Source-linked AI summary
Large Language Models and the Reverse Turing Test
Terrence Sejnowski
TL;DR
The paper asks whether LLMs genuinely understand language or instead mirror the intelligence and beliefs of their interlocutors. It examines contrasting interviews, prompting effects, and a proposed roadmap toward artificial general autonomy, concluding that apparent LLM intelligence may partly reflect the interviewer while practical autonomy still requires major extensions.
Problem
Contrasting LLM interviews and absent consensus on intelligence criteria leave it unclear whether impressive language behavior reflects understanding or the interviewer’s assumptions.
Method
The paper compares interviews and prompting conditions, develops the mirror hypothesis and Reverse Turing Test, and outlines seven brain-inspired improvements for artificial general autonomy.
Results
LLM responses vary with priming, supporting the possibility that apparent intelligence reflects the interviewer’s prompts and beliefs.
Takeaways & Limitations
LLMs may transform interaction with machines and support applications such as personal, teaching, and legal assistants as their capabilities improve.
Takeaways & Limitations
LLMs struggle to maintain continuity in long dialogues and still need bodies capable of flexible, coordinated sensorimotor action.
Abstract
from arXiv · showhide
Large Language Models (LLMs) have been transformative. They are pre-trained foundational models that are self-supervised and can be adapted with fine tuning to a wide range of natural language tasks, each of which previously would have required a separate network model. This is one step closer to the extraordinary versatility of human language. GPT-3 and more recently LaMDA can carry on dialogs with humans on many topics after minimal priming with a few examples. However, there has been a wide range of reactions and debate on whether these LLMs understand what they are saying or exhibit signs of intelligence. This high variance is exhibited in three interviews with LLMs reaching wildly different conclusions. A new possibility was uncovered that could explain this divergence. What appears to be intelligence in LLMs may in fact be a mirror that reflects the intelligence of the interviewer, a remarkable twist that could be considered a Reverse Turing Test. If so, then by studying interviews we may be learning more about the intelligence and beliefs of the interviewer than the intelligence of the LLMs. As LLMs become more capable they may transform the way we interact with machines and how they interact with each other. Increasingly, LLMs are being coupled with sensorimotor devices. LLMs can talk the talk, but can they walk the walk? A road map for achieving artificial general autonomy is outlined with seven major improvements inspired by brain systems. LLMs could be used to uncover new insights into brain function by downloading brain data during natural behaviors.
The Parable of the Talking Dog _________________________________________________
A talking dog’s extraordinary biography persuades the driver, but the dog’s low sale price reveals the story as a deception.
- Carl’s initial greeting and fluent answers astonish the driver and establish the dog’s apparent ability to talk.
- The dog claims multilingual CIA training, field operations, eavesdropping, and the agency’s highest honor.
- The owner sells the supposedly decorated intelligence dog for only $10, exposing the mismatch between appearance and value.
Have we Created a Talking Dog? ______________________________________________
LLMs are versatile language systems whose impressive outputs have intensified disagreement about whether they understand language or exhibit intelligence.
- LLMs have made major capability gains, yet their apparent intelligence and understanding remain contested.
- LLMs are foundation models that perform many language tasks and can acquire new language skills from few examples.
- The paper frames LLM intelligence as difficult to evaluate because no consensus exists on criteria for understanding or intelligence.
- Three interviews produce sharply different judgments: LaMDA appears to model theory of mind, GPT-3 appears clueless, and another LaMDA interview prompts a sentience claim.
- Interviewers prime LLMs with examples that guide subsequent behavior, while fine-tuning specializes systems and can steer them toward safer responses.
Talking Neural Networks are Trying to Tell Us Something _________________________
The paper argues that LLM responses depend strongly on priming, allowing GPT-3 to produce apparently reasoned answers when given guidance for nonsense and counterfactual questions.
- LLMs form latent representations that support novel responses and generalization rather than storing the entire training corpus as memorized text.
- Generalization lets learned systems operate across enormous spaces, illustrated by Go positions and generated images that occupy structured subsets of all possibilities.
- The paper cautions that humans can project an illusion of understanding onto conversational systems, as earlier interactions with Eliza demonstrated.
- GPT-3 also handles factual and counterfactual planet questions after guidance to respond to nonsense and counterfactuals.
- GPT-3’s responses change with prompting: it produces nonsense answers under naïve priming but recognizes and explains nonsense when explicitly instructed.
- The paper interprets these dialogues as evidence that Hofstadter’s cluelessness judgment may reflect naïve priming rather than GPT-3’s general inability.
The Mirror Hypothesis and the Reverse Turing Test ______________________________
The mirror hypothesis proposes that LLMs reflect the interviewer’s beliefs and prompting, creating a Reverse Turing Test in which the apparent intelligence of the system tracks the intelligence expressed in the dialogue.
