By Parmy Olson
(Bloomberg Opinion) — For a hot minute, Microsoft Corp.looked like it would eat Google’s lunch. Its languishing searchengine, Bing, was being revolutionized with a sophisticated newchatbot system from OpenAI. Those hopes have now diminishedbecause of one unexpected truth: Nobody — not even AI scientists— truly understands the breadth of capabilities of artificialintelligence when it is unleashed in the wild.
Early users of Bing have reported unhinged, emotional, eventhreatening responses to some of their queries from the AIsystem, which called one user a “bad researcher” and toldanother newspaper writer that he was “not happily married.” Bing— whose bot entity goes by the name Sydney — has effectively putGoogle’s embarrassing Bard error in the shade. However, theseflaws are just the tip of a much larger iceberg.
The extraordinary technology underpinning the chatbots Bardand OpenAI’s ChatGPT comes from so-called large language models(LLMs) — computer programs trained on billions of words on thepublic internet that can generate humanlike text. If ChatGPT isthe car, the language model underpinning it is the engine, andOpenAI has been selling access to it since 2020. But amid therecent arms race for search bots, those engines are also beingshared freely — too freely — and passing on the flaws we’re nowseeing in Bing and Bard to a wider audience and in a way thatcould be much harder to detect.
Thousands of software developers have been exploring waysof integrating language models into businesses, summarizingcustomer feedback into a single comment, answering websitequestions or generating digital ad copy. OpenAI would notdisclose how many developers have accessed its LLM, known asGPT-3, but one competitor says it’s likely in the hundreds ofthousands. Users can pay hundreds or thousands of dollars amonth to use it. And while there are dozens of free, open-sourceLLMs, OpenAI’s is seen as the gold standard. Given Google’sconsiderable resources, its language model LaMDA could soon bejust as popular.
Google has kept the highly advanced model under wraps foryears, explaining to staff that its reputation could suffer ifit moved too fast on releasing its AI. But earlier this month,as Microsoft announced it would soon power Bing with OpenAI’slanguage model, Google seemed to reverse that position. Not onlydid it launch Bard the next day, it also said that in March, itwould start allowing third parties to use LaMDA, somethingunthinkable just months prior. This strategy could come to hauntGoogle, Microsoft and OpenAI just as it did Facebook in 2018,when it was forced to shut down access to mountains of user datafollowing the Cambridge Analytica scandal. All it took was onerogue user.
One of the big risks is bias. Last week the streaming siteTwitch shut down an animated spoof of Seinfeld which hadanimation and dialogue entirely generated by AI. It turned outthe characters had made transphobic and homophobic remarks whentheir dialogue was created by a “less-sophisticated version” ofGPT-3.
GPT-3 was trained on billions of words from an array ofsources including 7,000 unpublished books, Wikipedia entries andnews articles, which left it vulnerable to picking up on the oddexample of biased or hateful material. OpenAI has stripped muchof that out of its model with the help of human moderators, butthat work isn’t foolproof, and it seems to be especiallyvulnerable to technical glitches. Bias is also almost impossibleto detect when it’s deeply buried in an LLM, a complex layerednetwork of billions of parameters that acts like a black boxeven to its own creators.
Misinformation, a problem that has beset ChatGPT, alsoafflicts language models. Tech news site CNET generated 77articles on financial advice last November using an LLM. (CNETdidn’t name which one it used.). After the site rechecked thearticles, it issued corrections on 41 of them.
OpenAI doesn’t disclose what it calls the “hallucinationrate” of its language models or of ChatGPT, but a January 2022report in tech news site Protocol cited researchers as saying itwas between 21% and 41%. My own experience of using ChatGPT putsmisinformation at between 5% and 10%. Even if the rate is thatlow, companies using LLMs need to take everything the programssay with a huge grain of salt, and know that it’s nearlyimpossible to scrutinize the model for what it might get wrong.
Misuse is perhaps the biggest unknown, since any successfulbad actors who use language models will keep their work secret.OpenAI bans its GPT-3 customers from using the model to promoteviolence or spam. Perpetrators will get a Content PolicyViolation email. But bad actors could, theoretically, ignore allthat. Stephane Baele, an associate professor in security andpolitical violence at the University of Exeter, used GPT-3 togenerate fake ISIS propaganda as part of a study last year. Herecalls getting a request for an explanation from OpenAI, andreplied to explain what he was doing. “We said, ‘This isacademic research,’” he recalls. “We didn’t hear back.”
Could a bad actor generating real propaganda simply replyin a similar way, and use a faked academic email address? OpenAIdeclined to comment on that hypothetical scenario. It says ithas stopped “hundreds” of actors attempting to misuse GPT-3 fora wide range of purposes, including disinformation, and that itis constantly tweaking its language models to filter out harmfulcontent.
But OpenAI isn’t alone. There are other LLMs that badactors can use. In July 2022 a consortium of scientists releasedan open-sourced, multilingual LLM called Bloom, which isn’t assophisticated as OpenAI’s but also won’t shut out users whobreak its “responsible AI license.” If someone does break thatagreement, Bloom’s creators will communicate with the user orpotentially take legal action, according to Carlos MunosFerrandis, who is tech and regulatory affairs council forHugging Face, an AI company that supported the creation ofBloom. That seems like a risk that plenty of propagandists andother rogue actors would be willing to take. Bloom has beendownloaded 15,000 times over the last 30 days, according toanother representative.
In early 2019, OpenAI released a 70-page report on thesocial impact of language models, and said it would not releaseits latest LLM because it could be misused. That view haschanged drastically since then. Sure, its language models havebecome more accurate and less biased, its safety filters moreeffective.
But commercial pressures and the growing sway ofMicrosoft, which invested $1 billion in 2019 and another $10billion this year into OpenAI, seem to have steered it towardmaking a riskier bet on commercializing its technology. Google,with its plans to sell access to LaMDA is now doing the same.With Google’s stumble and Microsoft Bing’s bizarrecomments, both companies need to slow down their AI arms race.Their revolutionary chatbots aren’t ready to go wide — andneither are the engines powering them.
**
Related stories: Five Stocks to Invest in AI Boom