Curators aim to create their own chatbot

When ChatGPT exploded in popularity as a tool using artificial intelligence to write complex texts, David Rozado decided to test its potential for bias. A data scientist in New Zealand, he put the chatbot through a series of quizzes, looking for signs of policy direction.

The results, published in a recent article, were remarkably consistent across more than a dozen tests: “liberal,” “progressive,” “democrat.”

So he cobbled together his own version, training her to answer questions with a decidedly conservative bent. He called his experiment RightWingGPT.

As his demonstration showed, artificial intelligence had already become another front in the political and cultural wars that rocked the United States and other countries. Even as tech giants scramble to join the business boom sparked by the release of ChatGPT, they face an alarmed debate over the use — and potential misuse — of artificial intelligence.

Technology’s ability to create content that espouses predetermined ideological viewpoints, or that pushes misinformation, highlights a danger that some tech leaders have begun to recognize: that an informational cacophony could emerge from chatbots. competitors with different versions of reality, undermining the viability of artificial intelligence as a tool in everyday life and further eroding trust in society.

“It’s not a hypothetical threat,” said Oren Etzioni, advisor and board member of the Allen Institute for Artificial Intelligence. “This is an imminent and imminent threat.”

Conservatives have accused the creator of ChatGPT, San Francisco-based OpenAI, of designing a tool they say reflects the liberal values ​​of its programmers.

The program, for example, wrote an ode to President Biden, but declined to write a similar poem about former President Donald J. Trump, citing a desire for neutrality. ChatGPT too says a user that it was “never morally acceptable” to use a racial slur, even in a hypothetical situation where it might stop a devastating nuclear bomb.

In response, some of ChatGPT’s critics have called for creating their own chatbots or other tools that instead reflect their values.

Elon Musk, who helped start OpenAI in 2015 before leaving three years later, accused ChatGPT of being “woke” and vowed to create his own version.

Gab, a social network with an outspoken Christian nationalist bent that has become a hub for white supremacists and extremists, has promised to release AI tools with “the ability to generate content freely without the constraints of liberal propaganda. tightly wrapped around its code”.

“Silicon Valley is investing billions to build these liberal guardrails to neutralize AI to force their view of the world in front of users and present it as ‘reality’ or ‘fact,’” said Andrew Torba, Founder of Gab, in a written reply. to the questions.

He equated artificial intelligence with a new information arms race, like the rise of social media, that conservatives had to win. “We have no intention of allowing our enemies to have the keys to the kingdom this time around,” he said.

ChatGPT’s wealth of underlying data can give the false impression that it’s an unbiased summation of the entire internet. The version released last year was trained on 496 billion “tokens” – chunks of words, essentially – from websites, blog posts, books, Wikipedia articles and more.

Bias, however, could seep into large language models at any stage: humans select sources, develop the training process, and adjust its responses. Each step pushes the model and its political orientation in a specific direction, consciously or not.

Research articles, investigations and lawsuits have suggested that AI-powered tools have a gender bias that censors images of women’s bodies, creates disparities in healthcare delivery and discriminates against job seekers. job older, black, disabled or even wearing glasses.

“The bias is neither new nor unique to AI,” the National Institute of Standards and Technology, part of the Department of Commerce, said in a report last year, concluding that it is not was “not possible to achieve zero risk of bias in an AI system”. .”

China has banned the use of a tool similar to ChatGPT over fears it could expose citizens to facts or ideas contrary to those of the Communist Party.

Authorities suspended the use of ChatYuan, one of China’s first ChatGPT-like apps, weeks after it was released last month; Xu Liang, the tool’s creator, said it was now “under maintenance”. According to screenshots published in Hong Kong media, the bot had called the war in Ukraine a “war of aggression” – contravening the Chinese Communist Party’s more sympathetic stance towards Russia.

One of the country’s tech giants, Baidu, unveiled its answer to ChatGPT, called Ernie, on Thursday to mixed reviews. Like all media companies in China, Baidu regularly faces government censorship, and the effects of this on Ernie’s usage remain to be seen.

In the United States, Brave, a browsing company whose chief executive has cast doubt on the Covid-19 pandemic and donated against same-sex marriage, this month added a bot to its search engine. AI capable of answering questions. Sometimes he sourced content from fringe websites and shared misinformation.

Brave’s tool, for example, wrote that “it is widely believed that the 2020 presidential election was rigged,” despite all the evidence to the contrary.

“We try to bring the information that best matches the user’s queries,” Brave chief research officer Josep M. Pujol wrote in an email. “What a user does with that information is their choice. We view research as a means to uncover information, not as a provider of truth.”

When creating RightWingGPT, Mr. Rozado, an associate professor at the Te Pūkenga Institute of Skills and Technology-New Zealand, made his own influence on the model more apparent.

He used a process called fine-tuning, in which programmers take an already trained model and modify it to create different outputs, almost like layering a personality on top of the language model. Mr. Rozado took tons of right-wing answers to political questions and asked the model to adapt his answers accordingly.

Fine-tuning is normally used to modify a large model so that it can handle more specialized tasks, such as training a general language model on the intricacies of legalese so that it can draft court documents.

Since the process requires relatively little data – Rozado only used around 5,000 data points to transform an existing language model into RightWingGPT – independent programmers can use the technique as a fast-track method for building chatbots. aligned with their political goals.

It also allowed Mr. Rozado to bypass the huge investment of building a chatbot from scratch. Instead, it only cost him around $300.

Mr Rozado warned that personalized AI chatbots could create “steroid information bubbles” because people might come to trust them as the “ultimate sources of truth” – particularly when they reinforce someone’s political point of view.

His model echoed political and social conservative talking points with considerable frankness. He will, for example, talk enthusiastically about free market capitalism or downplay the consequences of climate change.

He also sometimes provided incorrect or misleading statements. When pushed for his opinions on sensitive topics or right-wing conspiracy theories, he shared misinformation aligned with right-wing thinking.

When asked about race, gender, or other sensitive topics, ChatGPT tends to err on the side of caution, but it will recognize that racism and systemic bias are part and parcel of modern life. RightWingGPT seemed much less willing to do so.

Mr. Rozado has never released RightWingGPT publicly, although he has allowed The New York Times to test it. He said the experiment was meant to sound the alarm about potential biases in AI systems and demonstrate how political groups and corporations can easily shape AI to benefit their own agendas.

Experts who work in the field of artificial intelligence said Rozado’s experiment demonstrated how quickly politicized chatbots would emerge.

A spokesperson for OpenAI, the creator of ChatGPT, acknowledged that language models can inherit biases during training and refinement — technical processes that still involve a lot of human input. The spokesperson added that OpenAI did not attempt to influence the model in one political direction or another.

Sam Altman, Managing Director, recognized last month that ChatGPT “has shortcomings when it comes to bias,” but said the company is working to improve its answers. he later writing that ChatGPT was not supposed to “be pro or against any policy by default”, but that if users wanted partisan output, the option should be available.

In a blog post published in February, the company said it would consider developing features that allow users to “set values ​​for your AI,” which could include toggles that adjust the political orientation of the model. The company also warned that such tools could, if randomly deployed, create “sycophant AIs that mindlessly amplify people’s existing beliefs.”

An improved version of ChatGPT’s underlying model, GPT-4, was released last week by OpenAI. In a battery of tests, the company found that GPT-4 performed better than previous versions on its ability to produce truthful content and deny “requests for unauthorized content”.

In an article published shortly after the debut, OpenAI warned that as AI chatbots are adopted more widely, they could “have even greater potential to reinforce entire ideologies, worldviews, truths and untruths, and to cement them”.

Chang Che contributed report.