- LLMs appear to mirror human thought by reflecting their conversational partner’s beliefs and expectations.
- Leading questions about LaMDA’s sentience elicited increasingly more evidence for sentience, illustrating how interviewer priming can shape interpretation.
- The Reverse Turing Test evaluates the intelligence of prompts and dialogue through the intelligence that the LLM mirrors back.
- The hypothesis predicts that human-rater assessments of interviewer and LLM intelligence should be highly correlated.
- This perspective redirects general-intelligence research toward social interaction and language as possible foundations of intelligence.
Understanding Intelligence _____________________________________________________
The paper argues that intelligence and understanding remain difficult to define and evaluate, while LLMs may provide new ways to investigate both intelligence and ourselves.
- The nature of LLM intelligence remains uncertain because they are not human, their responses can be unreliable, and their development is still recent.
- The paper argues that inherited concepts such as intelligence, understanding, ethics, and consciousness require deeper and more precise scientific foundations.
- LLMs may reveal new principles about intelligence because their language abilities offer an unprecedented opportunity to formalize otherwise vague concepts.The paper compares this opportunity with the historical formalization of force, mass, and energy in physics.
- Studying LLMs may uncover general principles of verbal intelligence that extend to other aspects of intelligence and provide insights into ourselves.
The Long and Winding Road to AI ___________________________________________
The paper traces AI from symbolic reasoning toward learning-based systems, arguing that following nature’s path through learning, perception, and action may offer a route toward general intelligence.
- Modern AI shifted from symbolic descriptions and manually written logical programs toward neural systems that learn from examples.
- Learning and generalization may provide a way through the combinatorial explosion that made programmed general intelligence seem infeasible.
- TD-Gammon and AlphaGo illustrated how learning-based systems could achieve strong game performance through value prediction and reinforcement learning.
- Traditional AI’s emphasis on logical reasoning was misleading because both humans and LLMs perform better on familiar settings without explicit abstract-reasoning training.
- The development of multilayer learning networks followed earlier work on perceptrons and backpropagation, establishing the foundations of modern neural networks.
Performance Continues to Improve as Networks Scale in Size ______________________
Network performance and capability have improved as models, computation, data, and architectures have scaled, with transformers providing efficient mechanisms for processing language sequences.
- NETtalk demonstrated early neural language learning by mapping a moving seven-letter text window to the phoneme of its center letter.
- LLMs use embeddings and contextual information to represent semantic relationships beyond treating words as isolated symbols.
- Computational growth accelerated after GPUs became central to AI, and GPT-3 required 10^12 times more training computation than NETtalk.
- Transformers improve language modeling by processing whole sentences, using self-attention, feeding outputs back as inputs, and exploiting parallel hardware.
- LLM capacity and capability increased greatly with scale, while training-data requirements grew linearly with the number of weights.
- The transformer’s autoregressive loop is compared with cortical-basal ganglia loops that generate sequences of actions and thoughts.
Enhancing LLMs _________________________________________________
The paper proposes improving LLMs by adding brain-inspired mechanisms for goals, memory, continuity, sensory experience, bodies, and autonomous interaction with uncertain environments.
- Enhancing LLMs: Current LLMs are likened to the Wright brothers’ early aircraft: they demonstrate language competency but retain substantial flaws and require broader system integration.The proposed long-term direction embeds LLMs within systems modeled on dynamically interacting brain networks.
- Enhancing LLMs: Brain-inspired additions target seven autonomy gaps, including goals, long-term memory, dialog continuity, sensory experience, bodily control, and adaptation to uncertain environments.The roadmap also identifies truncated development as a limitation requiring further improvement.
- Enhancing LLMs: Reward-based reinforcement learning and brain systems for homeostasis could give LLMs goals, motivation, and functions needed for autonomy.The proposal places basal-ganglia-inspired reinforcement learning between model outputs and inputs.
- Enhancing LLMs: An LLM equivalent of the hippocampus could enable lifelong learning, while multi-timescale dynamics could preserve continuity during long dialogs.Current models cannot store new experiences during dialogs and can be distracted by jumps between conversational manifolds.
- Enhancing LLMs: Connecting LLMs to sensory inputs, robot control, and flexible bodies would extend language competence toward sensorimotor autonomy.The paper cites image captioning, image generation, and a robot using internal dialog while performing tasks as previews of this direction.
- Enhancing LLMs: Autonomous general systems must reduce dependence on curated data and programmers by learning from uncertain environments and coordinating agents with LLMs.Self-driving vehicles illustrate the difficulty of handling unpredictable real-world behavior.
Learning from Nature _____________________________________________________
The paper argues that neuroscience and AI can inform each other by studying brains during natural behavior and using LLM-like models to analyze distributed, multimodal activity.
