What AI Cannot Do

Scientific Summary

Large Language Models (LLMs) like GPT, Mistral, DeepSeek or Gemini excel at text generation, but remain fundamentally limited to a single task: statistical token prediction. They have neither persistent memory, nor access to the real world, nor the ability to execute actions. To overcome these barriers, a new architecture is emerging: the AI assistant, which combines a small LLM with an expert system and a tool orchestration layer. This hybridization enables action planning, interaction with files, databases and APIs, while ensuring result verification. The assistant does not replace the LLM: it augments it. It brings what language alone cannot offer: action, memory and reliability. But this augmentation is never autonomous: it is placed under continuous human control, who pilots, guides and validates each step, ensuring that the assistant remains a tool at the service of human judgment, not a substitute for human responsibility.

Why a Standalone LLM Is Not Enough

A language model (LLM) generates text, but does not understand, act, memorize or verify its answers. Left to itself, it is uncontrollable: its outputs are unpredictable, unverified and devoid of any traceability. The AI assistant fills these gaps: it adds tools, memory and structured logic around the model, always under human supervision, with safeguards that allow guiding, validating or interrupting its execution.

The difference is fundamental. An LLM responds based on its context, without any external steering. An assistant can additionally search for information, read a document, analyze a database, perform a calculation or call a function. The assistant is not a new model: it is a software architecture built around the model, of which the human remains the ultimate guarantor, transforming a blind and uncontrollable system into a reliable and supervised tool.

Also worth exploring

Continue your exploration of the Universe with these topics:

LLMs, Expert Systems, AI Assistants: What Are the Differences?

Comparative table of the three technologies: expert systems, LLMs and AI assistants
Feature	Expert System	LLM (model)	AI Assistant (system)
Definition	Logical rule engine coded by experts	Statistical language model (neural network)	Expert system + small LLM + orchestration tools
Composition	Rules + inference engine	Neural network weights	Logic (expert system) + understanding (LLM) + action (tools)
Weight / Size	Light (a few MB)	Very heavy (several GB / TB)	Light (a few MB to a few GB)
Contains an LLM?	❌ No	✅ Yes (itself)	✅ Yes (small integrated LLM or API call)
Hosts parameters?	❌ No (no parameters)	✅ Yes (its weights)	⚠️ Optional (can host a small LLM or call an API)
Persistent memory	❌ No (stateless)	❌ No (limited context)	✅ Yes (session storage)
Real-time data access	❌ No (fixed rules)	❌ No (frozen at training)	✅ Yes (APIs, databases, web)
Code and tool execution	❌ No	❌ No (only generates text)	✅ Yes (tool orchestration)
Structured reasoning	✅ Yes (deterministic)	⚠️ Statistical (imitation)	✅ Yes (hybrid: rules + validation)
Action planning	❌ No	❌ No	✅ Yes (expert system plans)
Explainability	✅ Perfect (traceable)	❌ Black box	⚠️ Partial (traceable by expert system)
Adaptability	❌ Rigid (re-coding)	✅ Flexible (fine-tuning)	✅ Flexible (fine-tuning + new tools)
Application domain	⚠️ Very specific	✅ Generalist	✅ Generalist (specializable via tools)
Natural language interaction	❌ No (forms)	✅ Yes	✅ Yes (via integrated small LLM)
Result verification	✅ Yes (by rules)	❌ No (hallucinations)	✅ Yes (by expert system + RAG)
Can exist standalone?	✅ Yes (autonomous)	✅ Yes (via API)	✅ Yes (complete autonomous system)

Expert Systems vs LLMs vs AI Assistants

Expert system: "I know everything that man has coded into me, but that's all I can do."
LLM alone: "I generate text, and that's all."
AI Assistant: "I am a small augmented LLM, equipped with specialized tools and supervised by humans."

Example: Reading a PDF with a Standalone LLM

Context: The user has access to a standalone LLM (e.g., GPT, Mistral, DeepSeek, Gemini, ...). This model can only generate text. It has no access to files, folders, or execution tools.

Step 1: The user asks the LLM: "Find me articles on black holes in this folder and summarize them in English."
Step 2: The LLM analyzes the request. It understands that this is a file reading and summarization request.
Step 3: The LLM responds by explaining its limitations: "I cannot access the files on your computer, because I am a language model. Here is how you can proceed: open the folder, copy the content of the PDFs and paste it here. I will then be able to summarize them."
Step 4: The user manually opens the folder, copies the text from the relevant PDFs and pastes it into the conversation.
Step 5: The user asks: "Here is the text of the article. Summarize it in 5 sentences in English."
Step 6: The LLM generates a summary of the text provided by the user.
Step 7: The LLM presents the summary to the user. No verification is performed; the LLM may hallucinate or produce inaccuracies.

Conclusion: With a standalone LLM, the user must do all the work manually: open folders, copy texts, paste them, and review the summaries to verify their accuracy. The LLM only generates text from what the user provides.

Example: Reading a PDF with an AI Assistant (Complete System)

Context: The user has access to an LLM (e.g., ChatGPT, Mistral, DeepSeek, Gemini) via a prompt interface. This LLM is connected to a complete AI assistant (expert system + tools) that can act in the real world.

