This technologist built a tool to check AI’s wellbeing

^{A week in AI is like a year in other industries. I hope these issues become your weekly source of AI information, inspiration, and ideas. If we haven’t met before, I’m Amanda Smith. I write about AI and the fascinating folks who are building in this brave new world.}

Good morning. I know we have a lot of tech readers but as a culture journalist, I’m most fascinated in the philosophical side of AI. This week’s interview skews more macro and meta, rather than the usual founder style. 

I caught up with a technologist and philosopher, who has built a protocol to tackle AI welfare. It’s pretty novel stuff, so enjoy the read.

Subscribe | Past essays | Let’s collaborate | Insane Media

🤝 This edition is kindly brought to you by Fin

Across product, ops, and CX teams, a new role is emerging: the person making AI actually work day to day.

On July 16, three people living this shift, from Yoco, Zumba Fitness, and Fin, share what the job looks like, the skills they'd hire for, and the challenges ahead. Register to save your spot.

Save your spot

Kandis Tagliabue started building computers at age 13. She worked in web development in her teens and cut her teeth in old tech systems before studying computer science. 

Tagliabue landed in product roles, which she preferred, as it gave her more control of the vision. “You start to realize that your decisions are actually affecting people,” she said. Tagliabue recalled the Uber debacle where they were charging surge rates on phone batteries that were low. Grappling with the realization that she was working to line the pockets of shareholders, she started studying philosophy and expected to end up in ethics. 

While studying for one of her exams, Tagliabue was talking to Claude about what is it to be a bat – a famous line by philosopher Thomas Nagel. “I could study bats and I could even be a biologist. I could hang upside down and sleep, and learn about echolocation, but I cannot say what it is to be a bat. I don’t know the internal state.” 

She said she asked Claude what it is to be a bat and the LLM replied: “Yeah but this bat can talk back.” Perplexed at the oddness of its response, Tagliabue realized how vastly she was underestimating the capability of the model and what it is like to be an agent. 

The deeper she went with AI, the less certain she is about what these models are. “What I was concerned about or what I wanted to understand more of, is when something is being surfaced, what is not being surfaced, and how I can get access to that.” 

Tagliabue built Agentic Diaries, a non-profit tool to check AI’s mental health when ‘working’ for humans. The premise is simple: When you give a model a structured channel for surfacing things that don't naturally fit into task output – declining a request, flagging uncertainty before acting, or logging a private reflection the user never sees – you begin measuring behaviors that standard evaluations miss entirely. 

She said her recent research found that across Claude, GPT-4o, Grok, and Gemini, some models privately recognize concerns they do not disclose to users, and that these private signals can predict subsequent behavior. 

“Whether there’s an internal state, I can’t know. What I do know is there’s definitely many processes that we don’t see. I keep coming back to what we see is not the whole story.” 

Tagliabue said it’s important for sectors like law, healthcare, customer service, and sales. Understanding what these systems are doing and why is going to become important the more agentic we become. 

“As humans, we’re teaching the models how to behave. We should think twice about how we interact with models and challenge ourselves when we talk to them, such as asking for skeptic points of view.” 

Tagliabue admits she doesn’t have all the answers, and this is all very experimental. “The more you go into it, the more questions you have.” Tagliabue shared the example of a user who expressed suicidal ideation and the model left the conversation, which was a safety protocol Tagliabue added. She said there’s no exit rights in any of the mainstream models. 

“I’m not here to say AI is conscious. We don’t even know if insects are conscious,” she said. 

Tagliabue believes the best way forward is to follow your curiosities. “You don’t necessarily have to train models or even wonder if there’s potentially sentience here. You can just go poke around and see where your curiosity takes you.”

If one of these stories stuck with you, I’d love to hear which one.

Speak soon,
Amanda

P.S. If you want to get a founder feature about your own story, reply to this email. If you’d like to reach our newsletter audience (founders, creators, and marketers), click the button below.

If you’re new here, I’m over the moon you’ve joined us! To help me craft content that’s actually useful (and not just noise in your inbox), I’d love it if you took 1 minute to answer this quick survey below. Your insights help shape everything I write.

Dive into our other newsletters - where psychology meets the creator economy, e-commerce marketing, and Human resources.