Digital Technology Risks Quietly Undermine China’s Breakthrough Innovation Efforts

Chinese listed companies that face greater exposure to digital technology risks file fewer patents in key core technology fields, according to a new study of Chinese A-share listed firms published in the journal Digital Economy and Sustainable Development. The research, which used a Chinese domestic large language model to measure risk exposure from annual report texts, found that the suppressive effect was strongest among profit-oriented firms, firms with weak industry reputation, and non-state-owned enterprises, while mature digital risk governance capabilities were associated with more key core technology innovation.

The study, led by Liangliang Han, Mingyu Liu, Hong Pan, Ge Zhang and Kangkang Wang, defines digital technology risk exposure (DTRE) as the extent to which a firm faces identifiable, material and potentially consequential digital-technology risks that demand immediate managerial response and governance optimization. The concept has two sub-dimensions: data risk exposure and cyber-security risk exposure. The authors distinguish it from the broader notion of uncertainty, drawing on the Knightian risk concept of identifiable tangible hazards with estimable probabilities, rather than ambiguous future states whose outcome probabilities cannot be estimated.

Measuring this exposure has long been a challenge. Traditional text analysis based on word-frequency statistics can only identify the disclosure of risk keywords in annual report texts, which the authors argue is difficult to reconcile with firms’ actual risk situations. To address this, the team followed a workflow that combines a large language model with a financial language model. They extracted candidate sentences associated with digital technology risks from the Management Discussion and Analysis sections of annual reports using a risk-keyword dictionary, yielding 227,200 candidate sentences. A random 10 percent sample by year, 22,720 sentences, was annotated with DeepSeek-V3.1-Terminus, producing 11,102 valid annotated sentences. These were used to fine-tune a FinBERT model, trained on 8,883 sentences and validated on 2,219 in an 8:2 stratified split. The fine-tuned model then scored all candidate sentences for negative or positive sentiment probability, and the scores were aggregated to the firm-year level. The DTRE indicator was calculated as the maximum negative sentiment probability among risk-exposure sentences minus the average positive sentiment probability among risk-prevention sentences, with boundary rules setting negative differences to zero.

The dependent variable, key core technology innovation (KCTI), was built from a catalogue of 1,047 key core technology keywords released by the National Manufacturing Strategy Advisory Committees. The authors matched these keywords with patent descriptions at the fifth-level codes of the International Patent Classification to identify patent applications in key core technology fields, then summed the annual number of such applications for each listed firm. The sample covered Chinese A-share listed firms from 2012 to 2024, excluding firms labeled ST and *ST, financial industry companies and observations with missing data, with all continuous variables winsorized at the 1st and 99th percentiles.

In the baseline regression, estimated with a Poisson Pseudo-Maximum Likelihood model that included firm and year fixed effects along with control variables such as firm size, leverage, revenue growth, return on assets, board independence, largest-shareholder ownership, Tobin’s Q and firm age, the coefficient on DTRE was -0.150 at the 5 percent significance level when controls were included, and -0.141 at the 5 percent level without them. Descriptive statistics showed a mean KCTI of 6.217 with a standard deviation of 17.203 and a range from 0 to 113, which the authors interpret as relatively low and uneven key core technology innovation among Chinese listed companies. The DTRE indicator had a mean of 0.060, a median of 0, a standard deviation of 0.231 and a maximum of 0.99954.

The finding survived a battery of robustness tests. A propensity score matching procedure, which split firms at the median of DTRE into treatment and control groups and used 1:1 nearest-neighbor matching, produced a coefficient of -0.094, significant at the 5 percent level. Alternative constructions of the DTRE indicator, including one using raw gap values without boundary truncation and one retaining only the maximum risk-exposure sentiment probability, preserved the negative sign and significance. Regression on the sub-sample of firms with non-zero exposure yielded a coefficient significant at the 1 percent level. Decomposing DTRE into its two dimensions showed that data risk exposure carried a coefficient of -0.332, significant at the 1 percent level, while the coefficient for cyber-security risk exposure, -0.045, was not statistically significant. Lagging DTRE by one year, excluding the 2020 to 2022 pandemic years, switching to an OLS model with a logarithmized dependent variable, and replacing the KCTI measure with an entropy-based breakthrough indicator all left the main conclusion intact.

The authors then probed two mechanisms. For the information-effect channel, they used the KV index, in which a larger value indicates poorer disclosure quality and more serious information asymmetry. DTRE had a coefficient of 0.015 on KV, significant at the 1 percent level, and a bootstrap test with a 95 percent confidence interval excluding zero supported the mediation. The theoretical account, grounded in threat-rigidity theory, holds that digital technology risks disrupt information flows within and between firms, tighten financing constraints, slow decision-making coordination and raise the proprietary costs of transparent disclosure, which weigh especially heavily on key core technology research with its long cycles, heavy resource commitments and high confidentiality demands.

The second mechanism concerned collaboration. The authors measured industry-research integration by the number of scientific papers published by listed firms in collaboration with universities and research institutes. DTRE had a coefficient of -0.028 on this measure, significant at the 5 percent level, with the bootstrap test again supporting the transmission. According to the authors’ reasoning, potential data security, operational stability and compliance risks raise the expected losses and risk perceptions of academic partners, while internal risk-control routines constrain the openness of technical resources and data interfaces, reducing the output of high-quality co-authored papers and joint research projects.

Heterogeneity analyses sharpened the picture. The inhibitory effect of DTRE was significant only among profit-oriented firms and not among growth-oriented firms, only among firms without solid industry reputation as classified by a median split of a multi-dimensional reputation index, and only among non-state-owned enterprises. The authors attribute these patterns to threat-rigidity responses: profit-oriented firms prioritize short-term performance stability, low-reputation firms lack institutional legitimacy and collaborative buffers, and non-state-owned firms face stricter market-based performance discipline than state-owned firms, which benefit from policy support and resource cushions that sustain research inputs.

Finally, the team exploited the staggered rollout of Data Management Capability Maturity Assessment Model (DCMM) certifications in China since July 2020 as a quasi-natural experiment. Using city-level certification status, assigned through geo-coding and reverse geo-coding, to construct treatment and control groups of listed firms, they found that certification was associated with significantly lower DTRE indicators, and a Callaway-Sant’Anna difference-in-differences estimation showed that the governance capability reflected by DCMM certification significantly promoted key core technology innovation. The authors conclude that effective digital technology governance can counteract the innovation-suppressing effects of digital risk exposure, and they suggest tiered incentives for vulnerable firms and shared regional data-security early-warning infrastructure, while cautioning that their evidence concerns Chinese A-share listed firms from 2012 to 2024 and the specific institutional setting of China’s digital economy.

Subject of Research: The relationship between corporate digital technology risk exposure and key core technology innovation among Chinese listed firms.

Article Title: Digital technology risk exposure and corporate key core technology innovation

Article References: Han, L., Liu, M., Pan, H., Zhang, G., & Wang, K. (2026). Digital technology risk exposure and corporate key core technology innovation. Digital Economy and Sustainable Development, 4(1), Article 12. https://doi.org/10.1007/s44265-026-00088-7

Image Credits: AI Generated

DOI: 10.1007/s44265-026-00088-7

Keywords: digital technology risk exposure, key core technology innovation, Chinese A-share listed firms, DeepSeek, FinBERT, annual report text analysis, threat-rigidity theory, information asymmetry, industry-research collaboration, DCMM certification, patent applications, difference-in-differences