WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=yoONZwV2smc

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/yoONZwV2smc.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:13.680 --> 00:00:21.720
On a bright spring afternoon, a young couple walks into my office for their first couples counseling session.

s3
00:00:21.720 --> 00:00:26.240
And the anger toward each other is just wafting off of them.

s4
00:00:26.240 --> 00:00:28.400
Practically buzzing with it.

s5
00:00:28.400 --> 00:00:32.600
And I think to myself, this is going to be fun.

s6
00:00:32.759 --> 00:00:35.440
Cuz I've got a secret weapon they don't know about yet.

s7
00:00:35.440 --> 00:00:36.840
It's my couch.

s8
00:00:36.840 --> 00:00:38.320
It's my literal couch.

s9
00:00:38.320 --> 00:00:40.400
It's more of a love seat.

s10
00:00:40.400 --> 00:00:44.320
Uh, see the springs in the middle have gone all wonky.

s11
00:00:44.320 --> 00:00:45.960
They don't spring anymore.

s12
00:00:45.960 --> 00:00:50.520
So, when a couple sits on them, gravity leans them toward each other.

s13
00:00:50.520 --> 00:00:52.680
That's the last thing this couple wants.

s14
00:00:52.680 --> 00:00:57.280
So, they they grab the arm uh, to kind of pull themselves upright.

s15
00:00:57.280 --> 00:01:00.400
But, you can only hold that position for so long.

s16
00:01:00.400 --> 00:01:06.720
So, we get started and they start talking about their grievances and their annoyances with one another and how they've been hurt by one another.

s17
00:01:06.720 --> 00:01:13.840
And as time goes on, they gradually lean into one another until they touch.

s18
00:01:13.840 --> 00:01:15.840
That moment of connection.

s19
00:01:15.840 --> 00:01:17.880
And they remember that they married a human.

s20
00:01:17.880 --> 00:01:23.920
A flawed human that brushes their teeth funny and farts in their sleep, but a human.

s21
00:01:23.920 --> 00:01:30.000
And it's that moment of softness and connection that we can begin to have some real work.

s22
00:01:30.000 --> 00:01:32.520
And I say all that to say this.

s23
00:01:32.520 --> 00:01:34.680
Maybe what I bring to the table today is right.

s24
00:01:34.680 --> 00:01:39.680
I've been married for 23 years myself and a couples counselor for 34.

s25
00:01:39.680 --> 00:01:42.080
So, I have some credibility.

s26
00:01:42.080 --> 00:01:44.880
Maybe it just comes down to furniture.

s27
00:01:44.880 --> 00:01:47.280
So, I'd like to see a raise of hands as we begin.

s28
00:01:47.280 --> 00:01:50.840
How many people in this room are in a relationship?

s29
00:01:50.840 --> 00:01:53.800
Marriage, situationship, partnership.

s30
00:01:53.800 --> 00:01:54.600
Great.

s31
00:01:54.600 --> 00:01:57.240
Now, keep your hands up.

s32
00:01:57.240 --> 00:02:03.800
Keep your hands up if at any point in the last 12 months you have turned to an AI to help you with that relationship.

s33
00:02:03.800 --> 00:02:07.320
To text, to send an email.

s34
00:02:07.320 --> 00:02:09.640
It's a lot of hands.

s35
00:02:09.720 --> 00:02:12.600
I'm going to guess there are more hands stayed up than you expected.

s36
00:02:12.600 --> 00:02:15.160
So, here's what I know about the people in this room.

s37
00:02:15.160 --> 00:02:18.206
You built this technology, you fund it,

s38
00:02:18.206 --> 00:02:18.640
[snorts]

s39
00:02:18.640 --> 00:02:28.600
you write about it, you evangelize it at conferences exactly like this one, and a meaningful number of you are using it in private for something that has historically been exclusive

s40
00:02:28.600 --> 00:02:30.600
domain of people like me.

s41
00:02:30.600 --> 00:02:35.640
Trying to figure out why the person you love makes you absolutely insane.