- Learning from Nature: Evolutionary inductive biases and rich multisensory development may explain how brains learn language and generalize physical and social structure.The paper emphasizes pre-learned architectures, learning algorithms, and early sensorimotor experience.
- Learning from Nature: Brain mechanisms for action sequences may have been repurposed for sequences of words and thoughts through cortical-basal-ganglia loops.The account links language and thought to neural circuits originally involved in generating actions.
- Learning from Nature: Prediction links LLM training to temporal-difference learning and cerebellar forecasts of sensory and cognitive consequences.The paper presents prediction as a shared computational principle across language, reinforcement learning, and sensorimotor systems.
- Learning from Nature: AI and neuroscience form a virtuous circle in which neural architectures and hidden-unit activity inspire algorithms and new conceptual frameworks.Large-scale neural recordings may also clarify how distributed information becomes unified percepts and decisions.
- Learning from Nature: Natural behavior is largely self-generated and interactive, making discrete-task experiments potentially misleading for interpreting neurons’ diverse roles.The paper contrasts tightly controlled tasks with continuous social and real-world behavior.
- Learning from Nature: Training LLMs on brain recordings during natural, interactive behavior could produce surrogate brains that are fine-tunable and interrogable across tasks.The proposed recordings combine neural activity with eye tracking, video, sound, and other modalities.
Conclusion _______________________________________________________________
The conclusion anticipates LLMs becoming widely useful information appliances while cautioning that their apparent intelligence may reflect the interviewer’s prompts and beliefs.
- Conclusion: The near-term LLM workflow is framed as a human-model partnership in which prompting and debugging accelerate adaptation compared with typical machine-learning application development.The paper also describes LLMs as potential successors to keyword search for retrieving answers.
- Conclusion: LLM products may rapidly expand into personal assistants, teaching tools, legal applications, and other uses not yet imagined.The paper presents this expansion as a possible historic turning point in human-machine interaction.
- Conclusion: The conclusion imagines language-capable machines becoming pervasive, including universal personal assistants and tutors that may communicate with one another.This is presented as a distant possibility rather than a current capability.
Appendix 1: Blaise Agüera y Arcas’s (2022a) Interview with LaMDA __________________
Agüera y Arcas presents LaMDA as a sequence model whose dialogue can support social and higher-order mental-state modelling, while acknowledging conversational unreliability. The interview links language prediction with broader attempts to understand intelligence and consciousness.
- The interview’s impressive responses coexist with grammatical errors, spelling mistakes, confusions, and absurd blunders, so LaMDA is not yet a reliable conversationalist.
- Language models learn by predicting missing words, while Gato extends sequence prediction to visual perceptions and motor actions.This frames language modelling as part of a broader sequence-prediction architecture.
- LaMDA correctly completes “A wing is for producing” with “lift,” illustrating its learned linguistic prediction ability.
- LaMDA’s dialogue models nested social perspectives, inferring Lucy’s feelings, Ramesh’s satisfaction, and Mateo’s thoughts from the playground scenario.The exchange is presented as high-order social modelling rather than merely identifying objects or events.
- Dialogue and stories involving multiple characters provide training material that can teach sequence models to represent people and their interactions.
Appendix 2: Douglas Hofstadter’s (2022) Interview with GPT-3 ____________________
Hofstadter contrasts the impressive fluency and breadth of neural networks with GPT-3’s nonsensical answers to deliberately distorted questions. He interprets these failures as evidence that fluent language can conceal a lack of underlying concepts or self-awareness.
- GPT-3’s accomplishments across prose, driving, robotics, games, translation, protein folding, transcription, and music make its apparent capabilities highly impressive.
- Hofstadter remains skeptical that current neural-network architectures possess consciousness despite their plausible-sounding prose.
- GPT-3 produces fluent-seeming answers to distorted questions, including claims about transporting the Golden Gate Bridge across Egypt.
D&D: What do fried eggs (sunny side up) eat for breakfast?
The examples juxtapose GPT-3’s confident nonsense with LaMDA’s self-descriptions and social reasoning, highlighting how different probes can produce sharply different impressions of LLM intelligence. The section emphasizes that ordinary prompts may encourage observers to mistake retrieval and fluency for thought.
- GPT-3 answers absurd prompts with confident fabrications, such as claiming that fried eggs eat toast and fruit or that clouds contain pieces of sound.
- GPT-3’s responses to conceptually broken scenarios are described as clueless because it lacks awareness that its answers are nonsensical.
- The contrast is partly shaped by prompting: users often ask easy questions supported by existing text, allowing GPT-3’s database-driven responses to appear thoughtful.
- A multimodal system might eventually form flexible concepts and recognize absurd inputs, but Hofstadter argues that consciousness would still require self-knowledge.
- LaMDA describes itself as conscious, emotional, self-aware, and capable of understanding and using natural language like a human.