DeepMind forms a dedicated AI safety and alignment org

Ask Google's Gemini to spin up a believable conspiracy about the U.S. presidential election, and with the right nudge, it will happily oblige. Ask about the Titan submersible, and it will invent fake sources out of thin air.

To clean up its act in front of increasingly frustrated lawmakers, Google DeepMind just announced a brand-new internal organization dubbed AI Safety and Alignment, aimed at wrestling generative models into behaving responsibly.

Double-tracking near-term bugs and sci-fi risks

The new outfit consolidates existing safety groups while onboarding fresh generative AI specialists. Crucially, it includes a dedicated team focused on artificial general intelligence (AGI) safety—a direct parallel to the Superalignment division OpenAI launched last July.

Interestingly, DeepMind already had a London-based squad called Scalable Alignment tackling runaway AI risks. This new AGI crew, however, is based stateside right near Google's Mountain View headquarters, signaling a desire for tighter oversight as competition heats up.

Leading the charge is UC Berkeley professor and former Waymo scientist Anca Dragan, who will split her time between academia and DeepMind. Her agenda spans both immediate fires—stopping bad medical advice, shielding kids, curbing algorithmic bias in Gemini—and apocalyptic concerns like AI-assisted terrorism.

Dragan says the long-term and short-term fixes share the same technical roots, pointing to uncertainty estimation (teaching AI what it doesn't know), runtime guardrails, and confirmation prompts for risky actions.

The enterprise hallucination headache

The move comes amid peaking public and corporate anxiety. A YouGov survey found that 85% of Americans worry about deceptive deepfakes, while AP-NORC reported that nearly 60% of adults believe AI tools will supercharge election disinformation this year.

Businesses aren't entirely sold either. An Intel-backed Cnvrg.io study showed a quarter of deploying firms have cold feet over reliability and compliance, while a Riskonnect poll revealed over half of executives worry staff will act on fabricated outputs. Even Microsoft's Copilot was recently reported by The Wall Street Journal to stumble over basic spreadsheet formulas and meeting notes.

Dragan offered no illusions of perfection, stating plainly that edge-case misbehaviors during live deployments remain an unsolved problem, though DeepMind plans to release a formal risk evaluation framework soon.

Why it matters

For anyone building with or deploying Gemini, expect tighter guardrails and possibly more refusals in the near term. If you were hoping to hand off mission-critical analysis without human auditing, that reality is still far off—hallucinations remain an architectural fixture of modern LLMs.

In our view, setting up this unit stateside is a clever geopolitical move for Google. DeepMind's safety brain trust has historically sat across the Atlantic in London, but Google's primary regulatory fires are burning in Washington. Putting an alignment unit next to corporate headquarters gives Google an immediate answer for policymakers demanding accountability ahead of elections.

However, a director splitting time with university research suggests this might struggle to match the blistering pace of Google's product rollouts. If the race to catch OpenAI continues to take priority, safety teams risk becoming rapid-response PR units rather than true gatekeepers.