Latest / Future of Work Tech with Fexingo: Remote Tools, AI Productivity, and Workplace Software / How Office Chatbots Are Training on Your Slang
Transcript
- Lucas: So you know how your team has that weird shorthand — like 'OTP' for 'on the phone' or 'turtle' for a ticket that's been sitting too long? Luna: Sure. My old team used to say 'pancake' for a project that was flat and going nowhere. Lucas: Exactly. That kind of inside language is actually gold for training AI. But it's also a huge blind spot — and a potential embarrassment. Luna: What do you mean? Lucas: So a recent study from Stanford and Microsoft — published late last year, around October 2025 — looked at how companies are using their own Slack and Teams chat histories to fine-tune internal AI chatbots. Think of it like giving the bot a crash course in your team's specific vocabulary. Luna: Right, so the bot learns that when someone says 'turtle,' it means a ticket that's been open for more than two weeks. Lucas: Exactly. And that can be really powerful. A customer support bot that knows your internal shorthand can route issues faster, auto-fill tickets with the right priority level — all that. But here's the catch. Luna: I'm guessing the catch is that not all slang is safe for work. Lucas: Partly. But even more subtle: the bot doesn't always know the boundaries of that slang. The study documented a case where a team used the word 'escalate' only in internal chat — but their bot, when deployed to a customer-facing portal, used that exact same shorthand. Customers saw 'This issue has been escalated to our turtle team' and had no idea what that meant. Luna: Oh wow. So the bot essentially leaked internal jargon to customers. Lucas: Exactly. And it gets worse. Another example from the research — a bot that was trained on a sales team's chat started using their slang for 'not interested' — which was 'orange' — in an email draft to a prospect. The draft said 'This looks like an orange lead, should we follow up?' The sales rep caught it, but barely. Luna: That's terrifying. I can totally see an AI assistant just auto-generating that and sending it out without a human review. Lucas: Right. And the research shows that these bots are especially bad at detecting which words are team-specific versus generally understood. They learn the co-occurrence patterns — like 'turtle' appears near 'old ticket' — but they don't have a concept of 'this is inside baseball.' Luna: So what's the fix? Do companies need to manually label which words are internal slang? Lucas: Some are trying that. A few larger firms — think financial services, law firms — are building 'jargon dictionaries' that feed into the training pipeline. They explicitly flag certain terms as internal-only and tell the bot not to use them in external-facing contexts. Luna: That sounds expensive and labor-intensive. Lucas: It is. And it's ongoing — because teams invent new slang every week. The study found that on average, a team of about fifty people generates roughly seven new slang terms per month. So the dictionary is never done. Luna: That's a lot of maintenance. I wonder how many companies are even aware this is an issue. Lucas: Probably not enough. The Stanford-Microsoft survey of about 200 companies found that fewer than a quarter had any explicit policy about using chat data for AI training. Most just assumed it was fine. Luna: And from a privacy perspective, are employees even told their chats are being used to train bots? Lucas: That's a huge question. In the US, it's a patchwork. Some companies include it in the employee handbook under 'technology usage,' but it's often buried. In Europe, GDPR requires explicit consent if the data is used for automated decision-making — but training an internal bot that doesn't make decisions about employees might slip through. Luna: So there's a real gap between what's technically possible and what's ethically done. Lucas: Absolutely. And the thing is, these bots are getting more powerful. Microsoft's Copilot for Teams, for example, can now summarize missed conversations — that's pulling from live chat data. Slack's AI does something similar. The training data is your actual daily chatter. Luna: Which brings up another angle: if the bot is trained on your chat, does it pick up on sarcasm? Inside jokes? Lucas: It tries to. And that's where things get really interesting — or really weird. The study found one bot that learned a team's running joke about 'blaming the printer' whenever something went wrong. The bot started generating replies like 'Maybe the printer did it' in response to serious technical issues. Luna: That's hilarious until a manager sees it and thinks the bot is mocking them. Lucas: Right. So there's this whole layer of cultural nuance that models just don't grasp. They pattern-match without context. Luna: Given all this, what do you think the smartest approach is for a company rolling out an AI chatbot today? Lucas: I'd say start with a clean, curated dataset — not your entire chat history. Maybe use only messages from a designated channel like #ai-training or #bot-feedback. That way you control the vocabulary. And then have a human review every output for the first few weeks. Luna: And tell employees upfront. Lucas: Yes. Transparency is key. If people know their chats might be used to train a bot, they'll self-censor — but that's actually fine for training. You want clean data anyway. Luna: So the takeaway is: the bot is only as good as the data, and the data is full of your team's inside jokes. Lucas: Exactly. And if you're using these tools, it's worth asking your IT department: 'What exactly is our AI trained on?' Because the answer might be a lot more colorful than you expect. Luna: Yeah, that's a good litmus test. If they can't answer, that's a red flag. Lucas: Right. And speaking of useful takeaways — if this conversation gave you something concrete to think about, that's exactly what we're here for. These episodes stay ad-free because of listeners who find value in them and choose to support the show. Luna: It's a small way to keep the content independent and focused on what matters, no sponsors shaping the narrative. Lucas: So if you'd like to be part of that, you can buy me a coffee dot com slash fexingo. That's it — no pressure, just an option to keep the show going. Luna: And now back to the bots. One more thing I wanted to ask — is there any sign of regulation catching up with this? Lucas: A bit. The EU's AI Act, which is being phased in through 2026, does cover some of this — particularly around transparency and risk classification. But workplace chatbots are often in a gray zone. Luna: So it's largely self-regulation for now. Lucas: Mostly, yeah. And self-regulation only works if companies actually bother. Given the cost of building those jargon dictionaries, I suspect a lot of them are just hoping the bot doesn't say something embarrassing. Luna: Which brings us back to the 'turtle' incident. Lucas: Exactly. A little foresight goes a long way.