Dr. Denise Herzing has spent 40 years listening to dolphins in The Bahamas, recording their clicks, whistles, and bursts in an effort to understand how they communicate. As a 12-year-old flipping through encyclopedias, she stopped at the whale and dolphin page and wondered: what's going on in their minds? Around the same time, Jane Goodall was studying chimpanzees, Dian Fossey was observing gorillas, and Jacques Cousteau had opened up the ocean world to new possibilities. Young Denise decided that's what she would do.
She studied marine biology and kept her focus on dolphin minds and dolphin communication. When she reached graduate school, she moved to Florida to work in The Bahamas. She was surprised to find no one was really working there, so she planted herself for what she thought would be 20 years. Four decades later, she's still there.
Dr. Herzing was recently named to the Time AI 100 list for her pioneering work using artificial intelligence to decode dolphin communication. Her research now powers DolphinGemma, a collaboration with Georgia Tech and Google DeepMind, and the first large language model trained on dolphin vocalizations.
The Technology That Had to Catch Up
From the beginning, Dr. Herzing knew she needed computational tools that didn't exist yet. She always had video and sound recording through commercial cameras, though the sound range was limited to human hearing. Dolphins vocalize into ultrasonic frequencies, so she was only capturing part of the picture. Recording high frequency sounds was possible but challenging for her specific situation. Her team swims with dolphins in open water, moving with them rather than staying tethered to a boat.
The technology evolved slowly. In graduate school, she used a cylinder device with carbon paper that rotated and sketched out sounds. Once everything became digital, the possibilities expanded, but manual work remained. Pulling sounds out of recordings. Matching them to behavior. Analyzing spectrograms one by one. Cameras got smaller. Sound recording improved. Now it's all GoPros and specialized recorders and 360-degree cameras.
But she always knew that someday there would be a computer she could throw her sounds into and make sense of patterns that human minds couldn't easily detect. She just had to wait for that technology to arrive.
Many associations sit in a similar position. Years of conference recordings gathering digital dust. Publications capturing decades of industry evolution. Member data spanning multiple platforms and eras. All of it waiting for tools that can reveal what's actually there.
How DolphinGemma Works
Dr. Herzing collaborates with Thad Starner's computer group at Georgia Tech and the team at Google DeepMind on DolphinGemma, a large language model trained on 8-10 years of dolphin vocalizations. The process works like this:
- Input a dolphin sound into the model
- The model generates what patterns might follow based on training data
- The team reviews underwater video to see what dolphins were doing when those patterns occurred
- They label patterns (A, B, BC) to track sequences over time
- The model surfaces patterns humans miss in manual review
DolphinGemma is generative AI, meaning it can take a sound and show what's likely to come next. A whistle followed by squawks. Particular burst patterns. The team integrates these with video footage to see what the dolphins were actually doing. They know the individuals, so they can identify patterns from specific mother-calf pairs or male coalitions.
The model runs on just 400 million parameters, remarkably small compared to commercial models with billions. Small enough to run on a device rather than requiring massive cloud infrastructure. The team now uses off-the-shelf equipment, including a Pixel phone, all in a small waterproof box they wear while swimming.
The challenge after identifying sound patterns will be interpreting them correctly. The team should be able to do much of that with their video archive, but questions remain about whether they're really getting it right. They might eventually need to play sounds back to dolphins and observe reactions.
Built for Sharing
DolphinGemma was designed from the start to be open source. The plan has always been for other researchers to plug in their own data and use the same architecture. Dr. Herzing studies spotted dolphins in The Bahamas, but the model could help researchers studying bottlenose dolphins, orcas, or any other species worldwide.
Open source matters in science because you need tools everyone can use and compare results across studies. Science advances through replication and comparison. If every researcher builds proprietary tools that only work with their specific setup, you can't compare findings or scale understanding.