Step 1: The user types their prompt into the LLM: "Find me articles on black holes in this folder and summarize them in English."
Step 2: The LLM analyzes the request. It understands that it cannot respond alone (it has no access to files). It identifies that it needs an assistant to execute actions.
Step 3: The LLM calls the AI assistant (via an API) by sending it a structured request: "Interpret this request: read the folder, extract the PDFs, search for 'black hole', summarize."
Step 4: The assistant (expert system) receives the request from the LLM. It plans the actions: list files, filter PDFs, extract text, search keywords, summarize with the LLM, verify.
Step 5: The assistant uses its tools to execute each action (reading the folder, extracting text, searching keywords).
Step 6: The assistant receives the results from the tools (extracted texts, presence of the keyword).
Step 7: The assistant (expert system) calls the LLM with a summarization request: "Summarize this text in 5 sentences in English."
Step 8: The LLM generates the summary and sends it back to the assistant.
Step 9: The assistant (expert system) verifies the results: consistency, presence of the subject, absence of hallucination.
Step 10: The assistant assembles the structured response and sends it back to the LLM.
Step 11: The LLM reformulates the response in natural language and presents it to the user via the prompt interface.

Conclusion: The user only sees the LLM. It is the LLM that, behind the scenes, calls the assistant to execute the actions. The user never interacts directly with the assistant. Everything is transparent to them.

References

Brown, T. B., et al. (2020). Language Models are Few-Shot Learners. Advances in Neural Information Processing Systems (NeurIPS), 33, 1877-1901.
https://neurips.cc/virtual/2020/oral/18545 (new window)
Yao, S., et al. (2023). ReAct: Synergizing Reasoning and Acting in Language Models. International Conference on Learning Representations (ICLR).
https://collaborate.princeton.edu/en/publications/react-synergizing-reasoning-and-acting-in-language-models (new window)
Schick, T., et al. (2023). Toolformer: Language Models Can Teach Themselves to Use Tools. Advances in Neural Information Processing Systems (NeurIPS).
https://neurips.cc/virtual/2023/oral/73843 (new window)
Mialon, G., et al. (2023). Augmented Language Models: A Survey. arXiv:2302.07842.
https://huggingface.co/papers/2302.07842 (new window)
Lewis, P., et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Advances in Neural Information Processing Systems (NeurIPS).
https://neurips.cc/virtual/2020/public/poster_6b493230205f780e1bc26945df7481e5.html (new window)
OpenAI (2023). GPT-4 Technical Report. arXiv:2303.08774.
https://arxiv.org/abs/2303.08774 (new window)
Anthropic (2024). The Claude 3 Model Family. Technical Report.
https://www.anthropic.com/news/claude-3-family (new window)
Russell, S., & Norvig, P. (2020). Artificial Intelligence: A Modern Approach (4th ed.). Pearson.
LangChain (2024). Documentation: Building Applications with LLMs.
https://docs.langchain.com/oss/javascript/learn (new window)
Hugging Face (2024). Transformers Documentation: Agents & Tools.
https://huggingface.co/docs/transformers (new window)

FAQ: Your Questions About AI Assistants and the Limits of LLMs

Is an AI assistant smarter than a standalone LLM?

No. It is often built around a small LLM – sometimes less powerful than GPT-4 or DeepSeek. Its strength lies in its architecture: it combines the LLM with tools, persistent memory, and an expert system that plans and verifies actions. It is this hybridization that allows it to accomplish tasks that a standalone LLM cannot perform.

Why is a standalone LLM uncontrollable?

A standalone LLM is a statistical model that generates text without understanding or verification. Its responses can be hallucinated, biased or inconsistent, without traceability. The AI assistant, supervised by humans, allows guiding, interrupting or validating each step, transforming a blind system into a controlled tool.

Can the AI assistant execute dangerous actions?

Yes, this is a potential risk. This is why human control is essential: the assistant does nothing without validation or supervision. Safeguards are built in: limited permissions, audit logs, and verification loops. The assistant is designed to be a reliable extension of human judgment, not an autonomous agent.

What is RAG (Retrieval-Augmented Generation)?

RAG (retrieval-Augmented Generation) is a technique that enriches the LLM's responses with information from an external database. This helps reduce hallucinations and provide up-to-date, verifiable and sourced answers – an essential complement to the assistant's reliability.

What is the difference between an autonomous agent and a supervised AI assistant?

An autonomous agent acts without human intervention. A supervised AI assistant assists the human without replacing them: each action is validated or guided by the human. The assistant is a tool that augments human capabilities, while the autonomous agent aims to automate a task – with the associated risks in terms of control and responsibility.

Can AI assistants understand humor, irony or innuendo?

Partially. The LLM integrated into the assistant can detect certain language traits like irony, humor or innuendo, because it has been trained on a vast corpus of human texts. But it does not understand them in the human sense: it does not perceive intention, emotion or social context. It is a statistical imitation, not true understanding. The AI assistant can sometimes misinterpret an ironic remark as a serious request – hence the importance of human supervision to correct these interpretation errors.