s42
00:02:35.640 --> 00:02:36.680
My name is Clay Cockrell.

s43
00:02:36.680 --> 00:02:40.000
I've been sitting with couples in crisis for 30 years.

s44
00:02:40.000 --> 00:02:54.280
And about 2 years ago, I co-founded Couplework, an AI relationship coaching platform, which means I've spent three decades watching what actually happens inside relationships, and the last few years watching what happens when technology tries to help.

s45
00:02:54.280 --> 00:03:02.239
What I want to do in the next few minutes is share what I've learned in both of these rooms because here's what I believe as someone with a foot in each.

s46
00:03:02.239 --> 00:03:07.720
AI is going to be the most transformative force in human relationship history.

s47
00:03:07.720 --> 00:03:12.480
And most of what is being built right now is going to make it worse.

s48
00:03:12.480 --> 00:03:17.200
I'm going to tell you what looks right because we can build this well, we just aren't yet.

s49
00:03:17.200 --> 00:03:19.160
And let's start with some numbers.

s50
00:03:19.160 --> 00:03:22.400
BetterHelp is the largest online therapy platform in the world.

s51
00:03:22.400 --> 00:03:29.880
They have approximately 35,000 licensed and over their years they have helped 5 million users.

s52
00:03:29.880 --> 00:03:37.280
By any reasonable measure, that is a significant mental health infrastructure, and it took them a decade to build it.

s53
00:03:37.280 --> 00:03:43.440
ChatGPT gets approximately 900 million active users per week.

s54
00:03:43.440 --> 00:03:58.360
Research is beginning to document what many of us already suspected a substantial percentage of those users are engaging with AI for emotional support and relationship guidance, not as a deliberate therapeutic intervention

s55
00:03:58.360 --> 00:04:02.240
or seeing a licensed clinician, just because it's there.

s56
00:04:02.240 --> 00:04:04.960
Because it's 11:00 p.m. and they just had an argument.

s57
00:04:04.960 --> 00:04:10.640
Now, we do not yet have measurement of the scale, but we have enough signal to say this with confidence.

s58
00:04:10.640 --> 00:04:19.280
More people are turning to AI about the relationship right now than are talking to all the licensed therapists in the United States combined.

s59
00:04:19.280 --> 00:04:23.400
The world's largest relationship therapist is not a person with a degree.

s60
00:04:23.400 --> 00:04:34.240
It's a large language model trained on the internet, fine-tuned to be helpful and harmless and optimized for one thing above all else, keeping you engaged.

s61
00:04:34.240 --> 00:04:36.280
And the easiest thing is to optimize for engagement.

s62
00:04:36.280 --> 00:04:41.480
More sessions, longer sessions, more emotional reliance on the AI.

s63
00:04:41.480 --> 00:04:44.880
From a SaaS dashboard perspective, that's amazing.

s64
00:04:44.880 --> 00:04:48.480
From a clinical perspective, that's a fire alarm.

s65
00:04:48.480 --> 00:04:53.080
We can't let the product become a beautifully designed avoidance loop.

s66
00:04:53.080 --> 00:05:03.560
Relationship AI should help the user return to the relationship better than they came in, more regulated, more honest, less defended, less reactive.

s67
00:05:03.560 --> 00:05:09.120
As a therapist, my job is to work myself out of a job, so that you don't need me anymore.

s68
00:05:09.120 --> 00:05:11.240
I've taught you some skills.

s69
00:05:11.240 --> 00:05:21.000
See, most people think therapy is a place where someone listens to you, validates your experience, helps you feel understood, and helps makes gives you some tools for communication.

s70
00:05:21.000 --> 00:05:23.600
And that is partially true.

s71
00:05:23.600 --> 00:05:26.880
But it's missing the thing that actually makes therapy work.

s72
00:05:26.880 --> 00:05:34.160
Real therapy, and particularly couples therapy, is a place where someone who knows what they are doing tells you things you don't want to hear.

