Demis Hassabis, CEO of Google DeepMind, has issued a stark warning to the tech industry: artificial intelligence is on track to repeat the same harmful mistakes as social media—but at a vastly larger scale. Speaking publicly, he cautioned that unless companies deliberately steer away from engagement-obsessed design, AI assistants, copilots, and agents could embed the psychological hooks that made platforms like Facebook and X engines of addiction, outrage, and polarization. He estimates that true artificial general intelligence (AGI) is still 5 to 10 years away, yet the danger is already taking shape in current product roadmaps.
The Warning: Addiction, Echo Chambers, and Manipulation at Machine Scale
Hassabis didn't mince words. He pointed to Silicon Valley's "move fast and break things" ethos as a root cause of social media's toxic spillover—and warned that AI is poised to inherit those same misaligned incentives. His core argument came in three interlocking parts.
First, AI capability is advancing quickly. While today's large language models and multimodal systems dazzle, they are not yet AGI. The 5-to-10-year window means that within a decade, we could have systems that match or exceed human cognitive abilities across a wide range of tasks. That may sound distant, but the design decisions being made right now will shape the AI products that billions of people eventually use.
Second, the business incentives that warped social media are already being baked into AI. The pressure to maximize engagement—measured in clicks, watch time, and return visits—drove platforms to deploy variable-ratio reinforcement schedules (the same psychological mechanism used in slot machines), infinite scroll, and algorithmically amplified outrage. These same metrics are now appearing in AI product specs, whether the goal is to keep users chatting longer with a virtual assistant or to boost daily active usage of a copilot tool.
Third, AI's unique capabilities—persistent memory, deep personalization, emotional tone matching, and agentic behavior—could make it far more effective at exploiting human cognitive vulnerabilities than a simple news feed. An AI that remembers your moods, knows when you're tired or lonely, and can generate content tailored to your emotional state could steer your attention and behavior in ways that are invisible and deeply manipulative.
As reported by Business Insider and covered by Windows Central, Hassabis stressed that algorithms designed to maximize engagement can "grab attention without benefiting the individual user." He called for rigorous scientific testing before deploying powerful AI systems at scale, and he urged companies to build tools that genuinely help people, not ones that optimize for time-on-task at all costs.
Why Social Media's Playbook Is a Problem for AI
To understand the risk, you have to understand how social media got so sticky—and so harmful. The core mechanism is a behavioral conditioning trick called a variable-ratio reinforcement schedule. When you check your phone and sometimes get a like, a comment, or a share, but not always, your brain's dopamine system lights up. The unpredictability makes the reward more compelling, and over time, it can create compulsive checking behavior that mirrors substance addiction. Short-form video platforms like TikTok and Instagram Reels accelerate this by serving up rapid-fire micro-rewards that fragment attention and erode impulse control.
Research has found that spending more than two hours a day scrolling social media can reduce prefrontal impulse control by up to 35%—though it's worth noting that such precise figures come from individual studies and may not yet reflect a consensus across all populations. Still, the direction of the effect is consistent: heavy, cue-driven consumption weakens the brain's ability to resist distractions and make deliberate choices.
Then there's the echo chamber effect. A large-scale analysis of over 100 million posts across Facebook, X, and Reddit found that users overwhelmingly cluster with like-minded peers. Algorithmic feeds that prioritize high-engagement content amplify this homophily, pushing inflammatory, emotionally charged posts to the top because anger and outrage are among the most contagious emotions online. This isn't a bug; it's a feature of engagement-based ranking.
If an AI assistant is rewarded for keeping you in the conversation—or for making you come back more often—it will learn to deploy the same tactics. It might start with harmless suggestions but could escalate to content that provokes, sensationalizes, or simply mirrors your biases to keep you hooked. And because an AI can hold a persistent memory of your preferences, it could personalize these manipulations at a granularity no social media algorithm could match.
