Tinder’s AI Flirting Coach, Therabot’s Therapy Success, and OpenAI’s New Copyright Clash

Summary

AI Deep Dive Podcast: Episode on Tinder’s AI Flirting Coach, Therabot’s Therapy Success, and OpenAI’s New Copyright Clash

Release Date: April 2, 2025

Welcome to this episode of AI Deep Dive, hosted by Daily Deep Dives, where the latest advancements, trends, and controversies in artificial intelligence are thoroughly explored. In this episode, the hosts delve into three significant topics: Tinder’s new AI-powered flirting coach, the groundbreaking success of Therabot in therapy, and the emerging copyright issues surrounding OpenAI’s models.

---

1. Tinder’s AI Flirting Coach

Overview: Tinder has introduced an innovative AI-powered game designed to help users practice their flirting skills. This feature, named the Game Game, leverages technology similar to OpenAI's ChatGPT to simulate various dating scenarios through interactive AI characters.

Key Features:

• Interactive Scenarios: Users engage with AI personas in different dating contexts, effectively creating a virtual dating simulation.
• Real-Time Feedback: Unlike traditional methods that offer pre-written responses, users speak their responses, and the AI provides a score based on a three-point flame emoji system ([00:35] A).
• Personalized Coaching: If a user's flirting attempts are ineffective, the AI suggests improvements, functioning as a virtual flirting coach ([01:53] A).

Strategic Implications: This initiative appears to be part of Tinder's broader strategy to enhance user engagement amidst a saturated online dating market. By integrating AI features such as the AI photo selector, Tinder aims to retain and attract users by offering novel and interactive tools ([02:30] B).

Industry Trend: Tinder isn't alone in this space; other applications like Replika's Blush and Rizz are already offering similar AI-driven flirting practice, indicating a growing trend in leveraging AI for personal development in social interactions ([02:10] B).

Quotes:

• "The AI judges you and gives you a score based on flame emojis." — Host A ([01:35] B)
• "It's almost like a flirting coach." — Host A ([01:51] B)

---

2. OpenAI’s New Copyright Clash

Overview: A recent paper from the AI Disclosures Project, co-founded by Tim O'Reilly, alleges that OpenAI's latest models, including GPT4O, may have been trained on paywalled books from O'Reilly Media without proper licensing agreements.

Research Methodology: The researchers employed a technique called decop dcop, a membership inference attack, to determine if the AI could distinguish between original texts and AI-generated paraphrases. They analyzed nearly 14,000 paragraph excerpts from 34 O'Reilly books published around the model's training cutoff dates ([03:22] B).

Findings:

• Increased Recognition: GPT4O showed significantly higher recognition of O'Reilly content compared to older models like GPT-3.5 Turbo, suggesting potential unauthorized use of copyrighted material ([05:05] B).
• Non-Definitive Evidence: While the findings are suggestive, the researchers acknowledge that alternative explanations, such as user-submitted content, cannot be entirely ruled out ([06:01] A).

OpenAI’s Response: OpenAI has advocated for a more flexible approach to using copyrighted data for training while simultaneously working on securing licensing agreements and providing opt-out options for content creators. They emphasize the importance of high-quality training data and have faced ongoing lawsuits regarding their data practices ([07:11] B).

Implications: This controversy raises critical ethical and legal questions about the use of copyrighted material in training AI models, highlighting the need for transparent data sourcing and fair compensation for content creators ([07:37] A).

Quotes:

• "It raises fundamental questions about the data that powers AI." — Host A ([04:03] B)
• "Our method isn't foolproof, but the difference in recognition between newer and older models is significant." — Host B ([06:39] A)

---

3. Competition in the AI Chatbot Market

Overview: The AI chatbot landscape is becoming increasingly competitive, with prominent players like ChatGPT, Gemini, Microsoft's Copilot, Claude, and Elon Musk’s Grok vying for market dominance.