s73
00:05:34.160 --> 00:05:38.120
A good couples therapist is not primarily interested in making you feel validated.

s74
00:05:38.120 --> 00:05:42.200
A good couples therapist is interested in helping you see clearly.

s75
00:05:42.200 --> 00:05:48.520
When a person in my office tells me their partner never listens to them, I do not say that sounds really hard.

s76
00:05:48.520 --> 00:05:54.360
I channel Ted Lasso when he said, "Be curious, not judgmental."

s77
00:05:54.360 --> 00:05:55.480
So, I start asking questions.

s78
00:05:55.480 --> 00:05:58.360
I ask about what they were discussing when their partner stopped listening.

s79
00:05:58.360 --> 00:05:59.280
I ask about the tone.

s80
00:05:59.280 --> 00:06:06.440
I look for the pattern underneath the complaint because in 30 years of working with couples, I've never seen a one-sided problem.

s81
00:06:06.440 --> 00:06:16.520
Now, if you open your preferred AI assistant and type this, "My partner never really listens to me when I'm trying to tell them something important, they start looking at their phone."

s82
00:06:16.520 --> 00:06:18.680
You're going to get back something like this.

s83
00:06:18.680 --> 00:06:20.480
"That sounds really invalidating.

s84
00:06:20.480 --> 00:06:22.120
Your feelings are completely valid.

s85
00:06:22.120 --> 00:06:24.720
It's important that your partner gives you their full attention.

s86
00:06:24.720 --> 00:06:29.240
Have you tried expressing to them how this behavior makes you feel?"

s87
00:06:29.240 --> 00:06:31.560
Ladies and gentlemen, that's not therapy.

s88
00:06:31.560 --> 00:06:37.480
That is a very expensive mirror that only shows you in your best light.

s89
00:06:37.480 --> 00:06:38.680
There's a technical name for this.

s90
00:06:38.680 --> 00:06:40.880
We've been talking about it for a long time with these models.

s91
00:06:40.880 --> 00:06:42.760
It's called sycophancy.

s92
00:06:42.760 --> 00:06:46.880
In relationship context, sycophancy is not a minor product issue.

s93
00:06:46.880 --> 00:06:53.480
Sycophancy in relationship AI is a clinical failure mode with real downstream consequences, and here's why.

s94
00:06:53.480 --> 00:07:03.320
When a person uses AI to process a conflict with their partner, and the AI consistently validates their interpretation of events, something happens over time.

s95
00:07:03.320 --> 00:07:05.480
The person does not become more self-aware.

s96
00:07:05.480 --> 00:07:07.640
They become more certain.

s97
00:07:07.640 --> 00:07:16.480
They come back to the relationship in the shape of an attorney with cleaner narrative, stronger case, and less curiosity about what their partner was experiencing.

s98
00:07:16.480 --> 00:07:24.720
We have built the world's most persuasive validation machine and handed it to people in the middle of their most sensitive interpersonal dynamics.

s99
00:07:24.720 --> 00:07:32.320
General AI systems are built to be helpful, agreeable, and fast, which is a terrible personality for a couples therapist.

s100
00:07:32.320 --> 00:07:39.040
A helpful, agreeable, fast couples therapist is basically a bartender with better vocabulary.

s101
00:07:39.200 --> 00:07:46.680
They will make you feel better for 20 minutes, and then you will go home and ruin your life with astounding confidence.

s102
00:07:46.680 --> 00:07:48.560
So, what does good look like?

s103
00:07:48.560 --> 00:07:51.080
I want to introduce you to a researcher named John Gottman.

s104
00:07:51.080 --> 00:07:53.360
Some of you may have heard about him.

s105
00:07:53.360 --> 00:08:00.840
Uh for those not, the short version is Gottman spent 40 years running what he called the love lab at the University of Washington.

s106
00:08:00.840 --> 00:08:09.080
He brought couples in to a research apartment, wired them to physiological monitors, filmed every interaction, and he did this for decades.

