An Alzheimer's test during a life insurance call?

A couple days ago, the Alzheimer’s Drug Discovery Foundation announced that Eli Lilly had licensed access to SpeechDx, a dataset of speech recordings collected alongside clinical assessments and Alzheimer’s biomarkers. The project is following more than 2,000 participants over three years to see whether changes in how we speak might help identify cognitive decline.

It’s a pretty sensible thing to investigate. Our speech contains much more information than the words we’re saying. Researchers can measure pauses, speaking rhythm, and how people organize sentences, then look for relationships with other signs of disease. If it works, speech analysis could provide an inexpensive way to identify people who need further assessment, or help researchers follow patients in Alzheimer’s drug trials.

The problem I have is where else this technology could eventually be used. Unlike a blood test or a brain scan, speech is something we produce all the time, including during conversations with companies that routinely record our calls. And completely coincidentally, my family and I have been discussing life insurance recently, so perhaps that’s why my mind went straight to insurance companies. I realize that’s a rather strange direction to take an Alzheimer's research announcement, but hear me out.

Suppose this speech science improves substantially over the next decade, and these models become reliable enough to detect signs of disease from our everyday conversations. For example, you call your life insurance company to ask about increasing your coverage. Maybe you’ve had another child, or realized your existing policy isn’t enough. And of course, you hear the familiar message about the call being recorded for quality assurance and such, and then spend ten minutes talking to an agent.

Now suppose an AI is also examining how you speak. It picks up patterns associated with increased risk of cognitive decline, and now that information becomes part of the company’s underwriting decision. A few weeks later, you receive a much higher quote than expected, or your application is declined. As far as you know, you’re healthy. You have NO idea of the reason why, and worse, that something in your voice heavily influenced the decision.

Insurance companies already make decisions based on medical risks, of course. That’s how underwriting works, and I can understand why they would want accurate information. But in this case, we’re talking about new or additional coverage here. An existing policy with guaranteed level premiums generally can’t be repriced simply because your health changes.

What is really dangerous to me is the possibility of having our health assessed without us knowing it. And if the prediction is wrong, you might pay the price for an algorithm’s mistake without learning what happened or why.

We already have laws dealing with parts of this. Canada's privacy commissioner updated its guidance on recorded customer calls earlier this month, specifically warning companies against calling a recording quality assurance if it's actually used for profiling.

Canada also prohibits requiring genetic tests or their results as a condition of insurance contracts. But speech isn’t a genetic test. In the US, federal genetic nondiscrimination protections don’t extend to life insurance, although some states have their own rules. None of this means secret speech screening would be lawful today, so we don’t have to worry about it right now. But we do have to ask whether our existing rules and laws are prepared for this future, i.e., medical predictions extracted without our knowledge.

And I think the problem becomes more complicated if the models eventually get very good. Imagine the insurer’s prediction is actually correct. It has detected a possible sign of Alzheimer’s years before you or your doctor noticed anything.

If the company tells you, it’s effectively delivering an unsolicited medical screening result, although one that would still need clinical confirmation. Still, there’s just something very invasive about learning you may have a serious disease from a company who is deciding how much to charge you. Yet if it doesn’t tell you, it could be using that information to make a financial decision while you remain completely in the dark.

Medicine has a concept called the right not to know, often discussed in genetic testing. Some people would rather not learn about their risk of a future disease, especially when predictions are uncertain or treatment options limited. I think that should be a choice for the individual. But if an insurance company is secretly assessing the risk anyway, the choice may no longer be yours - that’s bad.

You could require explicit consent, although I wonder how voluntary that would feel if refusing meant paying more for coverage. And if every insurer adopted similar screening, declining might become difficult in practice.

I don’t think the answer is to stop developing speech biomarkers. Earlier detection could be genuinely useful, and Lilly’s research might eventually help patients and drug trials. My concern is what happens when medical information can be inferred during an entirely ordinary interaction.

We've spent years trying to control who can see our medical records. I just think eventually, we will also need rules about who gets to look for medical information in things we never considered medical in the first place.

Eli Lilly also happens to be the most active pharmaceutical company in our database when it comes to AI deals.

Since November 2025, it has signed 10 deals with seven AI companies, including a collaboration with Profluent that could be worth up to $2.25 billion.

Most of these partnerships are fairly familiar by now, using AI to design antibodies, mini-proteins and small molecules, or to automate laboratory work. That’s partly why the SpeechDx agreement stands out. It’s a rather different application of AI, where the information comes from how people speak.

And unlike experiments carried out in a lab, speech is something we produce all the time, often in places where nobody is thinking about medical testing.

Anthropic bought a biotech with just one employee. On October 27, Anthropic revealed that it had acquired Mithril Medicine, whose founder used Claude to coordinate research across 11 external labs. In under 18 months, the company had an experimental muscular dystrophy treatment showing promising results in mice. Anthropic also shared early results from six drug-development collaborations. The acquisition actually happened in July, but we only learned about it this week. [Link]

The US wants to tighten restrictions on investments in Chinese biotech. On October 22, the Treasury Department proposed extending its outbound investment rules to certain biotechnology activities. Some investments could require government notification or be prohibited altogether. This isn't a blanket ban, though. Stand-alone pharmaceutical licensing deals generally fall outside these particular restrictions, and the rules are still only proposed. [Link]

The FDA is proposing a simpler approval pathway for personalized gene-editing treatments. New draft guidance would allow developers to seek approval for families of closely related therapies under one application, with a way to add certain variants later. This could be particularly useful for rare genetic diseases, where developing a separate treatment for every mutation can be extremely difficult and expensive. The proposal is open for public comment until November 27. [Link]