Geoffrey Hinton Leaves Google to Warn of AI Risks

Geoffrey Hinton, one of the key researchers behind the development of modern neural networks, has left Google after nearly a decade at the company. The 75-year-old British-Canadian scientist wants to be able to speak freely about the risks of generative artificial intelligence, just as tools such as ChatGPT and Bard have brought the technology to the general public.

Hinton announced his departure on Monday, although he left his position the previous week. Hinton told The New York Times that he had left the company so he could speak about the dangers of AI. In a post published afterward, he said Google had acted responsibly in developing the technology.

His departure matters because of who is making it. Hinton was one of the leading champions of neural networks: programs loosely inspired by how the brain works and capable of learning patterns from vast amounts of data. For years, they were a niche research field; today, they underpin conversational assistants, image generators, machine translators and much of the technology used for speech and computer vision recognition.

From deep learning to generative AI

In 2012, Hinton and his students Alex Krizhevsky and Ilya Sutskever showed that a neural network trained on massive image datasets could dramatically improve visual-recognition results. That work, known through the AlexNet system, accelerated the industrial adoption of so-called deep learning.

Google acquired DNNresearch the following year. The company had been founded by Hinton and two of his students, and the researcher joined Google. There, he combined his work at the company with teaching at the University of Toronto.

The difference now lies in the speed and scale. Language models do more than classify a photograph or suggest a word: they write text, program, answer questions and hold convincing conversations. They do not understand the world as a person does, but they can produce content persuasive enough to change how people get informed, work and make decisions.

Disinformation no longer needs major media outlets

Hinton’s most immediate concern is that AI will make it easier to produce fake content at scale. A manipulated image once required editing skills; a convincing video demanded even more resources. Generative systems lower that barrier, making it possible to create fake text, voices and images—and, increasingly, videos—at low cost.

The risk is not limited to a single forgery. A campaign can tailor thousands of messages to different audiences, imitate styles and respond in real time. That makes verification harder for journalists, platforms and citizens, especially amid elections and growing political polarization.

Hinton has also pointed to the potential impact on jobs. Automation had already transformed physical and repetitive tasks; generative models are now reaching parts of office work, including drafting documents, summarizing reports, handling queries, writing code and producing marketing materials. That does not mean all those professions will disappear, but companies and workers will have to decide which tasks to delegate to a machine and who is accountable for its errors.

A debate reaching Big Tech from within

Hinton’s departure does not amount to a total rejection of AI research. He helped build its foundations and acknowledges its potential benefits in areas such as health care, education and productivity. His warning is about the lack of time and adequate mechanisms to understand systems advancing rapidly and already being deployed to millions of users.

Google has presented Bard as its answer to ChatGPT, which OpenAI launched in November 2022. Microsoft, OpenAI’s main partner, has already integrated similar technology into its Bing search engine and office products. Competition is pushing companies to release tools before their limitations are fully understood.

The question is not whether AI should be stopped altogether, but under what rules it should be developed and distributed. Companies will have to improve safety testing, identify artificially generated content and explain more clearly what data their systems use. Governments and regulators, for their part, face the challenge of protecting the public without turning rules designed for traditional software into an ineffective or impossible-to-enforce brake.

The fact that a voice so closely associated with the birth of deep learning has publicly joined these warnings changes the tone of the debate. The discussion around AI is no longer just about what a chatbot can do, but who controls its deployment and what safeguards must be in place before its errors affect millions of people.