AI for Deep (and Accurate) Analysis
Gen AI gives answers based on probability, and it's tuned to give pleasing results. Without proper guardrails, those results can be anything from pandering to hallucinatory.
THOUGHTS
Posts tagged “trust”
7 posts
Want a better AI? Dial down the warmth.
AI feedback that feels warm and affirming can also make it more likely to agree when you're wrong. Two studies show how sycophancy can distort both AI accuracy and human judgment. The fix isn't a rude AI. It's asking for evidence, candor, and challenge instead of validation.
I (Nearly) Social Engineered My Way Into My Client's Amazon Account
A client had a unique password and two-factor authentication, but somebody still added an unfamiliar authenticator app to her Amazon account. Then I watched a support rep unlock it with only an e-mail address and phone number. You can get hacked even if your security is hardened, and paying attention can still catch it.
Don't Let Your AI Pretend to Be Human
A small business wanted to name its new AI receptionist after a real employee, and pass it off as human. That's the mistake. People sense pretty quickly when a voice isn't quite right, and the moment they realize they've been fooled, they feel tricked and lose trust. The fix isn't a better voice, it's transparency. So this week, if you're letting AI answer your phone, make it honest about what it is.
How Much of Your Information Are You Feeding the AI?
Those built-in AI assistants popping up in your email, your notetaker, your CRM are genuinely helpful, and they're free, which is exactly why I'd pause. When something digital is free, you're usually paying for it with something other than money, and here that something is your data. So this week, figure out whether your AI is training on what you feed it, and set your limits.
The Person on the Call Is Your Partner. (Except… It Isn't.)
You get a video call and it's your business partner — her face, her voice, the way she laughs at the end of a sentence — asking you to move money today. In early 2024, that exact scenario cost one firm $25.6 million, and the raw material to fake you or your partner is already out there. Here's the both/and: the deepfakes are astonishing, and the two defenses that beat them are almost stupidly simple, and very human.
A dangerous trend: using AI fakes to protest AI problems
I’m seeing a dangerous trend right now: using AI fakes to protest AI problems. It goes like this: someone shares a personal experience of AI’s biases, sometimes with utterly racist or misogynistic results. Often the post includes an expert offer to help people navigate this new world. And to illustrate the story, they include a …