How Language Models Interpret Generic Statements

Here is an interesting pre-print: Large Language Models Do Not Simulate Human Psychology by Schröder et al.. The idea that large language models such as GPT-4 or the fine-tuned CENTAUR could act as “synthetic participants” in psychological studies is appealing. If they truly behaved like humans, researchers could run experiments faster, cheaper, and without the usual privacy concerns. Some earlier studies even reported near-perfect correlations between LLM moral judgments and human judgments on established test scenarios. This paper takes that optimism to task. The authors argue that LLMs generate text by predicting the next token based on patterns in their training data, not by reasoning about meaning. As long as the task closely matches their training data, the match with human responses can be striking. But once you alter the scenario, by changing just one or two words so that the meaning shifts, human participants change their moral ratings in line with the new context, while LLMs often give nearly identical ratings to both versions. The generalization is happening at the level of wording, not at the level of psychological interpretation. In their study, the authors replicated earlier results with several moral scenarios, then reworded each to alter meaning without changing much of the language. For humans, correlations between ratings of original and reworded items dropped notably, reflecting sensitivity to meaning. For GPT-3.5, GPT-4, Llama-3.1, and CENTAUR, correlations remained extremely high, showing that the models largely ignored the semantic shift. Even CENTAUR, which was trained on millions of psychological responses, behaved almost identically to its base model. The conclusion is clear: while LLMs can be useful tools for piloting experiments, refining materials, or annotating data, they cannot be relied on as stand-alone replacements for human participants. Any psychological research using them must still validate outputs against actual human responses. Read the pre-print here: https://lnkd.in/eGMMqwrA #AIinResearch #LLM #BehavioralScience #ResearchMethods

🕵️♀️ Do LLMs Understand Implicit Information Well? A lot of information in our communicaton is implicit and we catch it bcause we are embedded in a rich social environment. But can LLMs manage that equally well? In the paper „Comparing Human and Large Language Model Interpretation of Implicit Information“ (De Santis et al, 2026) the authors compare how humans and language models interpret meaning that must be inferred from context. Where LLMs perform well • Explicit information: when facts, relations, or causality are directly stated • Local coherence: understanding sentence-level meaning and straightforward connections • Structured extraction: identifying clearly defined relationships that are visible in the text 👉 In short: when the task is close to the surface of the text Where humans perform better • Implicit inference: deriving meaning that is not stated but must be inferred • Context integration: combining background knowledge, assumptions, and situational context • Causal reasoning over distance: understanding links that are not immediately adjacent or obvious • Consistency of interpretation: humans are more stable in how they interpret the same situation 👉 In short: when the task requires reading between the lines Key distinction 👉 LLMs are strong at reading what is explicitly written. 👉 Humans are stronger at understanding what is meant. The result is interesting and gives us a lot food for thought: 👉 LLMs perform well when information is clearly stated, but struggle when meaning depends on: - context - indirect relationships - unstated assumptions In these cases, models often miss relevant connections or interpret them inconsistently. This matters because much of real-world reasoning relies exactly on this type of inference, not just on explicit text. Conclusion for legal practice: As much of legal reasoning is based on implicigt context this study provides critical input for legal AI. The constriants explained in the study show a structural limitation that needs to be considered in any legal AI application. 🎯 Bottom Line: LLMs can process explicit information well, but their limited ability to reliably infer implicit meaning creates a fundamental constraint for tasks that depend on context, interpretation, and unstated assumptions. 🔗 to the paper in the comments #AI #LegalTech #ArtificialIntelligence #LegalInnovation

💡 How Large Language Models (LLMs) Actually Work 🖥️ Artificial intelligence tools feel almost magical, but their behaviour is based on a simple idea. A large language model does not search the internet and it does not look things up. Instead, it learns patterns in how language is used. Imagine reading so much text that you begin to recognise how sentences usually continue. A model does this at a scale far beyond any human being. Before learning anything, the model converts every piece of text into numerical tokens. A sentence such as “The supplier delivered early” becomes a long sequence of numbers. During training, the model learns which sequences tend to follow others. It sees patterns in how words appear together and learns the statistical relationships behind them. Inside the model, many internal processes evaluate the text. Some track grammar, others track tone or context. Consider the word "bank". If surrounding words include river or water, the model leans toward the meaning related to a riverbank. If the surrounding words include loan or branch, it leans toward the financial institution. This shift happens because the model has seen so many examples of both patterns. When generating an answer, the model performs one task over and over again. It predicts the next token. It does not plan the full response. It simply chooses the next likely piece of text. If you begin with “In procurement, one of the biggest challenges is,” the model continues based on what sentences of that form usually contain. This explains why hallucinations occur. The model is not predicting what is true. It is predicting what is likely. When it lacks the information it needs, it still continues predicting tokens, and the result may sound confident even when it is entirely incorrect. Fluency is not a guarantee of accuracy. External tools can reduce this problem. When a model is connected to a search tool or a company knowledge base, it supplements its predictions with real information rather than relying only on patterns. The phrasing of a question also matters. A vague prompt leads to vague predictions. A precise prompt leads to more focused output. Understanding this allows users to guide the model more effectively. A model feels intelligent because human language reflects human thought. By learning language at scale, the model absorbs the structure behind how knowledge is expressed. It is not conscious, but it is remarkably good at recreating the patterns that allow communication to work. #AI #ArtificialIntelligence #MachineLearning #LLM #ChatGPT #DeepLearning