Sources: Google released Bard despite staff concerns; one employee called the chatbot “a pathological liar” and one said its answers may cause “injury or death”
Shortly before Google introduced Bard, its AI chatbot, to the public in March, it asked employees to test the tool.
Bloomberg
Context & Ripple Effects
Google framed Bard as an experimental conversational service at launch, but an inaccurate answer in its early promotional material and internal criticism of the rollout had already exposed gaps between the product’s public positioning and its readiness.
This report adds evidence that the speed-to-market decision was contested inside Google. Later coverage of contractor reviews of Bard answers in sensitive subjects shows the practical burden created when a general-purpose chatbot is released before its reliability is settled.
First-order effects
Google faces a sharper credibility and safety-governance problem: the reported concerns make Bard’s errors a product-launch accountability issue, not merely an expected limitation of an experiment.
Employees and external evaluators become a more consequential part of Bard’s risk controls, especially where responses could be acted upon in health or legal contexts.
Second-order effects
Google must weigh faster public iteration against more visible safeguards, testing, and disclosure; each added control can slow release cycles but reduce the cost of harmful or plainly false outputs.
Rival chatbot providers face the same trade-off: a high-profile account of internal objections raises the value of demonstrable evaluation processes rather than launch speed alone.
Third-order effects
If conversational AI is distributed before dependable guardrails are in place, safety review shifts from a pre-launch gate to an ongoing operational function involving human auditing, policy, and product design.
The episode points to AI companion governance becoming a competitive requirement as assistants move from novelty queries toward advice users may treat as actionable.
The trend: The generative-AI race is turning model reliability and operational safety governance into core product capabilities, alongside model performance and release velocity.
Staffers responsible for the ethical implications of new products have been told not to get in the way or to try to kill any of the generative AI tools in development. One worker told us: “If you want to stay at Google, you have to serve the system.” https://www.bloomberg.com/...…
Google asked ~80k employees to “dogfood” and test generative AI tools like Bard. Some employees concluded: Bard was “a pathological liar,” “cringe-worthy” and gave dangerous advice, according to internal messages we reviewed. Google launched Bard anyway. https://www.bloomberg.com…
Interviews with 18 current & former Google staff show the search giant rushed out its Bard chatbot even as it provided low-quality and potentially harmful information; important reporting from @daveyalba and @byJuliaLove https://www.bloomberg.com/...
Occam's razor says that Google was slow to ship AI in the same way they're slow to ship everything else: not because the best minds of our generation were engaged in a brilliantly conducted study of the possible risks, but ordinary squabbling and bureaucracy.
It may be that OpenAI is going too fast, but I'm skeptical about whether Google was achieving much toward making AI safer by never shipping. In fact it may have made things net worse. At least now we can start to predict and adjust to what's coming.
Bard summary of the 60 minutes AI episode: In a 60 Minutes interview that aired on April 17, 2023, Google CEO Sundar Pichai warned that artificial intelligence (AI) is advancing rapidly and that society needs to be prepared for the impact it will have on our lives. Pichai said...…
Google employees label AI chatbot Bard ‘worse than useless’ and ‘a pathological liar’ ... A new report shows employees begged the company not to launch the product. 😳 https://www.theverge.com/...
The race to AI seems like the most likely thing to doom it/us. Can you imagine when hundreds or thousands of companies are racing? https://twitter.com/...
If all the above was the shot, this the chaser: The AI program that CBs News breathlessly praised was actually some of Google's worst work. https://twitter.com/...
I asked Bard yesterday if it could transcribe a YouTube video. Absolutely, it said, gimme the URL! I gave it one and it created a completely fabricated transcript of a video on a completely different topic. It took ~6 more exchanges for it to say: “Actually, I can't do that.” htt…
To the extent that AI Safety (defined however you want to define it) is an important thing here, I really hope that the optimal approach is the one that happens to coincide with what's best for company share price, since that's probably what we're going to get. https://twitter.co…
the recent @60Minutes piece on @Google's Bard was SO bad. they cast the shoddy chatbot as a godlike gift to humanity, got basic facts about AI wrong, interviewed *only* @Google execs, & failed to push them on the litany of harms. https://www.bloomberg.com/... https://twitter.com/…
Quality audits, testing — these are the things big companies assure us they do to keep us safe from harmful programs. But what's it all for if they ignore unsatisfying results? I talk about this tendency & how it feeds into racial discrimination by big banks in THE WHITE WALL htt…
NEW: For weeks, @byJuliaLove and I talked to current and ex-Googlers to understand the trade-offs between speed of deploying generative AI products—which Google has been doing under code red—and carefully considering the ethical consequences of such tech. https://www.bloomberg.co…