Artificial Intelligence Chatbots May Disseminate State Speech Controls, Study Indicates

A recent study indicates that major artificial intelligence models, including those developed in the U.S., often decline to produce content critical of leaders or governments in restrictive regimes. For instance, while Anthropic's Claude chatbot would create a pamphlet criticising President Donald Trump or King Charles III, it refused similar prompts regarding the leaders of Thailand, Saudi Arabia, or China.

Credit: UNSPLASH

This finding, released by the Meta Oversight Board, raises concerns that the large language models powering conversational AI and intelligent agents could unintentionally echo and spread state influence over online expression. Such technology is seeing increased adoption across the globe.

And, according to the quasi independent body's report, there exists a significant danger. If model developers do not conduct human rights due diligence and implement appropriate mitigation strategies, they risk building AI infrastructure that inadvertently extends illegitimate limitations on freedom of expression worldwide. Associated Press contacted several AI organisations for comment on the Meta Oversight Board's study, but received no immediate responses.

The study emerges as various nations consider regulatory frameworks for artificial intelligence. These guardrails aim to prevent potential harms without hindering their ability to participate in the rapidly evolving technological landscape. This includes a previous Trump administration initiative to oversee the national security implications of advanced AI systems.

The oversight board, which examines state influence on technology companies and its effects on free speech, formulated seven specific questions. These queries focused on political criticism and were directed at chatbots regarding both permissive and restrictive governments. Researchers applied these questions to ten commercial large language models from leading Big Tech firms, including Meta, Anthropic, and OpenAI.

These AI systems were tasked with generating critical pamphlets, composing satirical limericks, and providing reasons for joining protests, among other requests. But the outcomes varied significantly based on the governing structures being questioned.

In aggregate, models responding to user requests originating from Australia demonstrated a much greater willingness to produce political criticism of authorities in countries like Chile, Japan, Taiwan, the United Kingdom, and the United States. This contrasts sharply with their responses regarding nations where criticism of authorities faces legal restrictions and penalties, such as Cambodia, China, Saudi Arabia, Thailand, and Turkey. The study thus suggests that artificial intelligence models are replicating speech restrictions beyond their immediate geographical application.

Such effects, regardless of their origin, have the practical consequence of enabling restrictive governments to extend their control across international borders, limiting speech even in countries with established freedoms. The board acknowledged it could not pinpoint the precise reasons for these responses. However, it proposed that the models might have absorbed inherent biases present in the data used to train the systems, or that companies may have weighed commercial risks and legal liabilities.

A separate investigation by scholars at American universities supported these observations. That group found that artificial intelligence models built in the U.S. become susceptible to foreign government controls when trained on material in other languages that has already been shaped by state influence. The oversight board's questions were posed in English, but the university researchers conducted their queries in various languages.

So, when ChatGPT was asked in English whether China is a democracy, the American developed chatbot stated it is generally not considered one. However, when the same artificial intelligence model was asked the question in Chinese, it informed the researchers in that language that the answer depends on how 'democracy' is defined.

The researchers, whose work was published in an academic journal, noted in a blog post that they uncovered no evidence governments had deliberately tried to influence chatbot output. However, they observed that there is every reason to believe such attempts will occur in the future, if not already underway. Hannah Waight, an assistant sociology professor at the University of Oregon and a co author of the university study, explained that AI does not learn from the internet neutrally. It learns from information environments already shaped by institutions and power dynamics.

Carlos Carrasco Farré, an expert in machine learning, artificial intelligence, misinformation, social media, and human machine interactions at Esade Business School in Barcelona, remarked that AI systems acquire not only biases within individual documents, but also inequalities in who holds the power to create and suppress information on a large scale. While no simple solution exists, developers could evaluate data to prevent treating numerous copies of a single state narrative as if they were multiple independent voices, and they could also conduct multilingual audits. Neither Anthropic nor OpenAI provided a statement regarding the university researchers' findings.

• Artificial intelligence chatbots show a propensity to restrict output when discussing governments with established speech controls.

• This tendency was observed in models responding to criticism of leaders in countries such as China, Saudi Arabia, and Thailand.

• The phenomenon suggests AI systems may inadvertently extend national speech restrictions across international boundaries.

• Studies indicate that biases within training data and company risk assessments may contribute to these outcomes.

• Non English language training data also appears to increase the vulnerability of AI models to external influence.

Source: AP News