Evidence, not impressions

What the research actually shows.

The findings are strong in some areas, contested in others, and absent in a few. Each result below keeps its source tier and its most important limitation.

Productivity

Productivity depends on context.

AI consistently increases output volume. Whether it shortens time to a correct outcome depends on task familiarity, codebase maturity, and downstream verification.

P

Peng et al. - controlled greenfield task

95 freelance developers built a JavaScript HTTP server. The AI group completed it in 71.2 minutes versus 160.9 minutes.

55.8% faster

Single bounded task; vendor-affiliated authors; wide confidence interval.

P

Cui et al. - enterprise field trials

Three randomized trials at Microsoft, Accenture, and a Fortune 100 manufacturer followed 4,867 developers over two to eight months.

+26.08% tasks

Measured throughput; researchers had no access to produced code quality.

PP

METR - experienced maintainers

16 experienced open-source maintainers completed 246 tasks in repositories they had worked in for an average of five years.

19% slower

Small, early-2025 snapshot. METR's February 2026 follow-up was judged an unreliable signal, because developers increasingly declined to work without AI. Its raw estimates point the other way, toward a speedup, and METR now believes developers are likely faster in 2026. Treat 19% as a finding about mature codebases in early 2025, not a current measure.

Security

Security is the clearest failure.

Different methods - generated scenarios, real GitHub snippets, exploit-based benchmarks, and model trend tests - converge on the same conclusion.

P

Pearce et al. - Copilot security scenarios

1,689 programs generated across 89 security-relevant scenarios.

~40% vulnerable

Early Copilot generation, but later research has not displaced the central result.

P

Fu et al. - snippets in real GitHub projects

733 AI-generated snippets were analyzed across 43 CWE categories.

29.5% Python

JavaScript measured 24.2%; snippet-level analysis cannot capture every architectural flaw.

P

BaxBench - real backend exploits

392 backend tasks across 14 frameworks tested whether generated implementations were correct and exploitable.

>60% failed

Incorrect or insecure; approximately half of correct solutions were still exploitable.

I

Veracode - three-year model trend

More than 150 models tested from 2023 through Spring 2026.

45-55% secure

Industry research, though the method and longitudinal comparison are published.

PP

Deng et al. - vibe-coded applications

10,517 confirmed applications identified; 200 deployed web apps audited with agents, exploit probes, and two human experts.

90% affected

Major 2026 preprint, not yet peer-reviewed; unusually rigorous audit design.

Human factors

Confidence can rise as security falls.

The human problem is not just trust. AI changes the work from producing to verifying - and verification requires skills that delegation can weaken.

Comprehension50% vs 67%

AI-assisted learners versus control in Anthropic's Trio experiment.

Best interaction86%

for generation followed by active interrogation of the code.

Worst interaction24%

for repeatedly asking AI to solve the errors it introduced.

Developer frustration66%

cite AI answers that are "almost right, but not quite."

Supply chain

Hallucinated packages become targets.

Package hallucination is repeatable. Attackers can register names that models reliably invent - a supply-chain technique called slopsquatting. The rate has fallen sharply on 2026 frontier models, but the attack surface has not closed.

USENIX 202519.7%

of recommended packages did not exist across 16 models tested in 2024.

Commercial models5.2%

still recommended nonexistent packages in that same study.

Repeatability43%

of hallucinated names recurred in all ten reruns.

2026 replication4.6-6.1%

across five frontier models, a large drop. But 53 names invented by all five remained registrable after disclosure.

More output can still produce more debt.

GitClear's observational analysis of 623 million changes reports less refactoring, more copy-and-paste, more short-term churn, and fewer cross-file calls. It is a strong directional signal - not causal proof - and should be read with that limitation intact.

Limits of the evidence

What we still do not know.

A credible account must show the gaps as clearly as the findings.

Well supported

  • Persistent security weakness
  • Access control dominance
  • Confidence-security mismatch
  • Stable package hallucination

Contested

  • Net productivity
  • Long-term skill atrophy
  • Value of larger models
  • Generalization across codebases

Under-evidenced

  • Runtime performance
  • Scalability of generated systems
  • Spec-driven security efficacy
  • Most vendor checklist claims