Associations face similar decisions. Do you build proprietary tools that only work with your specific data? Or do you invest in approaches that could benefit your entire industry? The collaborative model matters when you're trying to understand complex systems, whether dolphin pods or membership communities.
Dr. Herzing acknowledges she's on the user end. The Georgia Tech team and Google DeepMind are doing the actual AI work. Her role is providing data, domain expertise, and biological context that makes patterns meaningful. She sees the Time AI 100 recognition as a chance to showcase that AI has purposes beyond typical business applications. Science needs these tools, especially researchers with smaller data sets who need quicker ways to examine their findings.
Why Human Supervision Still Matters
Computers don't know what you're looking for. They find patterns, but humans verify those patterns actually matter. Dr. Herzing references a classic study where researchers tried to identify individual polar bears from photos. The AI kept identifying types of snow instead of bears. The computer found patterns, just not useful ones.
Or imagine analyzing conference recordings. AI might cluster pen-tapping sounds as significant data because they follow a rhythmic pattern. Someone coughing could show up as a meaningful cluster. Pattern recognition without domain expertise produces garbage.
The supervised approach has been critical throughout Dr. Herzing's machine learning work. In earlier projects, she had to look at sound clusters and verify them. These are similar sounds, call that cluster A. Here's cluster B. The computer pulls out patterns, but humans provide context that makes those patterns meaningful.
She mentions an interesting discovery from their CHAT system work (Cetacean Hearing and Telemetry). The underwater computer uses synthetic whistles outside the dolphin's natural repertoire to avoid accidentally saying something offensive in dolphin. While studying how dolphins mimicked these whistles, they discovered dolphins were mimicking in a way the team hadn't seen before in natural communication.
That discovery meant going back through basic sound data to look for those patterns. They might not have found what they wanted with CHAT results, but they discovered something unexpected about how dolphins might be communicating. Science often works that way. You look for one thing and find something else valuable. But you need human expertise to recognize when an unexpected pattern actually matters.
Beyond Marine Biology
Dr. Herzing sees AI as essential for analyzing complicated natural systems like ecosystems, climate patterns, and weather. These systems generate enormous data with intricate relationships that manual analysis can't fully capture. The computational power exists now to examine what was previously too complex to model.
She doesn't have answers about energy concerns around AI. The water needed for cooling, the electricity required. She acknowledges those are real issues needing national and global attention. Maybe AI itself can help solve those problems. How do you get more efficient with energy? Could quantum computing reduce requirements?
But she remains convinced AI will advance science in critical ways. Medical breakthroughs already demonstrate this. Researchers with smaller data sets need these tools.
Association data in some ways carries similar complexity to natural systems. Member behavior shifts over time in response to industry changes, economic conditions, generational preferences, and your own programming decisions. Engagement happens across events, publications, online communities, certification programs, advocacy efforts. Learning preferences vary by career stage, geographic region, organizational role.
These are complex interactions between dozens of variables playing out over months and years. Human analysis can't hold all those variables in mind at once while looking for patterns across thousands of member records. AI can.
Dr. Herzing spent four decades collecting data before the right computational tools arrived. Your organization has probably accumulated years of information about your members and your field. The question isn't whether that data matters. The question is whether you're ready to examine it with tools capable of revealing structure you couldn't see before.
She still has more data to feed into DolphinGemma. The model currently trains on 8-10 years of recordings, but she has decades more to add. Looking ahead, she'd like to see other researchers contribute their data sets. Maybe they can't get 40 years of recordings, but if DolphinGemma becomes efficient enough, perhaps a month or two of intensive data collection could yield insights.
The ultimate goal involves scaling beyond dolphins. The architecture might eventually incorporate terrestrial species. Birds, elephants, other long-lived social mammals. That was always part of the design vision.
The computational power exists. The architectures are getting more efficient. What you learn might surprise you, much like what Dr. Herzing continues to discover after 40 years of listening. Sometimes you just need better tools to hear what's been there all along.
.svg)
