Does AI Make Engineers Faster? The Research Behind "Still 4 Months"
By Matt Cullerton, founder of Mavric and Apso
A viral comedy sketch shows an engineer refusing to cut a four-month estimate no matter how a manager pitches AI. The four months is a joke premise with no source. The core point holds up better than you might expect: the only randomized trial on this question found AI tools made experienced developers 19 percent slower, and the 2025 DORA report ties AI adoption to less stable delivery.
What the sketch claims
A TikTok from @alberta.nyc (the creator Alberta Devor, who posts as Alberta Tech) plays both roles. A product manager asks how long an Android version of the app would take. The engineer says four months, then answers each follow-up:
- "What if you used AI?" gets "Still four months."
- Letting the AI build the whole thing and cleaning up afterward gets six months.
- Telling the AI it is an experienced software engineer gets "an over-engineered mess that will take a year to clean up."
- Shipping what the AI builds without looking at it gets "Then the app will be broken."
The caption reads "And if you disagree with 4 months, you don't work in corporate." No roadmap, sizing document or source for any of the numbers appears in the video.
How accurate are software estimates in general?
No published statistic says how long an Android port of an unspecified app takes. Researchers have measured how far estimates miss.
Magne Jørgensen and Kjetil Moløkken-Østvold reviewed that question in "How Large Are Software Cost Overruns? A Review of the 1994 CHAOS Report," published in Information and Software Technology volume 48 issue 4 in April 2006. They wrote: "All surveys of cost estimation accuracy in the relevant period and countries that we were able to find suggest average cost overrun in the range of about 30%." The surveys behind that range are Jenkins 1984 (23 software organizations in the USA, 34 percent), Phan 1988 (191 projects in the USA, 33 percent) and Bergeron 1992 (89 projects in Canada, 33 percent).
The same paper explains why the Standish Group's often-quoted 189 percent average overrun is a poor benchmark. The CHAOS report described its own sampling this way: "We then called and mailed a number of confidential surveys to a random sample of top IT executives, asking them to share failure stories." A sample built from failure stories does not represent typical projects.
Does AI make experienced developers faster?
In one randomized trial, no. Joel Becker, Nate Rush, Elizabeth Barnes and David Rein published "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" (arXiv:2507.09089) on 12 July 2025.
The design: "16 developers with moderate AI experience complete 246 tasks in mature projects on which they have an average of 5 years of prior experience. Each task is randomly assigned to allow or disallow usage of early 2025 AI tools."
The result: developers forecast that AI would cut completion time by 24 percent, and afterward they believed it had cut time by 20 percent. In fact, "allowing AI actually increases completion time by 19%--AI tooling slowed developers down."
So "Still four months" is conservative. For an experienced engineer in a familiar codebase, the trial points to a longer estimate. The study's limits are its size and scope: 16 developers, familiar open-source repositories and tools from February to June 2025.
What does cleaning up AI-generated code cost?
The six-month figure has no source. The cleanup cost it describes is measured in two places.
In the 2025 Stack Overflow Developer Survey, 66 percent of 31,476 respondents named "AI solutions that are almost right, but not quite" as a frustration, and 45.2 percent said debugging AI-generated code is more time-consuming.
GitClear's "The Maintainability Gap: 2026 AI Code Quality Research," published January 2026 and covering 623 million changes from 2023 to 2026, measured the shape of the problem:
- Block duplication rose 81 percent, from 40.3 to 73.0 duplicated blocks per million changed lines.
- Error-masking constructs rose 47 percent, and two-week code churn rose 15 percent.
- Refactoring fell to 3.8 percent of changed lines in 2026 year to date, from 21 percent in 2022.
The report's summary: "The throughput is real, but so is the debt it accrues." Neither source produces a six-month number. Both show that the "just kind of clean up any issues" step is the part developers report as expensive.
Does telling an AI it is a senior engineer change the output?
Not reliably, in either direction. This is the one place where the sketch's mechanism does not match the research.
Mingqian Zheng, Jiaxin Pei, Lajanugen Logeswaran, Moontae Lee and David Jurgens published "When 'A Helpful Assistant' Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models" (arXiv:2311.10054) on 16 November 2023. They tested 162 roles across 4 model families and 2,410 factual questions and found that "adding personas in system prompts does not improve model performance across a range of questions compared to the control setting where no persona is added." They concluded that "the effect of each persona can be largely random."
The sketch assumes the persona pushes output toward over-engineering. The paper finds it is not a dependable lever at all, though it measured factual accuracy rather than code architecture and predates current agentic coding tools.
What happens when teams ship AI code without review?
The 2025 DORA State of AI-assisted Software Development report, published 23 September 2025, surveyed nearly 5,000 technology professionals. It found that "90% of survey respondents report using AI at work," and that "30% report little or no trust in the code generated by AI."
It also split speed from stability. AI adoption now shows "a positive relationship" with delivery throughput, "However, AI adoption does continue to have a negative relationship with software delivery stability." The report's explanation: "Without robust control systems, like strong automated testing, mature version control practices, and fast feedback loops, an increase in change volume leads to instability." Its summary line is "AI doesn't fix a team; it amplifies what's already there."
METR found no speedup on individual tasks, while DORA found faster throughput at the organization level. Both support the engineer's last line, since the extra speed shows up as instability when nobody reviews the output.
Bottom line
The four-, six- and twelve-month figures are comedy, not data. The randomized evidence backs "still four months" for experienced developers, and DORA backs "the app will be broken" for unreviewed output. The persona joke is the weakest part, because research finds persona prompts have a largely random effect rather than a predictable one.
Sources
- Original TikTok from @alberta.nyc - the sketch.
- Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, arXiv:2507.09089 - the METR randomized trial.
- Personas in System Prompts Do Not Improve Performances of Large Language Models, arXiv:2311.10054 - the persona study.
- How Large Are Software Cost Overruns? Jørgensen and Moløkken - the review of estimate overruns and the CHAOS report.
- 2025 DORA report announcement, Google Cloud - adoption, trust, throughput and stability findings.
- Stack Overflow Developer Survey 2025, AI section - frustration and trust figures.
- The Maintainability Gap, GitClear - 2026 code quality research.
- Alberta Devor, Wikipedia - the creator's name and accounts.
Published October 12, 2026. Based on a TikTok by @alberta.nyc. Full video notes (markdown)