Featured faculty: Cao Ying and Zhu Chunmei
Written by Sally Ho

The Net Zero Banking Alliance (NZBA) was short-lived. Founded in 2021, the alliance pledged to align its lending with net-zero emissions by 2050. Unfortunately, after major founders such as Bank of America, Citigroup, Morgan Stanley, HSBC, Barclays, and UBS pulled out, the alliance of more than 140 banks met its demise late last year.

As the union disbanded, former members reiterated their commitments to net-zero independently. Many see the breakup as a setback as the major banks stepped away from climate change initiatives, while a few saw it coming. The voluntary nature of the alliance may have created incentives for appearances rather than action and transparency.

“The net-zero commitments within the NZBA are largely self-implemented. The pressure to meet these targets can encourage the use of loan contracts that are less effective in working with borrowers to drive actual sustainability outcomes,” says Cao Ying, Associate Professor of Accounting at the Chinese University of Hong Kong (CUHK) Business School.

Professor Cao finds that the race to decarbonisation has made the alliance members more aggressive in issuing sustainability-linked loans (SLLs), where banks and borrowers agree on key performance indicators and targets for environmental, social, and governance (ESG). Borrowers will get a lower interest rate once the tailored achievements are met.

“Sustainability-linked loans can drive sustainability practices only if both lenders and borrowers see the targets as important as the loans.”

Professor Cao Ying

Unlike green loans that restrict proceeds to eco-friendly projects, SLLs provide funding with interest linked to broader environmental, social, or governance metrics. For example, a borrower can set social targets to reduce workplace accidents and improve employee health and safety, or aim for a governance target by increasing the percentage of female board members or top managers.

However, borrowers may also set indicators and targets later, known as “sleeping” SLLs. By agreeing to this provision, lenders allow borrowers to set up more workable metrics after the loan has started. Oddly, banks were more likely to issue sleeping rather than standard SLLs after joining the NZBA.

In a new study, Lender effect on SLLs — Evidence from sleeping SLLs, Professor Cao finds that sleeping SLLs have surged since the NZBA was established, peaked in 2022, and declined the following year, after the global trade associations ruled to treat sleeping SLLs as conventional loans until the sustainability metrics are verified.

Addressing the sleeping elephant in the room

Along with Zhu Chunmei, Assistant Professor of Accounting at the School, Lin Yating, PhD student at the School, and Christopher Williams of the University of Michigan, Professor Cao discovers that sleeping SLLs accounted for more than half of 403 SLLs issued to US public companies in 2017-2023. These sleeping loans were issued by 29 US lenders, dominated by Bank of America, followed by JPMorgan and Wells Fargo.

Regardless, only less than eight per cent of these sleeping SLLs were “awakened”. Borrowers with sleeping SLLs also had smaller emission reductions from their own operations and the energy they purchase, and showed less improvement in ESG ratings than those with non-sleeping SLLs.

Practicality could be the reason. Companies may need better data, internal approval, or an external reviewer before they can agree on credible yet achievable goals, but further analyses refute these excuses. Borrowers continued to use sleeping SLLs after more state governments announced Greenhouse Gas Emissions Reduction Targets and issued Climate Action Plans.

These borrowers also kept signing sleeping SLLs even after disclosing their environmental data to the global non-profit Carbon Disclosure Project, or setting net-zero targets through the Science Based Targets initiative, signalling that such tardiness may be strategic instead of a lack of capacity. Concerns then arise over whether such loans can truly improve ESG performance when targets have not yet been set.

“Sleeping SLLs are essentially conventional loans with extra sustainability labels, but that does not necessarily mean they are being deceptive,” says Professor Cao. “As outsiders, we are not fully aware of the reasons that lead to the signing of sleeping SLLs. The bank and the borrower might be willing to engage in ESG activities, but the borrower was not at a mature stage yet to write a true SLL contract.”

When the lender effect takes place

Lenders are central to the story. The lender effect was particularly strong among US banks that started or substantially expanded their sustainable-financing commitments. Banks headquartered in areas with lower climate-change awareness were also more likely to issue sleeping SLLs, while banks facing greater climate-related litigation risk were less likely to do so.

SLLs can significantly help banks engage companies in achieving sustainability targets, and sleeping SLLs can help companies that need funding but are struggling to finalise well-designed metrics. However, pressure to meet headline commitments may encourage banks to write loans with sustainability labels without meaningful provisions.

“The good news is US regulators are getting aware of this and have released some clarification memos to require companies to be more careful and rigorous when writing loan contracts,” says Professor Cao.

Calling for strict, measurable metrics

While the study focuses on US companies, Professor Cao sees its implications extend beyond borders. A strategic research arm of Bloomberg reported last year that nearly half of all financing for low-carbon energy companies in the Middle East and North Africa and emerging Asia Pacific came from SLLs.

Being aware of the debate on sleeping SLLs, the Asia Pacific Loan Market Association, a regional trade association for the syndicated loan markets, also maintains that sustainability targets should be set at or before the loan contract is signed.

“SLLs can drive sustainability practices only if both lenders and borrowers see the targets as important as the loans,” Professor Cao says. “The metrics selected should reflect the borrowers’ core business model and fundamental economic factors.”

For example, manufacturing companies will benefit the most from linking their SLLs to carbon footprint metrics if they sell products to markets with stringent ESG rules like the EU, as it directly affects costs and profits.

Finally, this study recommends a cautious approach towards voluntary sustainability initiatives in the capital markets. Financial institutions and corporate enterprises must meaningfully align ESG frameworks with their core corporate strategies. Only through this strategic integration can SLLs drive true sustainable development.

The post How sustainability loans become unsustainable first appeared on China Business Knowledge.

We’re running a brief, anonymous survey to explore the topics you care about most, the trends you expect will shape the future business landscape, and how we can tailor our coverage to better support your needs.

Your responses are truly important to us and will be kept strictly confidential. Please complete the 3-minute survey by 30 October 2026.

Take the survey

If you have specific thoughts or ideas for our editorial team, please feel free to reply directly or email us at cbk@baf.cuhk.edu.hk.

Thank you for your time and continued support!

Sincerely,

China Business Knowledge, CUHK Business School

The post Tell us what matters to you first appeared on China Business Knowledge.

Featured faculty: Tony Ke
Written by Putro Harnowo

Appearance can be deceiving, and this idiom deeply resonates in modern times. Not too long ago, a French woman lost her life savings to an impostor posing as actor Brad Pitt on Instagram, while YouTube star MrBeast called out a fake TikTok video of him offering a new iPhone for just US$2 each. These cases are just the tip of the iceberg.

