A few years ago, recording a call came with a small ritual. Someone would say “do you mind if I record this?” and wait for the nods. Then the tools took over the ritual: Zoom or Teams would pop a banner saying the meeting was being recorded, and you clicked to continue. Today, I join calls where an AI notetaker is already sitting in the participant list before most of the humans arrive. The only thing anyone has to do is let it in. No introduction, no question, no pause.

I see this every day now, across customer meetings, partner calls and internal ones. I’m a strong believer in AI and I use these tools myself. I’m also running a company whose customers trust us with their IP and confidential conversations, and I’ve started to think we are drifting into something without having decided to.

“It was always possible” misses the point

The obvious counter is that anyone could record a meeting with a phone in their pocket for twenty years. True. But what changed isn’t the microphone. It’s three things around it.

Recording became the default, not a decision. The class action filed against Otter.ai last year alleges that when a host links their calendar, the notetaker can join a meeting without affirmative consent from anyone, including the host. Nobody chose to record. An integration did. A workforce survey this summer found roughly two thirds of people who had encountered an AI notetaker said consent was inconsistent or missing.

The consent moved from the people in the room to the account holder. The tools ask the person who owns the subscription. The rest of us are bystanders to someone else’s product. And a bot appearing in a participant list is not consent: as far as I can find, no jurisdiction treats that as legally sufficient notice.

The record became a product. An old audio file sat unplayed in a folder. A transcript is searchable, summarised, pushed into a CRM and in some cases used to train the vendor’s models. That’s a different thing to have floating around after a candid conversation.

Put together: the cost of recording fell to zero, and the cost of asking stayed exactly where it was. So asking lost. Consent didn’t become unimportant. It became friction, and products are built to remove friction.

The wearables make it silent

Meetings at least have a participant list. Wearables don’t.

Bee, now an Amazon product, is a bracelet that listens to every conversation you have and turns it into summaries, to-dos and a searchable timeline. Its privacy page is reassuring in one direction: audio is processed and deleted, and your conversations stay yours alone by default. Read that again. “Yours” means the wearer. The other half of every conversation isn’t the customer and isn’t asked.

Meta’s Ray-Ban glasses have had the louder backlash, partly because cameras trigger a reflex that microphones don’t. Users found they could cover the recording LED mid-capture, Meta patched it in August so the camera stops if the light is blocked, New York courts banned smart glasses, and a Comic-Con organiser banned them after guests complained about covert filming. That reaction is healthy. But notice it’s about the camera. There is no light to cover on an audio bracelet and nobody is running billboard campaigns about it.

Where I land, in two halves

Professionally, I think something like Bee would be useful. I’d happily have a device that remembers the commitments I made on a Tuesday call and drafts the follow-up. Being present in a conversation instead of typing is a real benefit, and people who are hearing impaired or forgetful describe these tools as an extended memory. I don’t want to pretend otherwise.

Personally, the same device worn in a restaurant, at a family dinner or in a corridor at a conference does something to the conversation before a word is said. If I assume anyone I talk to may be recording, I think twice before I speak. That’s not a privacy violation in the legal sense. It’s a tax on candour, and I’m not sure we’ve priced it.

The good news: a policy already covers the old ritual

We’re an ISO 27001 certified business, and our acceptable use policy already covers recording: no audio or video capture without the consent of the people involved, and only with tools the company has authorised. In practice that’s the old ritual written down, and it’s a better starting point than most firms I talk to have.

What it was never designed for is recording as a default. It doesn’t define what consent looks like when a bot joins before anyone speaks, whether every participant has to agree or just the host, which AI notetakers count as authorised, how long transcripts live, where they’re stored, or whether sending audio to a third-party AI processor is different from keeping a recording ourselves.

The legal side only adds pressure. In a dozen US states every party must consent to an audio recording, so one participant in California can make a recording unlawful even when the host is somewhere that only needs one party. For privileged conversations, a transcript held by an outside AI vendor may weaken that privilege. For a services firm, though, the customer’s reaction matters more than any of that.

So we’re building on the existing policy with a short recording and transcription standard: when recording is allowed, how consent is captured before it starts, which platforms are approved, which conversations are off limits, and retention and deletion rules for transcripts, including AI processing by outside vendors. What I haven’t solved is enforcement. A standard can stop my team from bringing a bot. It can’t stop the person across the table from wearing a bracelet. Policy sets the norm inside our walls; etiquette has to carry it outside them. That’s weaker than I’d like.

An open question, not a prediction

It would be easy to end with “soon every conversation will be recorded.” I don’t think the evidence supports that yet. Lawsuits are being filed, venues are banning devices, vendors are being forced to patch workarounds, and etiquette guides on how to ask before recording are appearing everywhere. Norms usually lag technology by a few years and then catch up.

The question I’m left with is whether they catch up before the default hardens. For now, I’d go back to the old ritual. At the start of a call, ask whether anyone is recording. It costs five seconds. That’s the whole point.

I find much of that picture credible. One question keeps getting in the way. If machines increasingly replace human labor, where does the purchasing power come from?

My thesis is simple. AI may solve the production side of abundance. I am less convinced that we have solved the distribution side.

I should say where I stand. I work with AI and spend much of my time looking for ways it can solve real business problems. I am optimistic about it. I believe it will create enormous productivity gains, and that AI developed securely and responsibly will do humanity considerably more good than harm. That view does not depend on whether we ever reach superintelligence, and neither does this argument.

