Category: 3. Business

The Suzuki Alto, one of Pakistan’s top-selling compact vehicles, experienced a major drop in sales last month. According to data from the Pakistan Automotive Manufacturers Association (PAMA), sales plummeted by 75%, falling from 9,497 units in June to just 2,327 in July.

The steep decline followed recent government changes to taxation. The General Sales Tax (GST) on cars up to 850cc was raised from 12.5% to 18%. Additionally, the government introduced a new NEV levy, applying 1% to vehicles up to 1300cc, 2% for 1301cc to 1800cc, and 3% for larger vehicles. The Alto, although below 850cc, was still impacted by these fiscal adjustments.

Reacting to the tax changes, Pakistan Suzuki Motors Company (PSMC) raised the price of the Alto to Rs. 3,326,450. The higher price point pushed the car out of reach for many budget-conscious buyers, particularly first-time car owners, a segment that traditionally makes up a significant portion of Alto’s customer base.

In an attempt to shield buyers from the tax hike, PSMC adopted an early invoicing approach in June, processing orders for July and August in advance. While this strategy boosted June’s numbers, it contributed to a significant decline in subsequent months.

Adidas executives visited a small Indigenous town in the mountains of southern Mexico on Thursday to offer an apology over a sandal-inspired shoe design that Mexico’s government had blasted as cultural appropriation.

The German sportswear company sent representatives from its Mexican unit to Villa Hidalgo Yalalag, a town in Oaxaca state, to deliver the comments in person after issuing a written apology last week.

The issue related to the Oaxaca Slip On, designed by Mexican-American designer Willy Chavarria, which locals say closely resembles their traditional handmade huarache sandals.

“We understand this situation may have caused discomfort, and for that reason, we offer a public apology,” Karen Gonzalez, head of Legal and Compliance at Adidas Mexico, told a few dozen people gathered at an outdoor sports field.

The event included traditional music and attendees in Indigenous attire.
Gonzalez said Adidas would in future seek collaboration with Villa Hidalgo Yalalag to ensure respect for its cultural heritage. The community is home to fewer than 2,000 people.

“Thank you very much for keeping your word,” said Mayor Eric Fabian. “(Our cultural heritage) is something we safeguard very carefully. Yalalag lives from its crafts,” he added.

The controversy drew national attention earlier this month when Mexican President Claudia Sheinbaum criticised Adidas and announced plans to explore legal avenues to protect Indigenous communities from alleged cultural appropriation by big companies.

Mexico has previously accused other big-name global fashion players of exploiting Indigenous designs without consent.

This study fills an existing gap by exploring the use of ChatGPT in the field of medical image-based questions and case-based teaching scenarios. Our findings demonstrate that ChatGPT exhibit the capability to accurately answer image-based medical questions, with the GPT-4o version demonstrating a numerically higher accuracy. Prompt engineering enables ChatGPT to assist in the design of lesson plans, showcasing significant potential in medical education. Nevertheless, human verification and correction remain essential. These findings may serve as a significant step toward expanding the practical use of advanced ChatGPT versions, which have already demonstrated great potential in medical fields.

Previous studies have widely assessed the performance of ChatGPT on medical questions, predominantly utilizing earlier versions such as GPT-3.5 and GPT-4, which yielded inconsistent accuracy levels [1, 3, 4, 13,14,15, 17, 26, 27]. However, these studies excluded image-based questions due to the limitations of earlier versions, hindering potentials of ChatGPT in real-world teaching contexts. In response, our study specifically targeted medical image-based questions to fill this gap. The accuracy of ChatGPT in our findings is consistent with previous studies [14,15,16,17], while the higher version, GPT-4o, exhibited greater accuracy (nearly 90%) [18]. While GPT-4o demonstrated higher accuracy than GPT-4 across the image-based items, this difference did not reach statistical significance, likely due to the limited sample size. Nonetheless, the consistent directional trend aligns with recent findings on the evolving capabilities of multimodal LLMs, and supports further investigation in larger-scale evaluations.

