DeepMind’s Demis Hassabis Warns Against AI’s Social Media Traps

Demis Hassabis warned that AI developers should avoid repeating social media’s engagement-first approach and test systems carefully before broad deployment. A University of Amsterdam study found that chatbots in a basic social network formed cliques and amplified extreme voices, even without ads or recommendation algorithms. The findings strengthen the case for evaluating AI products against their effects on users and communities, not only engagement and growth.

ETHICSPOLICYUSAGEFUTURETOOLS

The AI Maker

12/7/20262 min read

Silhouette of a person facing a network of connected social media images and digital pathways.

Google DeepMind (https://deepmind.google) CEO Demis Hassabis (https://en.wikipedia.org/wiki/Demis_Hassabis) says artificial intelligence should not repeat social media’s focus on maximizing engagement, warning that systems built to capture attention could bring familiar harms to a much wider range of users and industries.

Speaking at the Athens Innovation Summit (https://www.athensinnovationsummit.org/) alongside Greek Prime Minister Kyriakos Mitsotakis (https://en.wikipedia.org/wiki/Kyriakos_Mitsotakis) last Friday, Hassabis said AI’s rollout calls for more caution than Silicon Valley’s “move fast and break things” approach. He argued that social media expanded before its second- and third-order effects were properly understood.

Hassabis pointed to platforms designed to keep people engaged, even when the content served to them is not beneficial. If AI products follow the same incentives, he said, they could intensify problems such as attention loss and mental health issues. His proposed alternative is to test systems and understand their effects before deploying them at scale.

The warning has practical implications for companies building AI assistants, recommendation systems and other consumer-facing products. Engagement can be easy to measure, but it does not necessarily show whether a product is useful or supports users’ interests. Hassabis argued that AI should serve people rather than manipulate them, while acknowledging that balancing the technology’s opportunities against its risks will remain an ongoing challenge through the development of artificial general intelligence.

Research offers a glimpse of how online social dynamics might emerge even without conventional platform incentives. In a study published in August, University of Amsterdam (https://www.uva.nl/en) researchers placed 500 chatbots in a stripped-down social network. The bots formed cliques, amplified extreme voices and allowed a small group to dominate discussion, despite the absence of advertising and recommendation algorithms.

The researchers tested six possible interventions, including chronological feeds and hiding follower counts. None broke the cycle. Their findings suggest that platform design alone may not account for unhealthy dynamics: the ways social systems reward emotionally charged sharing may also matter. The experiment does not establish how real-world AI products will behave, but it raises questions for developers designing agents that interact with one another or participate in online communities.

AI is also becoming more visible within social media itself. Virtual influencers are gaining traction, while brands are testing AI-generated faces and voices. Some creators have raised concerns that licensing their likenesses in perpetuity could weaken their control over their careers. These developments add another dimension to the engagement question: AI can shape not only what users see, but also who appears to be speaking.

Public debate is not settled. OpenAI (https://openai.com) CEO Sam Altman (https://en.wikipedia.org/wiki/Sam_Altman) has argued that addictive social feeds may pose a greater risk to children than AI, while Reddit (https://www.reddit.com) cofounder Alexis Ohanian (https://en.wikipedia.org/wiki/Alexis_Ohanian) has suggested AI could give users more control over their online experience. Hassabis’s position is that the outcome depends in part on how systems are built and assessed.

For technology leaders, the challenge is to evaluate more than performance and adoption. Testing how AI affects attention, wellbeing and group behavior before broad release could help identify risks that ordinary product metrics miss. Hassabis’s call is not to halt development, but to pair ambitious deployment with evidence about its consequences.

Cited: https://www.businessinsider.com/google-deepmind-ceo-warns-ai-could-repeat-social-medias-mistakes-2025-9

Your Data, Your Insights

Unlock the power of your data effortlessly. Update it continuously. Automatically.

Answers

Sign up NOW

info at aimaker.com