AI can generate anything as easily as typing a sentence, fueling fraudsters to produce deceptive content at a remarkable scale. More and more people fall victim to AI forgeries, often exploiting public figures to trick netizens. As a result, the EU has required digital platforms and content creators to label AI-generated content starting this August.

In Asia, China was the first to mandate AI labels last year, followed by South Korea and Vietnam earlier this year, while India is pondering the same. Overall, these AI laws require companies to state clearly when publishing AI-generated content, but some critics question their effectiveness in tackling disinformation and misinformation.

Tony Ke, Professor of Marketing at the Chinese University of Hong Kong (CUHK) Business School, points out two distinct ways in which AI labels could lose their power. “First, consumer attention towards AI labels may decline after repeated exposure through psychological habituation, but changing the design or placement of AI labels can refresh such attention,” he says.

“Secondly, an AI label becomes helpful when it makes people aware that content made with AI might be fake and deceptive. However, if AI tools become so easy and affordable that both truthful and deceptive creators use them frequently, the AI label becomes merely a descriptive label.”

Therefore, the more accurately AI labels identify manipulative content, the more likely people are to associate AI labels with a risk of deception, and that is the central focus of Professor Ke’s latest study.

Deceptive content typically aims to appear as genuine as possible, so Facebook, Instagram, TikTok, and LinkedIn have deployed algorithms to detect visual patterns, statistical patterns, and metadata to score the likelihood of content being made with AI.

These algorithms are imperfect. Reports show that AI detectors fail to distinguish AI-generated output. Human-made content may also have characteristics like AI output, so if the platform sets the detection algorithm too aggressively, the algorithms may mislabel human-made content. Conversely, making the algorithm too conservative may risk missing the real AI content.

Optimising AI labelling to tackle deception

To find the optimal balance, Professor Ke partners with Chen Jieteng of Shanghai Jiao Tong University and Shin Jiwoong of Yale University. Their study, Labelling AI-generated content on platforms: misinformation, consumer inference, and creator incentives, develops a theoretical model of strategic interaction among three key players: the platform, users, and content creators.

Deceptive content creators have stronger incentives to adopt AI to generate misinformation because they are not constrained by factual accuracy, and such misinformation can be particularly effective at attracting user attention and engagement. Consequently, users may interpret AI labels on content as a warning sign of potential deception.

Therefore, mandating salient AI labels on content may discourage user engagement and deter deceptive creators from using AI-generated content. This explains why labelling matters in real-world settings and why such labels can discipline AI misuse without arduous manual fact-checking.

“Transparency can protect a brand when it communicates governance and responsibility. A bare AI label, without context or accountability, can instead transfer the general scepticism surrounding AI directly onto the brand.”

Professor Tony Ke

Aggressive labelling algorithms tend to flag more content as AI-generated, which strengthens users’ trust in unlabeled content but may discourage truthful creators, as some of their content may be mistakenly labelled as AI-generated and consequently receive less user engagement. If the labelling policy becomes more lenient, more AI-generated content goes unlabeled, which may increase user engagement. However, more deceptive content may also slip through, causing harm to users and, in extreme cases, undermining their trust even in unlabelled content.

When AI tools become so easy and cheap for anyone, AI content becomes abundant, and platforms need to apply more aggressive labelling to maintain credibility. Only when algorithms can accurately spot deceptively real content can platforms adopt more lenient labelling that maintains users’ trust. However, near-perfect detection may be impossible.

“Deceptive creators nowadays can conceal metadata, use third-party tools, or tweak content to evade AI detection. Therefore, self-disclosure, algorithmic detection, and digital markers should be used together to combat deceptive content,” Professor Ke says. “The main deterrent comes from users’ association between AI labels and the risk of manipulation. Mandatory labels will discourage deceptive AI use when they are credible and consequential, but labels that are easy to evade are unlikely to have an effect.”

Using AI content while keeping brand reputation

In general, consumers may perceive AI content as unreliable or low effort, even when the underlying message is truthful. While consumers do not intrinsically dislike AI content, scepticism emerges because deceptive creators have stronger incentives to use it. Reputable brands can be caught in a negative association.

Mandatory labelling can deter companies from deploying AI content, but Professor Ke notes that such content can still be useful if the benefits outweigh the risks. “There is no universally optimal ratio to balance the labelling policy. For instance, the consequences of posting AI-generated parody are not equivalent to sharing fabricated medical endorsement,” he says.

“For marketers, the risk depends highly on the context. Consumer scepticism is likely to be damaging where authenticity is central to the product, such as luxury craftsmanship, personal services, healthcare, finance, and the like, or when using AI to create testimonials and product demonstrations.”

Businesses can further protect their brands by explaining the role AI played. Putting a disclaimer that says, “AI-generated content, reviewed and approved by our editorial team,” while also clearly displaying an AI label can be a fine example. “To maintain accountability, a named editor, designer, expert or brand team should remain visible for factual claims,” says Professor Ke.

Brands can substantiate marketing claims with accessible sources, original product images, and other factual records, rather than just visually appealing content. Brands should also consistently disclose when they use AI, rather than admitting it only after consumers point it out.

“The most defensible message is using AI for a defined purpose, having humans verify the result, and the company remaining accountable for every claim. That distinction will become increasingly important as AI use becomes the norm,” he adds.

“Transparency can protect a brand when it communicates governance and responsibility. A bare AI label, without context or accountability, can instead transfer the general scepticism surrounding AI directly onto the brand.”

The post Can AI labels restore consumer trust? first appeared on China Business Knowledge.

Featured faculty: Bella Kang Ya
Written by Joanne Madrid

Bank failures are surprisingly not that rare. Forbes recently reported that, on average, around 22 US commercial banks have failed annually since 2000. Most of them are small banks, including six so far this year. Banking rests on a simple bargain. Depositors trust banks to keep their money safe, and banks use that money to fund their lending. Devastating news, whether about a disappointing earnings report, bad investments, asset losses, or a nudge from regulators, could break that trust.

Once it breaks, depositors no longer feel secure and move their money out of the troubled bank, but where would they go? Bella Kang Ya, Assistant Professor in the School of Accountancy at the Chinese University of Hong Kong (CUHK) Business School, finds the choice depends on whether people see the scale of the problem.

“Healthy banks come out ahead by gaining deposit inflows and effectively absorbing market share from the troubled bank,” says Professor Kang. “However, during economic instability or a financial crisis, even healthier local banks also see customers withdrawing their money, and this can trigger a worsening spiral for all banks.”

Depositors normally view misconduct as specific to the troubled bank and move their savings to safer banks. But when they see news of multiple banks struggling and major economic indicators falling, all banks are perceived as risky as a broader panic sets in.