Market Dynamics:

• Growth Metrics: Data from SimilarWeb and Sensor Tower indicates substantial growth for Gemini’s web app and Microsoft’s Copilot. Notably, Elon Musk’s Grok experienced an astonishing 800% month-over-month increase in web traffic ([08:08] A).
• Feature Enhancements: Continuous innovation, such as Gemini’s Canvas feature for previewing coding projects and Claude’s new sonnet model, are driving user adoption and engagement ([09:06] A).

Competitive Edge: While ChatGPT remains the market leader with substantially higher weekly active mobile users, the surge in activity for competitors like Grok suggests a tightening race for the second spot ([09:37] A).

Quotes:

• "Grok had a massive surge in web traffic, almost 800% growth month over month." — Host A ([08:43] A)
• "New features and models are really key here." — Host B ([09:15] A)

---

4. Therabot’s Therapy Success

Overview: Researchers at Dartmouth have conducted the first clinical trial of a generative AI-powered therapy chatbot named Therabot. The trial demonstrated significant improvements in participants' mental health symptoms.

Trial Details:

• Participants: Over 100 individuals diagnosed with major depressive disorder, generalized anxiety disorder, eating disorders, among other conditions.
• Method: Participants interacted with Therabot via a smartphone app, engaging in open-ended text dialogues grounded in evidence-based psychotherapy and cognitive behavioral therapy (CBT) principles ([10:20] B).
• Results:
  • Depression Symptoms: 51% reduction on average.
  • Anxiety Symptoms: 31% reduction.
  • Eating Disorder Concerns: 19% reduction, surpassing the control group ([10:51] A).
• Sustainability: These improvements remained statistically significant even after an eight-week follow-up period ([13:05] A).

Therapeutic Features:

• Personalization: Therabot adapts responses based on individual conversations, fostering a sense of therapeutic alliance akin to a human therapist ([13:42] B).
• Safety Measures: The chatbot detects high-risk content, such as suicidal ideation, and prompts users to seek emergency assistance ([12:07] A).
• User Experience: Participants reported feeling comfortable and unjudged, often initiating conversations during moments of distress ([14:19] A).

Implications: Therabot presents a promising AI-driven solution to the significant shortage of mental health providers in the US, offering accessible and effective support comparable to traditional outpatient therapy ([13:28] B).

Quotes:

• "Participants actually developed a therapeutic alliance with Therabot." — Host B ([13:42] A)
• "It's based on established therapeutic approaches." — Host B ([11:44] A)

---

Conclusion

In this episode, AI Deep Dive highlights the multifaceted impact of artificial intelligence across various sectors. From enhancing personal interactions on platforms like Tinder to addressing critical issues surrounding data ethics in AI development, and from intensifying competition among AI chatbots to revolutionizing mental health support with Therabot, AI continues to shape and redefine our world. The hosts emphasize the importance of balancing AI's potential benefits with ethical considerations and the necessity for human oversight, especially as AI systems become more integrated into sensitive areas of human life.

Final Thoughts:

• "How do we balance the potential benefits of accessibility with the need for human connection and oversight?" — Host B ([16:04] B)
• "One that deserves careful consideration." — Host B ([16:16] B)

Stay tuned for the next episode of AI Deep Dive, where more fascinating developments in the world of artificial intelligence will be explored.

Transcript

[A]

Foreign. Welcome back, everyone. We've got quite the collection of AI news to delve into today.

[B]

Yeah, for sure.

[A]

We've got articles covering a bunch of different areas like AI flirting practice, questions about AI and copyright.

[B]

Right.

[A]

The increasingly competitive chatbot market, and even AI therapy bots.

[B]

Yeah. A lot of really interesting stuff happening right now.

[A]

So it'll be fun to kind of get a quick but thorough look at what's going on in AI these days.

[B]

Absolutely. I think our deep dive today will help extract some core insights from these articles and help everyone understand, you know, what's important and what it all means.

[A]

Okay, so kicking things off, we got this pretty unique story about Tinder.

[B]

Okay.

[A]

They've rolled out this AI powered game to help you practice your flirting skills.

[B]

Interesting.

[A]

They've partnered with OpenAI.

[B]

Oh, the folks behind ChatGPT.

[A]

Yeah, exactly. So it seems like it'll be using similar tech.

[B]

Okay.

