Glonce

AI slop is changing how engineers review code

toolsThe Glonce editorsIEEE Spectrum🔥66Glonce rating66Impact65Novelty55Relevance90Surprise40Players65Credibility80
AI slop is changing how engineers review code

AI coding tools can now generate thousands of lines in minutes, but the resulting code can look clean while hiding faulty assumptions, security holes or subtle bugs, and fixing those problems risks erasing the productivity gains AI was supposed to deliver. A survey of more than 1,100 developers by Sonar, an AI code-verification startup, found that respondents estimated AI contributed 42 percent of the code they added to shared codebases, yet 96 percent did not fully trust that output to work correctly; 38 percent said reviewing AI-generated code takes more effort than reviewing a colleague's, and 61 percent said AI often produces code that looks correct but is unreliable. Investors are betting on the resulting gap: in August, AI code-review startup CodeRabbit raised $143 million at a $1.5 billion valuation, claiming more than 2 million reviews a week for 17,000 customers including Nvidia, Indeed and BMW Group.

At Synthesia, an AI video-generation company, 118 engineers went all-in on tools like Claude Code in November 2025, and pull requests were up 120 percent year over year as of August, with 95 percent of them containing AI-generated code, according to chief technology officer Peter Hill. A recurring problem is duplication: because the tools have limited context, Synthesia has found as many as 10 versions of the same function, forcing engineers to hunt down and remove the redundant ones before retraining the agent. Synthesia now uses AI agents to route higher-risk changes, such as code touching customer data or core business rules, to human reviewers, while lower-risk edits like error-message changes get less scrutiny; even so, fewer than 5 percent of changes bypass human review altogether.

Other companies are pushing the check earlier, into the planning stage. At Amazon Stores, senior principal engineer McLaren Stanley, whose 70-person team maintains the architectural backbone that supports more than 1,000 developers modernizing 17 years of code behind Amazon's mobile shopping app, now has engineers write detailed specifications before an agent starts generating anything. He recalls one missing instruction causing an agent to generate 25,000 lines of code in the wrong version of Swift; switching versions produced 600 errors the agent could not fix at once, so Stanley discarded the code, fixed the specification, and had the agent regenerate correct code in fifteen minutes. At Amazon Web Services, senior principal engineer David Yanacek says agents now run a first pass on generated code, testing whether it works, checking it against the original plan, and scanning for security flaws before a person looks at it.

At Bonterra, a nonprofit software provider with about 290 engineers, proposed changes tripled within three months of adopting AI, code entering review rose tenfold and review times tripled, making it impractical to inspect every line, CTO Tanuja Korlepra says. Bonterra's agents now compare code against the approved design, security rules, coding standards and accessibility requirements, then report a confidence score; a low score or flagged problem sends the change to a person, and code touching payments, personal data or other sensitive systems always gets human review. Temporal, an open-source developer platform, instead puts the burden on the person submitting code: under its "Send Back" policy, described by CEO Samar Abbas, engineers must explain in their own words the agent's design choices and how the code handles unusual conditions, or the reviewer rejects it outright.

The shift is also reshaping how junior engineers learn. JD Raimondi, chief AI architect at the consultancy Making Sense, warns that when machines produce more code than engineers can closely read, human approval can become "theater approval": an engineer confirms a feature works and approves it without understanding the underlying choices. Making Sense keeps junior staff involved in deciding why a customer needs a feature rather than just checking AI output. IBM is doing the opposite, giving new hires harder assignments sooner: general manager of automation and AI Neel Sundaresan says AI now helps junior engineers perform an estimated 70 to 80 percent of some tasks that once required a senior engineer, with juniors diagnosing and fixing failures before senior developers give final approval. Synthesia, which hires mostly mid- and senior-level engineers, pairs its less experienced staff with both a senior colleague and an AI agent. Korlepra sums up the stakes for the pipeline: "If the industry stops hiring juniors, the industry stops producing seniors."

Key facts

  • A Sonar survey of more than 1,100 developers found AI contributed an estimated 42 percent of code added to shared codebases, yet 96 percent of respondents did not fully trust AI output to work correctly.
  • AI code-review startup CodeRabbit raised $143 million at a $1.5 billion valuation in August, claiming over 2 million reviews a week for 17,000 customers including Nvidia, Indeed and BMW Group.
  • At Synthesia, pull requests rose 120 percent year over year as of August, with 95 percent containing AI-generated code, after 118 engineers adopted tools like Claude Code in November 2025.
  • At Amazon Stores, a missing specification caused an agent to generate 25,000 lines of code in the wrong Swift version, producing 600 errors; after the spec was fixed, the agent regenerated correct code in fifteen minutes.
  • At Bonterra, proposed changes tripled and code entering review rose tenfold within three months of adopting AI, so agents now pre-screen changes and route only low-confidence or sensitive ones to human reviewers.

Why it matters

AI coding tools have shifted the engineering bottleneck from writing software to reviewing it. Code that looks clean can hide faulty assumptions, security vulnerabilities or errors that surface only after deployment, and the article's own framing is blunt: fixing those problems could erase the productivity gains AI was supposed to deliver in the first place. How companies rebuild review, not how fast they generate code, will decide whether AI coding nets out as a genuine speedup.

Who it affects

Engineering organizations at scale feel it first: Synthesia (118 engineers, pull requests up 120 percent year over year), Amazon Stores and AWS, Bonterra (proposed changes tripled, review volume up tenfold in three months), and Temporal are all named restructuring review around AI volume. It also affects the code-review tooling market, where CodeRabbit's $143 million round points to investor interest in closing the trust gap, and it affects junior engineers specifically, whose traditional path into the profession, writing and being corrected on routine code, is being redesigned by employers like Making Sense and IBM.

How to use it

The practices described function as a playbook. Write a detailed specification before an agent starts generating code, since Amazon's McLaren Stanley found that a missing instruction cost 25,000 lines of wrong-version code and 600 unfixable errors, while a corrected spec regenerated correct code in fifteen minutes. Layer automated agents to do a first pass, testing functionality, checking plan conformance and scanning for security flaws, as AWS and Bonterra do, before routing low-confidence or sensitive results (payments, personal data, core business logic) to a human. Where volume makes full human reading impossible, shift the burden to the submitter: Temporal's "Send Back" policy requires engineers to explain the agent's design choices and edge-case handling in their own words, or the change gets rejected.

How solid is it

The reporting draws on named, on-record executives at multiple companies (Synthesia's Peter Hill, Amazon's McLaren Stanley and David Yanacek, Bonterra's Tanuja Korlepra, Temporal's Samar Abbas, Making Sense's JD Raimondi, IBM's Neel Sundaresan), which gives the specific figures and anecdotes direct attribution rather than secondhand paraphrase. The Sonar statistics, however, come from a company that sells AI code verification and rest on developers' self-reported estimates of how much of their code is AI-written, not a measured count, so they should be read as an industry pulse-check rather than a precise measurement.

Risks and caveats

Several figures lack full context: the article does not give a year for the CodeRabbit funding round or for Synthesia's pull-request growth, both dated only to "August", and it does not date the Sonar survey itself. The Sonar numbers also carry the built-in bias of being commissioned by a vendor with a financial stake in the code-review-tooling narrative. Separately, JD Raimondi's warning about "theater approval", engineers approving AI code they have not actually understood, points to a risk the described safeguards do not fully close, and the debate over how junior engineers build judgment when AI increasingly does the writing remains unresolved rather than solved by any single company's approach.

“I don't know if we ever get to the point where you can truly trust the agentic generation of code.”

— Peter Hill, Synthesia's chief technology officer