Reprinted from TidBITS by permission; reuse governed by Creative Commons license BY-NC-ND 3.0. TidBITS has offered years of thoughtful commentary on Apple and Internet topics. For free email subscriptions and access to the entire TidBITS archive, visit http://www.tidbits.com/ Slack AI Privacy Principles Generate Confusion and Consternation Adam Engst As I traveled home from Salt Lake City last Friday, controversy began to swirl around the popular group messaging system Slack, [1]which we use and cover regularly. A Slack online document titled '[2]Privacy principles: search, learning and artificial intelligence' went viral after it was discovered that Slack failed to use the Oxford comma in its title. Wait, no, that's wrong'the controversy came about after a link to the privacy principles was posted to Hacker News with the title '[3]Slack AI Training with Customer Data' and boosted on [4]X/Twitter, focusing on how the document includes an awkward opt-out-via-email provision if 'you want to exclude your Customer Data from Slack global models.' Well, heck, yes, who wouldn't want to exclude their data from Slack global models? One of Slack's selling points is its promises surrounding privacy and security, so confidential business and personal information can be discussed with impunity. If Slack were training large language models (LLMs) on customer data, anything you say in Slack could be ingested into the model, digested by the neural network, and spit back in who knows what situations and contexts. Maybe your competitors could even ask Slack AI to reveal your business secrets! Quick, panic on social media! If people actually read Slack's privacy principles document instead of just reacting to an incorrectly titled link or an out-of-context screenshot on X/Twitter, they would see that Slack isn't doing any of those things. Refer to unambiguous sentences like: We do not develop LLMs or other generative models using customer data. No Customer Data is used to train third-party LLM models. Slack does not train LLMs or other generative models on Customer Data, or share Customer Data with any LLM providers. That's not to say Slack's privacy principles are a paragon of clarity. The document uses a variety of terms without defining them (what's a 'global model,' and is it generative AI?), waits until the end to mention that Slack does use generative AI for the separate [5]Slack AI product offering (but not LLMs trained on customer data), and tends to mix what Slack doesn't do (develop LLMs using customer data) with what it does do (develop non-generative AI/ML models by analyzing customer data) in the same paragraph. The entire kerfuffle occurred because people didn't read carefully and posted out-of-context hot takes to social media. If only we had a tool that could help our social media-addled human brains extract meaning from complex documents... Hey, I know'I'll ask ChatGPT if Slack will be training generative AI models based on the confidential discussions in my Slack workspace. There you have it. ChatGPT: 1, random people on social media: 0. Unless, of course, ChatGPT is in cahoots with Slack AI, and that's just what they want us to believe. I also asked ChatGPT to tell me what people might find confusing about Slack's privacy principles and to rewrite it so there would be less chance for confusion. You know what? It did a pretty good job calling out Slack's definitional and organizational problems and creating a new version that relied heavily on headings and bullet lists to get the point across more directly. Paste in Slack's privacy principles and try it yourself sometime'chatbots can do a decent job of telling you what's wrong with a document. More seriously, there's an important point to make here. Even as we rely ever more on gadgets and services, society has lost a great deal of trust in the tech industry. This controversy arose because the suggestion that Slack was doing something underhanded fit a lot of preconceived notions. Slack didn't even get enough of the benefit of the doubt to cause people to read its privacy principles carefully. This lack of trust in how the generative AI sausage is made is understandable. OpenAI and other firms trained their LLMs on vast quantities of information from the Internet, and although that data may have been publicly available, using it to build and maintain billion-dollar businesses doesn't feel right without compensation or at least credit. Worse, will anything you say'or documents you submit'to an AI chatbot be added to future training data? Although the privacy policies for all the leading AI chatbots say they wouldn't think of doing any such thing, prudence would suggest caution when sharing sensitive health, financial, or business information with a chatbot. I fear there's no returning to the techno-optimistic days of yesteryear. Experts say a company can regain trust with clear and transparent communications, consistent competence, ethical practices, a positive social impact, and a serious attitude toward stakeholder concerns. That would be difficult for even a motivated company and seems impossible for the tech industry overall. But I'm not worried about Slack for the moment. References Visible links 1. https://tidbits.com/tag/slack/ 2. https://slack.com/intl/en-gb/trust/data-management/privacy-principles 3. https://news.ycombinator.com/item?id=40383978 4. https://x.com/QuinnyPig/status/1791220276350390575 5. https://slack.com/features/ai Hidden links: 6. https://tidbits.com/wp/../uploads/2024/05/Slack-fuss-ChatGPT.png .