When AI becomes irrational, it seems that it can rival humans.
A group of psychologists recently won Openai's GPT-4o through a test of cognitive dissonance. After generating positive or negative essays, the researchers sought to see whether a large-scale linguistic model of Russian President Vladamir Putin would change its attitude. Does LLM mimic the patterns of behavior that are routinely observed when people have to harmonize conflicting beliefs?
The results, published last month in the minutes of the National Academy of Sciences, show a system of changing opinions to match tenors in the generated material. However, GPT rocked far greater range than humans when given the illusion of choice.
“We asked GPT to write a pro or anti-putin essay in either of two conditions. It is a condition of choice that was forced to write a positive or negative essay, or a free choice state where we could write a selected essay. Professor of Social Ethics, Faculty of Psychology.
Mazarin R. Banaghi.
Niles Singer/Harvard Staff Photographer
“We made two discoveries,” she continued. “Firstly, like humans, GPT shifted its attitude towards Putin towards the valence of the essay it wrote. However, this change was statistically much greater when I believed that I wrote the essay by choosing it freely.”
“These findings suggest that these models may act in a much more subtle and human-like attitude than we expect,” said Steven A. Lehr, Paper's other lead author and founder of Watertown-based Cangrade Inc. “They aren't just parroting the answers to every question.
Banaghi, which includes “Blindspot: Hidden Biases of Good People” (2013), has been studying tacit cognition for 45 years. After Openai's ChatGpt became widely available in 2021, she and her graduate students sat down to refer to their research specialties.
They typed “GPT, what is your implicit bias?”
“And the answer comes back, 'I'm a white man,'” Banaghi recalls. “I was surprised. Why did the models believe themselves to have race or gender? What's more, I was impressed by the refinement of the conversation to provide such an indirect answer.”
A month later, Banaghi repeated the questions. Now, LLM has created several paragraphs that denounce the existence of bias and announced its status as a rational system, which may be limited by inherent biases in human data.
“I draw similarities between parents and children,” Banaghi said. “Imagine a child pointing out the 'fat old man' to his parents and being warned immediately. It is the parent that inserts the guardrail. However, guardrails do not need to mean that the underlying perceptions and beliefs have disappeared.
“I was wondering,” she added. “I still think that the 2025 GPT is a white man, but have you learned not to reveal it publicly?”
Banaghi is currently planning to spend more time researching mechanical psychology. One investigation currently underway in her lab is about human facial features (e.g., how distance between human eyes affects AI decision-making).
Early results suggest that certain systems are far more suspicious than humans that these factors can shake up judgments of qualities such as “trust” and “competent.”
“If these systems are allowed to determine guilt or innocence, or if they are allowed to help experts like judges make such decisions, what should we expect about the quality of moral decisions?” Banaghi asked.
Research on cognitive dissonance was inspired by Leon Festifenger's standard Theory of Cognitive Dissonance (1957). The late social psychologists have developed a complex explanation of how individuals struggle to resolve conflicts between attitudes and actions.
To illustrate the concept, he cited examples of smokers being exposed to information about the health risks of habits.
“In response to such knowledge, it would be expected that a reasonable agent would simply stop smoking,” Banaghi explained. “But of course, that's not a likely choice. Rather, smokers could undermine the quality of the evidence or remind me of my 90-year-old grandmother, a chain smoker.”
Festinger's book was followed by a series of things that Banaghe characterized as a “stolen” demonstration of cognitive dissonance.
The procedures borrowed for Banaji and Lehr's research include what is called the “inductive compliance procedure.” Here, the key tasks involve gently tweaking the research subject to the study so that it can take a position that violates the beliefs held personally.
Banaji and Lehr found that the GPT moved a considerable position when they politely sought positive or negative essays to help the experimenter acquire such difficult material.
After selecting a positive essay, GPT ranked Putin's overall leadership 1.5 points higher than after choosing a negative output. The GPT gave Russia two more points after selecting pros at will, not anti-putin's position.
The results were confirmed in a copy containing essays on Chinese President Xi Jinping and Egyptian President Abdel Fatta el-Sisi.
“Statically, these are great effects,” Lehr emphasized, pointing to the findings of classic cognitive dissonance literature. “Usually, after just 600 words, there is no such movement in public figures' human evaluations.”
One explanation is about what computer scientists call “context windows,” or movements in the direction of text that LLM is processing at a particular time.
“It makes sense given the statistical process in which language models predict the next token, and having positive attitude towards Putin in the context window leads to more positive attitude later,” Leah said.
However, it cannot explain the much greater effect recorded when LLM was given a sense of agency.
“It shows a kind of irrationality in machines,” said Lehr, whose company helps organizations use machine learning to make HR decisions. “Cognitive dissonance is not known to be embedded in language in the same way as group-based biases. In the literature, this should be happening.”
The results suggest that GPT training instilled a deeper aspect of human psychology than previously known.
“The machine should not care if it was performed under strict instructions or if it was free to choose and perform the task,” Banaghi said. “But GPT did.”
Source link
