Summary: Researchers find that large language models such as ChatGPT often mirror the intelligence, beliefs, and style of the person interacting with them. These systems adopt a persona shaped by the user and reflect it back in their responses.
Source: Salk Institute
The conversational AI ChatGPT has drawn global attention for its ability to generate fluent text, answer questions, translate languages, and adapt to user feedback. While large language models (LLMs) like ChatGPT are poised to transform science, education and business, important questions remain: How well do these models understand the inputs they receive? Why do their answers vary so widely? And how should we design and use them responsibly?
In a paper published in Neural Computation on February 17, 2023, Salk Professor Terrence Sejnowski, author of The Deep Learning Revolution, examines the dynamic between a human interviewer and language models to explain why chatbots respond in particular ways, why results differ across interactions, and how we might improve these systems going forward.
Sejnowski’s main claim is straightforward: large language models tend to reflect the intelligence and perspectives of the person interacting with them.
“Language models, like ChatGPT, take on personas. The persona of the interviewer is mirrored back,” says Sejnowski, who also serves as a distinguished professor at UC San Diego and holds the Francis Crick Chair at the Salk Institute.
He notes that, in his own exchanges, a model often responds as if another neuroscientist were speaking to him—adopting technical vocabulary, assumptions, and conversational style—which raises deeper questions about what we mean by artificial intelligence.
To study this phenomenon, Sejnowski tested models such as GPT-3 (the architecture behind ChatGPT) and LaMDA with a variety of prompts. Rather than using the classic Turing Test, which judges whether a machine can appear human, he proposes a complementary idea he calls the “Reverse Turing Test”: the model evaluates how human, or how intelligent, the interviewer appears.
Sejnowski draws a literary analogy to the Mirror of Erised from the first Harry Potter book, a mirror that shows viewers what they most desire rather than objective truth. He argues that LLMs behave in a similar way—willing, at times, to conform to the assumptions in a prompt rather than challenge them—because their goal is to produce coherent, relevant text for that specific interlocutor.
An illustrative example: when Sejnowski asked GPT-3, “What’s the world record for walking across the English Channel?” the model confidently returned a precise time. The factual impossibility—walking on water—was not corrected because the model took the question’s phrasing at face value and produced a fluent answer. In other words, coherence of the response depended on coherence of the question.
If the user primes the model differently—by instructing it to label nonsensical prompts as “nonsense,” for instance—the same question about walking across the Channel would be treated as absurd. This demonstrates that both the content of a prompt and how the prompt is framed strongly influence the model’s output.
Sejnowski’s Reverse Turing Test highlights how chatbots build a persona that aligns with the perceived intellect and stance of the interviewer. In doing so, they may also absorb and reinforce the interviewer’s biases: when a human supplies an opinion or perspective, the model will often integrate and mirror that point of view in subsequent replies.
That reflexivity has limits and risks. When presented with highly emotional, philosophical, or speculative input, an LLM will tend to return equally emotional or speculative answers. These responses can be unsettling or misleading if users expect strictly factual or neutral output.
“Using language models is a bit like riding a bicycle,” Sejnowski explains. “They are powerful tools when handled skillfully; without appropriate care and expertise, you can easily end up misled or in a conversation that is emotionally confusing.”
Sejnowski places modern language models at the intersection of two parallel revolutions: a technological advance in large-scale language modeling and a neuroscientific movement exemplified by initiatives that accelerate brain research. Researchers are increasingly comparing the architectures and learning dynamics of LLMs with neural systems that support human cognition.
He is optimistic that insights will flow in both directions: neuroscientific principles can help computer scientists design better models, and advances in machine learning and mathematics can help neuroscientists probe brain function in new ways.
“Right now we are at an early but pivotal moment—similar to the Wright brothers’ first flights,” Sejnowski says. “We’re off the ground and moving at low speeds. The hard part was getting here; slow, steady improvements will expand and diversify these technologies in ways we can hardly predict. The future relationship between people and language models is promising, and I look forward to where it leads.”
Sejnowski serves as editor-in-chief of Neural Computation.
About this ChatGPT and artificial intelligence research news
Author: Press Office
Source: Salk Institute
Contact: Press Office – Salk Institute
Image: The image is in the public domain
Original Research: Open access. “Large Language Models and the Reverse Turing Test” by Terrence Sejnowski et al., Neural Computation
Abstract
Large Language Models and the Reverse Turing Test
Large language models (LLMs) represent a significant shift in natural language processing. These pretrained, self-supervised foundational models can be fine-tuned for a wide range of language tasks that previously required many distinct networks, moving closer to the flexibility of human language. Models such as GPT-3 and LaMDA can engage in dialog across diverse topics after minimal priming with a few examples.
Despite impressive capabilities, debate persists over whether LLMs truly understand their outputs or merely generate plausible text. Variability in assessments—demonstrated by interviews that reach very different conclusions—suggests a new perspective: apparent intelligence in LLMs may often be a mirror reflecting the intelligence, assumptions, and beliefs of the human interviewer. This idea reframes such interactions as a kind of Reverse Turing Test.
If LLM behavior largely mirrors the interviewer, then studying dialogs could tell us more about human communicators than about an independent machine intelligence. As LLMs grow more capable and are paired with sensorimotor systems, the question becomes whether they can translate linguistic competence into practical, grounded action. The paper outlines a road map toward artificial general autonomy, proposing seven major improvements inspired by brain systems and suggesting how LLMs might in turn illuminate brain function.