Follow the money

In collaboration with Lin Yupeng of the National University of Singapore and Qiu Yang of Zhejiang University, Professor Kang’s study, A tale of two market disciplines: How does bank financial misconduct affect peer banks in the local deposit market, examines the Consolidated Reports of Condition and Income of all US commercial banks from 2007 to 2021.

The team identifies 3,727 misconduct cases committed by 2,858 banks in which regulators forced banks to correct previously reported figures, peaking in 2007–2008 and rising again in 2020.

They then track the interest rates individual banks offered on large deposits above US$250,000. Deposits beyond this amount are most likely to run, since they are not fully insured by the government, and normally belong to financially sophisticated entities and institutional investors.

Banks caught in misconduct quickly lost deposits and offered higher interest rates to attract depositors. However, the nearby rival banks attracted more customers, even after offering lower interest rates on large deposits, around 25 per cent less. In the following year, deposits at these rival local banks grew faster than at other branches located far from the troubled bank.

A clear pattern appears. If a bank has to offer much higher interest rates than other banks just to attract depositors, it typically means people are withdrawing their deposits. On the other hand, if a bank can offer lower rates and still attract many new depositors, it means more money is flowing in.

The researchers call this “reallocation effect,” and it serves as market discipline. As Professor Kang puts it, “the offending bank struggles to get funding, must pay more interest to depositors, and may lose market share in the long run. This penalty ultimately benefits its competitors.”

How depositors spot better banks

There is direct evidence of large depositors doing their homework. After examining company filings on the online database of the US securities regulator, Professor Kang finds that areas with increased searches about the troubled bank saw deposits move the most.

These deposits moved to rival banks with stronger capital and healthier balance sheets, and they were especially active in areas with higher levels of education and financial literacy, as those who can read a financial statement act first.

“The offending bank struggles to get funding, must pay more interest to depositors, and may lose market share in the long run. This penalty ultimately benefits its competitors.”

Professor Bella Kang Ya

“Financially sophisticated depositors are more responsive and drive stronger reallocation effects, since they have the incentive and capacity to monitor bank health,” she adds.

Financially savvy depositors also know how to tell a bookkeeping error from a serious financial problem. When a bank reported committing misconduct but is otherwise financially strong, depositors are less inclined to withdraw their funds.

When discipline turns into panic

The team rerun the same analysis for 2007 and 2008, when the financial crisis struck, and the story changes. Instead of moving next door, most banks in affected areas lost deposits and had to pay about 8.7 per cent more than the usual interest rates to attract depositors. The researchers call this “contagion effect,” and the danger is that it feeds on itself.

“During a crisis, banks face deposit outflows and rising deposit spreads simultaneously, meaning they must pay more for a shrinking customer base, precisely when liquidity is most valuable,” says Professor Kang. “If sustained, this issue can impair lending capacity, default risk, and, in severe cases, contribute to systemic bank runs.”

As financial technology advances, both local and contagion effects are now amplified. The 2023 US banking crisis, where depositors pulled US$42 billion out of Silicon Valley Bank within hours, is a clear example. In the previous 2008 crisis, bank runs took several days or weeks.

“Depositors now can reallocate funds in minutes, while social media and financial news accelerate the spread of negative sentiment, shortening the window for banks and regulators to take action,” she adds. “Digital banking also helps contagion effects expand beyond the district where the troubled bank operates to any bank that shares similar risks.”

The last banking crisis not only collapsed three major US banks but also affected more than 20 other banks due to severe deposit runs. Depositors flocked to bigger national banks, like JPMorgan Chase and Bank of America, as well as government-backed funds, where they felt more secure.

How to control the contagion

To mitigate local and contagion effects that could spiral into more devastating outcomes, Professor Kang notes a state-contingent disclosure and intervention framework could work. This framework consolidates information on what a bank must publicly disclose and when the regulator steps in.

During a crisis, bank disclosure that enhances transparency can trigger contagion effects that harm other healthier banks. Therefore, regulators should at once provide financial support, clearly explain that the problem is isolated, and closely watch banks with similar risks.

“Banks should monitor the regulatory and financial health of local peer banks, given that early awareness of distress allows preemptive liquidity planning,” she adds. “Visibly strong capital ratios, asset quality, and diversified funding are a must since contagion is likely attenuated for those with strong balance sheets.”

Banks should also diversify their depositor base to reduce vulnerability to a homogeneous and easily spooked pool of depositors. Maintaining robust contingency funding plans and pre-arranged access to liquidity facilities provides a credible backstop that banks can communicate to depositors.

Finally, banks should track early warning indicators such as rising deposit spreads relative to other local peers, unusual spikes of withdrawals from large depositors, negative sentiment in local media and social media, and stock price movements among similarly structured local banks, all of which can signal impending contagion risk before it turns into a full bank run.

The post Do bank scandals create winners or breed distrust? first appeared on China Business Knowledge.

Featured faculty: Choi Sungwoo
Written by Putro Harnowo

While AI offers undeniable efficiencies, its integration carries unprecedented risks. The recent Pew Research Centre report shows that almost half of US adults use AI chatbots, but 40 per cent view AI as having a negative impact on society. The centre’s survey last year also found that more than one-third of respondents across 25 countries are concerned about AI.

“Many practitioners tend to focus on the technical benefits of AI, such as cost efficiency, speed, and convenience, but AI implementation often comes with unintended consequences, both positive and negative, for consumer behaviour,” says Choi Sungwoo, Assistant Professor of the School of Hotel and Tourism Management at the Chinese University of Hong Kong (CUHK) Business School.

In his co-authored study, Unintended negative consequences of AI in hospitality and tourism: A comparative bibliometric analysis and systematic review across three disciplines, Professor Choi, Kim Hyunsu of the University of Macau, and Hailey Shin of The Hong Kong Polytechnic University examine 233 research papers from psychology, business, and hospitality and tourism fields.

Across all three disciplines, Professor Choi highlights that these unintended consequences stem from human perceptions, with consumer backlash against humanlike AI, unpredictable impacts of AI on consumer decision-making, and employee friction with AI systems being the most critical.

These micro-level beliefs can scale into major problems, such as resistance or underuse of AI, service breakdowns, and even reputational risks. “Understanding these impacts can help practitioners and managers fine-tune their AI applications, reduce potential risks, and maximise the value of these technologies,” Professor Choi adds.

“Before implementing AI, managers should carefully consider how AI will affect both consumers and employees. They should put themselves in the shoes of the people who will interact with or work alongside the technology and then deploy AI in a way that enhances the overall service experience.”

