The public debate on AI has been almost entirely monopolized by LLMs since 2022. ChatGPT, Claude, Gemini. People writing prompts, companies integrating language APIs. It’s a fraction of what AI does in the real world.

Your phone unlocks on your face thanks to an AI. Your Instagram feed is ordered by an AI. Fraud on your bank card is detected in real time by an AI. The car that warned you of a failing gearbox may have used a predictive AI. None of these AIs speak. None of them are LLMs.

Computer vision

Computer vision is the ability of a system to identify, locate and classify objects or patterns in images or videos. It is one of the most mature applications of deep learning.

What goes on in production: quality control on manufacturing lines (detecting visual defects at inhuman speeds), sorting of parcels in logistics, automatic reading of license plates, mammography and medical imaging analyses, worksite surveillance, biometric authentication.

A convolutional neural network (CNN) trained on millions of medical images can identify certain pathologies with an accuracy comparable to that of a senior radiologist on image types well represented in its data. Important words: “certain pathologies”, “well-represented image types”. Outside distribution, performance falls off.

The recommendation

Netflix offers series. Spotify creates playlists. Amazon suggests purchases. These are not LLMs. These are collaborative filtering and matrix factorization systems combined with deep learning, trained on billions of user interactions.

Recommendation is the AI that has had the longest direct economic impact. It optimizes a metric (clicks, viewing time, conversions). What it optimizes may diverge from what is good for the user. This is a property of the system, not a flaw to be corrected in the next version.

Anomaly prediction and detection

Bank fraud detection. Machine failure prediction. Credit scoring. Network intrusion detection. These systems have been in production for 15 years in sectors that never make the cover of Wired.

A fraud detection model analyzes dozens of features in real time (amount, location, time, history, device, browsing behavior) and produces a fraud probability score. If the score exceeds a threshold, the transaction is blocked or put under review. This system is not perfect. It has a false positive rate (legitimate transactions blocked) and a false negative rate (fraud not detected). The challenge is to optimize this compromise based on your actual data.

These systems make mistakes that have real consequences. Incorrect credit scoring affects a real person’s access to financing. An over-aggressive anti-fraud filter blocks legitimate customers. These biases are documented, regulated (the AI Act in Europe, the Fair Credit Reporting Act in the U.S.), and are not solved problems.

What a difference it makes to see the whole picture

Two practical implications.

To assess the real risks: AI risks are not just in LLMs. They also lie in the silent automated decision-making systems that affect rights and access without the people concerned being informed. The European AI Act provides a precise framework on this point.

**A computer vision project for quality control does not have the same risk profile, training costs or skill requirements as an LLM project. Confusing them because they’re both labeled “AI” leads to bad decisions.

Industrial, silent, non-LLM AI is the kind that has proven its economic value over decades. It deserves as much attention as the latest chatbot.