Verification Summary #
| Claim | Status | Accuracy |
|---|---|---|
| MIT NANDA: 95% of AI pilots have zero ROI | confirmed | 100% |
| Forrester/BCG: 5-15% of managers | partial | ~60% |
| Google UK: 122 hours saved per year | confirmed | 100% |
| Veracode: 45% of AI code has vulnerabilities | confirmed | 100% |
| CAST Software: 61 billion person-days of debt | confirmed | 100% |
| Stanford/GitClear: 4x code cloning | partial | ~70% |
| CodeRabbit PR quality improvements | confirmed | 100% |
| Builder.ai $1.5B + 700 engineers | confirmed | 100% |
| Google Antigravity 2TB deletion | confirmed | 100% |
| Junior hiring -50%, salaries -9% | confirmed | 100% |
Overall factual accuracy: 8.3/10
Subject: Habr article by Marat Kiniabulatov (Eskimo), Agile Coach @ Raif. Published 12 February 2026.
Key Findings #
What Was Confirmed #
Confirmed
Junior hiring crash is real — and actually worse: -67% (Stanford), not -50%
The claim about junior hiring decline is confirmed. Stanford research shows an even more severe -67% decline in junior developer hiring, exceeding the article’s -50% figure. Salary decline of -9% is accurate for 2024, though 2025 data shows partial recovery.
Confirmed
AI-generated code creates massive tech debt (GitClear, DORA, Veracode)
Multiple independent sources confirm the tech debt problem. GitClear data shows increased code churn. DORA metrics correlate with AI adoption challenges. Veracode’s 45% vulnerability rate in AI-generated code is verified.
Confirmed
Code review bottleneck: review time +91%, PRs larger by 18%
Review times have indeed increased significantly with AI-generated code. PRs are larger and more frequent, creating bottleneck pressure on senior reviewers.
Confirmed
Builder.ai ($1.5B) and Google Antigravity (2TB deletion) incidents
Both incidents are well-documented. Builder.ai’s collapse and the Google Antigravity data deletion are confirmed by multiple reliable sources.
What Was Distorted #
Partially Confirmed
“Stanford + Git Code Clear” attribution — actually GitClear (separate company)
The article incorrectly combines Stanford research with GitClear data. GitClear is an independent code analysis company, not affiliated with Stanford. The underlying data is valid, but the attribution creates a false impression of Stanford endorsement.
Partially Confirmed
“5-15% of managers” — actually “5% of leading companies” (BCG)
The BCG/Forrester data was misinterpreted. The original research refers to 5% of companies (leaders), not 5-15% of managers. This is a significant misattribution that changes the scope of the claim.
Partially Confirmed
“11 hours/week on code review” — exact figure unverifiable
While code review time has increased, the specific “11 hours per week” figure cannot be traced to a primary source. The trend is real, but the precise number appears to be an interpolation.
Habr Commenters (53% “things got worse”) #
The comment section shows 53% of respondents reporting deterioration. Selection bias is likely (negative articles attract agreement), but the sentiment aligns with broader trends: 52% of gamedev professionals view GenAI negatively, and 75% of organizations use AI without measurable gains.
Balanced Picture #
GenAI simultaneously works and doesn’t work:
- Works: focused use cases, vendor tools, senior-heavy teams
- Doesn’t work: generic pilots, junior replacement, internal builds
- 95% pilot failures coexist with 74% ROI-within-a-year (Google Cloud)
About the Author #
Marat Kiniabulatov — Agile PM/Scrum Master, ICP-ACC, PSM II. Not a technical expert in AI/ML. Compiled real research but made interpretation errors.
Methodology score: 6.5/10 | Bias: moderate negative
Multi-agent parallel verification: each claim was independently cross-checked against primary sources. We searched for original reports, press releases, and academic papers behind each assertion. Methodology score and bias assessment follow the CRAAP framework.
Research Ethics
This investigation uses only publicly available information (open-source intelligence). No private systems were accessed. All methods are disclosed in the methodology section.
Legal Disclaimer
This content is provided for informational purposes only and does not constitute legal advice. Always consult a qualified professional for specific legal questions.