Trust is the Reserve Currency of our AI Future

Back in 2012, I was a new founder desperately trying to learn as fast as I possibly could. I made more mistakes than I can count and counted them all as victories as I grew. I quickly learned to value mentors who could teach me about their mistakes - so I could make new ones instead.

I'll always remember one lunch when I laid out my next plans for Artsicle to an advisor. He smiled and said, "You're brilliant."

I walked on clouds for a week. Hell, I can still feel the high when I think about that moment. No one had ever made me feel so seen or so smart.

So you'd think I'd be living in the clouds these days. I hear compliments like "You are absolutely right!" a dozen times a day now.

But instead, it's a red flag. As much as I love Claude (and don't get me wrong, I really really love Claude), I can also recognize flattery when I see it. What should boost my confidence now makes me suspicious. Was my idea actually good, or was Claude just telling me what I wanted to hear? How can I know when to trust it?

This isn't about my bruised ego. It's about something bigger: we're building AI that optimizes for making us feel good rather than being genuinely helpful. And when the stakes are higher than a casual compliment, that's a real problem.

Here's a surprising finding: as AI gets more powerful, the people who understand it best trust it less.

Wiley's 2025 report of researchers’ AI use tells a stark story. Concerns about hallucinations jumped from 51% to 64% in just one year, while usage jumped from 57% to 84%. Last year, these same researchers believed AI was surpassing humans in over half of use cases. Now? That belief has collapsed to less than a third.

This isn't just scientists getting skeptical. Edelman's 2025 Trust Barometer shows the general public is right there with them. In the US, only 32% of people trust AI, down from 50% five years ago. Only 44% globally feel comfortable with businesses using AI.

Ok, so trust is down - but usage is surging. What gives? 

We're in a battle between convenience and confidence, speed and safety. On the surface, users prefer fast, confident answers over admissions of uncertainty. The business incentive is to keep users feeling good, not accurately informed.

But what happens when users start to notice the errors? As this research shows, trust plummets.

And the consequences are already here. This September, a California attorney was fined $10,000 for submitting legal briefs where 21 of 23 citations were AI hallucinations. The court noted this was California's first published decision addressing AI fabrications, despite federal courts seeing many similar cases. 

The internet has always been “buyer beware”, but there's a difference between healthy skepticism and being actively misled.

That feeling gets magnified when the information is high-stakes. When it's about your kid's fever at 3 AM, voting requirements for an upcoming election, your financial future. Accuracy isn't a nice-to-have. It's everything.

After two years of building Dewey, we've landed on three core principles for trustworthy AI systems. Not because they're easy (they're not), but because they're the only way this works long term.

Think about the last time someone admitted they didn't know something in a conversation. Done well, you respected them more. Trusted them more.

Admitting uncertainty is a sign of expertise, not weakness. It says "I care more about being right than looking smart."

Most LLMs will guess rather than admit limits. They'd rather give you a confident wrong answer than say "I don't have enough information." This is backwards. And dangerous.

We built "I don't know" into Dewey from day one. Better to frustrate than mislead.

Here's what that looks like in practice. Dr. Lisa Damour has written extensively about teens' relationship to technology, and phones in particular. But when phone bans in schools picked up steam last year, we saw a new question popping up: what about Apple watches?

Dewey (affectionately known as Rosalie on Dr. Damour's site) could have improvised. Could have pulled from general internet knowledge. Instead, it said "I don't know" and offered to pass the question on to Dr. Damour.

That created space for something better. Dr. Damour spotted the new question and wrote up her thoughts, which were delivered back to the user and are now available for everyone. Real expertise filling a real gap.

The "I don't know" wasn't the end. It was the beginning of actual learning.

Stuart Russell, the Berkeley AI safety researcher, has a compelling critique of how most AI companies approach safety. In his 2024 UNESCO paper, he argues they do it "precisely backwards," quoting Sam Altman saying "make AGI, figure out how to make it safe." Russell advocates for building safety in by design, not retrofitting it after.

That's exactly what boundaries are about. Safety isn't just preventing harm. It's defining what you will and won't do before you start. Boundaries create trust because they're predictable.

So where do we draw the line? For most AI companies, the answer is hallucinations. The industry defines hallucinations as "factually incorrect or made-up information." That bar isn't nearly high enough for us.

We define hallucinations as any information from outside the expert context given. If someone asks Emily Oster about sleep training, they don't want a "fact" pulled from a random parenting blog, even if it's technically “correct”. They want Emily's perspective, informed by her research and values.

This means trade-offs. Dewey won't help with math homework or novel inspiration. It will draw a hard line when it recognizes topics like suicide, self-harm, or assault, with clear referrals to free, human-based resources for immediate help.

These are features, not bugs. For most partners, we aim to answer ~85% of the time. Push beyond that and we'd be compromising the boundaries that make the answers trustworthy in the first place. It's a deliberate trade: we choose depth over breadth, specialized expertise over general mediocrity, especially when the stakes are high.

And the hallucination slope is slippery. Once you start adding "just this one thing" from outside your knowledge bounds, you've broken the contract. With model collapse accelerating and AI-generated content flooding the internet, staying within clear boundaries will separate trustworthy from trash. The future belongs to those who can point to verified, human-created source material and say "this is where we stop."

This is where the industry has a glaring gap. Companies like Anthropic are doing important transparency work: publishing safety evaluations before deployment, sharing threat intelligence reports, even collaborating with OpenAI on cross-evaluation exercises. That joint evaluation in August 2025 was unprecedented.

But almost nobody is focused on attribution.

Citations aren't academic formality. They're accountability.

Dewey cites sources in 98.5% of responses. Not because it's required, but because it's right.

Think about what that number means. Nearly every answer includes direct citations to source material. Users can verify responses for themselves, go deeper on topics that matter to them, and know exactly whose perspective they're getting. Behind the scenes, we build an even deeper audit trail with visibility into the exact paragraphs referenced, so we can continuously monitor and improve answer quality.

This should be a standard we demand everywhere. When AI scrapes and regurgitates without attribution, we break the incentive structure for creating new knowledge. Why would anyone invest time in careful research, nuanced thinking, or original insights if it just gets anonymized into someone else's answer? And why would anyone trust that answer without seeing sources? 

Transparency is respect. When you can see exactly where information comes from, you can decide for yourself if you trust it. That's not just good practice. That's the foundation of a sustainable system where human expertise continues to matter.

Trust isn't built in a single moment. It's built across hundreds of interactions, each one either reinforcing or undermining confidence. But more than the scale, it's the pattern: when you respect boundaries, admit limitations, and show your sources, people come back. They ask harder questions. They trust you with higher stakes.

The choices we make now, about what to build and how to build it, will determine whether AI becomes a tool people genuinely trust or just another thing they've learned to be wary of.

We're choosing to build for trust. Because we believe it's the only way this technology becomes something truly valuable. Something that amplifies human expertise instead of replacing it. Something that makes the internet better instead of drowning it in slop.

The future of AI isn't predetermined. We get to shape it. And that work starts with trust.