In addition to evaluating model performance, our findings highlight broader implications for the application of large vision-language models (LVLMs) in medical education—particularly in the areas of assessment design and instructional content development. LVLMs such as GPT-4o may assist in generating multiple-choice questions from image-based course materials, offering opportunities to streamline item development. Furthermore, our study demonstrates that GPT-4o possesses a notable capacity for logical reasoning and analytical processing when evaluating incorrect answers, suggesting its potential utility in developing explanations that reinforce clinical thinking frameworks. However, ensuring the clinical validity, cognitive appropriateness, and pedagogical alignment of AI-generated content remains a critical challenge, underscoring the need for expert oversight. While this study focused on interpreting images embedded within question stems, future work could explore more complex formats in which answer options themselves are visual (e.g., electrocardiogram tracings or radiographs). These tasks require fine-grained image discrimination and multimodal reasoning, areas in which current models still face limitations. As such, advancing the visual acuity and contextual understanding of LVLMs is essential to support their integration into high-stakes assessment environments.

Beyond assessment, LVLMs like ChatGPT show promise as instructional aids in case-based teaching. As demonstrated in exploratory prompts (Appendix 1 and 2), the model can support educators in structuring lessons around specific clinical concepts and generating adaptive instructional feedback [27, 28]. With its ability to present information logically and adjust content to learner needs, ChatGPT may enhance personalized learning and curricular efficiency [5, 28, 29]. Our findings, in conjunction with prior research [30, 31], support the growing interest in AI-assisted pedagogy. Nonetheless, limitations persist. While GPT-4o exhibits high coherence and insight in reasoning tasks, errors remain—particularly in nuanced clinical contexts involving medical images. These inaccuracies underscore the importance of human oversight and expert validation to ensure instructional reliability and clinical relevance [27, 32,33,34,35]. As such tools continue to evolve, their optimal use will likely depend on integration with faculty-led review and revision mechanisms, ensuring safety, accuracy, and pedagogical value in medical education.

This study has several limitations. First, all model responses were generated via the ChatGPT web interface under default settings rather than through the OpenAI Application Programming Interface (API). Although personalization was disabled and each question was submitted in a newly initiated session to minimize memory effects, API-based deployment would allow for greater control over system parameters and eliminate potential user-specific variability. Second, each question was evaluated only once per model. This approach, consistent with prior LLM evaluation studies, minimized contextual contamination from repeated prompts—particularly relevant in session-based environments—but precluded assessment of intra-model variability. Future studies should consider multi-sample testing under controlled API conditions to examine response stability and reproducibility. Third, the relatively small number of publicly available USMLE-style questions with image content (n = 38) limited the statistical power and generalizability of the findings. Expanding the question pool and including a broader range of visual modalities—such as radiographs and electrocardiograms—would enhance benchmarking rigor. Fourth, the study did not isolate the respective contributions of visual versus textual inputs to model performance. While examples in the Supplementary Appendix suggest engagement with image content, dedicated experimental designs are needed to disentangle multimodal reasoning pathways and assess their relative influence. Finally, the study focused exclusively on GPT-4 and GPT-4o, which were the most accessible and stable vision-capable models at the time of evaluation (late 2024 to early 2025). Comparative studies involving other LVLMs, such as Gemini, LLaVA, or DeepSeek, are warranted to explore model-specific strengths and inform future applications in medical education.

WASHINGTON — Now that Federal Reserve Chair Jerome Powell has signaled that the central bank could soon cut its key interest rate, he faces a new challenge: how to do it without seeming to cave to the White House’s demands.

For months, Powell has largely ignored President Donald Trump’s constant hectoring that he reduce borrowing costs. Yet on Friday, in a highly-anticipated speech, Powell suggested that the Fed could take such a step as soon as its next meeting in September.

It will be a fraught decision for the Fed, which must weigh it against persistent inflation and an economy that could also improve in the second half of this year. Both trends, if they occur, could make a cut look premature.

