Does AI help manufacturers make greener products? A 333-firm survey finds an association—not emissions proof
Managers at 333 Chinese manufacturers who were already using AI and pursuing green product development reported stronger innovation capabilities where AI capability was higher. The cross-sectional survey measured perceptions, not products’ environmental performance or causal impact.
The short version
- What happened
- The usable sample was 333 firms across five manufacturing sectors. Seven hundred questionnaires were distributed, 612 returned and 333 retained, equal to 47.6% of the distributed questionnaires.
- Who it affects
- Product and engineering teams need outcome-specific tests, not a generic claim that AI makes innovation greener.
- What remains unproven
- The survey was cross-sectional, so temporal order and causality cannot be established.
Keep in mind
The survey was cross-sectional, so temporal order and causality cannot be established.
Firms were purposively selected because they already used AI and pursued green product development.
Only 333 of 700 distributed questionnaires were retained, and one manager reported all focal variables for each firm.
The study measured perceived capability and innovation, not lifecycle emissions, energy, waste or independently verified products.
Findings from five Chinese manufacturing sectors may not transfer to other industries, countries or small non-adopters.
Source analysis · 1 linked source. Primary analysed source date: 7 November 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The answer: firms reporting stronger AI capability also reported stronger green product development
A peer-reviewed survey of 333 Chinese manufacturers found that managers who rated their firm's AI-powered capabilities more highly also reported stronger green new product development. The authors modelled two organisational pathways: exploration, meaning search and experimentation with new green knowledge, and exploitation, meaning refinement and reuse of existing green knowledge. Both statistically mediated the association between AI capability and the reported product-development outcome.
The estimated indirect effect through exploration was 0.234, with a bootstrapped 95% confidence interval from 0.139 to 0.354. The indirect effect through exploitation was 0.167, with an interval from 0.093 to 0.258. Although the exploration estimate was numerically larger, the contrast between pathways was 0.067 with an interval from -0.090 to 0.242, crossing zero. The data therefore do not support a confident claim that one pathway dominates the other.
This is evidence of an organisational association, not proof that buying AI makes a product greener. The design is cross-sectional, every focal variable was reported by one manager, and the outcome is a perceptual scale. Researchers did not inspect product specifications, lifecycle carbon, waste, pollution, energy use or market performance. A firm already committed to green innovation may invest in AI, and effective management may raise both ratings.[1]
Who was surveyed—and who was not
The researchers targeted Chinese manufacturing firms already engaged in green new product development and already using AI for low-carbon, green-decision or sustainability-related activity. That eligibility criterion makes the sample relevant to the proposed mechanism, but it excludes firms with no AI capability or no green programme. It cannot estimate how common these practices are across Chinese manufacturing or whether adoption would help a typical non-user.
Seven hundred questionnaires were distributed and 612 were returned, an 87.4% return rate. After screening, 333 were retained: 54.4% of the returned questionnaires and 47.6% of all distributed. The final sample comprised 77 electronics or semiconductor firms, 73 automotive firms, 67 machinery firms, 54 pharmaceutical or medical-device firms and 62 home-appliance or consumer-electronics firms.
One adult enterprise manager completed an anonymous, voluntary questionnaire for each firm. A single informed respondent can describe company practice, but cannot independently validate organisational outcomes. Firms willing and able to participate may differ from others, and the large reduction from returned to usable forms raises questions about how eligibility and quality screens shaped the sample. The paper reports the sectors clearly; readers should keep the purposive sampling frame visible.[1]
The distinction that matters
What the survey measured
The pathway is based on managers' reports. Environmental outcomes sit outside the measured chain.
Selected sample
333 firms already using AI and pursuing green products
Not a population estimate
Reported capability
One manager rates AI resources
Associated with two pathways
Reported process
Exploration and exploitation scales
Both indirect effects above zero
Not measured
No lifecycle or emissions audit
Environmental benefit remains unproven
What the variables mean in practice
AI-powered capability was treated as the firm's ability to mobilise AI-related resources and analysis. Green exploration captured searching for unfamiliar knowledge, experimenting and pursuing new environmental solutions. Green exploitation captured refining known approaches and using existing experience more efficiently. Green new product development was also measured through survey items rather than an audit of individual products.
