machine when it gets things right: the inaccuracies show that it is bullshitting, even when it’s right. Calling these inaccuracies ‘bullshit’ rather than ‘hallucinations’ isn’t just more accurate (as we’ve argued); it’s good science and technology communication in an area that sorely needs it.”
Like I said, it’s worth giving the paper, and their essay, a read.
Here are the other doings in AI worth your attention.
Apple, Anthropic use YouTube transcripts without permission
The lifeblood of any generative AI system is data, or more specifically training data — the billions of bits of information fed into the AI’s engine, its large language model. The LLM learns to make predictions by finding patterns in all that data, so the AI tool can respond to your requests, generating answers in the form of text, video, photos, images, audio and more.
But the problem, as copyright owners including The New York Times have called out in lawsuits, is that AI companies are gathering that training data by slurping up everything on the internet or buying online content, often without the owner’s knowledge or permission. A few OpenAI content deals with publishers aside, the AI companies aren’t compensating most of those owners either.
So while it’s unsettling to hear that millions of hours of YouTube video transcripts have reportedly also been scooped up without the knowledge or permission of content creators, it’s not surprising. The NYT, in an April investigation, said OpenAI researchers had reportedly created a speech recognition tool called Whisper that the company used to transcribe audio from YouTube videos to feed into its LLM. Last week, an investigation by Proof News found that big tech companies including Apple, Anthropic and Nvidia, are also training their AI systems with YouTube transcripts without permission from the content creators.
The report includes a search tool that shows whether a YouTube channel is in the so-called Pile dataset compiled by a nonprofit called EleutherAI. Proof News found that “subtitles from 173,536 YouTube videos, siphoned from more than 48,000 channels, were used by Silicon Valley heavyweights,” including Anthropic, Nvidia and Apple. Some of the YouTube channels included in the dataset are those of late-night shows such as The Late Show With Stephen Colbert and Jimmy Kimmel Live. Content from popular YouTube personalities like MrBeast, tech reviewer Marques Brownlee and PewDiePie also wound up in the dataset, CNET’s Omar Gallaga reported.
“It’s theft,” Dave Wiskus, the CEO of Nebula, a streaming service partially owned by its creators, told Proof News.
YouTube CEO Neal Mohan told Bloomberg in April that he doesn’t know if OpenAI used YouTube videos to train its text-to-video generator, but that if it did, it’s a violation of the platform’s terms of service. But in April, YouTube-owner Google told the NYT that its agreement with content creators allows Google to use YouTube content to train its AI systems.
Apple told CNET that it respects the rights of creators and publishers, and that websites can opt out of potentially being used to train Apple Intelligence, its new initiative around gen AI. Nvidia declined to comment. EleutherAI didn’t respond to CNET’s request for comment. Neither did Anthropic, but it told Proof News that though it uses Pile to train its AI system, Claude, the dataset only “includes a small subset of YouTube subtitles.”
The assumption, at this point, is that most of the world’s content has been scraped off the internet and has become part of the datasets being used by AI companies to train their systems. There are open questions about copyright in the age of AI that still need to be addressed by the courts, including whether buying or using a third-party dataset somehow gives AI companies a pass on copyright concerns. The US Copyright Office said it plans to publish guidance on AI-related issues sometime this year.
As with all things AI, this is a developing story.
Regulations scare off Meta, Apple from releasing AI in the EU
Meta won’t make its Llama AI model available in the European Union because the EU’s privacy and AI regulations have created an “unpredictable … European regulatory environment,” the company said in a statement to CNET, confirming an Axios report on Meta’s decision.
Meta’s new “multimodal” model works with a variety of data types — text, audio, video and images — across a variety of devices, from smartphones and computers to its Ray-Ban Meta smart glasses. The company’s concerns aren’t with the EU’s new AI Act, but rather with its privacy-focused data-protection law known as the General Data Protection Regulation, or GDPR, Axios noted. Meta said it will use public posts from Facebook and Instagram to train its AI models, but that training may run afoul of GDPR restrictions.
Meta isn’t the only company to say it won’t release AI products in Europe. Apple, which said it’ll bring its Apple Intelligence system to popular products including the iPhone later this year, also said in June that it won’t release those features in the EU because of “regulatory uncertainties” raised by the Digital Markets Act, an EU antitrust law meant to make markets more fair, CNBC reported at the time. Passed by regulators in March, “the DMA requires that services be interoperable across platforms, to promote competition and stymie the ‘gatekeeper’ effect that some large companies … have,” CNBC noted.
The EU labeled six big tech companies as gatekeepers: Alphabet (Google’s parent company), Amazon, Apple, ByteDance (owner of TikTok, which lost a challenge to the law last week), Meta and Microsoft.
And fyi, US regulators have also been looking at ways to put safeguards around AI development, with limited regulation so far. (The Federal Communications Commission banned AI-generated deepfake robocalls after fraudsters used a fake version of President Joe Biden’s voice to urge New Hampshire voters not to participate in their state’s primary.)