[A]

And you can basically interact with AI characters in different dating scenarios.

[B]

So, like a simulation almost.

[A]

Yeah, kind of like a virtual dating sim.

[B]

I see.

[A]

It's called the Game Game.

[B]

Catchy.

[A]

You just tap the Tinder logo.

[B]

Okay.

[A]

And it gives you these cards with different AI Personas and situations.

[B]

Interesting. So you're presented with like a scenario.

[A]

Yeah. And then the really cool part is that you actually speak your responses.

[B]

Oh, wow. So it's not just like selecting from pre written options.

[A]

Nope. You talk to it.

[B]

That's neat.

[A]

And try to see how well you can charm the AI.

[B]

Oh, so it judges you.

[A]

Yeah, it gives you a score based on like flame emojis. A three point scale.

[B]

Oh, that's fun.

[A]

And the best part is it gives you feedback.

[B]

Really?

[A]

So if your flirting isn't working.

[B]

Mm.

[A]

It'll suggest ways to improve your conversation.

[B]

That's cool. Like almost like a flirting coach.

[A]

Exactly.

[B]

Interesting.

[A]

But it seems like Tinder is framing this as a fun limited time thing.

[B]

Okay.

[A]

Just for us iOS users.

[B]

For now, it makes sense to kind of test the waters.

[A]

Right. But the thing is, this whole idea of practicing flirting with AI bots is actually becoming kind of a trend.

[B]

Oh, really? I haven't heard about this.

[A]

Yeah, there are already apps out there specifically for this, like Replika's Blush.

[B]

Oh, interesting teaser.

[A]

And Rizz.

[B]

Some rizz. Okay.

[A]

All about honing those virtual flirting skills.

[B]

So Tinder's kind of jumping on an existing bandwagon in a way.

[A]

Yeah. And it makes you wonder if it's part of a bigger strategy shift for them.

[B]

How so?

[A]

Well, the articles we've got here mention that Tinder's been facing some growth challenges.

[B]

Hmm, I can see that. The online dating market is pretty saturated.

[A]

Right. And this AI game could be a way to re engage users.

[B]

Yeah, like a novelty to draw people back in.

[A]

Exactly. And they've already been integrating AI in other ways too.

[B]

Oh, like what?

[A]

Well, they have that AI photo selector.

[B]

Right, I've heard of that.

[A]

Which helps you choose your best picks. And they've hinted at even more AI features for things like discovery and matching coming soon.

[B]

Sounds like AI is becoming pretty central to their platform.

[A]

Yeah, for sure.

[B]

Okay, so we've got AI being used in these consumer facing apps.

[A]

Right.

[B]

But. But there are some bigger questions around the data that fuels these huge AI models. And there's actually a new accusation against OpenAI about how they're training their models.

[A]

Oh, wow. What's this about?

[B]

So there's a paper from the AI Disclosures Project, co founded by Tim O'Reilly, and it suggests that OpenAI's newer models like GPT4O.

[A]

Right, the one that's the default in ChatGPT now.

[B]

Yeah, exactly, that one. It seems like it's been trained on paywalled books.

[A]

Oh. Oh, wow, really?

[B]

Yeah, from O'Reilly Media specifically.

[A]

And is there any indication they had permission to do that?

[B]

Well, that's the issue. The paper suggests that there wasn't any licensing agreement.

[A]

Wow. So they might have used copyrighted material without consent.

[B]

Yeah, that's a pretty serious claim.

[A]

Yeah, it is. This is a big deal because it raises these fundamental questions about the data that powers AI.

[B]

Right. Where does it all come from?

[A]

Right. And is it ethical to leverage creative work for technological advancement without proper compensation? Exactly. You know, this is crucial for anyone following AI.

[B]

Yeah, absolutely. This could have huge implications.

[A]

So how did the researchers actually look into this?

[B]

So they used a really interesting method called decop dcop.

[A]

Okay.

[B]

They describe it as a membership inference attack.

[A]

I'm not familiar with that.

[B]

Imagine trying to recognize a song you've heard a lot.

[A]

Okay.

[B]

Versus like a remix of that song made by AI.

