Human-AI Synergy

The adoption-rate gap isn't a failure to fix — it's the blueprint for who does what. Agents triage breadth; humans judge intent.

Why this, for you: you've wired up the AI reviewer and now its comments land in every PR — verbose, confident, and only sometimes right. The temptation is to treat that output as a to-do list, or to mute it entirely. Both waste the asset. The data says there's a third way: split the work by who's actually better at each half.

An empirical study of 278,790 code-review conversations across 300 GitHub projects measured exactly how AI and human reviewers differ. The headline numbers look like an indictment of AI review. Read them as role definitions instead, and the collaboration designs itself.

1 The measured gap is a design input

Human review suggestions are adopted 56.5% of the time. AI suggestions are adopted 16.6% — roughly one-third the rate, a 39.9-point gap. Over half of the unadopted AI suggestions are outright incorrect or were resolved by a different developer fix.

That gap is not a bug to close — it's a signal about role. AI review output requires triage, not equivalence to a human comment. Treating it as a checklist to clear works against you.

The two reviewers don't even cover the same ground. Over 95% of AI comments target one of two things: defect detection or code improvement. Human comments spread across understanding questions (17–31%), knowledge transfer (4–6%), testing feedback, and plain social communication. AI review structurally misses mentoring, architectural questioning, and team knowledge sharing — not because it's tuned wrong, but because that's not what it does.

2 AI first, human last

The role split the data supports is sequential, not parallel. The agent runs the mechanical first pass; the human makes the merge decision.

Stage	Owner	What it covers
First pass	AI agent	defect detection, consistent rule-checks, code-improvement flags
Triage	Developer	adopt / reject / fix differently — evaluate each suggestion on its own
Final review	Human	design judgment, knowledge transfer, architectural fit, merge call

The sequence works because the AI pre-triage reduces the human's cognitive load: known-pattern defects are handled first, freeing the reviewer's attention for the higher-level concerns only a human supplies. Agents pattern-match defect databases tirelessly; humans bring the judgment agents lack.

Watch the complexity creep

Even correct AI suggestions tend toward addition. Adopted AI fixes raise cyclomatic complexity (+0.085 vs -0.002 for humans) and code size (+0.216 vs +0.108 statements). Humans tend to simplify; agents tend to add. A technically right suggestion can still accrete technical debt — track the complexity delta of what you adopt.

3 Constrain the firehose, verify the firehose

AI agents emit 29.6 tokens per line of code reviewed versus 4.1 for humans — a 7x verbosity gap. Unconstrained, that becomes the same alert fatigue that signal-over-volume design exists to prevent. Two levers keep it productive:

Filter at the source — configure review agents with confidence thresholds and severity filters so the low-value 7x never reaches the developer's eyes.
Verify with a second agent — a second AI that validates the first's findings filters incorrect suggestions before triage, cutting the human's load further (the committee pattern, applied to review).

But correlated failure is the catch: when the reviewing agent shares the generating agent's training distribution and no executable specification anchors the review, the two reason alike and miss the same classes of error — failures echo rather than cancel. Mitigate with executable specs or cross-family agent panels.

4 When the split backfires

AI-first, human-last adds a triage cost to every PR. That cost isn't always worth it:

Low-adoption repos — where AI suggestion quality sits below ~10% adoption, the triage burden can exceed the defect-catch benefit. Measure adoption per-repo before committing to the pattern.
Reviews of AI-generated code need 11.8% more rounds than human-written code, and 85–87% of AI-initiated reviews end with no follow-up discussion — AI review is a broadcast, not a conversation.
No filter tooling — without confidence/severity filters, the 7x verbosity degrades trust over time and pushes effective adoption below the 16.6% baseline.

# Review-agent config — the human-last split, made operational confidence_threshold: 0.7 # drop low-confidence noise at the source severity_filter: [high, med] # tame the 7x tokens/LOC firehose verifier_agent: cross-family # second pass; avoid correlated failure # Output is a TRIAGE QUEUE, not a to-do list. Human makes the merge call.

↪ Your win: split by comparative advantage

Read the 16.6% vs 56.5% gap as a role spec — AI output is a triage queue, never a checklist to clear.
AI first, human last — agents triage breadth (defects + improvements); humans judge intent, design, and merge.
Track the complexity delta — even correct AI fixes add cyclomatic complexity; humans simplify.
Filter the 7x verbosity — confidence thresholds and severity filters, or alert fatigue kills trust.
Measure adoption per-repo — below ~10%, the triage cost can exceed the defect-catch benefit.

Retrieval practice — recall, don't peek

Question 1The 16.6% vs 56.5% adoption gap should be read as…

Question 2In the sequential model, the human reviewer owns…

Question 3Compared with human suggestions, adopted AI suggestions tend to…

Question 4A second verifier agent can fail to help when it…

Question 5 · spaced recall from an earlier lessonTo keep plausible-but-false findings off the human's desk, every finding should first be…

Ask me anything. Want help setting confidence/severity thresholds for your review agent, or a way to measure per-repo adoption before you commit to AI-first? Next in Part 3: PR Slicing & Descriptions — authoring AI PRs so the human half of this split can actually review them.