In early February 2023, Google held a press event in Paris to introduce Bard, its conversational AI product. The announcement had been hastily scheduled. The pressure on Google to demonstrate competitive capability following Microsoft’s integration of ChatGPT into Bing had become impossible to ignore. The event was meant to reset the narrative.
It did not. The promotional materials shared in advance of the event contained a factual error. A demonstration of Bard answered a question about the James Webb Space Telescope by attributing a discovery to it that had been made by a different telescope years earlier. The error was identified within hours by astronomers who knew the actual history. The reaction in the market was severe. Alphabet stock fell roughly seven percent on the day of the event, wiping approximately one hundred billion dollars from the company’s market capitalisation in a few hours.
The size of the market reaction was disproportionate to the size of the error. Factual mistakes in language model outputs were widely understood to be a feature of the technology in early 2023. ChatGPT made similar errors regularly. The fact that Bard had made one was unsurprising on its own. What spooked the market was the timing and the framing. Google had presented the demo as evidence that it was competitive with OpenAI. The error suggested that Google had moved fast enough to make mistakes, and that confidence in its measured approach to AI safety was being sacrificed for speed.
The internal Google discussions during this period were reportedly difficult. The company had been a leader in AI research for years and had built much of the foundation that companies like OpenAI were now leveraging. Being publicly cast as the slower follower in AI products was an unfamiliar position. The pressure to ship faster conflicted with the cultural disposition toward more measured release, and the Bard demo was a visible casualty of that tension.
What the episode confirmed was that the AI product cycle had moved into a phase where competitive pressure was overriding established release practices. Companies that had previously released products only when they were ready by their own internal standards were now releasing them when the market demanded a response. The error rate in production AI products would rise as a result, with effects that would compound through 2023.