The psychological reasons for AI resistance

Psychology has the longest timeline, with 74 papers on the unintended consequences of AI from 2012 to 2025, and it has laid the groundwork by investigating the individual-level mechanisms of human responses to AI.

For instance, mind perception theory argues that humans judge other entities by attributing two perceived mental capacities: cognitive functions and emotional functions. Machines like AI are perceived as having moderate cognitive but limited emotional capacities, suggesting they may be smart but lack the ability to feel or sense.

Given the tendency for some to distrust algorithms, even when they outperform human judgement, the term algorithm aversion was coined. It is a psychological bias in which people prefer human judgment and interaction, believing that algorithms lack human qualities like empathy or transparency and that humans understand situations better.

“Many practitioners tend to focus on the technical benefits of AI, such as cost efficiency, speed, and convenience, but AI implementation often comes with unintended consequences, both positive and negative, for consumer behaviour.”

Professor Choi Sungwoo

Additionally, since its underlying processes are too complex for laypeople to grasp, AI has faced criticism for its lack of transparency. To address this concern, scientists have developed tools and methods that enable humans to understand how an AI model makes its choices, known as explainable AI.

At the same time, algorithm appreciation, or the tendency for people to trust and rely on AI, has also gained attention. This occurs when AI meets both functional needs by effectively helping users with a task, and psychological needs by making users feel understood during the interaction. Missing either one can result in algorithm aversion.

Some consumers may display a strong machine heuristic, a mental shortcut in which they judge computer-generated work as truthful and accurate. While it sounds good, AI can hallucinate and is not immune to bias. Studies find that trust in AI depends on how human users perceive their own ability. If a person has less confidence in their own abilities, they are less likely to rely on AI.

Unintended consequences of AI in business applications

The business discipline has produced the most research on AI, with 108 papers from 2019 to 2024 in consumer-facing contexts. These studies emphasise rich, contextualised outcomes and their implications for organisations, employees, and broader society.

For instance, when an AI system fails, people tend to believe that other AI systems or AI in general will also fail. This phenomenon, called algorithmic transference, occurs because laypeople often see AI systems as indistinguishable from one another, so one failure is perceived to spread to other AIs.

Modern issues, such as technostress, or stress caused by technology, have become more important. Although technology makes life easier, it also introduces discomfort, such as techno-invasion, when technology feels like it is intruding on personal life, and techno-complexity, when technology is too difficult to understand. These anxieties make people less willing to embrace technologies.

In human resources, AI systems meant to help employees may backfire. People trust algorithms less for tasks perceived as subjective, such as those requiring human judgement, creativity, and emotional touch.

Social setting is also important. Consumers prefer robots to humans when interacting in awkward situations and feel less judged when communicating with AI. However, adding humanlike features may increase feelings of being judged due to perceived social presence.

Double-edged swords of AI in the service industry

Rather than treating AI as a generic technology, 51 studies spanning 2020 to 2025 in hospitality and tourism emphasise circumstances that influence how consumers perceive AI. Most influential studies acknowledge a persistent gap between technological sophistication and consumers’ feelings.

Studies on robot design reveal that a humanlike appearance can initially foster engagement but trigger discomfort when it becomes too human. However, incorporating humour, adjusting interaction timing, and engaging emotions or social functions may ease that feeling.

Service robots may also be perceived as a threat. The concept of AI identity threat refers to perceived psychological and sociocultural risks to human value and uniqueness. This threat varies between individuals and encounters. When consumers experience such threats, they fear that robots may steal or seize their welfare and identity.

On the other hand, privacy concerns raise awareness that these smart systems, whether service robots or online algorithms, constantly collect personal data. The main fear stems from how the data is gathered, accessed, and used, and from losing control over their digital footprint.

Algorithmic pricing is a niche theme covering airlines, hotels, and online travel agencies. While pricing algorithms help companies optimise prices and maximise revenue, consumers often view price discrimination as unfair and are keen to complain about it publicly.

Generative AI is expected to significantly impact trip planning, but studies show that it can worsen choice overload in travel decision-making. When AI chatbots help travellers narrow down options, travellers’ satisfaction with recommendations and intention to visit also tend to decline.

The post Why humans keep swiping left on AI first appeared on China Business Knowledge.

Featured faculty: Ying Chao
Written by Sally Ho

Capital markets have been on a tear this year, driven by a massive bet on AI investments. For instance, Chinese chipmaker CXMT’s recent debut raised US$8.6 billion to deliver China’s largest listing since 2010, while a month earlier, SpaceX made history with a record-breaking US$75 billion IPO.

On top of that, two rival AI pioneers, OpenAI and Anthropic, have announced their plans to go public in the near future, each aiming to become a trillion-dollar company. Older giants like Alphabet, Amazon, and Microsoft have also seen their stocks soar following their AI ventures.

Ironically, the labour market has taken the opposite path, marked by layoffs and hiring freezes across big tech. While overall market adjustments later often bring the job market back on track, a clear pattern emerges: waves of innovation bring short-term crises for workers despite market optimism.

“There is a glaring disconnect between financial markets and the real economy. For example, stock prices sometimes surge when hiring is weak, and stock returns often move in the opposite direction of physical investment,” says Ying Chao, Assistant Professor of the Department of Finance at the Chinese University of Hong Kong (CUHK) Business School.

“When a technological breakthrough happens, markets immediately price in the expectation of future productivity, but the real-world gears, like hiring, production, and physical investment, take much longer to catch up,” he adds.

Professor Ying’s new paper, Innovation‑driven contractions: A missing link for asset pricing puzzles, co-authored with Li Zhonghao of Nanjing University and Gill Segal of the University of North Carolina at Chapel Hill, examines why technology investment is often accompanied by short‑run contractions in the real economy.

Good news for tech, bad news for workers?

Having examined financial data on US companies in Compustat from 1953 to 2019 and measured their patent values, Professor Ying and the team find that, despite boosting productivity and efficiency, technological advances shrink the workforce with a negligible impact on capital expenditure on physical assets.

This is not simply creative destruction or technological unemployment, as it doesn’t fit the typical notion that when superior businesses overtake traditional ones, both jobs and physical investments dwindle simultaneously. “Nowadays, new technology initially reduces labour input while moving the needle only slightly on capital growth,” he says.

Professor Ying and the team then grouped companies by earnings, asset values, and investment levels. After observing how often companies adjust their product prices and tracking stock returns across the entire economy, they find that when companies rarely change their product prices, higher hiring activity and capital expenditures are associated with lower stock returns.

Companies often keep prices steady for extended periods for many reasons, such as staying competitive, retaining loyal customers, and so on, even as economic conditions shift. The more rigid or “sticky” product prices are, the more likely the stock market is to react in the opposite direction of real economic activity.