Trump has urged Powell to slash rates, arguing there is “no inflation” and saying that a cut would lower the government’s interest payments on its $37 trillion in debt.

Powell, on the other hand, has suggested that a rate cut is likely for reasons quite different than Trump’s: He is worried that the economy is weakening. His remarks on Friday at an economic symposium in Grand Teton National Park in Wyoming also indicated that the Fed will move carefully and cut rates at a much slower pace than Trump wants.

Powell pointed to economic growth that “has slowed notably in the first half of this year,” to an annual rate of 1.2%, down from 2.5% last year. There has also been a “marked slowing” in the demand for workers, he added, which threatens to raise unemployment.

Still, Powell said that tariffs have started to lift the price of goods and could continue to push inflation higher, a possibility Fed officials will closely monitor and that will make them cautious about additional rate cuts.

The Fed’s key short-term interest rate, which influences other borrowing costs for things like mortgages and auto loans, is currently 4.3%. Trump has called for it to be cut as low as 1% — a level no Fed official supports.

However the Fed moves forward, it will likely do so while continuing to assert its longstanding independence. A politically independent central bank is considered by most economists as critical to preventing inflation, because it can take steps — such as raising interest rates to cool the economy and combat inflation — that are harder for elected officials to do.

There are 19 members of the Fed’s interest-rate setting committee, 12 of whom vote on rate decisions. One of them, Beth Hammack, president of the Federal Reserve’s Cleveland branch, said Friday in an interview with The Associated Press that she is committed to the Fed’s independence.

“I’m laser focused … on ensuring that I can deliver good outcomes for the for the public, and I try to tune out all the other noise,” she said.

She remains concerned that the Fed still needs to fight stubborn inflation, a view shared by several colleagues.

“Inflation is too high and it’s been trending in the wrong direction,” Hammack said. “Right now I see us moving away from our goals on the inflation side.”

Powell himself did not discuss the Fed’s independence during his speech in Wyoming, where he received a standing ovation by the assembled academics, economists, and central bank officials from around the world. But Adam Posen, president of the Peterson Institute for International Economics, said that was likely a deliberate choice and intended, ironically, to demonstrate the Fed’s independence.

“The not talking about independence was a way of trying as best they could to signal we’re getting on with the business,” Posen said. “We’re still having a civilized internal discussion about the merits of the issue. And even if it pleases the president, we’re going to make the right call.”

It was against that backdrop that Trump intensified his own pressure campaign against another top Fed official.

Trump said he would fire Fed Governor Lisa Cook if she did not step down from her position. Bill Pulte, a Trump appointee to head the agency that regulates mortgage giants Fannie Mae and Freddie Mac, alleged Wednesday that Cook committed mortgage fraud when she bought two properties in 2021. She has not been charged.

Cook has said she would not be “bullied” into giving up her position. She declined Friday to comment on Trump’s threat.

If Cook is somehow removed, that would give Trump an opportunity to put a loyalist on the Fed’s governing board. Members of the board vote on all interest rate decisions. He has already nominated a top White House economist, Stephen Miran, to replace former governor Adriana Kugler, who stepped down Aug. 1.

Trump had previously threatened to fire Powell, but hasn’t done so. Trump appointed Powell in late 2017. His term as chair ends in about nine months.

Powell is no stranger to Trump’s attacks. Michael Strain, director of economic policy studies at the American Enterprise Institute, noted that the president also went after him in 2018 for raising interest rates, but that didn’t stop Powell.

“The president has a long history of applying pressure to Chairman Powell,” Strain said. “And Chairman Powell has a long history of resisting that pressure. So it would be odd, I think, if on his way out the door, he caved for the first time.”

Still, Strain thinks that Powell is overestimating the risk that the economy will weaken further and push unemployment higher. If inflation worsens while hiring continues, that could force the Fed to potentially reverse course and increase rates again next year.

“That would do further damage to the Fed’s credibility around maintaining low and stable price inflation,” he said.

The detailed operation of conventional Bandgap Reference (BGR) Core circuits, including the Current Mirror BGR (CM-BGR) and the Cascaded Current Mirror BGR (Cascaded CM-BGR), is discussed in this section. These architectures are analyzed in terms of their temperature compensation mechanisms, process variation tolerance.

Conventional BGR core

Integrated circuits (ICs) must operate reliably in harsh environmental conditions, ranging from hot desert temperatures to sub-zero polar temperatures. To operate stably under such conditions, a Bandgap Voltage Reference (BGR) Core has been designed to generate a temperature-independent reference voltage, as shown in Fig. 1. The BGR circuit is essential to maintain stable operation by demonstrating process independence, stable operation over different semiconductor fabrication processes, voltage independence, minimizing variations due to supply fluctuations, and temperature independence, to operate reliably over a broad temperature range, usually from − 40 to + 125 °C¹⁴. The basic principle behind a BGR circuit is the generation of two voltages with opposite temperature coefficients to attain thermal stability. One such voltage is the Complementary to Absolute Temperature (CTAT) voltage, derived from the base-emitter voltage (V_BE) of a bipolar junction transistor (BJT). The V_BE voltage is of negative temperature coefficient, reducing by about − 2 mV/°C with a rise in temperature¹⁵. This CTAT voltage, from a diode-connected BJT, forms the foundation for temperature compensation in BGR circuits to provide a stable reference voltage over changing environmental conditions¹⁶.

So, V_BE is negative temperature co-efficient (− 2mv/°C).

The second one is PTAT Voltage (Proportional to Absolute Temperature). Derived from the thermal voltage (V_T = kT/q), which increases linearly with temperature at a rate of approximately + 0.087 mV/°C. By scaling this voltage appropriately, its temperature coefficient can be matched to that of VBE. A PTAT voltage generator achieved by subtracting the VBE of two BJTs operating with a current density ratio N⁷.

So, V_T is positive temperature co-efficient 0.087mv/°C)

The sum of these voltages, appropriately weighted, yields a temperature-independent reference voltage.

The output reference voltage is given by.

VREF = VBE + V_T⋅ln (N) = constant, which independent of PVT variations (when slope of CTAT = PTAT).

where ln (N) is a scaling factor determined by the ratio of slopes(m = 2/0.087 = 23). So, N should be very large in millions of transistors should be connected in the 2^nd stage.

Robust cascaded current mirror-based Bandgap Reference (BGR) circuits

A conventional current mirror and a stable cascaded current mirror-based BGR circuit are analysed in this section for improving stability against variations of Process, Voltage, and Temperature (PVT)^13,14,15. A single-stage BGR circuit based on current mirrors can be seen in Fig. 2a. The circuit consists of a PMOS current mirror (M1-M2) and a transistor-based BGR core (Q1-Q5). Temperature compensation is achieved by combining proportional-to-absolute-temperature (PTAT) and complementary-to-absolute-temperature (CTAT) voltages through the resistor R_BGR.

The cascaded current mirror-based BGR¹⁷ with its robust structure shown in Fig. 2c is comprised of two additional current mirror stages (M11-M22 and M33-M44) to achieve better current replication. As a result, the output reference voltage (V_BGROUT2) has enhanced immunity to supply variations.

Figure 2b, d illustrate the impact of R_BGR variation on V_BGROUT1 and V_BGROUT2. Similarly, the Single-stage current mirror and the Cascaded current mirror have linear relationships between V_BGR and R_BGR with varying sensitivity slopes, at 37.236 μV/Ω and 42.246 μV/Ω, respectively¹². Due to the greater slope obtained in the cascaded current mirror structure, it may be possible to obtain proportional to absolute temperature compensation (PTAT) for smaller R_BGR values.

The stability of BGR across an operating temperature range is shown through Fig. 2b, d. Figure 2e also shows how V_{BGR_OUT1} and V_{BGR_OUT2} with varying supplies (V_DD) compared to each other. A stable bandgap reference (BGR) topology, by virtue of using cascaded current mirrors, has less sensitivity to the supply voltage noise. Supply voltage is changed from 0 to 4 V, simulations are carried at various process corners as well as over a − 40 °C and 125 °C temperature range making use of the 65 nm CMOS process.