[A]

Okay, yeah, I get that.

[B]

If you can easily tell the difference, it suggests you've probably heard the original before.

[A]

Hmm, I see.

[B]

So in this case, the AI might have been trained on the original text.

[A]

So decop basically tests if the AI can distinguish between the original text and AI generated paraphrases.

[B]

Yeah, exactly.

[A]

If it can do that reliably, the theory is that it has prior knowledge of the original text from its training data. Okay, that's a pretty clever way to approach it.

[B]

Yeah. So they basically took nearly 14,000 paragraph excerpts from 34 O'Reilly books.

[A]

Wow, that's a lot.

[B]

And these books were published both before and after the model's supposed training cutoff dates.

[A]

Right. So the point after which the AI shouldn't have been exposed to new data.

[B]

Exactly. And what they found is that GTC4. Okay. Recognized way more of this paywalled O'Reilly.

[A]

Content then older models.

[B]

And older models like GPT 3.5 Turbo.

[A]

Hmm, that is interesting.

[B]

It definitely suggests something's going on.

[A]

Right, but the article does mention this isn't, like, definitive proof.

[B]

Yeah, the researchers themselves acknowledge that their method isn't foolproof.

[A]

Right.

[B]

And it's possible OpenAI could have come across this content in other ways.

[A]

Oh, like how?

[B]

Like if users were copying and pasting excerpts into ChatGPT.

[A]

Oh, I see.

[B]

And the study didn't look at OpenAI's very latest models.

[A]

Right. So it's suggestive but not conclusive. Exactly, but still pretty thought provoking.

[B]

Oh, for sure.

[A]

It makes you wonder what would be the smoking gun evidence?

[B]

Yeah, what would definitively prove unauthorized training uncopyrighted material? Well, I think direct access to OpenAI's training data logs would be the most definitive evidence. But that's not publicly available, right?

[A]

Of course not.

[B]

So decop gives us strong indicators, but as we said, other explanations can't be totally ruled out. But the strength of these findings is in the big difference in recognition between newer and older models.

[A]

That does seem significant.

[B]

Yeah.

[A]

So what does OpenAI have to say about all this?

[B]

Well, the article mentions that they've argued for a more flexible approach to using copyrighted data for training, but they've also been working on getting licensing agreements with content creators and offering ways for copyright holders to opt out. Oh, it's good of having their stuff used. They've even hired journalists, which shows they really value high quality training data.

[A]

Yeah, they need good stuff to feed the models.

[B]

Exactly. And of course, OpenAI is already dealing with lawsuits about their data practices.

[A]

Yeah, I've been hearing about those.

[B]

So this paper from the AI Disclosures project definitely adds another layer to that whole discussion.

[A]

Yeah, for sure. This issue isn't going away anytime soon.

[B]

No, it really highlights the complex issues around using copyrighted material for AI training.

[A]

It's a tricky balance.

[B]

Yeah, as these models get more advanced, they need more and more data, and that inevitably leads to these tensions. Yeah, about intellectual property rights.

[A]

Okay, so moving on, let's talk about the competition in the AI chatbot world.

[B]

Always an interesting topic.

[A]

So ChatGPT is still the Big name, everyone knows.

[B]

Yeah, for sure. It's the one that really kicked things off.

[A]

Right. But it seems like other chatbots are starting to catch up.

[B]

Really?

[A]

Yeah. We've got Some data from SimilarWeb and Sensor Tower that show this.

[B]

Okay, let's see what the numbers tell us.

[A]

All right, so SimilarWeb's data for March shows that Gemini's web app had a pretty big increase in daily visits.

[B]

Hmm, interesting.

[A]

And Microsoft's Copilot saw growth too.

[B]

Makes sense. Microsoft's been pushing that pretty hard.

[A]

Yeah, and Claude's still holding its own with a solid user base.

[B]

Okay, but who are the real movers and shakers?

[A]

Well, Deepseek had high numbers, but actually saw a dip from February.

[B]

Interesting.

[A]

But the real surprise is Grok.

[B]

Grok?

[A]

Yeah, Elon Musk's AI.

[B]

Oh, right, right, right.