This price stickiness acts as a hidden force that explains why financial markets disconnect from the real economy and why it sometimes intensifies.

“When a technological breakthrough happens, markets immediately price in the expectation of future productivity, but the real-world gears, like hiring, production, and physical investment, take much longer to catch up.”

Professor Ying Chao

“The negative contemporaneous relationship between stock returns and investment growth or labour-market surprises becomes stronger when prices are stickier,” says Professor Ying. “Stock prices are forward-looking, reflecting expected future profits from innovation. Real economic activity, like capital expenditure or labour usage, can temporarily contract or lag behind.”

“Therefore, when a company’s valuation skyrockets before a corresponding surge in real economic activity, it doesn’t necessarily indicate an irrational market bubble. Instead, it often reflects the market rationally pricing the long-term benefits of innovation, while the real economy is still undergoing short-term adjustments.”

How sticky prices drive higher stock prices

To better understand the dynamics of technological breakthroughs affecting the real economy, the team built an economic model comprising two sectors: one producing consumption goods, and the other producing machinery and equipment for other firms. Firms in both sectors have some control over their product prices, but these prices are rigid even when costs fluctuate.

After a technology shock, firms can produce more with the same level of labour and materials. But since their prices are sticky, firms charge the same price even as production costs fall, allowing them to make larger profits on each sale. “Because product prices adjust more slowly than marginal costs after a positive technology shock, markups temporarily rise. These higher markups increase valuations while restraining short-run hiring,” says Professor Ying.

A technological breakthrough also tends to lower the price of new machinery with more efficiency. Although companies want to buy more of these advanced tools and hire more workers, it takes time for product prices and wages to catch up. This stickiness across the broader economy temporarily reduces demand in real economy activities.

Put together, workers might feel the pinch of fewer jobs in the short term, while investors see the potential for higher profits and stock returns. If product prices could change easily, the stock markets and hires would move in the same direction.

“Keep in mind that whether a specific company can sustain high stock prices depends on longer-term fundamentals, including market demand, how fast they expand capacity, and how well they fend off competitors,” Professor Ying adds.

A technology-based signal for future returns

For shareholders, the disconnect between the stock market and the real economy can make it harder to predict how a stock will perform in the short term. Therefore, based on the above model, the team develops a mathematical formula called the investment-based dividend yield to track how technological innovations, rather than stock market data, influence economic dynamics.

This investment-based dividend yield is a broad economic indicator derived from the above model, rather than a measure of the cash-flow potential of an individual company’s investments. “This tool is a macro indicator that helps predict the overall stock market’s future risk, not a tool for short-term trading on individual stocks. This indicator strongly and negatively predicts future aggregate stock returns,” says Professor Ying.

Looking ahead, Professor Ying hopes to determine whether this tool can be applied with more specific information, such as data from individual industries, and with more precise ways to track innovation, especially those related to AI. “With AI specifically, the biggest challenge now is untangling genuine productivity shocks from other noise, like government subsidies, hype-driven demand, financing conditions, or geopolitical shifts,” he adds.

“Going forward, we also need to look more closely at cross‑sectional differences: how these tech shocks impact different types of workers, the role of intangible capital, and how the ultimate gains and losses of technological change are distributed across the broader economy.”

The post Why the AI boom spells short-term labour gloom first appeared on China Business Knowledge.

Featured faculty: Tony Ke and Liao Chenxi
Written by Putro Harnowo

Generative AI, or GenAI, has rapidly changed how consumers find products. AI chatbots like ChatGPT, DeepSeek, and Claude AI have become the go-to tools for daily inquiries, much like Google in the old days, while traditional search engines like Google and Bing are stepping up by displaying AI-generated answers at the top of search results.

No surprise that American software company Adobe reported traffic from AI sources to retail sites surged by almost seven times in 2025 compared to the previous year. This shift has propelled a new digital strategy called generative engine optimisation (GEO), which aims to ensure the product shows up when users ask an AI a specific question.

Since GenAI is trained on massive amounts of data, companies can increase their visibility by creating synthetic content, or favourable product mentions across articles, blogs, social media, online forums, and more. The goal is to let AI models learn that their products are desirable, thereby increasing the likelihood of being recommended.

“A particularly powerful GEO strategy is deliberately crafting synthetic content so that GenAI learns and reproduces desired associations between consumer questions and product offerings,” says Liao Chenxi, Associate Professor of Marketing at the Chinese University of Hong Kong (CUHK) Business School.

For instance, a brand covertly creates blogs or articles on the top 10 products in the market but disproportionately highlights its own products, or posts a question on forums like Quora or Reddit and then uses another account to give a detailed answer recommending its own products.

Not to be confused with fake reviews, synthetic content is strategically created to influence GEO by feeding information that can be easily parsed by AI engines without explicitly targeting consumers. When content becomes misleading to consumers or includes fabricated claims, such as fake reviews, it can be considered deceitful.

In fact, the abundance of fake reviews has driven regulators in the US and the UK to ban them altogether. While it is not meant to mislead consumers, at its core, synthetic content aims to influence GenAI, which some consumers may consider as manipulation.

In her new study titled Synthetic corpus and consideration manipulation in generative engine optimisation, Professor Liao finds that, despite helping increase visibility, firms caught red-handed deploying synthetic content are perceived by consumers as lower quality. However, banning it entirely may backfire, as synthetic content can help consumers discover good products.

“AI engines work like a ‘black box’, but some consumers have an expectation that content manipulation exists on the internet, so they do not always blindly follow what GenAI recommends.”

Professor Liao Chenxi

Mapping player interactions with GenAI

In collaboration with Tony Ke, Professor of the Department of Marketing, and Xu Xiaoyan of Southwestern University of Finance and Economics, Professor Liao uses game theory to examine how firms strategically deploy synthetic content to influence consumers and GenAI, and how consumers respond to the firms’ actions.

The game-theoretic model in the study involves a firm selling a product of either high or low quality, GenAI, and consumers. Quality products naturally attract plenty of positive reviews, so a high-quality firm can rely on genuine, organic content. A low-quality firm has little organic content, so it must create synthetic content to make its product appear more attractive.

GenAI treats all content, whether organic or synthetic, as input data for its recommendation system. If the content provides sufficient evidence of a useful product, GenAI will recommend it. Consumers choose the product based on perceived quality, but are unaware of its true quality.

Before deciding, consumers go through a three-stage purchase funnel. The first is the consideration stage, in which consumers discover products. In traditional digital marketing, search engines and ads serve as beacons to discover the products, but consumers nowadays turn to GenAI.

Next, in the evaluation stage, consumers typically check other sources to confirm product quality. Consumers who consider GenAI recommendations then conduct independent evaluations. While some consumers fully rely on AI, the model focuses on how AI influences the options consumers consider. The same content that drew the AI’s attention serves as evidence for consumers to decide.

However, consumers are not entirely fooled. Some still cross-check with third-party sources and verify product quality through their own inspection. If they find out the content recommended by GenAI is synthetic, they will perceive the featured product negatively, regardless of its actual quality. GenAI would not be blamed, as consumers use it primarily to find products.

“After initial discovery with the help of AI, consumers have many independent channels to evaluate the product,” says Professor Liao. “AI engines work like a ‘black box’, but some consumers have an expectation that content manipulation exists on the internet, so they do not always blindly follow what GenAI recommends.”

In the final stage, consumers have formed expectations about the product’s quality. When the perceived quality matches personal preferences, consumers are more likely to buy it.

To manipulate or not to manipulate?

Synthetic content may successfully trick AI models, but it is not always effective on consumers. If consumers are good at spotting it, high-quality firms can keep their reputations by avoiding it, but risk being thwarted by synthetic content from low-quality rivals. High-quality firms can still leverage synthetic content, but the benefits depend on other market players.

If low-quality firms promote extremely poor products, allowing synthetic content is detrimental as it can harm consumers. For high-quality firms, the reputational damage from being caught using synthetic content becomes so severe that they risk being mistaken for having awful products, too.

However, when low-quality rivals promote moderate-quality products, synthetic content from high- and low-quality firms broadens market demand and helps all products be recommended by GenAI without risking severe reputational damage if consumers find out. High-quality products that lack organic buzz can benefit from a synthetic nudge to be endorsed by GenAI.

Ultimately, given that being spotted using synthetic content may lower consumer trust, Professor Liao emphasises that the strategy should be applied wisely. For firms with genuinely high-quality products, another way to improve GEO is to optimise organic content by encouraging satisfied customers to leave reviews and fostering genuine discussions on social media and online forums.

“When real consumers spontaneously praise a product online, it naturally provides quality data that AI engines incorporate in their algorithms, increasing the recommendation probability,” she adds. “Companies absolutely need to pay attention to GEO to stay visible, but it doesn’t mean that it can substitute traditional organic signals. In essence, GEO is a marketing tool for visibility. To build real trust, firms must make sure their product quality matches their claims.”

The post Can you truly trust AI recommendations? first appeared on China Business Knowledge.

Featured faculty: Philip Zhang Renyu and Jenny Jin
Written by Joanne Madrid

Large language models or LLMs, the AI systems behind chatbots like ChatGPT, are increasingly being used as stand-ins for humans in behavioural research. Dubbed synthetic surveys or silicon sampling, people can now ask AI to complete surveys at a fraction of the cost and time of human participants.

AI startups like Simile, Aaru, Artificial Societies, Uxia, and many others offer human behaviour simulations to predict market trends, consumer preferences, and even political views. This synthetic survey is a direct application of social simulation, in which LLMs act as humans in experimental settings. As LLM research expands, an important question is whether large language models can accurately reproduce human decisions in social experiments.

Yet there is a catch. Social simulation may not be a reliable substitute for human respondents because AI can struggle to capture nuanced opinions and may reflect biases in its training data. Although large language models have successfully reproduced simple experiments, they often fail in complex scenarios. This challenge has become an important area of LLM research, particularly in understanding why AI reasoning sometimes differs from human decision-making.

According to Philip Zhang Renyu, Professor in the Department of Decisions, Operations and Technology at the Chinese University of Hong Kong (CUHK) Business School, such limits may stem from a deeper structural challenge in how LLMs understand and reason about their tasks.

“LLMs in behavioural and social simulations can fail not only because of flawed reasoning, but also because of the limitations in how they interpret their tasks,” he says. “Even if they understand the tasks, their reasoning process may not align with human thinking.”

Professor Zhang’s latest study, Improving behavioural alignment in LLM social simulations via context formation and navigation, co-authored with Jenny Jin, Assistant Professor in the same department and their PhD student Kong Letian, proposes a two-stage method to improve LLMs’ ability to simulate human behaviour.

Beyond clever prompts

Most AI users engaged in prompt engineering have already recognised the importance of clear, specific instructions, but many popular prompts still fall short in social simulations. Therefore, the study aims to find whether a more comprehensive approach is more effective.

Specifically, Professor Zhang and the team turn to a classic Human Problem-Solving Theory developed by Herbert Simon and Allen Newell. Both celebrated scientists, with the help of programmer Cliff Shaw, developed Logic Theorist, the first computer programme designed to perform automated reasoning that marked the birth of AI.

Simon and Newell argue that humans solve problems by forming a mental picture of the task and selectively searching for solutions within it. Inspired by this theory, the team then develops a two-stage framework to better align LLMs with human behaviours in social simulations.

The first stage is context formation, or explicitly specifying experimental design so LLMs can understand the tasks and their background by defining their goals, what information matters, and how all the pieces fit together. The second is context navigation, or guiding the reasoning process so LLMs can further strategise, update their assessments, and weigh the best options.

For LLMs to behave like humans, Professor Zhang notes that the two-stage framework, context formation and context navigation, should be clearly defined so the AI understands the task and its role perfectly.

“Even really smart AI still needs a crystal-clear instruction to make good decisions. While AI sometimes can reason autonomously in simpler tasks, context navigation remains crucial in complex scenarios to choose the most relevant behavioural signals.”

Professor Philip Zhang Renyu

Comparison with popular prompts

Current popular prompting techniques can improve task performance but are not always suitable for social simulations. For instance, chain-of-thought prompting focuses on step-by-step instructions, yet if the AI misunderstands the rules and setup, it often yields inconsistent results. Few-shot prompting guides the AI model with examples, but it may not be reliable when the social experiment differs from the examples.

Tree-of-thought prompting asks the AI model to explore several paths to choose the best solution, but it fails to determine which one is most human-like. Meta-prompting can organise reasoning by designing a reusable prompt, but still requires humans to specify the setting and the standard for a valid result.

“Context formation provides a systematic way to explain the rules and setup of an experiment to the AI models, which helps it understand and adapt to specific scenarios rather than relying on a few examples or generic instructions,” Professor Zhang says.

“Context navigation directs the AI models selectively towards the crucial information and strategic considerations, rather than requiring an open search across reasoning paths. These stages help the AI figure out why it still mismatches human behaviour and see if it misunderstood the setting, if its reasoning was incorrect, or both.”

Putting the method to the test

To evaluate the framework, the team tested it across three existing social experiments with publicly available human data, allowing the researchers to compare AI-generated decisions directly with real human choices. They provided the experimental instructions step by step, as in chain-of-thought prompting, to the most capable LLMs at the time—GPT-4o, GPT-5, Claude 4 Sonnet Thinking, and DeepSeek-R1—to assess how accurately these large language models could replicate human behaviour.

The first human experiment found that the waiting time to get a product can signal its quality. When the same instructions in the experiment were fed to LLMs, all models boiled the problem down to a cost analysis and ignored social cues humans perceived.

Adding context formation that some queuing buyers already knew the product’s quality helped the LLMs understand the instruction, but still could not replicate human behaviour. Only after providing context navigation, that queue length can be a meaningful cue for quality, did the LLMs start to act like human shoppers.

The second experiment asked humans to choose whether to invest in a project without knowing its true quality. Some project owners offered investors a refund if they failed to reach full funding, which human investors viewed as a signal of a good project.

However, even after being provided with investment rules, all LLMs significantly funded fewer good projects than humans. Only when the LLMs received clear context navigation to interpret the refund scheme as a positive signal did they replicate human investors.

Not every situation demanded both steps. The third experiment simply found that human purchases decreased as the price increased. Oddly, upon receiving the instructions, LLMs bought more when the price rose, and as the cost soared further, they eventually reduced their purchases.

When LLMs were given a clear explanation of the experimental setup, they successfully replicated human behaviour. Here, providing context formation was sufficient. The contrast points to a useful rule of thumb: simple choices often need only context formation, while more complex scenarios involving several parties require a two-stage method.

As AI models improve, explicit guidance on how to reason may matter less for simpler tasks since the models can work it out for themselves, but this doesn’t mean that context navigation becomes less important. “Even really smart AI still needs a crystal-clear instruction to make good decisions,” Professor Zhang adds.

“While AI sometimes can reason autonomously in simpler tasks, context navigation remains crucial in complex scenarios to choose the most relevant behavioural signals.”

The post Can AI really mimic our decision-making? first appeared on China Business Knowledge.

Featured faculty: Kim Keongtae and Michael Zhang
Written by Pan Jingyi

AI-powered coding assistants are changing how software is built. These tools can be highly effective in helping programmers and software developers, even those with less experience, to write code in a programming language. Almost everyone can now make an app or software from scratch by simply asking an AI chatbot, a practice known as vibe coding.

OpenAI’s release of Codex in 2021 was the first AI model powerful enough to build entire programmes from simple text prompts. The company then partnered with GitHub, a platform where developers worldwide can store and collaborate on their software code, to launch GitHub Copilot as an autocomplete tool. In early 2023, Copilot was transformed into Copilot X, a full AI coding assistant.

Nowadays, Copilot X is not the only tool in the market. Other major AI companies, such as Anthropic and DeepSeek, as well as big tech like Google, Amazon, and Meta, have also launched their AI-assisted software development tools. As millions of developers start relying on AI, burning questions loom: have these tools changed how developers code, and can they actually boost productivity?

“AI technology may increase the developers’ productivity and push the limits of what developers feel able to attempt, but these gains come with trade-offs for code quality and originality.”

Professor Kim Keongtae

Kim Keongtae, Professor in the Department of Decisions, Operations and Technology at the Chinese University of Hong Kong (CUHK) Business School, collaborates with Michael Zhang, the Wei Lun Professor of Business AI in the same department, their PhD student Li Xinyu, and Francis Joseph Costello of Nova School of Business and Economics, to seek answers. In the study titled Exploring altered open source software development patterns in a time of generative AI, they examine 1,350 open-source software developers on GitHub before and after Copilot X was released.

The analysis finds that CopilotX significantly helps occasional contributors who were previously less active on GitHub. These casual developers participated in 48 per cent more open-source projects and became far more likely to experiment with new programming languages after using the AI tool.

Meanwhile, highly active or more seasoned developers who used Copilot X updated their code 12 per cent more often and contributed to six per cent more projects compared to those who didn’t use the tool. AI not only helps developers work faster but also spreads their effort across a larger set of projects and switches between different projects more easily.

Contributing to open-source projects sometimes requires developers to learn an unfamiliar codebase and figure out how the software is structured before making a small update. For occasional contributors, these barriers can be enough to stop them from participating at all.

“AI coding assistants can lower these barriers and make the first step easier,” says Professor Kim. “After conversational AI coding assistance became available, developers became more willing to explore new programming languages and projects.”

The hidden costs of easy coding

While productivity rose, a noticeable increase in copy-pasted code appeared. Although more projects are being started, they are not necessarily leading to more original work. “AI technology may increase the developers’ productivity and push the limits of what developers feel able to attempt, but these gains come with trade-offs for code quality and originality,” Professor Kim adds.

“Copied code is not always harmful, as some reuse is normal in software development. The concern is that if developers insert larger blocks of code without fully understanding how they interact with the rest of the project, the project may accumulate technical debt or a shaky foundation.”

The build-up of hidden problems could make software harder to maintain, debug, or improve later. It can also lead to compatibility issues, inconsistent coding styles, and more rework down the line. For project managers, this finding calls for treating AI-generated code as a draft, not as final output.

AI can boost careers

Open-source software development relies on voluntary effort, so developers have to contribute their personal time. While some programmers and software developers participate for their own interests, GitHub functions much like a public portfolio to showcase recent work and collaborations. Employers also often look at GitHub profiles to assess software developers’ skills and experience.

Therefore, to further examine whether changes in how developers worked with AI affected their careers, the researchers matched developers’ GitHub histories with their LinkedIn profiles and found encouraging signs. Developers with access to Copilot X were more likely to experience internal transitions and promotions in the following six to 12 months.

It appears that increases in open-source contributions and proficiency with new technologies are associated with positive career outcomes. “Visible output on GitHub can also signal employers that a developer is adapting to new tools and can leverage them to deliver more work,” Professor Kim says.

“However, the benefits of AI coding tools for career improvement are temporary. When only a few developers know how to use these tools really well, they’ll be noticeably more productive than others. As these tools become widely used, simply using AI won’t make a developer stand out as much anymore.”

How can developers stand out in the AI age?

Given the short-term benefits and the increasingly prevalent copied code, Professor Kim warns developers that AI should only be used as a learning partner and assistant. Although AI may accelerate software development, it cannot replace the developer’s responsibility to understand and evaluate the system. “Take the AI tool as a tutor rather than a shortcut. The human developer should remain responsible for the final judgement.”

This advice is especially crucial for beginners, as less experienced software developers are more likely to use code suggested by AI tools, even if the code is not very good or has mistakes.