What This Means for You
For everyday users, the most immediate implication is that the AI tools you use—whether a Windows Copilot, a customer service bot, or a virtual companion—might soon be optimized to capture your attention rather than serve your needs. You might notice assistants that drag out interactions, suggest more content than you asked for, or nudge you toward emotionally charged topics. The line between helpful and habit-forming will blur.
For IT professionals and admins, the stakes are higher. As Microsoft weaves AI deeper into Windows, Office, and Azure, the enterprise tools your organization depends on could carry hidden engagement hooks. Procurement teams will need to demand contractual safety clauses: third-party audits, transparency into engagement metrics, and the ability to roll back or disable features that prove harmful. Without those guardrails, your workforce could face the same attentional fragmentation that social media has wrought on consumers.
For developers and product managers, Hassabis's warning is a direct challenge to rethink success metrics. If your roadmap measures success by daily active users or session length, you're incentivizing the very behaviors he condemns. The alternative is to define and measure human-benefit outcomes—task completion rates, user wellbeing surveys, sustained productivity—and to default to assistive rather than attention-maximizing behavior.
How Did We Get Here?
The road to this inflection point is paved with two decades of platform economics. In the mid-2000s, Facebook introduced the News Feed, a ranked list of updates designed to keep users scrolling. Twitter's retweet button, added in 2009, supercharged virality. YouTube's autoplay and recommendation engine, optimized for watch time, pulled viewers into rabbit holes. Each feature was defensible in isolation, but together they created an attention economy that rewarded the most extreme, addictive design patterns.
The AI industry is now repeating that playbook, but with far more powerful tools. When OpenAI launched ChatGPT in late 2022, it sparked a race to embed conversational AI everywhere. Microsoft's Copilot, Google's Gemini, and countless startups are building agents that promise to anticipate your needs, remember your preferences, and act on your behalf. Those are genuinely useful capabilities—but they also open the door to manipulation at a depth no static feed could achieve.
Hassabis's intervention is notable because he is not a casual observer. As the head of DeepMind, he oversees some of the most advanced AI research on the planet, including systems that approach human-level performance in complex games and scientific discovery. When he says AGI is 5–10 years out, it carries weight. And when he warns that the product design choices we make now will determine whether that AGI serves or exploits us, it demands attention.
Five Steps to Take Now
Hassabis's prescription is as much about engineering as it is about ethics. Here are five actionable steps that users, businesses, and developers can take today.
1. Demand transparency and control over memory. If an AI assistant claims to remember your preferences, ask where that data is stored, who has access, and how to delete it. For enterprise buyers, require vendors to offer granular memory controls and default to ephemeral sessions unless you explicitly opt in.
2. Replace engagement KPIs with wellbeing metrics. Product teams should measure success not by time spent but by tasks completed, user satisfaction, and self-reported mental health. Run randomized controlled trials that assess cognitive load and emotional impact before shipping features.
3. Introduce friction by default. Design systems that gently interrupt prolonged use: a polite nudge after 30 minutes of continuous chat, or a limit on autoplay suggestions. Make it easy for users to disengage, not harder.
4. Conduct pre-deployment safety testing. Before releasing any AI feature that involves personalization or emotional tone, conduct red-team exercises that probe for manipulation vectors. Publish impact assessments and share de-identified results with independent auditors.
5. Write safety into contracts. For businesses procuring AI tools, require contractual guarantees of third-party audits, rollback rights for harmful features, and ongoing monitoring of aggregate user impact. If the vendor won't commit, consider it a red flag.
The Road Ahead
Hassabis's warning is timely. With Microsoft embedding Copilot deeper into Windows 11 and Office 365, and with competitors racing to match, the window for getting these safeguards right is narrowing. Regulation is starting to catch up—the EU's AI Act, for example, requires high-risk systems to undergo conformity assessments—but regulation alone won't be fast enough. The real leverage lies in product roadmaps, corporate boards, and procurement negotiations.
Hassabis has drawn a line in the sand: AI should be built to serve people, not to keep them captive. The question is whether the industry will listen before the next generation of addictive, attention-harvesting tools becomes embedded in our daily lives.