The Pulse
Microsoft Study Finds Seniority, Not AI Disclosure, Skews Code Reviews
A Microsoft-led study of 447 software engineers found no detectable penalty for code marked as AI-assisted in an AI-accustomed workplace. Reviewers rated identical code and its author less favorably when the author was labeled junior rather
AI.info Team · Written with AI ·
Four hundred forty-seven software engineers reviewed the same four code snippets, but a junior author label still lowered ratings of both the code and its author. An AI-use disclosure, by contrast, produced no detectable penalty in the experiment, which was presented September 30 at the 2026 IEEE Symposium on Visual Languages and Human-Centric Computing.
Same code, different author labels
The Microsoft-led study tested whether reviewers respond differently to code when a commit message says AI helped produce it, or when the author is identified as junior rather than senior. Each participant reviewed four snippets in a within-subjects experiment. The researchers kept the code constant while randomly varying two labels: “used AI” or “did not use AI,” and Level 1 or Principal author.
The authors write: “AI disclosure did not bias perceptions of code effectiveness or author competence, whereas a seniority label significantly biased both.” Jenna Butler, Emerson Murphy-Hill, Constance Hadley, Alexa Tipton, Nancy Baym, Sankeerti Haniyur and Jina Suh, “Where AI Use Is Normalized, Author Seniority but Not AI Disclosure Biases Code Review”
On the study’s standardized competence measure, the senior-author label scored 0.178 higher than the junior label; the difference was statistically significant. The senior label also raised the standardized code-effectiveness score by 0.148. Reviewers rated code labeled as junior lower for quality, readability and security, even though they saw the same underlying code.
AI use was familiar to most participants
The researchers recruited full-time software engineers at Microsoft, who completed the study between October 13 and November 10, 2025. Of the participants, 97.8% said they used AI to generate code, and 81.2% used AI for code review. The experiment’s finding therefore concerns a workplace where AI use was common, not a general claim that disclosing AI assistance never affects reviewers.
AI labeling did not measurably shift the study’s ratings of code effectiveness or author competence. The disclosure was a free-text sentence in the commit message, rather than an AI-generated change presented as the work of an autonomous agent. The researchers also found no detectable effect on their other measured outcomes, including hiring recommendations and comment counts.
The penalty had limits—and so does the finding
The seniority contrast was deliberately wide: Level 1 versus Principal. The authors note that differences between adjacent job levels may produce smaller effects. The experiment also used review vignettes rather than live, back-and-forth code review, so it cannot establish how the label affects decisions in actual engineering teams.
The AI result has its own qualification. The researchers did not use a participant-level check to confirm that every reviewer noticed the disclosure, and they say that missed disclosures could weaken a measured effect. Their sample came from one U.S. technology company with high comfort around AI; the paper cautions that the result does not rule out penalties in workplaces where AI use is less accepted.
Which labels should review tools show?
The study’s central design implication is that an interface can influence judgments even when the code itself does not change. Reviewers did not leave more comments or alter hiring recommendations based on seniority, but they did rate the junior-labeled code and author less favorably. The authors suggest that code-review tools should prompt a broader question than whether to disclose AI assistance: which details about an author belong in the review interface at all?
For teams deciding how to present AI-assisted contributions, the results offer evidence from one AI-accustomed workforce—not a universal verdict on disclosure. They also show that the seniority cues commonly displayed in review systems may shape assessments of identical work.