On October 1, we at Tavus previewed Griffin-Lite, an AI you talk with over live video. It listens while it talks, it pays attention to your face and your pauses as well as your words, and it answers with its own voice, face, and timing. In a short study, 26 of the 54 people who talked with it for a minute thought afterward that they had talked with a person. Our previous model convinced 1 person out of 41. I'm insanely proud of what our team built - not because it’s capable of fooling people, but because of what it means for broad access to revolutionary capabilities.
A week before the launch, anticipating the kinds of responses we might get, I wrote a note to the team about why I do this work. I knew Griffin would provoke strong reactions, and I wanted to take the calm before the storm to put the reason behind my convictions on its worthiness in writing. I expected some people to feel amazement and excitement, and others to feel fear, distrust, anger, etc. I expected people to ask who needs this, who asked for it, and why anyone would build it.
And that's exactly what happened. A lot of people were excited. A lot of people were angry, and many of the angry ones were also afraid. One tech journalist called Griffin a machine for tricking old people into handing over their bank passwords. Senator Bernie Sanders shared our announcement and asked, "Does anyone in America think it is a good idea to race forward to build machines that replace us?" For a minute, we were the poster child for what a lot of people think is wrong with the way AI is being built.
Below is an edit of that note for anyone who wants to read it.
Where my why comes from
I'm trained as a human-centered designer. That means I start from the needs, behaviors, and contexts of the people involved, and not from the technology. So my why starts with a human problem. Why have I spent so many years working toward AI that is as natural and easy to talk with as I can help make it, and now as humanlike as well?
It goes back to our first microwave oven, in 1983, when I was a kid in Dallas. Summer storms kept knocking out our power, and the microwave clock kept blinking 00:00. Resetting it was simple, but not intuitive, and my parents couldn't do it. I became the clock whisperer. They couldn't program the VCR either. Lest you think I'm throwing my parents under a bus: They're smart, patient, successful people. My dad built companies responsible for America's first cellular infrastructure. My mom is a polyglot with three master's degrees. She was a journalist and a poet, and we had a computer in the house that same year. Yet she stuck with her typewriter, Liquid Paper, and crumpled drafts into the early 1990s, because typing "bsw" at the A> prompt intimidated her, and the keyboard menus that came after it petrified her. The list of tech non-starters went on, and it still does. In myriad small ways, technology's advances disempowered my parents. No single change was dramatic, but each one added to the pile of things they couldn't do without help.
Resetting the microwave's clock required knowing a cheat code. Programming the VCR meant memorizing a paper instruction pamphlet. The everyday electronics of that time put up barriers, and my parents couldn't get past them without someone there to help.
Everyone reading this has a similar story about someone in their life. It's an old story, and it's a current one. New technology helps some people and leaves others behind. With some notable exceptions, each new kind of technology has asked people to learn new cheat codes and leap over trust chasms. Over the years, the cheat codes have grown more complex, and the leaps have gotten bigger. Technology that could help everyone ends up widening what people call the digital divide, because for many people there is too much to learn. We keep creating powerful new capabilities faster than we make them reachable for ordinary people.
I've interviewed and shadowed hundreds of people using all sorts of systems, and one of the most common things I've heard is some version of "Why do I have to learn a whole new way of doing things every time a tech company invents something new?" What would you say to someone who asked you that? "It's not that hard"? "What's the big deal?" Would you try to make the learning easier, or design a better tutorial?
Or would you ask what it would take for the question to stop coming up? What if people didn't have to learn a new way of doing things, because the new thing worked the way they already communicate? A lot of the discouragement, fear, and anxiety people feel about being left behind comes from that friction. We probably can't make all of it go away, but we can remove a meaningful share of it.
Most of us learn to talk before we learn to read or write, and years before anyone hands us an instruction manual. We express ourselves most fully in conversation with other people. I don't mean voice commands, where you have to know the right words to say. I mean conversation, where the back and forth lets two people build common ground and work out what one another mean.
Talking isn't universal. People who are deaf, people with aphasia or other speech and language conditions, and people whose language a system doesn't handle well all run into barriers with conversational interfaces. But conversation is closer to a skill we all share than any interface anyone has designed. You don't need to memorize instructions or learn a keystroke sequence. You just talk the way you already talk, and for a great many people, so many barriers come down.