The authors used confirmatory factor analysis, hierarchical regressions and PROCESS Model 4 with 5,000 bootstrap resamples. Controls included industry, firm age, size, AI investment and ISO 14001 status. The full model explained 54% of variance in the reported green-product-development measure. That is a substantial within-sample association, but explained variance in a shared survey does not establish environmental impact or the direction of causality.
Because predictor, mediators and outcome came from the same questionnaire at the same time, common-method bias can strengthen relationships. Statistical checks can detect some problems but cannot replace independent outcomes or time order. A longitudinal design would measure AI capability first, observe changes in exploration and exploitation, and then verify product results. A controlled comparison is difficult in firms, but matched panels and objective records would move the evidence closer to a causal claim.[1]
What managers can reasonably take from the study
The useful management implication is not 'install a model and products become sustainable'. It is that AI may be more valuable when embedded in both search and execution. Exploration could involve screening materials, analysing design alternatives or identifying unfamiliar technologies. Exploitation could involve finding repeated defects, optimising known processes or reusing validated knowledge. Each requires people, reliable data and authority to change product decisions.
A business case should define an environmental outcome before deploying a tool. Examples include verified material intensity, energy consumption during use, repairability, hazardous substances, waste, supplier emissions or lifecycle carbon. Teams should compare the AI-supported process with the previous workflow, record cost and error, and test for rebound effects. A faster design cycle that creates more short-lived products is not automatically greener.
Workers need a clear account of how recommendations affect engineering, procurement and quality roles. If a model ranks options, staff should be able to inspect inputs, challenge unsafe suggestions and record why a design changed. Suppliers may be asked for new data, creating costs and confidentiality questions. The survey does not measure workforce effects, implementation failures or whether smaller manufacturers can afford the data infrastructure assumed by the capability scale.[1]
Why the environmental claim remains unproven
The phrase 'green new product development' can describe an organisational process without demonstrating a net environmental improvement. A product may use less energy but require scarce materials; a recyclable component may raise manufacturing emissions; an efficiency gain may increase total consumption. Lifecycle boundaries and counterfactuals determine whether a change is genuinely beneficial. None was measured here.
Nor did the study compare AI's own energy, computing or data costs with any environmental savings. Those costs need not outweigh benefits, but they belong in the denominator when firms make sustainability claims. Marketing language should not convert a survey association into evidence of lower emissions. Verification requires product-level measurements, consistent baselines and disclosure of material trade-offs.
The result is geographically and institutionally specific. China's manufacturing scale, policy environment, supply chains and digital infrastructure may shape both AI investment and green innovation. The five sectors are economically important, yet the sample does not cover every industry or region. Replication in other countries should preserve sector and firm-size detail rather than reporting a single global average.[1]
What would change the assessment
Confidence would rise with multi-respondent surveys separating the people who report AI capability, innovation processes and outcomes. Longitudinal data could establish whether capability precedes change. Administrative records, patents, product launches and independently audited environmental indicators would reduce reliance on perception. Researchers should publish attrition reasons and compare retained firms with excluded respondents.
The most useful test would link a defined AI intervention to product-level lifecycle outcomes and a credible comparator. A stepped rollout, matched firms or difference-in-differences design could examine whether outcomes change more where the capability is introduced. Pre-specified measures should include failures and costs as well as successful projects, with funding and vendor relationships disclosed.
For now, the paper supports a bounded conclusion: among selected Chinese manufacturers already pursuing both AI and green product development, reported AI capability was associated with reported innovation through both exploratory and exploitative pathways. It gives managers a vocabulary for organisational design. It does not prove causality, show that exploration is superior, or establish that any product reduced environmental harm.[1]