Bold and Responsible: Google's James Manyika on How to Build AI That Scales
At the NDTV Ind.ai Summit, Google's James Manyika made a clear point for builders: being bold and being responsible with AI are not opposites. "We actually don't see these as different... both are true and important to work on."
He pointed to India and the Global South as proof that ambition matters, citing "the scale of the possibilities... at this population scale." At the same time, he backed strong guardrails and smart policy. "AI is too important not to regulate and is also too important not to regulate well."
Stuart Russell's warning-humanity still lacks a full answer to what happens if machines start "thinking"-got acknowledgement, not dismissal. Manyika's stance: face real risks head-on, and don't stall the useful deployments that improve people's lives.
What "bold" means for builders
- Target population-scale problems: language access, healthcare triage, learning support, citizen services, agricultural advisories, and SME digitization.
- Design for constraints: low bandwidth, low-cost devices, intermittent connectivity, multilingual UX, and voice-first flows.
- Optimize for cost-to-serve: retrieval-first patterns, caching, smaller distilled models, and on-device inference where practical.
- Measure real outcomes: task success rate, time saved, cost per resolved task, and user trust scores-not vanity metrics.
What "responsible" looks like in practice
- Adopt clear principles and an internal review path for sensitive use cases. See Google's approach: AI Principles.
- Risk-tier your systems (e.g., advisory vs. medical triage) and match controls to risk: stronger review, tighter monitoring, and stricter access for higher-risk features.
- Privacy by design: data minimization, regional storage when required, consent flows, and secure retention policies.
- Red-teaming and adversarial testing before launch; repeat on every major model or prompt change.
- Guardrails at multiple layers: prompt hardening, tool-use gating, content filters, rate limits, and abuse detection.
- Auditability: model cards, data lineage, decision logs, and reproducible prompts for high-stakes actions.
Regulation with a two-sided view
Manyika argued for rules that reduce risk and also keep space for useful innovation. That means compliance that helps teams ship safer systems faster-not paperwork theater.
- Use a risk framework to drive controls, not opinions. A good anchor: NIST AI Risk Management Framework.
- Map features to applicable laws and standards early (e.g., EU AI Act categories, sector rules). Build "policy as code" checks into CI/CD.
- Prepare for audits: document datasets, model choices, eval results, and known limitations; keep incident playbooks current.
Engineering checklist for LLM and generative systems
- Data and security: classify inputs/outputs, scrub PII, license-check training and RAG corpora, encrypt at rest/in transit, enforce RBAC.
- Models and prompts: keep a versioned prompt library, use toolformer-style patterns for safe tool use, add function-call schemas, and constrain outputs.
- Evals: golden datasets per use case; track helpfulness, factuality, bias, toxicity, and refusal quality; run evals on every release.
- Safety gates: jailbreak tests, prompt-injection defenses, URL and file sanitization, and allow/deny lists for tools and connectors.
- Monitoring: latency, cost per call, drift, hallucination rate, safety events, and user feedback loops; set SLOs and automated rollback.
- Incident response: kill switches, data purge routines, alerting on abuse patterns, and clear on-call ownership.
Building for India and the Global South
- Multilingual by default: support major local languages and code-mix; prioritize ASR/TTS quality and low-latency translation.
- Offline and low-bandwidth modes: compress models, quantize where sensible, cache retrieval, and fall back gracefully.
- Device diversity: progressive enhancement for high-end phones; functional parity for low-end Android hardware.
- Cost controls: price caps per session, batch heavy jobs, choose smaller models with smart retrieval, and monitor GPU burn against value delivered.
Key takeaways for IT and development teams
- Be ambitious about impact at scale-don't sandbag useful applications because risk feels abstract.
- Pair every big bet with a concrete safety plan: risk tiering, evals, guardrails, monitoring, and audit trails.
- Treat regulation as design input, not an afterthought; good rules can help teams ship the right things faster.
- Invest in the platform: LLMOps, prompt/version management, policy as code, and an eval pipeline are compounding advantages.
Resources
- Google AI Principles - reference for principle-driven development.
- NIST AI RMF - practical risk controls and governance patterns.
- AI for IT & Development - courses and resources to build useful AI while keeping operations accountable.
As Manyika put it, you don't have to choose. Be bold because it helps people. Be responsible because people count on you.
Your membership also unlocks:
