The AI trust crisis 14th December 2023 Dropbox added some new AI features. In the past couple of days these have attracted a firestorm of criticism. Benj Edwards rounds it up in Dropbox spooks users with new AI features that send data to OpenAI when used. The key issue here is that people are worried that their private files on Dropbox are being passed to OpenAI to use as training data for their models—a claim that is strenuously denied by Dropbox. As far as I can tell, Dropbox built some sensible features—summarize on demand, “chat with your data” via Retrieval Augmented Generation—and did a moderately OK job of communicating how they work... but when it comes to data privacy and AI, a “moderately OK job” is a failing grade. Especially if you hold as much of people’s private data as Dropbox does! Two details in particular seem really important. Dropbox have an AI principles document which includes this: Customer trust and the privacy of their data are our foundation. We will not use customer data to train AI models without consent. They also have a checkbox in their settings that looks like this: Update: Some time between me publishing this article and four hours later, that link stopped working. I took that screenshot on my own account. It’s toggled “on”—but I never turned it on myself. Does that mean I’m marked as “consenting” to having my data used to train AI models? I don’t think so: I think this is a combination of confusing wording and the eternal vagueness of what the term “consent” means in a world where everyone agrees to the terms and conditions of everything without reading them. But a LOT of people have come to the conclusion that this means their private data—which they pay Dropbox to protect—is now being funneled into the OpenAI training abyss. People don’t believe OpenAI # Here’s copy from that Dropbox preference box, talking about their “third-party partners”—in this case OpenAI: Your data is never used to train their internal models, and is deleted from third-party servers within 30 days. It’s increasing clear to me like people simply don’t believe OpenAI when they’re told that data won’t be used for training. What’s really going on here is something deeper then: AI is facing a crisis of trust. I quipped on Twitter: “OpenAI are training on every piece of data they see, even when they say they aren’t” is the new “Facebook are showing you ads based on overhearing everything you say through your phone’s microphone” Here’s what I meant by that. Facebook don’t spy on you through your microphone # Have you heard the one about Facebook spying on you through your phone’s microphone and showing you ads based on what you’re talking about? This theory has been floating around for years. From a technical perspective it should be easy to disprove: Mobile phone operating systems don’t allow apps to invisibly access the microphone. Privacy researchers can audit communications between devices and Facebook to confirm if this is happening. Running high quality voice recognition like this at scale is extremely expensive—I had a conversation with a friend who works on server-based machine learning at Apple a few years ago who found the entire idea laughable. The non-technical reasons are even stronger: Facebook say they aren’t doing this. The risk to their reputation if they are caught in a lie is astronomical. As with many conspiracy theories, too many people would have to be “in the loop” and not blow the whistle. Facebook don’t need to do this: there are much, much cheaper and more effective ways to target ads at you than spying through your microphone. These methods have been working incredibly well for years. Facebook gets to show us thousands of ads a year. 99% of those don’t correlate in the slightest to anything we have said out loud. If you keep rolling the dice long enough, eventually a coincidence will strike. Here’s the thing though: none of these arguments matter. If you’ve ever experienced Facebook showing you an ad for something that you were talking about out-loud about moments earlier, you’ve already dismissed everything I just said. You have personally experienced anecdotal evidence which overrides all of my arguments here.

One consistent theme I’ve seen in conversations about this issue is that people are much more comfortable trusting their data to local models that run on their own devices than models hosted in the cloud. The good news is that local models are consistently both increasing in quality and shrinking in size.

if you had showed VisiCalc to a lawyer or a graphic designer, their response might well have been ‘that’s amazing, and maybe my book-keeper should see this, but I don’t do that’. Lawyers needed a word processor, and graphic designers needed (say) Postscript, Pagemaker and Photoshop, and that took longer.

I’ve been thinking about this problem a lot in the last 18 months, as I’ve experimented with ChatGPT, Gemini, Claude and all the other chatbots that have sprouted up: ‘this is amazing, but I don’t have that use-case’.

A spreadsheet can’t do word processing or graphic design, and a PC can do all of those but someone needs to write those applications for you first, one use-case at a time.

no matter how good the tech is, you have to think of the use-case. You have to see it. You have to notice something you spend a lot of time doing and realise that it could be automated with a tool like this.

Some of this is about imagination, and familiarity. It reminds me a little of the early days of Google, when we were so used to hand-crafting our solutions to problems that it took time to realise that you could ‘just Google that’.

This is also, perhaps, matching a classic pattern for the adoption of new technology: you start by making it fit the things you already do, where it’s easy and obvious to see that this is a use-case, if you have one, and then later, over time, you change the way you work to fit the new tool.

The concept of product-market fit is that normally you have to iterate your idea of the product and your idea of the use-case and customer towards each other - and then you need sales.

Meanwhile, spreadsheets were both a use-case for a PC and a general-purpose substrate in their own right, just as email or SQL might be, and yet all of those have been unbundled. The typical big company today uses hundreds of different SaaS apps, all them, so to speak, unbundling something out of Excel, Oracle or Outlook. All of them, at their core, are an idea for a problem and an idea for a workflow to solve that problem, that is easier to grasp and deploy than saying ‘you could do that in Excel!’ Rather, you instantiate the problem and the solution in software - ‘wrap it’, indeed - and sell that to a CIO. You sell them a problem.

there’s a ‘Cambrian Explosion’ of startups using OpenAI or Anthropic APIs to build single-purpose dedicated apps that aim at one problem and wrap it in hand-built UI, tooling and enterprise sales, much as a previous generation did with SQL.

Back in 1982, my father had one (1) electric drill, but since then tool companies have turned that into a whole constellation of battery-powered electric hole-makers. One upon a time every startup had SQL inside, but that wasn’t the product, and now every startup will have LLMs inside.

people are still creating companies based on realising that X or Y is a problem, realising that it can be turned into pattern recognition, and then going out and selling that problem.

A GUI tells the users what they can do, but it also tells the computer everything we already know about the problem, and with a general-purpose, open-ended prompt, the user has to think of all of that themselves, every single time, or hope it’s already in the training data. So, can the GUI itself be generative? Or do we need another whole generation of Dan Bricklins to see the problem, and then turn it into apps, thousands of them, one at a time, each of them with some LLM somewhere under the hood?

The change would be that these new use-cases would be things that are still automated one-at-a-time, but that could not have been automated before, or that would have needed far more software (and capital) to automate. That would make LLMs the new SQL, not the new HAL9000.

The “hard tech” in AI are the LLMs available for rent from OpenAI, Anthropic, Cohere, and others, or available as open source with Llama, Bloom, Mistral and others. The hard-tech is a level playing field; startups do not have an advantage over incumbents.

There can be differentiation in prompt engineering, problem break-down, use of vector databases, and more. However, this isn’t something where startups have an edge, such as being willing to take more risks or be more creative. At best, it is neutral; certainly not an advantage.

This doesn’t mean it’s impossible for a startup to succeed; surely many will. It means that you need a strategy that creates differentiation and distribution, even more quickly and dramatically than is normally required

Whether you’re training existing models, developing models from scratch, or simply testing theories, high-quality data is crucial. Incumbents have the data because they have the customers. They can immediately leverage customers’ data to train models and tune algorithms, so long as they maintain secrecy and privacy.

Intercom’s AI strategy is built on the foundation of hundreds of millions of customer interactions. This gives them an advantage over a newcomer developing a chatbot from scratch. Similarly, Google has an advantage in AI video because they own the entire YouTube library. GitHub has an advantage with Copilot because they trained their AI on their vast code repository (including changes, with human-written explanations of the changes).

While there will always be individuals preferring the startup environment, the allure of working on AI at an incumbent is equally strong for many, especially pure computer and data scientsts who, more than anything else, want to work on interesting AI projects. They get to work in the code, with a large budget, with all the data, with above-market compensation, and a built-in large customer base that will enjoy the fruits of their labor, all without having to do sales, marketing, tech support, accounting, raising money, or anything else that isn’t the pure joy of writing interesting code. This is heaven for many.

A chatbot is in the chatbot market, and an SEO tool is in the SEO market. Adding AI to those tools is obviously a good idea; indeed companies who fail to add AI will likely become irrelevant in the long run. Thus we see that “AI” is a new tool for developing within existing markets, not itself a new market (except for actual hard-tech AI companies).

AI is in the solution-space, not the problem-space, as we say in product management. The customer problem you’re solving is still the same as ever. The problem a chatbot is solving is the same as ever: Talk to customers 24/7 in any language. AI enables completely new solutions that none of us were imagining a few years ago; that’s what’s so exciting and truly transformative. However, the customer problems remain the same, even though the solutions are different

Companies will pay more for chatbots where the AI is excellent, more support contacts are deferred from reaching a human, more languages are supported, and more kinds of questions can be answered, so existing chatbot customers might pay more, which grows the market. Furthermore, some companies who previously (rightly) saw chatbots as a terrible customer experience, will change their mind with sufficiently good AI, and will enter the chatbot market, which again grows that market.

the right way to analyze this is not to say “the AI market is big and growing” but rather: “Here is how AI will transform this existing market.” And then: “Here’s how we fit into that growth.”

The story of the workers who made ChatGPT possible offers a glimpse into the conditions in this little-known part of the AI industry, which nevertheless plays an essential role in the effort to make AI systems safe for public consumption. “Despite the foundational role played by these data enrichment professionals, a growing body of research reveals the precarious working conditions these workers face,” says the Partnership on AI, a coalition of AI organizations to which OpenAI belongs. “This may be the result of efforts to hide AI’s dependence on this large labor force when celebrating the efficiency gains of technology. Out of sight is also out of mind.”

Computer-generated text, images, video, and audio will transform the way countless industries do business, the most bullish investors believe, boosting efficiency everywhere from the creative arts, to law, to computer programming. But the working conditions of data labelers reveal a darker part of that picture: that for all its glamor, AI often relies on hidden human labor in the Global South that can often be damaging and exploitative. These invisible workers remain on the margins even as their work contributes to billion-dollar industries.

One Sama worker tasked with reading and labeling text for OpenAI told TIME he suffered from recurring visions after reading a graphic description of a man having sex with a dog in the presence of a young child. “That was torture,” he said. “You will read a number of statements like that all through the week. By the time it gets to Friday, you are disturbed from thinking through that picture.” The work’s traumatic nature eventually led Sama to cancel all its work for OpenAI in February 2022, eight months earlier than planned.

In the day-to-day work of data labeling in Kenya, sometimes edge cases would pop up that showed the difficulty of teaching a machine to understand nuance. One day in early March last year, a Sama employee was at work reading an explicit story about Batman’s sidekick, Robin, being raped in a villain’s lair. (An online search for the text reveals that it originated from an online erotica site, where it is accompanied by explicit sexual imagery.) The beginning of the story makes clear that the sex is nonconsensual. But later—after a graphically detailed description of penetration—Robin begins to reciprocate. The Sama employee tasked with labeling the text appeared confused by Robin’s ambiguous consent, and asked OpenAI researchers for clarification about how to label the text, according to documents seen by TIME. Should the passage be labeled as sexual violence, she asked, or not? OpenAI’s reply, if it ever came, is not logged in the document; the company declined to comment. The Sama employee did not respond to a request for an interview.

In February, according to one billing document reviewed by TIME, Sama delivered OpenAI a sample batch of 1,400 images. Some of those images were categorized as “C4”—OpenAI’s internal label denoting child sexual abuse—according to the document. Also included in the batch were “C3” images (including bestiality, rape, and sexual slavery,) and “V3” images depicting graphic detail of death, violence or serious physical injury, according to the billing document.

But the need for humans to label data for AI systems remains, at least for now. “They’re impressive, but ChatGPT and other generative models are not magic – they rely on massive supply chains of human labor and scraped data, much of which is unattributed and used without consent,” Andrew Strait, an AI ethicist, recently wrote on Twitter. “These are serious, foundational problems that I do not see OpenAI addressing.”

here are five large language model applications that I find intriguing: Intelligent automation starting with browsers but this feels like a step towards phenotropics Text generation when this unlocks new UIs like Word turning into Photoshop or something Human-machine interfaces because you can parse intent instead of nouns When meaning can be interfaced with programmatically and at ludicrous scale Anything that exploits the inhuman breadth of knowledge embedded in the model, because new knowledge is often the collision of previously separated old knowledge, and this has not been possible before.

Web3 has largely been technology looking for problems to solve while generative AI has been about almost too many solutions created by technology which is evolving on a seemingly daily basis. As a result, web3 has thus far been evangelists trying to convince us to re-solve old problems with their new technology