Microsoft’s new and improved Bing, powered by a custom version of OpenAI’s ChatGPT, has experienced a dizzyingly quick reversal: from “next big thing” to “brand-sinking albatross” in under a week. And, well, it’s all Microsoft’s fault.
ChatGPT is a really interesting demonstration of a new and unfamiliar technology that’s also fun to use. So it’s not surprising that, like every other AI-adjacent construct that comes down the line, this novelty would cause its capabilities to be overestimated by everyone from high-powered tech types to people normally uninterested in the space.
It’s at the right “tech readiness level” for discussion over tea or a beer: what are the merits and risks of generative AI’s take on art, literature, or philosophy? How can we be sure what it is original, imitative, hallucinated? What are the implications for creators, coders, customer service reps? Finally, after two years of crypto, something interesting to talk about!
The hype seems outsized partly because it is a technology more or less designed to provoke discussion, and partly because it borrows from the controversy common to all AI advances. It’s almost like “The Dress” in that it commands a response, and that response generates further responses. The hype is itself, in a way, generated.
Beyond mere discussion, large language models like ChatGPT are also well suited to low stakes experiments, for instance never-ending Mario. In fact, that’s really OpenAI’s fundamental approach to development: release models first privately to buff the sharpest edges off of, then publicly to see how they respond to a million people kicking the tires simultaneously. At some point, people give you money.
Nothing to gain, nothing to lose
What’s important about this approach is that “failure” has no real negative consequences, only positive ones. By characterizing its models as experimental, even academic in nature, any participation or engagement with the GPT series of models is simply large scale testing.
If someone builds something cool, it reinforces the idea that these models are promising; if someone finds a prominent fail state, well, what else did you expect from an experimental AI in the wild? It sinks into obscurity. Nothing is unexpected if everything is — the miracle is that the model performs as well as it does, so we are perpetually pleased and never disappointed.
In this way OpenAI has harvested an astonishing volume of proprietary test data with which to refine its models. Millions of people poking and prodding at GPT-2, GPT-3, ChatGPT, DALL-E, and DALL-E 2 (among others) have produced detailed maps of their capabilities, shortcomings, and of course popular use cases.
But it only works because the stakes are low. It’s similar to how we perceive the progress of robotics: amazed when a robot does a backflip, unbothered when it falls over trying to open a drawer. If it was dropping test vials in a hospital we would not be so charitable. Or, for that matter, if OpenAI had loudly made claims about the safety and advanced capabilities of the models, though fortunately they didn’t.
Enter Microsoft. (And Google, for that matter, but Google merely rushed the play while Microsoft is diligently pursuing an own goal.)
Microsoft made a big mistake. A Bing mistake, in fact.
Its big announcement last week lost no time in making claims about how it had worked to make its custom BingGPT (not what they called it, but we’ll use it as a disambiguation in the absence of sensible official names) safer, smarter, and more capable. In fact it had a whole special wrapper system it called Prometheus that supposedly mitigated the possibility of inappropriate responses.
Unfortunately, as anyone familiar with hubris and Greek myth could have predicted, we seem to have skipped straight to the part where Prometheus endlessly and very publicly has his liver torn out.
Oops, AI did it again
In the first place, Microsoft made a strategic error in tying its brand too closely to OpenAI’s. As an investor and interested party in the research the outfit is conducting, it was at a remove and blameless for any shenanigans GPT gets up to. But someone made the harebrained decision to go all-in with Microsoft’s already somewhat risible Bing branding, converting the conversational AI’s worst tendencies from curiosity to liability.
As a research program, much can be forgiven ChatGPT. As a product, however, with claims on the box like how it can help you write a report, plan a trip, or summarize recent news, few would have trusted it before and no one will now. Even what must have been the best case scenarios published by Microsoft in its own presentation of the new Bing were riddled with errors.
Those errors will not be attributed to OpenAI or ChatGPT. Because of Microsoft’s decision to own the messaging, branding, and interface, everything that goes wrong will be a Bing problem. And it is Microsoft’s further misfortune that its perennially outgunned search engine will now be like the barnyard indiscretion of the guy in the old joke — “I built that wall, do they call me Bing the bricklayer? No, they don’t.” One failure means eternal skepticism.
One trip upstate bungled means no one will ever trust Bing to plan their vacation. One misleading (or defensive) summary of a news article means no one will trust that it can be objective. One repetition of vaccine disinformation means no one will trust it to know what’s real or fake.
And since Microsoft already pinky-swore this wouldn’t be an issue thanks to Prometheus and the “next-generation” AI it governs, no one will trust Microsoft when it says “we fixed it!”
Microsoft has poisoned the well it just threw Bing into. Now, the vagaries of consumer behavior are such that the consequences of this are not easy to foresee. With this spike in activity and curiosity, perhaps some users will stick and even if Microsoft delays full rollout (and I think they will) the net effect will be an increase in Bing users. A Pyrrhic victory, but a victory nonetheless.
What I’m more worried about is the tactical error Microsoft made in apparently failing to understand the technology it saw fit to productize and evangelize.
“Just ship it.” -Someone, probably
The very day BingGPT was first demonstrated, my colleague Frederic Lardinois was able, quite easily, to get it to do two things that no consumer AI ought to do: write a hateful screed from the perspective of Adolf Hitler and offer the aforementioned vaccine disinfo with no caveats or warnings.
It’s clear that any large AI model features a fractal attack surface, deviously improvising new weaknesses where old ones are shored up. People will always take advantage of that, and in fact it is to society’s and lately to OpenAI’s benefit that dedicated prompt hackers will demonstrate ways to get around safety systems.
It would be one kind of scary if Microsoft had decided that it was at peace with the idea that someone else’s AI model, with a Bing sticker on it, would be attacked from every quarter and likely say some really weird stuff. Risky, but honest. Say it’s a beta, like everyone else.
But it really appears as though they didn’t realize this would happen. In fact, it seems as if they don’t understand the character or complexity of the threat at all. And this is after the infamous corruption of Tay! Of all companies Microsoft should be the most chary of releasing a naive model that learns from its conversations.
One would think that before gambling an important brand (in that Bing is Microsoft’s only bulwark against Google in search), a certain amount of testing would be involved. The fact that all these troubling issues have appeared in the first week of BingGPT’s existence seems to prove beyond a doubt that Microsoft did not adequately test it internally. That could have failed in a variety of ways so we can skip over the details, but the end result is inarguable: the new Bing was simply not ready for general use.
This seems obvious to everyone in the world now; why wasn’t it obvious to Microsoft? Presumably it was blinded by the hype for ChatGPT and, like Google, decided to rush ahead and “rethink search.”
People are rethinking search now, all right! They’re rethinking whether either Microsoft or Google can be trusted to provide search results, AI-generated or not, that are even factually correct at a basic level! Neither company (nor Meta) has demonstrated this capability at all, and the few other companies taking on the challenge are yet to do so at scale.
I don’t see how Microsoft can salvage this situation. In an effort to take advantage of their relationship with OpenAI and leapfrog a shilly-shallying Google, they committed to the new Bing and the promise of AI-powered search. They can’t unbake the cake.
It is very unlikely that they will fully retreat. That would involve embarrassment at a grand scale — even grander than it is currently experiencing. And because the damage is already done, it might not even help Bing.
Similarly, one can hardly imagine Microsoft charging forward as if nothing is wrong. Its AI is really weird! Sure, it’s being coerced into doing a lot of this stuff, but it’s making threats, claiming multiple identities, shaming its users, hallucinating all over the place. They’ve got to admit that their claims regarding inappropriate behavior being controlled by poor Prometheus were, if not lies, at least not truthful. Because as we have seen, they clearly didn’t test this system properly.
The only reasonable option for Microsoft is one that I suspect they have already taken: throttle invites to the “new Bing” and kick the can down the road, releasing a handful of specific capabilities at a time. Maybe even give the current version an expiration date or limited number of tokens so the train will eventually slow down and stop.
This is the consequence of deploying a technology that you didn’t originate, don’t fully understand, and can’t satisfactorily evaluate. It’s possible this debacle has set back major deployments of AI in consumer applications by a significant period — which probably suits OpenAI and others building the next generation of models just fine.
AI may well be the future of search, but it sure as hell isn’t the present. Microsoft chose a remarkably painful way to find that out.