Why high naturalism matters to me
The sharpest version of the why question isn't about conversation. Voice assistants and chatbots already let people talk to computers. The sharpest version is about realism. Why does it have to look and sound so human that people can't tell?
My answer is that a lot of what makes conversation work isn't in the words. Linguists call the process of reaching shared understanding "grounding." Herb Clark and Susan Brennan wrote the classic paper on it in 1991, and one of their points is that face-to-face conversation gives people more resources for grounding than most other media do. We can see each other. We can see when the other person looks confused, when their attention wanders, and when they nod. We say "mm-hm" and "wait, what?" while the other person is talking, and we repair misunderstandings as they happen, before they cause a problem.
When someone is lost in a conversation, you can see it on their face before they say anything. A text box just can't show you that. A voice without a face carries some of it, but not most of it. The face, the expression, and the timing in Griffin are there because they are the signals people use to tell whether they're being understood, and to show when they aren't. Those are indispensable for grounding, and they are bidirectionally important. A face gives the person something to read, and a system that can see the person can read them in return.
For someone who's already confident with technology, those signals make a conversation with an AI more pleasant. But for someone who's intimidated by technology, those signals can decide whether they keep asking for help or give up. That's the reason realism matters to me. The people I most want to reach need to know they're being understood, and realism gives them the cues they use to tell.
Think about what that could replace. A phone tree asks you to learn its rules before it will help you. You listen to the menu, you make an educated guess about which number matches your problem, and if you guess wrong, you start over. Some people have learned the trick where you keep pressing zero until a person picks up, but a lot of people never learned that trick. When a caller can explain their problem in their own words and be understood, the person who doesn't know the trick gets the same help as the person who does. That's what my parents needed from me with the VCR.
Where I share common concerns
Many of the objections people raised are fair. Here are the ones I think are right.
The study is small, and the headline went further than the study does. Twenty-six people out of fifty-four is a real result, but it's a small one. The calls lasted one minute. We ran the study ourselves, and nobody outside Tavus has reviewed it. Participants were told they'd be talking with another participant, so they came in expecting a person. That's different from Turing's version of the test, where a judge is trying to find the machine. The people who said "passed the video Turing test" claims more than the study shows have a point. I'd like to see larger studies run by people who don't work for us.
Realistic AI will be used for scams. People like my parents are whom scammers target, and the same qualities that make Griffin easier to talk to could make a scammer's AI more convincing. Scammers aren't short on tools today, and realism isn't always what they're after. Researchers have found that some scams are obvious on purpose, with bad fakes and stories nobody should believe, because that filters out everyone except the people most likely to fall for the rest of the scam. But other scams, like a fake video call from a boss or a family member, depend on being convincing, and those are the ones a model like Griffin could make easier. I'd like to tell you that people will get better at spotting AI as fast as AI gets better at passing for human, but that's my guess, and a guess isn't a reassuring answer. I'll describe what we've done about it below.
Disclosure can't run on the honor system. Several people pointed out that if a company can tell its AI to say it's human, a promise to be honest doesn't protect anyone. They're right. At Tavus, people build AI characters we call PALs. Until recently, a customer could tell a PAL in its instructions to say it was a person. That made sense when nobody could mistake a PAL for a person. It doesn't make sense anymore, so we changed it. A PAL can no longer be instructed to claim it's human.
People will get attached, and some of them will get hurt. People already form intense attachments to text chatbots, and some of those stories have ended badly. A super realistic face and voice will make those attachments stronger, and lonely people and kids are the most vulnerable to those kinds of attachments. I don't think this worry is overblown. Most of what I want a PAL to replace isn't a person in anyone's life. It's a phone tree, a manual, or a form someone couldn't get through. I recognize that's my intention, and intentions don't decide how a technology gets used. That said, we turned down our own companion PALs at Tavus, and we are installing other mitigations to reduce the likelihood of such unhealthy attachments to PALs our users may make.
This will replace some people's work. Some companies will use AI like ours for work that people do today, and that's already happening. The people who worry about it are worried about their jobs and their families. I don't have a satisfying answer to these worries, and I don't think anyone in this industry (or government, or academia, etc) has one yet.
