Google asks its own employees to rewrite Bard's bad answers

What do you do when your flagship AI gets basic facts wrong in its public debut? If you are Google, you make your own employees fix the homework.

Following a disastrous rollout that wiped out nearly 9% of the company's market value, Google is turning to its massive workforce to clean up the mess. In an internal email, search boss Prabhakar Raghavan asked staffers to help rewrite incorrect Bard answers, explaining that the AI learns best by example.

This follows CEO Sundar Pichai's earlier call for employees to spend two to four hours a day testing the chatbot, warning that the AI race is going to be a long journey.

The official guide to fixing Bard

Internal sentiment has been brutal since the botched demo, with employees openly blasting leadership for a rushed and comically short-sighted launch. Now, executives are handing out specific do's and don'ts for teaching the AI.

Staffers are urged to tackle topics they know deeply. When rewriting, Google wants answers to sound polite, casual, approachable, and delivered in a neutral, first-person voice.

Then come the guardrails. Workers must avoid stereotypes based on race, nationality, gender, religion, or politics. Crucially, they are strictly forbidden from making Bard sound human, implying it has feelings, or claiming it has human experiences.

If the AI starts hallucinating medical, financial, or legal advice, or generates abusive responses, employees are told not to rewrite it at all. Instead, they just hit thumbs-down and send it to safety teams.

To sweeten the deal, contributors will receive a digital Moma badge on their internal profiles. The top 10 rewrite contributors in Raghavan's division will even get a face-to-face listening session with leadership to vent about Bard's biggest flaws.

Why it matters

Google's all-hands-on-deck cleanup proves just how frantic the scramble has become after OpenAI took the world by storm.

For regular users, this is a clear reality check. If thousands of elite engineers are currently spending their afternoons fixing elementary errors, you definitely shouldn't trust Bard with critical research or real work just yet. It is still very much in beta, whatever marketing says.

Strategically, tapping high-salary tech workers to do manual data annotation is a clever way to generate premium alignment data that cheap crowdsourced workers simply cannot provide. It will almost certainly accelerate Bard's short-term tuning.

At the same time, we think this exposes Google's vulnerability. Burning thousands of high-value engineering hours on manual sentence rewrites shows how caught off-guard the search giant really was. Patching bad answers by hand cannot solve fundamental hallucinations, and it reveals how far Google is lagging behind OpenAI's polished deployment pipeline.