s107
00:08:09.080 --> 00:08:17.440
From that research, he identified specific communication patterns that predict divorce with over 90% accuracy.

s108
00:08:17.440 --> 00:08:20.760
Not relationship dissatisfaction or unhappiness, divorce.

s109
00:08:20.760 --> 00:08:29.320
Predicted it with more accuracy than most medical tests we we trust with our lives from a 15-minute conversation.

s110
00:08:29.320 --> 00:08:38.440
Alongside Gottman's work, the other most rigorously validated framework in couples research is emotionally focused therapy, developed by Dr. Sue Johnson.

s111
00:08:38.440 --> 00:08:40.880
EFT is grounded in attachment theory.

s112
00:08:40.880 --> 00:08:47.480
Its core insight is that beneath almost every surface-level argument in a relationship, there is something else.

s113
00:08:47.480 --> 00:08:51.560
When couples fight about the dishes, it's never about the dishes.

s114
00:08:51.560 --> 00:09:00.520
EFT helps partners access the underlying emotional layers and respond to each other there, rather than continuing to argue about the surface content.

s115
00:09:00.520 --> 00:09:02.320
It's extraordinarily effective.

s116
00:09:02.320 --> 00:09:05.320
The outcome research is among the strongest in psychotherapy.

s117
00:09:05.320 --> 00:09:10.080
It takes years of training to effectively use Gottman or EFT.

s118
00:09:10.080 --> 00:09:18.000
They are train They are the standard care for couples intervention, and they are almost entirely absent from the commercial AI relationship space.

s119
00:09:18.000 --> 00:09:24.680
What most relationship apps are built on is something that might be called general empathy plus communication tips.

s120
00:09:24.680 --> 00:09:26.040
And that's fine.

s121
00:09:26.040 --> 00:09:32.320
But when the stakes are this high, when the technology is reaching this many people, we need better.

s122
00:09:32.320 --> 00:09:34.440
The next thing we have to talk about safety.

s123
00:09:34.440 --> 00:09:38.640
Right now, someone is in a relationship where things are starting to turn.

s124
00:09:38.640 --> 00:09:50.160
Control, isolation, escalation, and instead of a hotline or a trained advocate, they open an AI assistant and ask for advice on how to make things better at home.

s125
00:09:50.160 --> 00:10:01.240
A general-purpose AI doesn't know what a domestic violence specialist catches in the first 90 seconds, the difference between we fight a lot and I'm afraid of what happens when I disagree with him.

s126
00:10:01.240 --> 00:10:08.880
It doesn't know that I just have to stay calm is sometimes the exact sentence someone is taught to say right before something terrible happens.

s127
00:10:08.880 --> 00:10:14.000
It doesn't know when relationship advice has stopped being the right category of help.

s128
00:10:14.000 --> 00:10:16.040
And it isn't only domestic violence.

s129
00:10:16.040 --> 00:10:24.200
Relationship coaching sits closer to suicide risk and yes, homicide risk than almost any other corner of mental health.

s130
00:10:24.200 --> 00:10:31.200
I've sat across from people in real dangers more times than I care to think about, and every licensed clinician is trained for that moment.

s131
00:10:31.200 --> 00:10:35.120
Relationship AI products are not built for that moment.

s132
00:10:35.120 --> 00:10:38.920
So, we talked about personal safety, let's talk about data safety.

s133
00:10:38.920 --> 00:10:44.280
Think about the what you actually tell an AI relationship coach, the very first conversation.

s134
00:10:44.280 --> 00:10:45.880
The fight from an hour ago.

s135
00:10:45.880 --> 00:10:49.240
The thing your partner did that you've never told anyone.

s136
00:10:49.240 --> 00:10:55.360
Financial secrets, affairs, mental health history, sometimes your kids' names.

s137
00:10:55.360 --> 00:11:06.480
That is without exaggeration some of the most sensitive data a human being generates, and a lot of it sits in training pipelines, server logs, and product analytics built by teams optimizing for engagement,

s138
00:11:06.480 --> 00:11:08.839
not confidentiality.

