The Effort Dividend

For most of human history, the quality of what you produced was evidence of how hard you worked, and how hard you worked was evidence of how much you cared. Neither assumption survives the AI disruption.

In June 2026, the open-source Ladybird browser project stopped accepting pull requests from the public (via). The project maintainers' reasoning was very clear. Whether code was typed by hand is beside the point, they wrote - what matters is "who will answer for the consequences" once a change enters the codebase. Most people read it as an open-source governance story. They shouldn't. Ladybird has said out loud what every organisation that evaluates knowledge work is about to discover.

Effort was the gold standard backing the currency of trust. The code repository with a full suite of tests. The polished slide deck. The typo-free tender document. Each signalled the work put in - nobody spent a weekend on a pull request, or six years on a professional qualification, as a prank. The work was too expensive to fake, so it paid out trust automatically. For decades we collected the effort dividend. AI has not counterfeited this currency. It has done something worse: given everyone a central bank. A currency anyone can print is not forged. It is just worthless.

Two things broke at once. The signal, which let strangers trust your work, and the formation - the way doing the work built the skills and judgement behind it. The trust problem might be repaired. The harder question is what happens when people stop learning by doing.

The first attempts to counteract this are understandable but flawed. If effort can no longer be observed, measure something adjacent to it. Meta, Microsoft and Salesforce built internal leaderboards tracking employee AI token consumption and got exactly the behaviour you'd expect - engineers running agents that burn tokens for no reason. I have written before about AI vendors' alarm at insufficient token consumption. The leaderboard is the vendor's metric, adopted by the customer: consumption as proof of engagement. Goodhart's Law is usually stated as a warning that "a measure made into a target ceases to be a good measure." Evidence of work put in was Goodhart-resistant because it was genuinely costly.

Other reserves have been proposed to back the failing currency of effort. Evidence of independent thought is one. When one open source maintainer planted a fake instruction saying contributions marked with a tell-tale emoji would be fast-tracked, 21 of the next 40 submissions included it - proof that AI agents follow rules without fully understanding them. Taste over volume rewards judgment rather than output, and conveniently privileges the people who already possess it. Direct accountability, Ladybird's answer, works precisely because the project is now a closed community of people known to one another. Notice the pattern. Every countermeasure solves the signal problem by shrinking the circle of who is permitted to contribute.

Visible effort was always partly theatre. Renaissance courtiers even had a word for making hard things look effortless: sprezzatura. But sprezzatura concealed effort that was really there. AI lets you skip much of the effort entirely. The same mask now hides an empty room.

But effort was never only a signal. It was formative. The open-source developer who spent a weekend on a new feature in 2020 could tell you why every choice was made, because that was how they built the expertise. The person who runs an agent in 2026 and clicks "send" cannot answer for the consequences - not necessarily because they are acting in bad faith, but because they never built the understanding. Accountability can be assigned by policy, judgment cannot. No shortcut. You have to have done the work.

Even scrutiny itself has degraded. When elite consultants were tested inside their own domain, they often failed to detect incorrect AI outputs, even when warned that the system could be wrong. The fluency of the answers appeared to make errors harder to spot rather than easier. We read polished prose as competence. We always have. But that reflex is now a liability, and it feeds the same verification gap that keeps turning up everywhere else - scrutiny is the real work and it needs a significant budget.

The comprehensive bid proposal, the immaculate CV, the ridiculously beautiful fifty-page design report - every one of these artefacts traded on the effort dividend, and every one of them is now practically free to produce. Trust has come off the gold standard, and no replacement reserve has been agreed. In open-source, the permissionless model that built much of today's infrastructure is starting to fray, project by project. The only honest replacement - accountability earned slowly by known experts - is expensive and does not scale.

None of this means giving up on AI. For the signal problem: make scrutiny a named, budgeted line in the plan of work, not an assumed background activity - and defend it to clients as the deliverable rather than an overhead. For the expertise-building problem: let people draft with the tool, then require them to verify and defend the result, because verification is where judgment will now get built. The agent can produce the artefact. It cannot do the understanding for you.

Ladybird's repository is still public. Anyone can read it. To contribute, you must first be known - the oldest trust mechanism there is, and the one AI cannot yet counterfeit. But it locks out everyone who has not yet had the chance to become known. The effort dividend worked because strangers could spend it. We need to work out what strangers spend now.