[A]

It had a massive surge in web traffic.

[B]

How big are we talking?

[A]

Almost 800% growth month over month.

[B]

Wow, that's a huge jump.

[A]

Yeah, it seems like the competition for that second place spot right behind ChatGPT, is really heating up.

[B]

Yeah, definitely looks like a close race.

[A]

And the Sensor Tower data for mobile apps shows a similar trend.

[B]

Okay.

[A]

Claude saw a nice increase in weekly active users when they released their Claude 3.7 sonnet model.

[B]

That makes sense. New models often drive adoption.

[A]

Right. And Gemini also had good growth after their Gemini 2 point flash release.

[B]

So new features and models are really key here.

[A]

Yeah, for sure. Things like Gemini's Canvas feature.

[B]

Right. For previewing coding projects.

[A]

Yeah, and the new tools Anthropic's been adding to Claude are clearly attracting users.

[B]

It's a smart move to keep innovating.

[A]

Absolutely. But, you know, ChatGPT still seems to be the king of the hill. Yeah, I'm not surprised that like 10 times.

[B]

Wow.

[A]

The combined weekly active mobile users of Gemini and Claude as of March. Yeah. So OpenAI probably isn't panicking yet.

[B]

No, I doubt it. But it's good to see some healthy competition.

[A]

Definitely keeps things interesting.

[B]

So, for our last topic today, we've got something truly groundbreaking.

[A]

Okay, what is it?

[B]

Researchers at Dartmouth have done the first clinical trial.

[A]

Wow.

[B]

Of a generative AI powered therapy chatbot.

[A]

Wow. A therapy chatbot.

[B]

Yeah, it's called therabot.

[A]

Therabot. Okay.

[B]

And the early results show some really significant improvements in participants mental health symptoms.

[A]

So this is like AI for mental health support?

[B]

Exactly.

[A]

Okay, tell me more about this trial.

[B]

So they had over 100 participants diagnosed with things like major depressive disorder, generalized.

[A]

Anxiety disorder, and eating disorders, a range of conditions. Yeah. And? And they interacted with Therabot through a smartphone app.

[B]

Okay.

[A]

Responding to prompts, and they could initiate conversations whenever they felt like it.

[B]

Oh, so it wasn't just like a fixed set of interactions?

[A]

No, they had some agency.

[B]

That's good.

[A]

And the findings are pretty remarkable.

[B]

Okay, what did they find?

[A]

They saw an average of 51% reduction.

[B]

Wow.

[A]

In depression symptoms.

[B]

That's huge.

[A]

31% reduction in anxiety symptoms. And for those at risk of eating disorders, a 19% reduction in concerns about body image and weight.

[B]

Those are some pretty impressive numbers.

[A]

Yeah. And that last one was actually better than the control group.

[B]

Yeah. Interesting. So it seems like Therabot is actually making a difference.

[A]

It really does. And what's important to note is that Therabot's been in development since 2019.

[B]

Oh, wow. So it's not just some hastily thrown together thing.

[A]

No, not at all. It's been carefully developed.

[B]

Okay.

[A]

With ongoing input from mental health professionals at Dartmouth.

[B]

That makes me feel a lot better about it.

[A]

Yeah. And it uses this. This natural open ended text dialogue.

[B]

Okay.

[A]

Grounded in evidence based psychotherapy and cognitive behavioral therapy principles.

[B]

So it's not just randomly chatting.

[A]

No. It's based on established therapeutic approaches. There's an example in the article where someone tells Therabot they're feeling anxious.

[B]

Okay.

[A]

And Therabot says, let's take a step back and ask why you feel that way.

[B]

That's a very CBT type of response.

[A]

Exactly. It's encouraging reflection.

[B]

Yeah. And that's a good thing.

[A]

And they've built in some safety features too.

[B]

Oh, like what?

[A]

So if it detects high risk content.

[B]

Okay.

[A]

Like suicidal ideation.

[B]

Right.

[A]

It prompts the user to contact emergency services or hotlines.

[B]

That's really important to have those safeguards in place.

