From the Lab’s Biggest AI Skeptic to Its Advocate

Jessica Armand's PhD at Columbia examined how cancer cells decide to divide, specifically the G1-to-S phase transition, and what gets expressed differently in a cancer cell to drive it through to replication. Her thesis project focused on CDK4/6 inhibitors, the first-line targeted therapy in breast cancer. Within a few years, patients typically become resistant to it. Armand spent five years working out what creates that resistance and proposing combination therapies to prevent it.

Her motivation sharpened during a session at the San Antonio Breast Cancer Symposium that included patient testimonials. What patients said they cared about was quality of life during treatment. Two years on a well-tolerated targeted therapy is a different two years than those spent unable to get out of bed or attend a child's soccer game.

She was also, by her own admission, the lab's leading AI skeptic. The models available during her PhD returned black-box answers with no reasoning and no sources. She watched other scientists take those answers at face value. She would ask why they chose a next step, and they'd say a generic LLM told them to. It made her deeply wary of the whole category.

When Armand joined Causaly, former colleagues were surprised. You were the number one AI curmudgeon, they told her. You were constantly telling us not to use it. Her answer was that it wasn't the right tool for research.

You now spend a lot of your time onboarding scientists. What's the typical mindset of the people you engage with?

Scientists are deeply skeptical creatures by nature, and I love that about us. Some people show up to onboarding excited. Others show up with a healthy skepticism and want to be convinced.

I'm always prepared to do some convincing, which is ironic because a year ago I would have needed to be convinced too.

What earns trust is transparency. Every piece of information comes with an inline citation that takes you straight back to it's original source. When I'm on a call with a skeptical scientist, that makes all the difference. We're helping them navigate an amount of information that can't be digested manually anymore, without summarizing the literature and asking them to take our word for it or removing the guesswork by making the decision for them.

What's the moment where it clicks?

Inline citations, and one other thing. In BioGraph you can click into a given relationship, say the role of IL-6 in lung cancer. Expand it and you get a relationship summary, showing that 500 papers say it up-regulates and 50 say it down-regulates.

Seeing that contention on screen is a huge aha moment. We surface the disagreement instead of breezing past it to declare it up-regulates, because we want you to know the opposite claim exists. If it interests you, filter down to the papers that say down-regulate, investigate them, and draw your own conclusion.

There's a consistency argument too. Ask ten scientists to go check the literature on IL-6 in lung cancer without a tool like this, and ten different answers come back. Now everyone's looking at the same landscape, and the conversation shifts to what the 500 and the 50 mean.

You've said "faster" is the wrong goal. Why?

Done badly, AI adoption just gets you wrong answers faster.

What I see when I check in with users a month or two after onboarding is confidence in decision-making. You could spend six months reviewing the literature on IL-6 in lung cancer and still not be sure you have the right relationship, asking yourself whether you need another month. That's what tells you where to progress and where not to.

I say this as someone with a specific fear about it. In my second year of my PhD, I got scooped. Someone else found what I found and published first, so I had to start over. Everyone doing a PhD carries that anxiety, needing enough literature review to know a target is worth pursuing but not so much that someone else gets there first. This happens in research often, it’s a bit of an academic boogeyman. The threat here is that if you move too quickly through the literature review, you may miss something crucial. You can put years into a question, and discover it was answered and published before you started. Once that's the case the work is redundant, and there's nowhere to publish it.

You waste years of your life because you knew the 500 up but you didn't know the 50 down. Years into a project, and you're back at square one.

What happens when scientists use a general-purpose LLM instead?

I'm always surprised to hear that anyone is having a good time with one. During my PhD, there was not a single thing a generic LLM could have told me that I would have believed if it didn't link me to a PubMed article.

Think about the actual workflow. You go to a black-box model, you get an answer, and then you still have to run the manual literature search yourself to find something that validates it. It removed maybe two percent of the work. It isn't showing you empirical data for why the response is correct.

I'll give it credit on the computational side. It was helpful when I was learning sequencing pipelines and for code generation. For scientific thinking or scientific direction, it disappointed me.

Where is the field still stuck as it relates to AI?

On the idea that AI is here to replace expertise. It isn't, and the framing gets in the way.

The bottleneck moved. It used to be data generation. With the democratization of sequencing and multi-omics, generating data got cheap and fast. Now the bottleneck is interpretation, keeping up with what exists. Human cognition cannot match the pace of research output. That's simple arithmetic.

There's something almost serendipitous about the timing. Sequencing became foundational, and a few years later these tools arrived to help scientists keep up with the knowledge that sequencing produces.

We need the human in the loop. We aren't making the final decision. We're putting the information in front of you so your expertise has everything it needs to proceed with confidence.

Where do you hope this ends up, in oncology specifically?

Anticipating tolerability. Predicting and preventing adverse events requires an unbelievably complicated network of organ systems, PK/PD, and tolerance, etc. all interacting.

In my own research I could look at one drug, one axis of one pathway, and answer 'go or stop'. I was working in two dimensions: does it hit the target, and is it safe enough to continue. Whether that go-or-stop decision was well tolerated cognitively, or what it meant for someone's quality of life, sat miles above what I had the tools to consider.

My hope is that we get to build on the Z-axis, not only the X and the Y, so that while you're making the stop-or-go call, you can also address the higher order impacts like quality of life. That's what patients told us at that conference they wanted, and it's still the thing I most want the science to reach.

‍