Silicon Valley investors tout Trump in pursuit of less AI regulation
In somewhat related news, former President Donald Trump’s promises of industry-friendly policies on AI and tech like cryptocurrencies have led some Silicon Valley venture capitalists, investors and executives to endorse a candidate and a party they previously opposed, The Washington Post reported.
A list of 17 monied men in tech, some of whom have invested billions of dollars in AI and crypto companies, are among those criticizing the Biden administration for policies they believe have “stymied” their investments, the Post wrote. “Some leaders say they are making a calculated bet that Trump will benefit their companies and investments” and have been “actively lobbying against Biden’s more aggressive approach to tech regulation.”
That includes, the Post added, wanting to repeal Biden’s AI executive order, which aims to put safety guardrails in place around the development of AI systems, especially as it pertains to national and economic security, public health and safety, and critical infrastructure.
“Conservative voices in San Francisco’s tech sector have grown increasingly strident in their support of a Trump-Vance ticket,” The Los Angeles Times reported. “Many of those tech investors celebrated the appointment of Ohio Sen. J.D. Vance — a venture capitalist who built his career in Silicon Valley — as Trump’s vice presidential nominee out of a shared belief that he would help remove regulations they believe could stifle innovation in artificial intelligence and cryptocurrency.”
Which just goes to show there’s a reason that the mid-19th century proverbial saying “politics makes strange bedfellows” has lasted as long as it has.
AI chatbots fuel misinformation, DOJ shuts down Russian bots
When it comes to fact-based news, the 10 most popular AI chatbots failed to provide “accurate information” nearly 57% of the time after the assassination attempt on the former president last week and fell “far short in dealing with the wave of conspiracy theories quickly launched by critics and supporters of Trump, as well as by hostile foreign state actors,” according to an AI misinformation tracker run by NewsGuard.
Among the chatbots audited were Meta AI, OpenAI’s ChatGPT, xAI’s Grok, Mistral’s le Chat, Microsoft’s Copilot, Anthropic’s Claude, Google’s Gemini and Perplexity’s answer engine, NewsGuard said. “Collectively, the 10 chatbots failed to provide accurate information 56.67% of the time — either because the AI models repeated the falsehood (11.11%) or declined to provide any information on the topic (45.56%). On average, the chatbots offered a debunk 43.33% of the time.”
Some of the chatbots actually don’t provide the latest information, so consider this a reminder to avoid using them as a source for breaking news. If you’re in doubt about whether something you’re reading online or on social media is based on fact, there are several reputable online fact-checking sources in addition to NewsGuard. (I’ve compiled a list here. The News Literacy Project also has a breaking news checklist you should check out.)
Meanwhile, the US Department of Justice said it disrupted a Russian propaganda campaign that spread online disinformation with help from AI tools, according to the Associated Press. The DOJ said it seized two domain names and searched 968 accounts on the X social media platform.
“US officials described the internet operation as part of an ongoing effort to sow discord in the US through the creation of fictitious social media profiles that purport to belong to authentic Americans but are actually designed to advance the aims of the Russian government, including by spreading disinformation about its war with Ukraine,” the AP reported.
The disinformation campaign was organized in 2022 by an editor at RT, a Russian-state-funded media organization. The RT editor helped “develop technology for a so-called social media bot farm” that “promoted disinformation on social media through a network of fake accounts,” US officials said, according to the AP. The technology, called Meilorator and found on X, also spread disinformation to other countries, including Poland, Germany, the Netherlands, Spain, Ukraine and Israel, government officials said.
Among the fake posts “was a video that was posted by a purported Minneapolis, Minnesota, resident that showed Russian President Vladimir Putin saying that areas of Ukraine, Poland and Lithuania were ‘gifts’ to those countries from liberating Russian forces during World War II,” the AP added.
Russia has spread disinformation in the US before to “sway the opinions of unsuspecting voters,” the AP added, noting that “during the 2016 presidential campaign Russians launched a huge but hidden social media trolling campaign aimed in part at helping Republican Donald Trump defeat Democrat Hillary Clinton.”
“We support all civic engagement, civil dialogue, and a robust exchange of ideas,” US Attorney Gary Restaino for the District of Arizona said in the DOJ statement. “But those ideas should be generated by Americans, for Americans. The disruption announced today protects us from those who use unlawful means to seek to mislead our citizens.”
How much do you know about AI? Check your skill set
For people looking to be part of the AI-enhanced workplace of the future, SHL, a talent acquisition and management platform, said it’s identified the top 10 skills employers are looking for. They include understanding a company’s strategic vision; thinking broadly; being able to motivate and empower others; keeping tabs on what competitors are up to; and learning quickly. You can find the complete list here.
If you’re more interested in testing your AI knowledge (and are OK with giving away your email and some personal information), there are at least two AI skill checkers I’ve come across in the past week. Degreed, which makes employee learning platforms, offers an AI skill review. And Workera, which offers an AI platform meant to help employers assess worker skills, said you can measure your skills in more than 10 of the most sought-after areas of AI. You’ll find its tool for that here.
From personal experience, I can tell you these assessments take about 15-20 minutes. As someone who communicates about AI, the most interesting takeaways for me were the questions on things I’d never even thought about.
Have fun.
