Hiltzik: CNET’s chatbot stunt shows the limits of artificial intelligence

We’ve all been trained by decades of science fiction to think of AI as a threat to our future of work. The idea is: if an AI robot can do a job like a human — cheaper and less opposition between people — who needs a human?

CNET technology news site has attempted to answer this question calmly and discreetly. For months, the site used an artificial intelligence engine to write articles for CNET Money’s personal finance page. Articles covered topics such as “What is compound interest?” and “What happens when you bounce a check?”

At first glance and for beginners in the financial field, the articles seemed convincing and informative. CNET continued this practice until early this month, when it was brought up by the website Futurism.

Close examination of the work produced by CNET’s artificial intelligence makes it seem less like a sophisticated script generator and more like a plagiarism machine, casually pumping out plagiarized works.

– John Christian, Futurism

But as Futurism has identified, bot-written articles have major limitations. For one thing, many of them are riddled with bugs. On the other hand, a lot of plagiarism spreads – in some cases from CNET itself or its sister websites.

John Christian of Futurism put the error problem outright in an article in which he said the problem with CNET’s article-writing AI is that it’s “kind of idiot.” Christian followed up with an article revealing numerous cases ranging “from verbatim copying to mild alterations to significant paraphrases, all without properly crediting the original”.

This level of misconduct will result in the expulsion of a human student or journalism class.

We’ve written before about the unrecognized frontiers of new technologies, especially those that seem almost magical, like applications of artificial intelligence.

To quote Rodney Brooks, a roboticist, entrepreneur, and robotics entrepreneur I wrote about last week, “There is a real social media cottage industry on two sides; one gushes over the outstanding performance of these systems, perhaps chosen, and the other shows how incompetent they are at very simple things, Again. The problem is that you as a user don’t know in advance what you’re going to get.”

This brings us back to CNET’s article writing bot. CNET didn’t specify what specific AI app it was using, though the timing suggests it’s not ChatGPT, the AI ​​language generator that has created buzz among techs and concerns among educators for its apparent ability to produce written work that can be hard to distinguish as non- mankind.

CNET didn’t make the AI ​​contribution to its articles particularly clear, appending only a small print line, “This article was assisted by an AI engine and reviewed, verified, and edited by the editorial team.” Over 70 articles have been attributed to the “CNET Money Staff”. Since Futurism was revealed, the subtitle has been changed to simply “CNET Money”.

Last week, according to the Verge, CNET executives told employees that the site would temporarily stop publishing AI-generated material for the time being.

As Christian Futurism has established, errors in bot articles have ranged from basic misdefinitions of financial terms to gratuitous oversimplifications. In an article about compound interest, the CNET bot originally wrote, “If you deposit $10,000 into a savings account that earns 3% compound interest annually, you’ll earn $10,300 at the end of the first year.”

This is wrong – the annual earnings will be only $300. The article has since been corrected to read “You will earn $300, which added to the base amount will bring you $10,300 at the end of the first year.”

The bot also initially described the interest payment on a $25,000 car loan at 4% as “a flat $1,000…a year.” It’s auto loan payments, like mortgages, that get fixed — interest is charged only on outstanding balances, which shrinks as payments are made. Even on a one-year auto loan at 4%, the interest would only come to $937. For long-term loans, the total interest paid decreases each year.

CNET corrected that as well, along with five other errors in the same article. Put it all together, and the website’s assertion that its AI bot is “fact-checked and edited by our editorial team” starts to sound a little weak.

Bot impersonation is much more impressive and provides an important clue as to how the program works. Christian found that the bot appeared to have copied text from sources including Forbes, Balance, and Investopedia, which all operate the same field of personal financial advisory as CNET Money.

In those cases, the bot used similar masking techniques as the human impersonators, such as simple paraphrases and word swaps. In at least one case, the bot was stolen from Bankrate, a sister publication to CNET.

None of this is particularly surprising since one of the keys to the job of language bots is their access to a huge volume of human-generated prose and poetry. They may be good at finding patterns in the source material that they can replicate, but at this stage of AI development they are still co-opting human brains.

The remarkable coherence and argumentative strength of the outputs of these programs, including ChatGPT, seems to have more to do with their ability to select from human-generated raw materials than with any ability to develop and articulate new concepts.

In fact, Christian wrote, “A close examination of the work produced by CNET’s artificial intelligence makes it look less like a complex text generator, and more like a plagiarism machine, casually pumping out plagiarized works.”

It’s hard to say where we stand on the continuum between robot-generated inconsistency and true creative expression. The most complex language bot at the time, known as GPT-3, had obvious limitations, Jeff Shaten, a professor at the University of Washington and Lee University, wrote in a September article.

“She gets bogged down by complex writing assignments,” he wrote. “She can’t craft a novel or even a decent short story. Her attempts at science writing…are funny. But how long before the ability existed? Six months ago, GPT-3 suffered from rudimentary queries, and today it can write a sensible blog post discussing ‘the ways in which An employee can get a promotion from a reluctant boss.”

It is possible that those who need to judge written work, such as teachers, may find it more difficult to distinguish between AI-generated material and human output. A professor recently reported catching a student submitting a paper written by a bot the old-fashioned way – it was very good.

Over time, the confusion about whether something is human or robotic may depend not on the abilities of the robot, but on the abilities of the humans responsible.

Related Posts