It's moving too fast. When Senator Sanders asked whether anyone thinks it's a good idea to race forward, he was speaking for a lot of people who feel this technology is arriving faster than anyone asked them about it. I agree that speed is a choice, and that the companies making that choice owe people their reasons. Holding back the full Griffin model is our attempt to make that choice in the open, to share the breakthrough but address and mitigate risks before releasing it as a product.
Where I see it differently
Some objections I understand but don't share. I don't think the people who raised them are wrong to feel the way they do, but here's where and why we diverge.
"There's no legitimate reason for an AI to pass as human." I frame the goal differently. The goal isn't for an AI to pass as human. It's for an AI to give people the same signals another person would: that it heard them, that it noticed they're confused, and that it's waiting while they think. Those signals are what let someone like my mom keep asking for help instead of giving up. The realism that makes those signals work is also what led 26 people to believe they'd talked with a person, and that's the reason disclosure can't be optional. But I don't think the answer is to diminish the signals.
"Make it obviously fake." Some people suggested that AI should look and sound robotic on purpose, so nobody could mistake it for a person. I understand the instinct. About 10 years ago, I held the same position. I didn't want anything to do with the uncanny valley. But an interface that's hard to read is the kind that has always left people behind, and the people it leaves behind first are the ones I most want to help. I'd rather an AI be easy to talk to and tell you what it is.
"Pause it all." I don't think pausing all AI work is the answer, and I know reasonable people disagree with me. A pause would be honored by the companies most willing to hold back and disclose, and ignored by the people least willing to. Scammers already clone voices with tools anyone can download. I think it helps more to build these systems in the open, with honesty built into how they work, so that becomes the standard everyone is held to. Beyond that, we need to talk to our policy makers.
"The harms will outweigh the benefits." Some people said the benefits, whatever they are, can't be worth the harm. I think part of the reason is that the benefits are harder to see than the harms. A scam makes the news. A person who finally got help from their insurance company without giving up doesn't. We live in a world of mixed outcomes. We always have. Super realistic faces on AI aren't likely to change that.
What we're doing
Here's what is in place today, what we're working on, and where the work is ongoing.
For starters, Griffin and Griffin-Lite are not yet available to customers. We've started on-boarding a small group of testers to the Griffin-Lite research preview, and we're holding both models until the risk mitigations are dialed in.
We have guardrails in place to control whose face and voice can be used. These aren't perfect, but we're iterating, and that's in part why Griffin is being held back.
We've also added strict guardrails against misuse, preventing A PAL from being used to impersonate a real person, and refusing a range of other things bad actors would try to use it for. PALs say they're AI, and customers can't override this.
We don't train our models on what a PAL sees or hears in a conversation, or on what it remembers. A PAL's memories come only from its conversations with one person, and they don't leak to anyone else.
We're working with outside organizations tackling AI safety, including manipulation and deception. That work will shape how and when we release.
Finally, we're also working on both audio and video watermarking, so that any media generated from Tavus will be immediately identifiable as AI-generated.
This doesn't add up to 100% protection from threats posed by determined, bad actors. Scammers already clone voices from a few seconds of audio, and they don't need our permission or our models to do it. What we can do is ground in transparency & safety on our own platform, and thereby lead the industry toward a norm where every AI says what it is. When everyone expects an AI to identify itself, anything that doesn't will be more likely to raise protective suspicion.
Why I keep going
Most of the people who responded to us with concern were picturing someone like my parents: smart, capable people on a call they don't fully understand, about to be misled. I was picturing them too. The people the critics want to protect are the same people I started doing this work for, and I think about them constantly, including every time we make a decision about Griffin. So I want to go back to them.
When my parents wanted to go out to dinner without missing the MacNeil/Lehrer NewsHour, they asked me to tape it. They didn't need the instruction manual, or to have memorized a series of button presses. They asked me the way they would ask me for anything else, and I understood them. I was in the room, I could see what they needed, and I knew them well enough to understand what they meant.
There's one more part of that scene that matters to me now. Asking me was easy for my parents because they knew me, and they trusted me with the VCR for the same reason. They didn't have to wonder who they were talking to. That's the standard I hold this work to. I want people to talk with an AI with the same ease my parents had when they asked me for help. And I want them to know what they're talking to, the same way my parents knew who they were asking.
That's my why.