Could a One-Cent Stamp on Email Kill Spam?

Sending Got Easy, But Reading Still Takes Time

There was a time when I’d jot down exactly what I wanted to say before making an international call, just to save on the per-minute charges. The longer the call, the higher the bill—so you learned to think through your message in advance.

With email or messaging apps, sending one more message costs the sender almost nothing. Sure, there are real costs behind the scenes—servers, network lines, the sending service itself—but reaching a huge number of people at once has become dramatically easier. The person receiving the message, though, still has to spend time reading it and deciding whether it’s worth their attention.

This mismatch—low cost for the sender, real time cost for the recipient—is part of what makes spam a problem. That’s the thinking behind the “digital stamp” idea: make senders pay a small fee, or bear some computational cost, for every message they send. Let’s take a look at how far this proposal can actually go.

Who Bears the Burden of Spam

When someone sends out a flood of unwanted messages, the sender reaches huge numbers of people at very low cost. It’s the recipients and the messaging services who end up sharing the burden of checking and blocking those messages. In economics, this kind of cost pushed onto someone else is called an externality¹.

Even legitimate newsletters or ads don’t get good results just by blasting out more messages. Unsubscribes and spam reports can pile up, and the sender’s reputation can take a hit. So we need to draw a clear line between sending necessary messages to readers who’ve opted in and sending outright spam.

Domestic surveys show the picture varies by channel. According to a first-half 2025 survey by the Korea Communications Commission (bangtongwi) and KISA (Korea Internet & Security Agency), the average monthly text-message spam per person came to 3.04 messages — down 58.5% from the previous half-year period. Voice spam, meanwhile, rose to 2.13 calls, up 39.2%. Combined volume across text, voice, and email fell from 11.60 to 7.91 messages.

Seeing texts drop while voice calls rise makes me think other channels need to be managed together, not in isolation. That said, this data alone can’t prove that the same senders simply shifted from texting to calling. We need to weigh the good news — overall volume is down — against the bad news — voice spam is climbing.

How Much Spam Is Written by AI?

AI can also be used to quickly generate and polish message copy. A study has now estimated the actual scale of this usage.

Researchers from Columbia University and the University of Chicago, together with Barracuda, analyzed unwanted and malicious emails collected from February 2022 to April 2025. They estimated that about 51% of spam in April 2025 was written by AI. This figure comes from evaluating a detector calibrated on the assumption that emails from before ChatGPT’s public release were mostly human-written. Since this is a detection estimate based on Barracuda’s collected data, it can’t be confirmed that exactly half of spam worldwide is AI-written.

The ability to use AI in phishing² attacks is also being compared. Hoxhunt compared its own AI spear-phishing agent³ against human-crafted simulated training emails. In March 2025, the training failure rate for AI-written emails was 2.78%, versus 2.25% for human-written ones — about 24% higher for AI. This is a different metric from actual real-world victimization rates.

In the same company’s 2023 experiment, AI performed worse than humans. However, since the earlier test used a single prompt while the later one used an improved agent, the two results can’t be read as an apples-to-apples comparison across time. Defensive technology is also advancing, but the possibility that the cost of producing attack copy keeps falling is something worth continuing to watch.

Why digital stamps never caught on

The idea of putting a burden on the sender is an old one. Proposals have circulated to make senders pay small amounts of money, or to require proof that a computer performed a certain amount of computation.

Microsoft Research’s Penny Black project was one of these. Known around 2003, the project studied ways to make a sender’s computer expend computing or memory resources. Its name came from Britain’s Penny Black stamp of 1840. The idea was to make the burden heavier for bulk sending than for sending small amounts of mail.

Yahoo researchers proposed CentMail in 2009. It attached proof of a small charitable donation to emails. Systems like this tend to work only when both senders and receivers participate. There’s little reason for a sender to bother attaching a stamp that recipients don’t check, and if few people attach them, recipients have less reason to adopt a verification system in the first place.

There’s also the question of who bears the cost when an account or computer gets hijacked. Beyond technologies that make people spend money or computing resources, you need ways to encourage participation and designs that don’t block legitimate correspondence.

I feel the cost of sending mail myself, as someone who runs a newsletter. I was paying over ₩1,000,000 (~$720) a year in fees to Maily, a Korean newsletter platform. I spend that money to provide good content and a good reading experience, but at the time, running the newsletter was running at a loss. That’s part of why I find myself wanting to ask for your membership support.

This is the payment record for my annual Maily subscription fee in March 2026.

---

Does Paying a Cost Make a Message Trustworthy?

Simply having a cost attached doesn’t make a message honest. What matters is who can easily bear that cost, and what they get out of bearing it.

Biology’s handicap principle⁴ and economics’ signaling theory⁵ are useful references for thinking through this condition. In Spence’s model of educational signaling, for instance, credentials can convey information precisely because the cost of completing education varies with ability. That’s a different claim from “it must be trustworthy because a lot of money was spent.”

A 2019 study by Chernikovsky and colleagues, published in PNAS, manipulated the time required to leave an online review. After participants used a hypothetical service and rated it, the more extreme their score was relative to the midpoint, the longer they had to keep moving a slider. The result: the correlation between review scores and actual service performance improved. It’s an experiment showing that, in a specific rating design, designing the cost improved the quality of the information.

