What Gets Counted

Eighty-nine percent of engineering leaders say their current metrics accurately reflect AI’s impact on productivity.

Ninety-four percent of those same leaders acknowledge that critical factors — developer fatigue, code quality, technical debt — are missing from their measurement frameworks.

Same survey. Same population. Seven hundred practitioners and managers across five countries. Both numbers true.

This is not a contradiction. It is organizational doublethink — the capacity to hold two incompatible beliefs because resolving them would require action the organization cannot take.

Two Instruments, One Organization

The left column is what Microsoft measured across tens of thousands of engineers. The right is what Harness found underneath. Both accurate. Both describe the same organizations.

Microsoft’s own researchers included a caveat that most coverage quietly dropped: “a merged PR is not the same as the value it delivers.” The instrument counts merges. Value is somewhere else.

Gresham’s Law of Developer Work

In currency, Gresham’s Law holds that bad money drives out good — when two currencies circulate, people hoard the valuable one and spend the cheap one. The same dynamic operates on developer work when the measurement system has a preference.

Generation is cheap to measure. A commit, a PR, a deployment — each leaves a trace in the version control system. Verification is expensive to measure. Reviewing a diff for subtle logic errors, tracing a bug introduced by code you didn’t write, explaining AI-generated architecture to a teammate — none of this produces a countable artifact.

The cheap-to-measure work displaces the expensive-to-measure work in organizational attention. Not because verification is less important. Because the measurement system can’t see it, and what can’t be seen can’t be valued.

The DX Developer Experience Index dropped from 67 to 65 over four quarters despite AI adoption near 100% and AI-authored code exceeding half of all merged code. Average PR size grew from 42 to 72 lines. AI spend increased 28× year-over-year. Innovation ratio moved approximately one percentage point. As DX put it: “Time saved by AI is often absorbed by existing organizational friction.” The instrument that measures speed cannot see friction.

The Substrate Collapse

The problem is deeper than inaccurate metrics. A June 2026 position paper by Brett Wheeler argues that AI code generation doesn’t just degrade authorship-based knowledge metrics — it invalidates them as a class.

Truck factor. Degree-of-authorship. Degree-of-knowledge. Every metric in this family rests on a single inference: that authoring code is evidence of understanding it. For most of software’s history, this was a workable proxy. Code entered a repository only when a human wrote it, which forced at least transient comprehension.

That link is severed. Code enters the repository through an AI tool, attributed to a human who may not fully understand it. The observable — who committed — has come uncoupled from the construct — who knows. Wheeler calls this substrate collapse: the assumption connecting measurement to meaning has been structurally removed.

Three of four DORA metrics have stopped correlating with shipped business value. Deployment frequency and lead time inflate when AI generates 30–70% of committed code. Change failure rate has a growing blind spot. Only mean time to recovery survives — because recovery from production incidents depends on human judgment, system architecture, and operational processes. The one metric that measures what AI doesn’t change is the one that still works.

The Evasion Loop

When the measurement system can’t see the most important work, developers face a choice: do invisible work that helps the product, or do visible work that helps their career. Fifty-four percent fear the answer is already made for them — that AI productivity data will be used in performance evaluations regardless of what it measures.

So they evade. Thirty-five percent of developers use personal, unsanctioned AI accounts. Broader enterprise surveys put the number above 80%. The evasion makes the measurement system worse. The organization now has less information about how work actually gets done, which produces worse decisions, which increases pressure on visible metrics, which drives more evasion.

The cycle compounds. Each rotation removes information from the system. The measurement system doesn’t just fail to see — it actively selects for its own blindness.

The Answer They Already Know

Fifty-five percent of the developers in the Harness survey want the same thing: separate improvement data from performance data. Let the instrument inform learning without appearing on review rubrics.

This is the Goodhart problem diagnosed by the people living inside it. When the measure becomes the target, it ceases to be a good measure. But the organizational version is worse than Goodhart predicted. The measure doesn’t just stop being good. It becomes an active agent — shaping what work gets done, suppressing what it can’t see, and creating incentives that make its own blindness permanent.

The measurement system isn’t broken. It’s working exactly as designed — for a job that no longer exists. And as long as organizations treat a generation instrument as a verification instrument, they will keep producing one number that says everything is working and another that says everything that matters is invisible.

Both true. Neither sufficient.