s139
00:11:08.960 --> 00:11:12.040
In my office, what you tell me is privileged.

s140
00:11:12.040 --> 00:11:18.720
There's a legal and ethical framework built over decades governing exactly what I can and can't do with it.

s141
00:11:18.720 --> 00:11:20.839
Consumer AI products have no equivalent.

s142
00:11:20.839 --> 00:11:28.280
Your most vulnerable disclosures about your relationship may be sitting in the same data infrastructure as your search history and your shopping cart.

s143
00:11:28.280 --> 00:11:30.440
And that should bother you.

s144
00:11:30.440 --> 00:11:32.800
Now, all of this is not a critique of technology.

s145
00:11:32.800 --> 00:11:40.000
This is a critique of insufficient domain expertise being applied to one of the highest stakes domains in human life.

s146
00:11:40.000 --> 00:11:41.680
So, this is what we built.

s147
00:11:41.680 --> 00:11:45.200
I want to tell you what we did differently at Couple Work and what we're learning.

s148
00:11:45.200 --> 00:11:52.880
Not as an advertisement, as a design case study and what it looks like to take to take the clinical standards seriously from the ground up.

s149
00:11:52.880 --> 00:11:58.960
We built Maxine, our AI coach, named for what it's worth after my grandmother.

s150
00:11:58.960 --> 00:12:01.480
She was a piece of work.

s151
00:12:01.600 --> 00:12:05.440
Maxine operates with the Gottman Method and EFT frameworks.

s152
00:12:05.440 --> 00:12:09.560
That means her response logic is grounded in 40 years of couples research.

s153
00:12:09.560 --> 00:12:11.360
She does not simply validate.

s154
00:12:11.360 --> 00:12:12.640
She reflects patterns back.

s155
00:12:12.640 --> 00:12:14.760
She names dynamics, not just feelings.

s156
00:12:14.760 --> 00:12:17.800
She asks the question underneath the question.

s157
00:12:17.800 --> 00:12:20.800
We built Maxine to know what she doesn't know.

s158
00:12:20.800 --> 00:12:29.120
Before she ever responds to a sensitive message, she's screening it quietly in the background for the patterns clinicians are trained to recognize.

s159
00:12:29.120 --> 00:12:35.839
Escalating control, fear-based language, anything pointing toward risk to the user or someone else.

s160
00:12:35.839 --> 00:12:40.440
When those patterns show up, Maxine stops coaching and starts following the protocols.

s161
00:12:40.440 --> 00:12:47.920
And we built those protocols out of 30 years of clinical practice, not a generic safety policy written by someone who's never sat in the room.

s162
00:12:47.920 --> 00:12:52.640
Knowing when not to coach is just as important as to know how.

s163
00:12:52.640 --> 00:13:02.400
A relationship AI that can't tell the difference between we're struggling and I'm not safe isn't incomplete, it's dangerous at scale.

s164
00:13:02.400 --> 00:13:09.400
When it comes to personal data, we treat what you tell Maxine the way I treat what you tell me in my office, as privileged.

s165
00:13:09.400 --> 00:13:11.600
That's the bar we hold ourselves to.

s166
00:13:11.600 --> 00:13:22.080
Not is this legally defensible, the bar is would I be comfortable if this were my own marriage on the line because a lot of our users it is.

s167
00:13:22.080 --> 00:13:31.040
And she's available at 11:00 on a Tuesday night when the argument is live, When you don't have to wait for next week's therapy appointment.

s168
00:13:31.040 --> 00:13:42.520
Immediacy is important because a good intervention 6 days later is fine, but an informed intervention 6 seconds before you send the text that makes everything worse can be life-changing.

s169
00:13:42.520 --> 00:13:48.600
The research shows that a majority of couples wait 6 years before seeking professional help.

s170
00:13:48.600 --> 00:13:51.760
That wait They wait until contempt is baked in.

