Research & Evidence Library

Do not generalize this bounded laboratory writing result to all complex, iterative professional workflows.

For tasks outside the capability frontier ("jagged frontier"), AI users were 19 percentage points less likely to produce correct solutions.

Do not imply that customer-support chat gains apply uniformly to independent knowledge workers or consultants.

Broader detectable work-pattern and overall productivity changes across enterprise metrics were statistically limited.

Completed tasks are not identical to delivery time or software quality. Individual company trials were noisy and varied, and the study was provider-affiliated.

Evaluated one constrained greenfield task with visible tests using an early Copilot version. This provider-affiliated laboratory result cannot be generalized to production repositories.

The early-2025 slowdown occurred in familiar mature codebases. METR's February 2026 update noted later signals pointing toward speedups, but reported that participant and task selection, a lower pay rate, and concurrent-agent time measurement made the magnitude unreliable—considering the later estimate a weak signal rather than a dependable current effect size.

Observational event study using internal Microsoft telemetry and public GitHub data; authors disclose past Microsoft Research affiliations. Shows that activity gains attenuate before production release.

Covers older tool generations and static-analysis criteria across mixed assistants; does not establish current Copilot vulnerability rates, but confirms why security and defect review belong in workflow costs.

Evaluated 2023 systems, not current product versions. Fluent prose with inline citations can still contain unsupported claims; verification remains an essential cost.

Observed in a bounded experimental comparison of earlier tools; does not prove a current subscription ROI or universal productivity effect.

Evaluated one visual ideation experiment; not a test of Canva, Canva Pro, or commercial production speed. Supports charging for human review rather than assuming more generations equal more value.

Controlled experiment with five condition-blinded evaluators drawn from the research group being advertised; exploratory subgroup findings. Not a Canva Pro field study; supports measuring approved, useful designs rather than novelty.

Short-story writing experiment, not visual design, Canva or subscription ROI. Shows that individual assistance can coincide with collective standardisation.

Studies were mostly from Indonesia and of moderate quality. Shows educational utility of Canva, but does not isolate Canva Pro, prove causality across all contexts, or justify paid subscriptions for educators eligible for Canva Education.

Provider-produced, model-estimated observational research. Anthropic explicitly notes this method cannot count human validation or accuracy checks outside the conversation and may overstate current productivity effects. It is not a Free-versus-Pro experiment and must not be used as a causal productivity multiplier.