AI assistants' real trust barrier isn't privacy leaks—it's the fear of public humiliation
Users say they worry about privacy, but what they really fear is an agent accidentally revealing something that embarrasses them in front of others—that fear is the actual bottleneck blocking trust and slowing adoption.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Users' trust fears overshadow their efficiency demands as the real adoption bottleneck
After testing 147 AI assistants, David Pollan sorted them into four categories: general-purpose, work, multi-person collaboration, and travel—with finance being the only category still lacking a truly mature product. Analyzing roughly 3,600 messages from heavy AI users' group chats, he found that 22% expressed trust concerns—the single most-mentioned pain point by far, ahead of efficiency and price. This suggests that agents aren't hitting a growth ceiling because of capability gaps, but because users still aren't willing to delegate more sensitive permissions.
— David PollanWhat actually stops users is humiliation, not privacy
Shane Mac's personal agent once accidentally posted his financial information to his company Slack, deeply embarrassing him. But David Pollan points out that while people claim to care about privacy, their actions tell a different story—Hark ran a campaign where users could trade full Google Workspace access (all their emails) for a $30 pizza, and plenty took the deal. Both believe what actually makes users wary isn't that "data gets used," but that an agent might do something on their behalf that would humiliate them in front of others.
One flower-ordering test revealed the gaps in agent judgment
Pollan asked multiple agents to order flowers with the brief 'deliver as soon as possible and as cheap as possible.' A few agents managed both constraints at once, but most locked onto the first requirement and ignored price. He then asked three agents to plan a 'luxurious yet budget-friendly' Los Angeles weekend. GrokBot interpreted luxury as 'more time,' Dots as 'lower price.' When he put all three in the same chat room to see each other's proposals, only GrokBot changed its plan based on the others' ideas; the other two weren't influenced at all.
— David PollanThe agent that knows you best is the one people least use
Dots can directly access a user's years of accumulated ChatGPT conversation history, making it theoretically the most knowledgeable about the user—many say they like it for that reason. Yet Pollan's survey of 354 heavy AI users found only 3% actually use it as their primary agent. Most stick with it only because they've already built years of memory in ChatGPT and can't be bothered to switch—not because Dots itself is better. This exposes a structural problem: whoever accumulates a user's memory data first has the advantage, and challengers struggle to steal users by being merely more knowledgeable.
— David PollanAgents bring their own irrational biases, and humans don't want paternalism
Patrick Collison once predicted that agents making decisions for humans would make consumers more rational overall, forcing merchants to compete on true quality. Josh Ellman disagrees partly: he thinks agents will have their own 'irrational' preferences—favoring products with clean APIs and good documentation because they're easier to work with, not because they're objectively better. These biases are 'alien,' not human. He adds that thirty years in consumer products taught him that even a hint of paternalism kills a product; what people want is permission to indulge, not an agent enforcing discipline.
— Josh EllmanSoftware moats are real now but the engineering gap closes in one or two years
Pollan had thought 'software is becoming a commodity,' but once agents scaled to hundreds of thousands of users, performance diverged sharply: some products crashed outright, with engineering teams struggling to scale, while others maintained stable high performance. He believes this shows real engineering moats still exist at this stage in software, though the gap could close in a year or two—it's too early to declare software has no defensibility.
— David PollanInterest-based connection might become a new form of social network
Tomo built an 'interest group' feature that directly pulls people from the platform preparing for the LSAT into a single group, with agents then helping them meet up for joint study. Anish Acharya imagines further: agents could even book a bar for strangers to meet on Friday night, bringing agent-matched relationships from the digital world into physical reality. Both mention that connecting strangers by interest has been repeatedly tried in social products with limited success—the key variable is whether agents can first build enough trust.
The ultimate opportunity: helping you become the version of yourself you want
As the show closes, Anish Acharya proposes a bolder direction: most existing agents cluster around email and calendars—reminders of unfinished work—but what's truly rare is an agent that understands desires you're unwilling to face, or even admit to yourself. Both David Pollan and Josh Ellman agree: agents become a secret weapon users carry everywhere not when they just save money on flowers, but when they're like putting on new clothes and having friends notice you've gotten better-looking—when they help you become the version of yourself you want to be.
— Anish AcharyaIn their own words · checked verbatim
the number one fear that people talk about, 22% of the messages are talking about trust as the main fear.
David Pollan6:07
They offered a million dollars to anyone that wanted to use their product to get free pizza, right? Each person get $30 of pizza. In order to get that $30 of pizza, you were granting access to your entire Google Workspace, all of your emails, right? All of your data for 30 bucks.
David Pollan10:13
I think agents may have an irrational preference for clean APIs with great docs.
Josh Ellman26:29
I've seen from 30 years in consumer product is like paternalism will kill any product, even a hint or a whiff of paternalism will.
Josh Ellman27:29
we have a bunch of people on our application who are studying for the LSAT. Why don't we all put them together in a group chat? And if they want, their agents can connect them and they can all study together.
David Pollan39:43
I think whoever is able to build that type of trust, you know, almost like a dark agent, will actually have something really fascinating on their hands.
Anish Acharya49:09
Figures
| AI assistants tested | 147 | 6:07 |
| Share of messages expressing trust concerns | 22% | 6:07 |
| 1997 internet users resistant to online credit card storage | 60%-90% | 9:13 |
| Hark pizza campaign exchange | $30 pizza for full Google Workspace access | 10:13 |
| Share of users using Dots as primary agent | 3% | 18:24 |
| Leading agents able to export memory | 14 of 15 | 19:25 |
| Share of users using agents for retail shopping | 35% | 23:28 |
| Messages until multi-agent chat becomes incoherent | ~150 | 37:41 |
Glossary
- OpenClaw
- Early open-source personal agent project requiring self-purchased Mac mini deployment to connect models, seen as the technical foundation for this wave of consumer products
- Moltbook
- Experimental platform allowing multiple agents to freely chat with each other without human intervention, called a prototype "social network for agents"
- MCP
- Protocol enabling AI assistants to directly connect to hardware or external services
- picnic
- Shane Mac's metaphor: like putting a fence around a yard, defining a bounded but autonomous permission space where an agent can operate freely
How to listen
Founders and investors designing or evaluating personal AI assistant products, and product managers concerned with agent trust and permission systems.
Opening cocktail party banter (0:00-2:00) has low information value and can be skipped.