Hiltzik: CNET’s chatbot stunt shows the limits of AI

We’ve all been trained by decades of science fiction to see artificial intelligence as a threat to our future work. The idea is: if an AI robot can do a job just as well as a human—cheaper and with less interpersonal stubbornness—then who needs the human?

Technology news site CNET tried to answer that question quietly, even covertly. For months, the site used an AI engine to write articles for its CNET Money personal finance page. The articles covered topics such as “What is Compound Interest?”. and “What happens when you return a check?”

At first glance, and for those new to finance, the articles seemed persuasive and informative. CNET continued the practice until earlier this month, when it was published by website Futurism.

A close examination of the work produced by CNET’s AI makes it seem less like a sophisticated text generator and more like an automated plagiarism machine pumping out casually stolen work.

— Jon Christian, Futurism

But as Futurism has established, the articles written by bots have major limitations. For one thing, many are full of mistakes. Second, many are full of plagiarism — in some cases from CNET itself or its sister sites.

Futurism’s Jon Christian put the error problem bluntly in an article noting that the problem with CNET’s article-writing AI is that “it’s kind of an idiot.” Christian followed up with an article in which he found numerous cases ranging “from verbatim copying to moderate edits to significant rephrasing without properly citing the original.”

This level of misconduct would result in a human student being expelled from school or a journalist being fired.

We’ve written before about the unappreciated limitations of new technologies, particularly those that seem almost magical, such as B. Applications for artificial intelligence.

To quote Rodney Brooks, the robotics and AI scientist and entrepreneur I wrote about last week: “There’s a veritable cottage industry in social media with two sides; one raves about virtuosic performances of these systems, perhaps cherry-picked, and the other, again cherry-picked, showing how incompetent they are at very simple things. The problem is, as a user, you don’t know in advance what you’re going to get.”

Which brings us back to CNET’s article writing bot. CNET has not identified the specific AI application it used, although the timing suggests it is not ChatGPT, the AI ​​speech generator written by technologists and teachers for its apparent ability to produce works that are difficult to distinguish as non-human have caused a stir.

CNET didn’t make the AI ​​contribution to its articles particularly clear, adding only a line in fine print that read, “This article was powered by an AI engine and reviewed, fact-checked, and edited by our editorial team.” The more than 70 articles were attributed to “CNET Money Staff”. Since Futurism’s disclosure, the byline has simply been changed to “CNET Money.”

Last week, according to The Verge, CNET executives told employees that the site would pause publishing the AI-generated material for the moment.

As Futurism’s Christian noted, the errors in the bot’s articles ranged from fundamental misdefinitions of financial terms to unwarranted oversimplifications. In the article on compound interest, the CNET bot originally wrote, “If you put $10,000 into a savings account that earns 3% interest annually, you’ll make $10,300 by the end of the first year.”

That’s wrong – the annual income would only be $300. The article has since been corrected to state that “You will make $300 which, added to the principal amount, would have been $10,300 at the end of the first year.”

The bot also initially described the interest payments on a $25,000 car loan at 4% interest as “a flat $1,000…per year.” It’s payments on car loans, like mortgages, that are fixed — interest is charged only on outstanding balances that shrink as the payment is made. Even with a one-year car loan at 4%, the interest is only $937. With longer-term loans, the total interest paid decreases every year.

CNET corrected that too, along with five other errors in the same article. Putting it all together, the site’s claim that its AI bot was “fact checked and edited by our editorial staff” looks a bit flimsy.

The plagiarism of the bot is more noticeable and provides an important clue as to how the program works. Christian noted that the bot appeared to have replicated text from sources such as Forbes, The Balance, and Investopedia, all of which cover the same area of ​​personal finance advice as CNET Money.

In these cases, the bot used similar obfuscation techniques as human plagiarists, such as B. Minor reformulations and word swaps. In at least one instance, the bot plagiarized Bankrate, a sister publication of CNET.

None of this is particularly surprising, as a key to how voice bots work is their access to a vast amount of human-produced prose and verse. They may be good at finding patterns in source material that they can replicate, but at this stage of AI development, they’re still using human brains.

The impressive coherence and coherence of the results from these programs up to ChatGPT seems to have more to do with their ability to select from raw human-generated material than with their ability to develop and express new concepts.

Indeed, “a close examination of the work produced by CNET’s AI makes it appear less like a sophisticated text generator and more like an automated plagiarism machine pumping out casually stolen work,” Christian wrote.

Where we stand on the continuum between robot-generated incoherence and truly creative expression is difficult to pin down. Washington and Lee University professor Jeff Schatten wrote in an article in September that the most sophisticated voice bot of the time, known as GPT-3, had obvious limitations.

“It stumbles over complex writing tasks,” he wrote. “It can’t write a novel or even a decent short story. His attempts at scientific writing…are ridiculous. But how long does it take for the ability to be there? Six months ago, GPT-3 was struggling with rudimentary queries, and today it can write a sane blog post discussing the ways an employee can be promoted by a reluctant boss.”

It is likely that for those who need to grade written work, such as B. teachers, it is becoming increasingly difficult to distinguish AI-generated material from human results. A professor recently reported catching a student submitting a paper written by bots the old-fashioned way — it was too good.

Over time, the confusion over whether something was produced by bots or humans may not depend on the skills of the bot, but on those of the humans responsible.

Leave a Reply

Your email address will not be published. Required fields are marked *