What I am questioning is narrower: the economic mechanism that takes us from greater productivity to universal abundance. I see three separate problems in it, production, distribution and transition, and only the first looks close to solved.

Production: the part I find convincing

The evidence on cost is already striking. Stanford's AI Index reports that the price of querying a model at GPT-3.5 level fell from $20 per million tokens in November 2022 to $0.07 by October 2024, a drop of more than 280 times in under two years.

If robotics follows a similar curve, the cost of producing many goods and services could fall dramatically. I take that seriously. Production increasingly looks like a technology problem, and technology problems tend to get solved.

Distribution: cheaper is not the same as accessible

Consider a car manufactured almost entirely by robots. It still needs steel, aluminum, lithium and copper. It needs energy, land, factories, machinery, transportation infrastructure and capital. Automation may reduce many of those costs dramatically, but the underlying resources remain scarce, and someone owns them.

The question is not whether the car becomes cheaper. It almost certainly does. The question is why dramatically lower production costs would make it effectively free, or accessible to someone with no income.

Fuel makes the same point. According to the U.S. Energy Information Administration, crude oil accounted for about 51% of the average U.S. retail price of gasoline in 2025. Automation can reduce the cost of extraction, refining and transportation. It does not remove the economic value of a scarce natural resource, and the resource still has an owner.

That leads to what I think is the deeper issue. As the contribution of human labor declines, ownership matters more: ownership of AI models, robots, compute, energy generation, land, factories, natural resources, intellectual property, and the companies and capital behind them.

There is also a demand side to this. Businesses need customers. If wages shrink as a share of income, the people expected to consume what the machines produce need some other source of purchasing power. I can imagine several: dividends, public transfers, broader ownership. Each needs someone to design, fund and agree on it.

I do not mean this as an argument for or against any economic system. It is a design question. If productive assets generate more of society's income while wages generate less, how does economic participation evolve?

The IMF estimates that almost 40% of global employment is exposed to AI. Exposure cuts both ways, since AI is expected to complement some jobs and replace others. But the IMF has also warned that AI could worsen inequality without policies that address how the gains are shared.

Some AI leaders say much the same. One has called fair distribution more of a political question, and another has argued that people will need an ownership stake in what AI creates. So the gap is recognized. What I haven't yet seen is a convincing explanation of how it closes.

Transition: the mortgage problem

Even if a new economic model eventually emerges, people live inside the existing one. It contains mortgages, loans, retirement savings, pensions, taxes and wages, and many of those are long-term commitments written in fixed amounts.

Take someone five years into a 30-year mortgage. They owe a fixed payment for another 25 years.

Suppose AI causes prices and interest rates to fall sharply. Very low mortgage rates are not hypothetical. In 2019 a Danish bank offered a 10-year fixed mortgage at a nominal rate of minus 0.5%, although fees meant borrowers could still pay back more than they borrowed. Research on negative-rate environments also found pressure on bank margins and profitability, which suggests there are limits to how far lenders can follow rates down.

The larger point is simpler. Zero or negative interest rates do not eliminate the principal. $300,000 remaining over 25 years at 0% interest is still $1,000 a month.

Refinancing also normally requires income or creditworthiness. That is an obvious problem if the borrower's employment income is the thing that has disappeared.

This is not a niche issue. U.S. households held about $13.1 trillion in mortgage debt in mid-2026, according to the New York Fed.

So what happens to existing debt during the transition to an economy where human labor becomes substantially less valuable? I don't know. I haven't seen the abundance thesis address it.

The strongest argument against my concern

Previous technological revolutions also displaced workers. They went on to create new industries, new jobs, new wealth and new consumer demand. Research led by MIT's David Autor found that about 60% of U.S. employment in 2018 was in job specialties that did not exist in 1940.

AI may follow the same pattern. That is probably the strongest argument against worrying about a long-term income problem, and I take it seriously. But the abundance thesis, as I understand it, goes further than earlier waves of automation. If AI and robotics can eventually perform not only physical labor but a very large share of economically valuable cognitive work, then the historical solution, people moving into new kinds of productive work, cannot simply be assumed to hold forever.

Perhaps entirely new forms of human work will emerge. I am not claiming that mass technological unemployment is inevitable. But if the destination really is an economy in which machines do most economically valuable work, the original question returns: how is purchasing power distributed?

Why does abundance feature so prominently in the AI narrative?

The technology industry is in the middle of an extraordinary infrastructure buildout across compute, data centers, energy, semiconductors, models, networks and robotics. The four largest U.S. cloud and AI platform companies have together guided to more than $700 billion of capital expenditure in 2026, up from roughly $410 billion in 2025.

When companies and investors commit sums like that, the long-term economic promise naturally matters. The abundance thesis may be a sincere technological forecast, an economic vision, part of the investment case for AI infrastructure, reassurance about the disruption AI may bring, or some combination of these.

I do not know the motives of individual leaders, and I do not want to speculate about them. The more interesting question is whether the economics behind universal abundance have been worked through.

Better off is a smaller claim than abundance for everyone

"Better off" is a smaller claim than "abundance for everyone," and I am comfortable making the first. I believe AI will make us enormously more productive.

I believe it can make humanity better off. What I have not yet seen is the economic bridge between extraordinary productivity and universal abundance.

Perhaps there is an answer. I may be missing part of the economic model, and this remains an open question for me.

Production may increasingly become a technology problem. Distribution will remain a human one. If someone has worked through that math convincingly, I genuinely want to read it.