Figure 3 shows the effect of startup resistance variations on the BGR circuit output and the startup transistor biasing characteristic^13,16. In Fig. 3a, the BGR output voltage (V_BGROUT1) is studied versus various startup resistance (Rs) values. For lower Rs values (regions S1 and S2), the output is shifted away from the expected value and remains at around 3.2 V, which shows proper startup functioning, due to improper biasing of NM4. Thus, it is operated in a linear region as given in Fig. 3c. Nevertheless, as Rs increases above a critical value (around 20kΩ), the BGR output begins producing a constant 1.3 V (from Fig. 3b) (regions S3 to S5), which indicates that the startup circuit successfully biases the BGR properly at high resistance values, when NM4 is biased to run in the saturation region. Figure 3c shows the biasing voltage (V_GS) of the startup transistor NM4 versus different Rs values. At the beginning, NM4 is in the linear and saturation regions, allowing proper circuit startup. However, as Rs increases above a certain value, NM4 switches to the cutoff region, disabling the startup circuit and causing startup failure. This analysis verifies that choosing an optimum startup resistance is crucial in ensuring reliable BGR operation.

As shown in Fig. 4a, without the startup circuit, BGR fails to initialize correctly and stays in a metastable state, while with the startup circuit, it reaches the correct operating voltage of 1.3 V, as shown in Fig. 4b. A startup circuit is required to pull a small current at the beginning of operation in order to force the BGR into its correct operational state in Fig. 4c. Once the startup circuit is turned on, the startup transistor (NM4) pulls down on the V_X4 node, which turns on the PMOS load transistors (M1, M4), taking the BGR from the zero-current state into the active operating region. During this stage, the biasing voltage VNM4 for the NM4 is very high which guarantees its conduction. The startup circuit must turn off once BGR settles so that unnecessary power dissipation does not take place when NM5 is turned on by a high potential at V_X3, pulling down the N4 node, consequently turning NM4 off and ensuring that the startup circuit is disabled after initialization. After stabilization of the BGR core, the startup circuit acts as a normal turn-off circuit that prevents any unnecessary power from being consumed, as is seen from the startup transistor (I_NM4) reverting to zero at 120 µs.

Figure 5 presents the output occurrence graphs of CM-BGR and Cascaded CM-BGR under a supply sweep at 4 V. The results indicate that the Cascaded CM-BGR is able to maintain its target output voltage of 1.05 V, thus demonstrating its stability against supply changes. On the other hand, the CM-BGR demonstrates large deviations in its output voltage, tending towards higher supply levels in the samples. Such deviations indicate that the CM-BGR is more sensitive to supply changes, thus less stable.

BGR circuit design using operational amplifier

Operational amplifiers are used in applications where there is high gain and high speed. A differential input and differential output multi-stage configuration makes the circuit highly stable. With this configuration, differential signals are amplified and common-mode signals and noise are rejected simultaneously¹⁸. Differential inputs, V+ and V −, are connected to transistors M9 and M10 as shown in Fig. 6. A differential voltage can be translated into a differential current by these transistors, which form a differential pair. Transistors M1, M2, M3, and M4 also form the current mirror circuit, which provides the active load impedance of the differential pair.

The first stage consists of the differential pair (M5, M6) and the current mirror load (M1, M2). The gain of this stage is

A2 is the gain of the second stage (Common Source Amplifier)

Figure 7a, b illustrates how the phase response of the circuit changes when frequency is altered. Phase margin, or the phase shift from − 180°, is highly significant in establishing the stability of the amplifier.

At differential gain, differential gain is the difference between the Vout gain and the gain of V+, V−, with intermediate nodes (Vy1, Vz1) considered. Ripples and peaks at high frequencies indicate parasitic effects or insufficient compensation. A sudden voltage gain vs. frequency spike in Fig. 7c might indicate resonances or noise coupling within the circuit. A plot of CMRR (Fig. 7d) illustrates the difference between the differential signal and the interference resulting from common-mode signals. The higher the CMRR, the higher the rejection by a differential amplifier of common-mode signals and noise.

