Threat Picture

Latest Cybersecurity News

Yoshua Bengio's Search for AI We Can Keep Under Control

A leading figure in teaching computers to learn from data, he founded LawZero to explore a different design: AI built to explain and predict without pursuing goals of its own.
By Charles Joseph · Updated
Pen-and-ink watercolor portrait of Yoshua Bengio with curly gray hair, a white beard and a dark shirt.
Yoshua Bengio. Portrait adapted from a photograph by Xuthoria (CC BY-SA 4.0).
Share
Share
Copy Link

Yoshua Bengio is a computer scientist who helped advance deep learning, a way for computers to learn patterns through many layers of calculations. He founded the Montreal research institute Mila and helped develop methods that let computers learn useful relationships between words.

At LawZero, the nonprofit he later founded, he's pursuing a different challenge: using one AI to check whether another AI's planned actions could cause harm. His proposed checker would predict consequences without being rewarded for making a particular outcome happen.

The project, called Scientist AI, builds on a career spent studying both the possibilities of learning systems and the reasons they struggle.

A Learning Problem in Montreal

Bengio earned his computer-science doctorate at McGill University in 1991. After postdoctoral research at MIT and Bell Laboratories, he joined the Université de Montréal in 1993 and began building a research group there.

His interests centered on neural networks, systems of connected calculations that learn from examples. Training adjusts the numbers controlling those connections, helping a network recognize useful patterns. With many layers of these calculations, the approach is called deep learning.

Some tasks required remembering information across a sequence. In a sentence, for example, a word near the end may depend on something introduced much earlier.

Recurrent neural networks try to handle sequences by passing information from one step to the next. But teaching a network what to preserve over many steps proved difficult.

In a 1994 paper, Bengio, Patrice Simard and Paolo Frasconi examined that problem. The signal used to adjust the network during training had to travel through all the intervening steps, making distant connections harder to learn.

The analysis helped explain why giving a system a form of memory wasn't enough. Researchers also needed training methods that could teach it to keep the right information and use it at the right time.

Sponsored

Teaching a Model About Words

A language model learns to estimate which words are likely to follow others. But language has far more possible combinations than any collection of training text can contain.

In a 2003 paper, Bengio, Réjean Ducharme, Pascal Vincent and Christian Jauvin explored a way to share what a model learned across similar words. Their system assigned each word a list of numbers, learning those numbers while also learning to predict word sequences.

Consider “cat” and “dog.” The words can play similar roles in a sentence. Learning that relationship could help a model judge a new sentence about a dog after seeing a similar sentence about a cat, even if it hadn't encountered that exact combination before.

The lists of learned numbers, often called embeddings, gave the computer a way to work with those relationships. Through learning from examples, it could make useful predictions about text it hadn't already seen.

Letting a Translation Refer Back

Translation added another difficulty. A system had to preserve the meaning of an input sentence while producing words in a different language, often in a different order.

Early neural approaches could squeeze the source sentence into a fixed-size numerical summary, then use that summary to generate a translation. Packing all the relevant information into one representation created a bottleneck, especially for longer sentences.

In work first published in 2014, Dzmitry Bahdanau, Kyunghyun Cho and Bengio proposed letting the translation system consult different parts of the source as it worked. Their paper on jointly aligning and translating appeared at the ICLR conference in 2015.

As the system produced each translated word, it assigned different weights to positions in the original sentence. Relevant source words could contribute more strongly to the next step.

This use of attention let the model select useful information as it worked. It no longer had to depend entirely on one compressed summary.

The team tested the method on English-to-French translation. It was another example of changing how information moved through a learning system to make a difficult task more manageable.

A Laboratory Grows Into Mila

The Montreal group grew alongside this research. What began as Bengio's laboratory in 1993 became the foundation for Mila, an institute bringing together researchers working on language, vision, generative models and other areas of machine learning.

Mila's institutional history describes a broader university collaboration established in 2017, followed by nonprofit status in 2018. The institute gave researchers a shared base for developing ideas, training students and working across disciplines.

He remains a professor at the Université de Montréal and is Mila's scientific advisor. His current leadership of LawZero is a separate role, as its co-president and scientific director.

In 2019, the Association for Computing Machinery announced that Bengio, Geoffrey Hinton and Yann LeCun would share the 2018 Turing Award. It recognized their contributions to deep neural networks.

The three researchers had advanced different parts of that work, helping turn it into a practical approach to recognizing images and processing language.

Sponsored

A Change in the Research Question

Bengio dates a major shift in his priorities to 2023. In his first-person account of founding LawZero, he connects concern about advanced AI with the future facing his children, grandchild and students.

He wanted to concentrate on reducing the possibility of catastrophic harm. The technical question became whether researchers could build powerful systems whose design gave them fewer reasons to manipulate people or pursue unwanted objectives.

LawZero launched in June 2025, after being incubated at Mila. The organization says its nonprofit structure is intended to help protect that research mission from commercial pressure.

Its Scientist AI proposal focuses on explaining and predicting. A system following this design would assess what is likely to be true, or what might follow from an action, without being trained to seek a particular outcome in the world.

In Bengio's proposed design, a system asked whether an action could cause harm would gain nothing from making that harm happen. It would also have no reward for preventing someone from checking its answer.

Bengio is working on the public evidence question as well. He chairs the International AI Safety Report, a collaborative assessment of capabilities, risks and possible safeguards, and co-chairs a separate UN scientific panel with journalist Maria Ressa.

How the Proposed Checker Would Work

A March 2026 paper by Damiano Fornasiere, Oliver Richardson, Gaël Gendron, Iulian Serban and Bengio describes a design that pairs two functions. One generates possible explanations; another estimates their probabilities.

The proposed system would also distinguish a statement someone made from an established fact. Evidence that a person reported an event would count as evidence of the report, without automatically establishing that the event happened.

In July 2026, LawZero published a formal safety argument for the approach. The argument depends on stated assumptions, which the team still needs to examine through experiments.

Accurately predicting what would happen under a different choice remains an open problem. An honest model can also make errors, and a person could misuse its answers even if the model has no objective of its own.

LawZero wants to use the predictor to check another AI's plans before it acts. That would make the accuracy of its answers more than a research score: an error could let a harmful action through. Testing those judgments is part of the work still ahead.

QUOTE:

"Amateurs hack systems, professionals hack people."