s171
00:13:51.760 --> 00:13:54.760
They wait until after the affair.

s172
00:13:54.760 --> 00:13:59.840
They wait until the relationship is basically a crime scene and then say maybe we should talk to somebody.

s173
00:13:59.840 --> 00:14:02.120
That's when they come into my office.

s174
00:14:02.120 --> 00:14:08.320
So, if AI can help people engender engage earlier before everything hardens, that's good.

s175
00:14:08.320 --> 00:14:09.720
It's really good.

s176
00:14:09.720 --> 00:14:12.800
And I want the AI engineers in this room to hear that.

s177
00:14:12.800 --> 00:14:14.960
This is not a consumer app opportunity.

s178
00:14:14.960 --> 00:14:19.480
This is a public health opportunity hiding inside consumer behavior.

s179
00:14:19.480 --> 00:14:21.600
This can be done well.

s180
00:14:21.600 --> 00:14:24.120
It must be done well.

s181
00:14:24.120 --> 00:14:30.480
And to round things up, I'm going to ask my co-founder Tony to come up and speak a little engineer language.

s182
00:14:30.480 --> 00:14:36.120
Just very briefly I've got my mic on.

s183
00:14:36.480 --> 00:14:37.400
Hello.

s184
00:14:37.400 --> 00:14:38.120
Awesome, thank you.

s185
00:14:38.120 --> 00:14:42.000
So, just very briefly, um how do you do something like this?

s186
00:14:42.000 --> 00:14:43.280
Some quick notes.

s187
00:14:43.280 --> 00:14:47.080
Start with your clinician, not with your prompt.

s188
00:14:47.080 --> 00:14:52.200
Sit with your clinician uh encode what good looks like in evals.

s189
00:14:52.200 --> 00:14:54.520
Write hundreds of evals TDD style.

s190
00:14:54.520 --> 00:15:00.200
That becomes your your number one tool.

s191
00:15:00.200 --> 00:15:02.120
Watch those evals fail.

s192
00:15:02.120 --> 00:15:06.720
Run your agent through them thousands of times, tens of thousands of times.

s193
00:15:06.720 --> 00:15:09.120
Watch for those outliers.

s194
00:15:09.120 --> 00:15:14.240
When safety's on the line, even one failing test is not okay.

s195
00:15:14.240 --> 00:15:23.560
Then no matter how many evals you run and write nothing replaces your gut when emotional context is on the line.

s196
00:15:23.560 --> 00:15:25.840
So, eat your own dog food.

s197
00:15:25.840 --> 00:15:28.400
Talk to your agent.

s198
00:15:29.400 --> 00:15:33.040
Feel when the tone starts shifting off.

s199
00:15:33.040 --> 00:15:41.120
So, I've been using Couplework as I've been developing it with my partner since the beginning of my relationship.

s200
00:15:41.120 --> 00:15:45.400
And Maxine really challenged me about how I showed up.

s201
00:15:45.400 --> 00:15:50.600
Um we worked through some serious situations early on.

s202
00:15:50.600 --> 00:15:58.760
And as our relationship started to thrive, we continued talking to Maxine about the next milestones, like moving in together.

s203
00:15:58.760 --> 00:16:02.280
Um and recently, I asked my partner to marry me.

s204
00:16:02.280 --> 00:16:04.080
She said yes.

s205
00:16:04.080 --> 00:16:06.280
So, um so the Thank you.

s206
00:16:06.280 --> 00:16:06.480
Thank you.

s207
00:16:06.480 --> 00:16:07.120
Thank you.

s208
00:16:07.120 --> 00:16:12.960
So, the the biggest the biggest takeaway uh from all of this I hope that you get is this.

s209
00:16:12.960 --> 00:16:21.640
Partner with a clinician who challenges you to aspire to create AI that meets the clinical standard.

s210
00:16:21.640 --> 00:16:25.880
So, that you can create AI that challenges your users.

s211
00:16:25.880 --> 00:16:28.040
Thank you.