11 AI questions UC Berkeley experts are watching in 2026
AI went from toy to infrastructure fast. It's now funding stock runs, writing code and papers, and quietly mediating how people work, learn, and relate. With that scale come risks we can't shrug off: privacy, bias, political conflict, and a flood of convincing fakes. Here's what leading UC Berkeley researchers say to watch this year - and how science and research teams can act on it.
1) Will the AI bubble burst? - Stuart Russell
Data center buildouts are now rivaling the largest tech projects in history. Yet revenue lags, model performance feels flat, and there are real limits to what current LLMs can learn efficiently. If the bubble pops, the fallout could be ugly; if it doesn't, the path likely runs straight through far more capable systems - and control gets harder.
- Track fundamentals: unit economics, energy constraints, and model capability per dollar, not hype.
- Invest in evaluation science and safety tooling alongside capability work.
- Scenario-plan for both contraction (compute scarcity) and breakthrough (alignment and governance bottlenecks).
2) Can we trust anything anymore? - Hany Farid
High-quality deepfakes are becoming routine, cheap, and fast. The asymmetry is brutal: it takes minutes to fabricate, days to debunk, and the damage lingers. Journalism, courts, and elections are all exposed.
- Adopt provenance standards like C2PA across your media workflows.
- Stand up internal verification pipelines and "pre-bunking" playbooks for likely hoaxes.
- Budget for human review and crisis response - detection alone won't save you.
3) AI-enabled discoveries that benefit people - Jennifer Chayes
Personalized agents and model-assisted science are accelerating discovery across labs and industries. The upside is real - if we center responsible use and inclusive access. The task is to translate capability into measurable human benefit.
- Embed ethics, security, and data governance in project kickoff, not post hoc.
- Prioritize reproducibility: notebooks, datasets, model cards, and ablations by default.
- Co-design with domain experts and end users; define "benefit" up front.
4) Privacy risks created by chatbot logs - Deirdre Mulligan
People pour intimate details into chatbots. Those logs can surface in lawsuits, law enforcement requests, or vendor audits. The risk is bigger than a privacy policy - it's your exposure map.
- Choose vendors with strict retention controls and enterprise-grade privacy terms.
- Minimize data: redact, truncate, or run on-prem where feasible.
- Publish clear guidance for staff on sensitive prompts and export controls.
- Prepare legal processes for data requests and incident response.
5) Relationship chatbots and human isolation - Jodi Halpern
Companion bots are moving from teens to toddlers without real guardrails. Dependence is rising, while social skills and empathy may erode. The downstream effects touch mental health and civic life.
- Support longitudinal studies on developmental and mental health impacts.
- Push for age gating, usage friction, and duty-of-care standards.
- Design for augmentation of human ties, not replacement.
6) Can robots learn useful manipulation? - Ken Goldberg
Claims that humanoids will "soon" replace skilled workers gloss over a stubborn fact: dexterous manipulation is hard, and robot data is scarce. There's a yawning gap between LLM-scale corpora and what's available for real-world hands.
- Invest in shared teleoperation datasets, sim2real pipelines, and self-supervised tactile learning.
- Benchmark on practical tasks with clear success metrics, not sizzle reels.
- Set timelines that reflect data bottlenecks and safety requirements.
7) Progress on worker technology rights - Annette Bernhardt
Unions and advocates are pushing policy on electronic monitoring and algorithmic management. Expect movement on guardrails for surveillance, automated firing, discrimination, and rights to appeal. Human say in high-stakes decisions is becoming table stakes.
- Map every automated decision that affects employment or access to services.
- Institute bias testing, explanations, and human-in-the-loop for critical calls.
- Document data sources, features, and error handling - assume audit.
8) Weaponization of AI against workers - Nicole Holliday
Speech-scoring systems rate "charisma" and "rapport," often without consent or clarity. These models tend to penalize neurodivergent speakers, second-language English, and stigmatized dialects. It's a discrimination risk wrapped in dashboards.
- Ban automated scoring for performance management unless methods, data, and error rates are disclosed and validated.
- Require informed consent and accessible opt-outs.
- Provide accommodations and alternative assessments by design.
- Monitor litigation trends and update policy accordingly.
9) AI's effect on political conflict - Jonathan Stray
"Politically neutral AI" sounds good until you try to define or test it. Bias can drift or hide in subtle behavioral changes. The open questions: should an AI persuade at all, and what does enforceable neutrality look like?
- Set explicit policies on persuasion, advocacy, and election content in your models and apps.
- Build evaluation suites for political topics and viewpoints; log changes and run regression tests.
- Red-team for subtle style shifts, agenda setting, and selective omission.
10) More sophisticated deepfakes - Camille Crittenden
High-end manipulation is going mainstream at scale: political disinfo, abuse material, and fraud are the predictable outcomes. New authenticity requirements help, but they won't close the gap alone. Media literacy and technical provenance have to work together.
- Combine watermarking, provenance, and forensic checks in one pipeline.
- Train staff and users to spot and report suspicious content quickly.
- Align with the NIST AI Risk Management Framework for governance and controls.
11) Intelligence limits and the search for truth - Alison Gopnik
There may be no single "general intelligence." Useful progress could come from agents that experiment, engage the world, and learn for the sake of truth, not human scoring. Curiosity-driven systems might be the next step forward.
- Explore intrinsically motivated RL and grounded, tool-using agents.
- Reward discovery and predictive accuracy over human preference alone.
- Publish open benchmarks that require exploration and causal reasoning.
What this means for research leaders
Expect more convincing fakes, more scrutiny of data practices, and louder debates over what AI should be allowed to decide. Build verification layers, tighten privacy discipline, and keep humans in the loop for high-stakes calls. Be conservative on robotics timelines, aggressive on documentation, and honest about limits.
If your team needs structured upskilling to meet these standards, you can browse role-specific programs here: AI courses by job role.
Your membership also unlocks:
