Relying on tech giants for both answers and assistance, rather than just information, could entrench them into our lives more deeply than ever
Seeing the new artificial intelligence-powered chatbots touted in dueling announcements this past week by Microsoft and Google drives home two major takeaways. First, the feeling of “wow, this definitely could change everything.” And second, the realization that for chat-based search and related AI technologies to have an impact, we’re going to have to put a lot of faith in them and the companies they come from.
When AI is delivering answers, and not just information for us to base decisions on, we’re going to have to trust it much more deeply than we have before. This new generation of chat-based search engines are better described as “answer engines” that can, in a sense, “show their work” by giving links to the webpages they deliver and summarize. But for an answer engine to have real utility, we’re going to have to trust it enough, most of the time, that we accept those answers at face value.
The same will be true of tools that help generate text, spreadsheets, code, images and anything else we create on our devices—some version of which both Microsoft MSFT -0.20%decrease; red down pointing triangle and Google have promised to offer within their existing productivity services, Microsoft 365 and Google Workspace.
These technologies, and chat-based search, are all based on the latest generation of “generative” AI, capable of creating verbal and visual content and not just processing it the way more established AI has done. And the added trust it will require is one of several ways in which this new generative AI technology is poised to shift even more power into the hands of the biggest tech companies.
Generative AI in all its forms will insinuate technology more deeply into the way we live and work than it already is—not just answering our questions but writing our memos and speeches or even producing poetry and art. And because of the financial, intellectual and computational resources needed to develop and run the technology are so enormous, the companies that control these AI systems will be the largest, richest companies.
Annual global corporate investment in artificial intelligence, by type
Note: Inflation adjusted. In constant 2021 dollars.
Source: NetBase Quid via AI Index Report (2022
OpenAI, the creator of the ChatGPT chatbot and DALL-E 2 image generator AIs that have fueled much of the current hype, seemed like an exception to that: a relatively small startup that has driven major AI innovation. But it has leapt into the arms of Microsoft, which has made successive rounds of investment, in part because of the need to pay for the computing power needed to make its systems work.
The greater concentration of power is all the more important because this technology is both incredibly powerful and inherently flawed: it has a tendency to confidently deliver incorrect information. This means that step one in making this technology mainstream is building it, and step two is minimizing the variety and number of mistakes it inevitably makes.
Trust in AI, in other words, will become the new moat that big technology companies will fight to defend. Lose the user’s trust often enough, and they might abandon your product. For example: In November, Meta made available to the public an AI chat-based search engine for scientific knowledge called Galactica. Perhaps it was in part the engine’s target audience—scientists—but the incorrect answers it sometimes offered inspired such withering criticism that Meta shut down public access to it after just three days, said Meta chief AI scientist Yann LeCun in a recent talk.
Galactica was “the output of a research project versus something intended for commercial use,” says a Meta spokeswoman. In a public statement, Joelle Pineau, managing director of fundamental AI research at Meta, wrote that “given the propensity of large language models such as Galactica to generate text that may appear authentic, but is inaccurate, and because it has moved beyond the research community, we chose to remove the demo from public availability.”
On the other hand, proving your AI more trustworthy could be a competitive advantage more powerful than being the biggest, best or fastest repository of answers. This seems to be Google’s bet, as the company has emphasized in recent announcements and a presentation on Wednesday that as it tests and rolls out its own chat-based and generative AI systems, it will strive for “Responsible AI,” as outlined in 2019 in its “AI Principles.”
My colleague Joanna Stern this past week provided a helpful description of what it’s like to use Microsoft’s Bing search engine and Edge web browser with ChatGPT incorporated. You can join a list to test the service—and Google says it will make its chatbot, named Bard, available at some point in the coming months.
But in the meantime, to see just why trust in these kinds of search engines is so tricky, you can visit other chat-based search engines that already exist. There’s You.com, which will answer your questions via a chatbot, or Andisearch.com, which will summarize any article it returns when you search for a topic on it.
Even these smaller services feel a little like magic. If you ask You.com’s chat module a question like “Please list the best chat AI-based search engines,” it can, under the right circumstances, give you a coherent and succinct answer that includes all the best-known startups in this space. But it can also, depending on small changes in how you phrase that question, add complete nonsense to its answer.
In my experimentation, You.com would, more often than not, give a reasonably accurate answer, but then add to it the name of a search engine that doesn’t exist at all. Googling the made-up search engine names it threw in revealed that You.com seemed to be misconstruing the names of humans quoted in articles as the names of search engines.
Andi doesn’t return search results in a chat format, precisely because making sure that those answers are accurate is still so difficult, says Chief Executive Angela Hoover. “It’s been super exciting to see these big players validating that conversational search is the future, but nailing factual accuracy is hard to do,” she adds. As a result, for now, Andi offers search results in a conventional format, but offers to use AI to summarize any page it returns.
Andi currently has a team of fewer than 10 people, and has raised $2.5 million so far. It’s impressive what such a small team has accomplished, but it’s clear that making trustworthy AI will require enormous resources, probably on the scale of what companies like Microsoft and Google possess.
There are two reasons for this: The first is the enormous amount of computing infrastructure required, says Tinglong Dai, a professor of operations management at Johns Hopkins University who studies human-AI interaction. That means tens of thousands of computers in big technology companies’ current cloud infrastructures. Some of those computers are used to train the enormous “foundation” models that power generative AI systems. Others specialize in making the trained models available to users, which as the number of users grows can become a more taxing task than the original training.
The second reason, says Dr. Dai, is that it requires enormous human resources to continually test and tune these models, in order to make sure they’re not spouting an inordinate amount of nonsense or biased and offensive speech.
Google has said that it has called on every employee in the company to test its new chat-based search engine and flag any issues with the results it generates. Microsoft, which is already rolling out its chat-based search engine to the public on a limited basis, is doing that kind of testing in public. ChatGPT, on which Microsoft’s chat-based search engine is based, has already proved to be vulnerable to attempts to “jailbreak” it into producing inappropriate content.
Big tech companies can probably overcome the issues arising from their rollout of AI—Google’s go-slow approach, ChatGPT’s sometimes-inaccurate results, and the incomplete or misleading answers chat-based Bing could offer—by experimenting with these systems on a large scale, as only they can.
“The only reason ChatGPT and other foundational models are so bad at bias and even fundamental facts is they are closed systems, and there is no opportunity for feedback,” says Dr. Dai. Big tech companies like Google have decades of practice at soliciting feedback to improve their algorithmically-generated results. Avenues for such feedback have, for example, long been a feature of both Google Search and Google Maps.
Dr. Dai says that one analogy for the future of trust in AI systems could be one of the least algorithmically-generated sites on the internet: Wikipedia. While the entirely human-written and human-edited encyclopedia isn’t as trustworthy as primary-source material, its users generally know that and find it useful anyway. Wikipedia shows that “social solutions” to problems like trust in the output of an algorithm—or trust in the output of human Wikipedia editors—are possible.
But the model of Wikipedia also shows that the kind of labor-intensive solutions for creating trustworthy AI—which companies like Meta and Google have already employed for years and at scale in their content moderation systems—are likely to entrench the power of existing big technology companies. Only they have not just the computing resources, but also the human resources, to deal with all the misleading, incomplete or biased information their AIs will be generating.
In other words, creating trust by moderating the content generated by AIs might not prove to be so different from creating trust by moderating the content generated by humans. And that is something the biggest technology companies have already shown is a difficult, time-consuming and resource-intensive task they can take on in a way that few other companies can.
The obvious and immediate utility of these new kinds of AIs, when integrated into a search engine or in their many other potential applications, is the reason for the current media, analyst and investor frenzy for AI. It’s clear that this could be a disruptive technology, resetting who is harvesting attention and where they’re directing it, threatening Google’s search monopoly and opening up new markets and new sources of revenue for Microsoft and others.
Based on the runaway success of the ChatGPT AI—perhaps the fastest service to reach 100 million users in history, according to a recent UBS report—it’s clear that being an aggressive first mover in this space could matter a great deal. It’s also clear that being a successful first-mover in this space will require the kinds of resources that only the biggest tech companies can muster.
Write to Christopher Mims at [email protected]
Appeared in the February 11, 2023, print edition as 'What Does New AI Need? Your Trust'.
- alf9872000 and Karlston
-
2
Recommended Comments
There are no comments to display.
Join the conversation
You can post now and register later. If you have an account, sign in now to post with your account.
Note: Your post will require moderator approval before it will be visible.