How Sahara AI Helped Power Snap's Breakthrough in AI Chatbots That Feel Human

AI can write production code, pass the bar exam, and reason through graduate-level math. It still can't text like a friend for three weeks without starting to sound like a bot. For Snap, that gap is a real product problem. Snapchat runs one of the largest consumer chatbots on the planet, with more than 150 million people sending over 10 billion messages to My AI, in front of an audience that grew up texting and can often identify a fake-sounding bot right away.

Every major AI lab is chasing AI that people can have active, ongoing conversations with without it feeling like you're chatting with a bot. Personal agents that know you. Companions that remember what you told them last month. Assistants that respond like they actually understand you. Whoever gets there first owns the next era of consumer AI.

The issue is most chatbots learn conversation from other chatbots. Nearly all conversational AI trains on synthetic dialogue, where one model writes thousands of pretend chats and another model studies them. It’s fast and cheap, but the end product often sounds too templated and nothing like a real conversation.

Fixing this takes real, continuous conversations between real people, captured daily over weeks, and that data didn't exist anywhere. Every vendor Snap approached turned the project down. Sahara AI took it on. As a result, Snap now has REALTALK, the longest real-world human-to-human conversation framework ever collected for AI research, and the industry has its first direct measurement of how far today's chatbots are from real human conversation.

Why Snap Chose Sahara AI

Snap didn't need a labeling vendor. They needed a partner who understood what an AI has to learn to feel human, and that understanding comes from building AI, not annotating it.

Sahara AI has spent years building and deploying agentic AI in live production environments for some of the world's largest enterprises. We know firsthand how agents fail. They forget what users tell them. They flatten emotionally over long interactions. They lose their personality from one conversation to the next. Because we build the agents ourselves, we understood what this data needed to capture before collection started.

The Project Every Other Vendor Turned Down

Snap's researchers needed matched pairs of real people talking naturally, every single day, for three straight weeks, with enough structure behind it to measure what an AI can remember and how well it reads emotion.

That meant recruiting participants against tight criteria. Pairing them so conversations felt comfortable and real rather than forced. Keeping ten separate conversations alive daily for 21 days, each hitting targets for message volume, image sharing, and callbacks to earlier topics. Monitoring every conversation every day and correcting the moment a pair drifted off track.

Every data company that reviewed the project declined, and Snap's research stalled without it. In their words, "We were unable to move forward with research until we gathered the dataset that satisfied all our stringent criteria and was large enough in volume."

Sahara AI took the project on and delivered.

Capturing Real Relationships at Research Scale

Sahara AI recruited ten participants, all native English speakers aged 18 to 25, and paired each person into two separate conversations with two different partners. Because every speaker appears in two chats, the data shows how the same person shifts tone, closeness, and emotional register depending on who they're talking to. Researchers need that to study whether an AI can hold a consistent personality across different relationships.

Over 21 days, Sahara AI held every pair to a detailed playbook modeled on how friends actually text. Casual small talk. Personal stories, real and invented. References to time and place, like "last Friday" or "when I was ten." Images woven in naturally. Topics deliberately revisited across days so they deepened instead of resetting. Sahara AI monitored every conversation daily for the full three weeks to keep each one on standard.

The final corpus contains nearly 160,000 words of authentic human conversation across roughly 9,000 exchanges, more than 300 shared images, and 21 daily sessions per pair. Each conversation runs around 14x longer than the human-collected benchmarks the industry has relied on until now. The work also included the annotation layer that makes the data testable, with 728 memory questions that can only be answered by connecting details scattered across weeks of dialogue and 600 documented life events recording what was actually happening in each participant's life.

This all came together to become REALTALK, the longest continuous record of real human conversation ever captured for AI research, and the first benchmark that lets researchers test an AI directly against genuine human interaction, from how well it holds a personality to how much it remembers across weeks. 

The Future of more "Human" AI

For the first time, real human conversation could be measured directly against AI-generated conversation. The results explain exactly why chatbots feel fake.

The synthetic conversations opened at maximum closeness and stayed there, like two strangers pretending to be best friends from the first message. They remained relentlessly empathetic even when the topic was negative, and their emotional range stayed narrow across weeks. The human conversations built intimacy slowly, moved through a much wider band of emotion, and were full of the small things real people do to stay connected, like asking follow-up questions, reflecting, and checking in.

Memory was the other gap. Even when given the entire conversation history, today's top models reached only moderate accuracy on questions that required recalling and connecting things said earlier. Holding the thread of a relationship over weeks is still unsolved across the entire industry.

That is the frontier Sahara AI builds on. An agent that genuinely feels like yours has to remember what you told it and respond like it actually knows you, whether it's managing your workflows or, like Sorin, tracking your portfolio and adapting to how you invest. Snap is already putting the data to work. In their words, "Using real conversations, we are able to design better LLM dialogue agents and create better and more realistic synthetic conversational datasets." There's also active interest in extending the work, with longer conversations and continued human annotation.

This Is What Sahara AI Does

Snap is one of the biggest names in social media. When they needed to solve a problem nobody else could, they came to Sahara AI, the same way MIT did for OSGym and Microsoft Research did for MATHVISTA. Our data service capabilities excel because we build agentic AI, not the other way around. We know how agents work, how they break, and what they need to learn from, so we know what's actually worth collecting.

• Global Reach: 200,000+ pre-vetted contributors across 35+ countries, covering 45+ languages and dialects.

• Multi-Modality Coverage: Text, image, video, and audio collection and annotation.

• Programs Nobody Else Will Run: Long, high-touch data collection with daily monitoring, participant matching, and strict standards held over weeks.

• AI + Human Synergy: Combined AI-driven and human-in-the-loop workflows for speed without giving up quality.

Microsoft, Amazon, Snap, and MIT trust Sahara AI when accuracy, speed, and dependability are non-negotiable.

About Sahara AI: Sahara AI is the agentic AI company dedicated to making AI more accessible and equitable. We build the core protocols, infrastructure, and applications that let personal agents anticipate and execute on your behalf. For this to work, infrastructure has to be trustworthy: verifiable execution, enforceable usage policies, and automatic value distribution across every tool, model, and service an agent touches. Sahara is building a growing suite of agent-powered applications on top of this foundation, including Sorin, your personal agent for global digital markets. Our solutions currently power AI agents and high-quality data for consumers, Fortune 500 enterprises, and leading research labs, including Microsoft, Amazon, MIT, Motherson, and Snap.