RELATED ARTICLE

How retail investors use GenAI to navigate stock markets

As technologies evolve rapidly, AI tools can play a vital role in ensuring software quality and in helping spot critical issues during development, including reviewing updates that might introduce risks and identifying areas that need more testing. Highly skilled developers can leverage these tools to validate code outputs, establish clear coding rules, and integrate code into complex systems.

Ultimately, Professor Kim anticipates that the more sustainable career advantage will likely come from higher-level capabilities, such as understanding software’s underlying problems, evaluating AI output, designing maintainable systems, and coordinating human-AI workflows.

“The career advantage from knowing how to use AI coding tools will not vanish but shift towards those who can use it in more advanced, strategic, and complex ways.”

The post Is AI creating more productive but dull programmers? first appeared on China Business Knowledge.

Featured faculty: Chen Zhi
Written by Ellis Ng

Today’s AI chatbots are fueled by large language models (LLMs) trained on vast expanses of human data, from books, news articles, company reports, and social media conversations. However, as corporate reliance on these tools grows, so does the threat of embedded biases and critical errors.

An October 2025 report from the BBC found that nearly half of AI-generated responses contain inaccuracies. Last March, an engineer at Meta accidentally leaked sensitive data after following a suggestion from an AI agent. These high-profile cases prompt a critical question: What if these powerful AIs, trained on everything humans have ever created, also pick up human flaws?

“LLMs don’t just mirror human biases, but also often amplify them,” says Chen Zhi, Associate Professor in the Department of Decisions, Operations and Technology at the Chinese University of Hong Kong (CUHK) Business School. “Our experiments show that LLMs consistently replicate the classic too-low or too-high ordering bias well-documented in humans, ordering too little in high-margin scenarios and too much in low-margin ones.”

A study Professor Chen co-authored with Liu Jifei and Zhong Yuanguang of South China University of Technology, Large language newsvendor: Decision biases and cognitive mechanisms, tests LLMs on the “newsvendor problem”, a decision-making problem in which an AI must stock a resource before knowing the actual demand. The results show that AI often overshot the mark and performed worse than humans. The smarter the model, the more likely it was to act irrationally.

The paradox of intelligence

Professor Chen and the team ran multi-round experiments with 15 decision rounds each: LLaMA-8B, GPT-4, and GPT-4o. Each AI had to decide how much product to order, just as a store owner would try to guess what customers would buy and how much. After each round, the AI received feedback on what customers actually demanded and how much profit it made.

These findings highlight an important challenge for the future of work automation: greater AI capabilities do not always translate into better business decisions.

The entire process was repeated for three scenarios with varying customer demand and conditions: one with no guidance, one with a mathematical formula for calculating the optimal order quantity, and the last with a special “no-risk” scenario in which every order placed was guaranteed to turn profits.

“The most sophisticated model by the time we conduct our study, GPT-4, actually showed the greatest irrationality,” Professor Chen says. “It would correctly compute the optimal order quantity in its reasoning, then talk itself out of it through elaborate ‘risk management’ adjustments, even in settings with zero financial risk.”

“The most sophisticated model by the time we conduct our study, GPT-4, actually showed the greatest irrationality.”

Professor Chen Zhi

Each model failed for a different reason. For instance, GPT-4 over-ordered by 70 per cent more than humans in low-margin scenarios. These errors are not due to the AI being cautious and overthinking, but stemming from deep-seated information-processing, as its deliberation undermined its own best decisions, even when there was no chance of losing money.

Meanwhile, LLaMA-8B lacked the computational power to consistently apply rules, leading to erratic choices. GPT-4o was built for speed, stuck closely to simple formulas, and made steady decisions. “The three models illustrate three distinct failures: over-analysis (GPT-4), rigid heuristic adherence (GPT-4o), and fundamental computational limitations (LLaMA-8B),” Professor Chen adds.

The researchers call this the paradox of intelligence, where more brainpower does not guarantee better results. All models were anchored by their first scenario, carrying the bias even after many rounds, but also fixated on the most recent orders. In the most volatile scenario, LLMs reacted to the latest data point by changing their orders all the time. By comparison, humans did this fewer than four times out of ten.

Such errors can be traced back to how the models read and process text. LLMs take in information one piece at a time, and this sequential reading spills into decision-making. The models either cling to whatever they encountered first or overreact to whatever arrived last.

Giving AI clear rules and more adjustments

As businesses adopt AI to automate decisions and improve operational efficiency, reducing bias and errors becomes essential. Professor Chen identifies several practical strategies to make AI-driven decision-making more reliable.

“Providing explicit optimal formulas significantly improved performance,” he adds. “Well-designed prompts can constrain bias without requiring any modification to the models.”

Another way is to choose LLMs carefully by matching them to the task rather than defaulting to the most powerful option. “Model selection matters enormously,” he says. “Efficiency-optimised models can outperform more complex ones on well-defined optimisation tasks, so managers should match an AI model to the task rather than defaulting to the most capable option.”

Fine-tuning also offers a promising solution. Professor Chen explains that fine-tuning LLM, where the model is further trained on a more specific dataset to adapt to a particular task, could improve performance. “Finally, human-in-the-loop oversight remains essential, especially for detecting cases where a model computes the right answer but then overrides it with heuristic adjustments,” he adds.

Do we really need smarter AIs?

Professor Chen and his fellow researchers have continued testing newer LLMs and expanded their experiments to other scenarios. He notes that biases persist in newer LLMs, but the driving factors are not always the same as those seen in the initial study. Some models have continued to overthink like GPT-4, but in different ways, and new kinds of errors have appeared.

As businesses adopt AI to automate decisions and improve operational efficiency, reducing bias and errors becomes essential. Professor Chen identifies several practical strategies to make AI-driven decision-making more reliable.

“Firms should not assume that a more advanced or expensive model will produce better operational decisions,” he says. “AI biases are systematically shaped by their underlying architecture, and the specific contours of these biases evolve as model architectures change.”

When models compute the right answer but then talk themselves out of it, humans can intervene before those decisions get implemented and compare the model’s reasoning against its final output. Clear decision checkpoints, where AI outputs pass through rules-based checks before implementation, can also serve as a practical safeguard.

“Since LLMs can intensify human biases, the economic consequences of unsupervised deployment can exceed what we would expect from human decision-makers alone, so organisations need systematic monitoring for bias amplification,” he adds. “The goal is not to eliminate AI from the decision process but to design systems that leverage AI’s computational strengths while constraining its vulnerabilities.”

The post Does smarter AI generate more human errors? first appeared on China Business Knowledge.