Humans Are Bad at Game Theory

I run a large part of my company through an AI assistant. It books my demos, chases follow-ups and drafts my outreach. The obvious next move is to get my peers on it too. Assistants that can talk to each other schedule faster, share context and skip the email ping-pong.

Most people don't make that move. They keep the tool quiet. Telling a peer feels like giving away a head start.

That instinct is the first mistake. The second one runs the other way: once a tactic works, everyone copies it at scale, and the tactic dies. Automated outreach is the clearest case. If everyone's agent sends a thousand perfect cold emails, nobody reads cold email.

Both mistakes have the same root. We treat games as zero-sum when they aren't, and we ignore the commons when we're in one.

1. We would rather be ahead than be better off

In 1998, Sara Solnick and David Hemenway asked people to choose between two worlds. In one, you earn more than everyone around you. In the other, everyone is richer, including you, but you earn less than the people around you. Half the respondents said they would take a world with 50 percent less real income as long as their relative position was high. (source)

That is the status trap in one line. Absolute gain loses to relative rank.

It shows up directly in how people treat AI at work:

• 52% of people who use AI at work are reluctant to admit using it for their most important tasks, and 78% bring their own AI tools to work (Microsoft and LinkedIn, 2024 Work Trend Index). (source)
• 48% of desk workers would be uncomfortable telling their manager they used AI for common tasks. The top reasons: it feels like cheating (47%) and fear of looking lazy (46%) (Slack Workforce Index, fall 2024). (source)

So the tool spreads, but quietly. Everyone has a private edge and nobody shares the playbook. The team, the company and the network all lose the gains that come from using it together.

2. Why it feels zero-sum, and when it actually is

Economist Fred Hirsch called some goods "positional": their value depends on how much you have compared with others. (source) A top college seat is positional. So is being the only founder in the room who replies in five minutes.

Here is the catch. Not every tool is a status race. There are two different games, and we keep confusing them.

Stag hunt (coordination): value grows when others join. Going alone is safer but smaller. AI example: assistants that coordinate with other people's assistants. Rational move: invite peers early. What humans do: hoard.

Prisoner's dilemma (commons): each player gains by defecting, and when all defect, all lose. AI example: agents blasting one shared channel like inboxes, phone lines, LinkedIn and job portals. Rational move: don't race to the bottom. What humans do: race.

The stag hunt, first described by Rousseau, has two stable outcomes: hunt the stag together (bigger payoff) or hunt hares alone (safer). (source) People pick the hare because they don't trust others to show up. An assistant that coordinates with other assistants is a stag hunt. It gets more useful with every peer who joins. Keeping it secret is picking the hare.

Products that grow through invites know this. Dropbox paid both sides: for every friend who installed it, you both got 500 MB of free space. (source) It turned a sharing decision that felt like a loss into a visible gain for both people. Meta launched Muse this month as "a personal AI agent" that "actually does the work." (source) The assistant race is now mainstream, and the invite question will come up for everyone.

3. The other mistake: when everyone automates, the channel dies

The flip side is just as real. Some edges only exist because few people have them. Once everyone copies the move, the edge is gone, and the channel is often worse than before.

Banner ads. The first banner ad, on HotWired in 1994, had a 78% clickthrough rate. By 2011, Facebook ads were at about 0.05%. Growth writer Andrew Chen has a rule for this: over time, every marketing channel ends up with poor click rates. (source)

Cold email. Average reply rates fell from 8.5% in 2019 to about 5% in 2025. The median sender in one 65-million-email analysis gets 0.48%. (source) Gmail now requires senders to keep reported spam rates below 0.3%, with stricter rules for anyone sending over 5,000 messages a day, in effect since February 2024. (source) The inbox providers are fencing off the commons.

Phone calls. Robocalls did to the phone what spam did to email. Eight in ten Americans say they don't generally answer calls from unknown numbers. Only 19% do. (source) A cold call now mostly reaches voicemail.

Job applications. LinkedIn is processing about 11,000 applications a minute, up 45% in a year, driven by AI tools that write and submit applications at scale. (source) Recruiters answered with AI screeners, and candidates now use AI to answer AI. LinkedIn's own survey found 38% of US job seekers applying to more jobs than ever but hearing back less often. (source) Both sides spend more and match worse.

Trading. Finance has measured this precisely. Across 97 published stock-return signals, returns were 58% lower after the research was published, as traders piled in. (source) An edge that everyone knows is not an edge.

If everyone automates outreach, outreach itself dies. The channel was never the asset. Scarcity of attention was.

4. Same root, two failures

These look like opposite mistakes: one hoards, the other floods. Both come from reading the game wrong.

• In a coordination game, we act like it's zero-sum. We hide the tool that gets better when shared.
• In a commons, we act like we're alone. We scale the tactic that only works while it's rare.

In both cases, status wins over reasoning. Being first feels better than being better off, even when being first is temporary and being better off is not.

5. What to do about it

1. Ask which game you're in. If a tool gets more useful when the other side uses it too, it's a stag hunt. Invite early, because the network is the product.
2. Treat channel edges as temporary. A new outreach hack will decay. Use the window to build relationships that outlast it.
3. Move from volume to trust. When every inbox is full of machine-written mail, what gets through is a warm intro, a referral, or an agent that talks to someone's agent with both people's permission.
4. Build for the team, not the lone hero. Pick tools and habits that make a whole team better together, instead of one person keeping a private hack.
5. Make sharing pay both sides. Dropbox didn't ask people to be generous. It made the invite a win for the sender too. Tools and teams can do the same.

Assistants will make everyone faster. The people who win will not be the ones who kept theirs secret the longest. They will be the ones who noticed which games reward going together, and which ones punish everyone for going at once.