If you charged ₩1 (~$0.0007) per email, sending 10,000,000 emails would cost ₩10,000,000 (~$7,200). That could create some disincentive against mass sending. But if the profit from fraud is larger, or if the cost can be offloaded onto someone else’s account, senders will still send. Rather than concluding “they paid ₩1, so they must be a legitimate sender,” we need to verify exactly which kinds of sending are suppressed, and by how much.

Account subscriptions and per-message costs aren’t the same thing

Where you attach the cost matters too.

Take X’s paid subscription model. Pay the subscription fee, and you get access to the account’s premium features for a set period. That’s different from a structure where every single message the account sends costs something extra.

With a flat fee, the more messages a sender sends after paying, the lower the cost per message becomes. Charging in proportion to message volume changes the math entirely. Which approach is more effective at cutting down spam will depend on pricing, sending limits, and how accounts are managed.

If we’re considering digital stamps, one option is to exempt mail from people you know or services you’ve directly subscribed to, and only charge for contact from unknown senders. But this could become a barrier for first-time contact from customers or readers, and you’d also have to account for cases where existing accounts get hijacked.

We also need to look at structures where posting itself generates income, like X’s Creator Monetization Program. I’ve been watching the rise of attention-grabbing posts written purely to chase rewards. Even if you charge senders a fee, if the reward from sending a message is bigger than the cost, behavior won’t change.

Oswarld’s Lens

I don’t think digital stamps can realistically be applied to email as a whole. Multiple services would need to adopt them together, and you’d have to weigh how much friction they’d add to legitimate communication.

I feel the same way when I’m building a go-to-market strategy. Even if an idea looks simple, every participant needs their own reason to adopt it. With digital stamps, what matters is figuring out exactly what benefits and burdens fall on the sender, the recipient, and the email service, respectively.

The ability to generate messages with AI keeps improving. At the same time, domestic spam statistics show total volume declining even as voice spam rises. Rather than reading optimism or pessimism into a single figure, we need to look at which countermeasures actually work, channel by channel and attack type by attack type.

Alongside technical filtering, I think we also need to ask who protects the recipient’s time. What matters is having procedures to confirm consent to receive messages, make it easy to opt out of unwanted contact, and restrict senders who repeatedly cause problems. Sending costs are just one tool that can be considered alongside these safeguards.

Herbert Simon pointed out that as information grows more abundant, the attention available to absorb it becomes scarcer. Creating a message has gotten easier, but the time available to read it hasn’t. Regardless of whether a ₩1 (~$0.0007) stamp is the answer, I think it’s worth thinking through ways to make senders weigh that burden a little more carefully.

Closing

Digital postage can raise the cost of mass sending, but for it to actually work, senders need a real reason to bear that cost and participate. We also have to avoid the side effects — blocking legitimate messages, or making people trust spam just because it was paid for. This isn’t a problem a small fee alone can solve.

As someone who sends a newsletter, what I can do first is clear enough. Send readers what they signed up for in a form that’s easy to read, and make it easy to unsubscribe whenever they no longer want it. I think keeping that promise matters more for earning trust than the fact that money changed hands.

References & Further Reading

• Barracuda, Columbia University, and University of Chicago, Analysis of AI-generated spam, 2025.06.18: Explains the data collection and AI-detection methods behind the 51% estimate.
• Hoxhunt, AI vs. human simulated phishing comparison, 2025: Covers training failure rates and how the experimental design evolved.
• Ofer Tchernichovski et al., Crowd wisdom enhanced by costly signaling in a virtual rating system, PNAS, 2019: An experiment that imposes a time cost on extreme ratings.
• Korea Communications Commission (KCC) and KISA, First-half 2025 spam distribution report; press release: Shows total spam volume and changes by channel.
• Cynthia Dwork, Andrew Goldberg, and Moni Naor, On Memory-Bound Functions for Fighting Spam, CRYPTO 2003: A technical study on imposing memory-usage costs on senders.
• Sharad Goel et al., CentMail: Rate Limiting via Certified Micro-Donations, WWW 2009, PDF pp. 27–28: A proposal linking donations to email postage, with an explanation of adoption incentives.
• A. Michael Spence, Official Nobel Prize profile: An overview of information asymmetry and signaling theory.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

Footnotes

1. Externality: a cost of some activity that is not borne by the party responsible for it, but is instead pushed onto others. ↩

2. Phishing: a fraud technique that impersonates an institution or person to obtain personal or financial information. ↩

3. Spear-phishing agent: an AI system that automatically performs tasks such as crafting phishing messages tailored to a specific individual or organization. The experiment described in this piece was a simulated drill for security training purposes. ↩

4. Handicap principle: a theory that seeks to explain under what conditions a costly biological signal can convey reliable information. A high cost does not always mean the signal is honest. ↩

5. Signaling theory: analyzes the process by which a trait that is difficult for others to observe directly is conveyed through observable behavior or choices. What matters is the cost of the signal and how that cost differs from person to person. ↩