Humanizing AI: keeping technology human-centric in the age of machines

Executive overview

AI has made rapid gains in cognitive intelligence but remains emotionally blind — it processes what we say, not how we say it or what we feel. Rana el Kaliouby argues the next frontier is emotional and social intelligence (EQ) in machines, not just IQ. Without it, AI will keep optimizing for the wrong things.

The fix isn't just technical. It requires better benchmarks, more diverse builders, and users who vote with their wallets for human-centric AI.

Core insight: we only build what we measure for — and today's AI benchmarks measure IQ, not EQ.

The EQ gap in AI

• Only 7% of human communication is words; 93% is nonverbal — facial expression, tone, gesture
• AI today is almost entirely focused on what is said, not how or why
• Humanoid robots can fold laundry but can't interact naturally with people — teams optimized for function, not human fit
• All major AI benchmarks are IQ-focused; there are no widely adopted EQ benchmarks
• Progress on AGI requires marrying cognitive and emotional intelligence

Fact or fiction: five AI myths

• AI bubble — mostly fiction. Pre-product companies raising at billion-dollar valuations are a red flag, and circular capital flows between hyperscalers are worth watching. But real products transforming real industries are early and very real.
• Robots taking over — partly true, differently than feared. Robots will absorb repetitive and dangerous jobs (e.g. ship welding, where human supply is already short). Sci-fi takeover scenarios are not the risk.
• AI is bad for creators — fiction. AI democratizes creation and lowers the floor; the premium shifts to human originality, lived experience, and distinctive perspective.
• AI will outsmart humanity — fact, and probably fine. Let AI be more intelligent; humans should double down on intuition, embodied wisdom, and emotional depth — what Ariana Huffington calls "the GPS of the soul."
• AI is a boys club — fact. Women are underrepresented as founders and are receiving less funding. The economic gap will widen if this isn't corrected.

Human skills that rise in value

• Collaboration — with both humans and machines
• Communication — original, human voice stands out as AI-generated text becomes identifiable
• Critical thinking — still essential for evaluating AI outputs
• Creativity — the ceiling rises even as the floor goes up

AI-native devices and world models

• Smartphones are pre-AI devices; the next form factor (glasses, wearable pin, or unknown) must be perceptual, conversational, empathetic, contextual, and ambient
• World models are to physical AI what large language models are to text — they encode real-world physics and spatial understanding, not just internet text
• Training data for world models comes from humans wearing cameras in real environments (kitchens, streets, workplaces)
• Unlocking physical AI — robotics, AI-native hardware — requires world models

AI therapy and companions

• AI can provide real value as a 24/7 support presence, especially at 2am when no human is available
• Should not replace human relationships or human therapeutic oversight
• Guardrails are severely underdeveloped; cases of self-harm linked to AI chatbots are a documented risk
• Every deployed model should be tested against AI safety benchmarks before release
• Bots deployed for narrow purposes (shopping, customer service) will inevitably receive emotional conversations — guardrails must account for this

Keeping up as workflows change

• Organizations should encourage teams to experiment, accept mistakes, and lean into new tools
• The risk group is the middle layer — senior leaders and AI-native new hires are adapting; mid-level workers face the most disruption
• AI agents are already filling junior roles: Rana's fund uses an AI chief of staff named Blue for research, CRM updates, and routine tasks
• All workflows are being reimagined as human-AI collaboration, not human-only or AI-only

Separating signal from noise in AI investing

• Defensibility is now time-bounded — a moat that exists today may be obsolete after the next model release
• Key questions: how defensible is the technology in one to five years, not just now?
• Complexity of the problem and quality of IP or data moats matter more than current traction
• Red flags: pre-product, pre-revenue companies at inflated valuations; founders who haven't thought about ethics or safety at all

What individuals can do

• Vote with your feet — choose AI tools from companies that take ethics, bias, privacy, and safety seriously
• Stay curious and playful; experiment even when outcomes are imperfect
• Be vocal about demanding transparency on how models are built, validated, and deployed
• Nurture in-person human connection alongside AI adoption — both are necessary