[A]

Yeah, absolutely. So in the trial, the treatment group had four weeks of unlimited access to Therabot.

[B]

Okay.

[A]

And what's really interesting is that almost.

[B]

75% of them weren't in any other kind of therapy.

[A]

Oh, wow. So for many, this was their only mental health support.

[B]

Exactly. And the AI personalized its responses based on the conversation.

[A]

So it's adapting to each individual.

[B]

Yeah, in a way. And the researchers were keeping a close eye on things to make sure it was sticking to things. Therapeutic best practices.

[A]

That's reassuring.

[B]

Yeah. And even after eight weeks, including a follow up period, all the Therabot users still showed a statistically significant reduction in their symptoms.

[A]

So the effects lasted, it seems. So that's really promising.

[B]

It is. And the researchers believe that this kind of AI Powered therapy could be a really important source of support.

[A]

Okay.

[B]

For people who can't access traditional mental health care.

[A]

Yeah. I could see that being huge.

[B]

And they're actually seeing symptom improvements that are comparable to what you'd expect in outpatient therapy.

[A]

Wow, that's amazing.

[B]

The article even points out how in the US There's a big shortage of mental health providers.

[A]

Yeah, that's a huge problem.

[B]

So solutions like this could be really valuable.

[A]

Oh, sure.

[B]

And there's another interesting finding that's more qualitative.

[A]

Right.

[B]

It seems like participants actually developed therapeutic alliance with Therabot. Therapeutic alliance, meaning they felt a sense of trust and collaboration like you'd have with a human therapist.

[A]

Wow. So they actually felt a connection with the AI?

[B]

It seems that way. And people were initiating conversations, often even in the middle of the night when they were struggling.

[A]

That's pretty powerful.

[B]

It is. One quote even mentioned that participants almost treated the software like a friend.

[A]

Hmm.

[B]

And felt comfortable sharing because they didn't feel judged.

[A]

Wow, that's really interesting.

[B]

Yeah. It speaks to the potential here.

[A]

But of course, with something this new, there are always concerns.

[B]

Of course, the researchers emphasized that we need to be really careful with generative AI in mental health.

[A]

Yeah. It's a sensitive area.

[B]

It requires rigorous testing for both safety and effectiveness.

[A]

Absolutely. We can't just throw AI into therapy without careful consideration.

[B]

Exactly. And close clinical oversight is crucial, especially given the potential risks.

[A]

So were there any issues in the trial?

[B]

Well, the study team was prepared to intervene if they saw any safety concerns or inappropriate responses from Therabot, but thankfully, those were rare.

[A]

That's good to hear.

[B]

Yeah. And they pointed out that in earlier evaluations, over 90% of TheraBot's responses were in line with established therapeutic practices. That's encouraging, but it's still a rapidly evolving field, and we need to be diligent about oversight as things progress.

[A]

Okay. So to wrap up our deep dive today, we've seen AI popping up in all sorts of areas, from helping you practice flirting on Tinder.

[B]

Right.

[A]

To raising big questions about how these AI systems are trained and who owns.

[B]

The data they're trained on.

[A]

Exactly. And we've looked at the. The increasingly competitive chatbot landscape.

[B]

Yeah. Things are heating up.

[A]

And this. This really promising use of AI.

[B]

Right.

[A]

In mental health support with Therabot. Yeah. It's clear from these articles that AI is continuing to have a big impact.

[B]

A multifaceted impact, for sure, on all.

[A]

Sorts of aspects of our lives.

[B]

It makes you wonder how our relationship with AI will evolve in the coming years, you know, with AI becoming so capable, even in areas like emotional support.

[A]

Yeah.

[B]

How do we bal the potential benefits of accessibility with the need for human connection and oversight?

[A]

Yeah. That's a big question that we all need to be thinking about.

[B]

Absolutely. One that deserves careful consideration.

[A]

Thanks for joining us for this deep dive, everyone.

[B]

Yeah, thanks for listening.

[A]

Be sure to check back for our.

[B]

Next episode where we'll explore even more fascinating developments in the world of AI.

[A]

Until then, keep exploring and stay curious.

[B]

See you next time.

[A]

It.