A BGR circuit generates stable voltage that is independent of temperature, supply voltage, and process variations. Q1, Q2, Q3, Q4, Q5 and Q6 generates the CTAT voltages (VBE1, VBE2). R1 determines the current I1, which is proportional to the voltage difference VBE1–VBE2. R2 scales the PTAT current to generate the required voltage at the output. The operational amplifier enforces a virtual short condition, ensuring that the voltages at its inputs are equal¹⁹.

Operational Amplifier-based Bandgap Reference (BGR) circuit functions based on the production of a process-insensitive and temperature-stable reference voltage^7,20,21. The operational principle of this circuit is the integration of two voltage terms with opposite temperature coefficients: the base-emitter voltage (VBE) of a bipolar junction transistor (BJT), with a negative temperature coefficient, and thermal voltage (Vt)⁷. The nomenclature is derived from the resistor network that exhibits a positive temperature coefficient. The operational amplifier provides for proper biasing by equating the voltages at its input terminals (V_BE1, V_X), thus ensuring the desired current flow through the BJT network as shown in the Fig. 8a. The current is copied across multiple transistors (M0, M1, M2) to provide a voltage proportional to absolute temperature (PTAT), which is then added to the complementary-to-absolute-temperature (CTAT) results, producing an effectively temperature-insensitive output voltage. The resistive ratio determines the PTAT voltage value, allowing precise reference voltage adjustment, typically to 1.2 V. MOSFET-based current sources offer a stable bias condition, with minimal supply voltage sensitivity. Figure 8b shows the relation between R_BGR and slope of V_BGROUT3 for the range of temperature − 40 °C to 125 °C. R_BGR is adjusted to 139KΩ bring PTAT strength which equals to CTAT. The simulation results demonstrate the range of R_BGR for which slope of V_BGROUT3 is constant²².

Sub-BGR circuit design using operational amplifier

The circuit shown in Fig. 9a is a sub-BGR, which produces constant reference voltage through current summing technique^5,9. The sub-BGR is achieved using bipolar junction transistors (BJTs) (Q1–Q5) to produce base-emitter voltages (V_BE) with negative temperature coefficient (CTAT). The R1 resistor is used to produce a current with positive temperature coefficient, since the differential CTAT voltage (V_BE1–V_BE2) produces a voltage drop across it, which essentially produces a proportional-to-absolute-temperature (PTAT) current. Since V_X is a CTAT voltage, the voltage drops across R2 produces a CTAT current. The sum of PTAT and CTAT currents flow through M1 to make the resulting overall current constant²³. M2 mirrors the constant current to resistor R0, where the combined CTAT and PTAT components produce a stable reference voltage at V_BGROUTC1. A PMOS current mirror (M1) supplies stable bias currents to facilitate proper circuit operation. An operational amplifier keeps equal voltages (V_BE1 and V_X) at its inputs to facilitate proper generation of the PTAT current. The PTAT current, produced by the difference between the base-emitter voltages (V_BE1 and V_BE2) of the BJTs and scaled by resistor R1, is summed up with the CTAT current at the output node to facilitate temperature-compensated reference current²⁴. This current, when passed through RBGR, produces a stable output voltage (V_BGROUTC1).

Figure 9b plots the output voltage (V_BGROUTC1) variation with temperature for different R_PTAT values, which reflects the effect of R_PTAT on temperature stability and the operation of the bandgap reference. The ideal bandgap reference should have a stable output voltage irrespective of the temperature change, with proper temperature compensation¹⁹. When R_PTAT is scaled PTAT behavior is dominated and slope of output voltage (V_BGROUTC1) is also changes, thus R_PTAT is determined based on the V_BGROUTC1 slope.

A century ago, the County Limerick village of Kilteely had seven pubs but one by one they shut. This year, it braced to lose the last.

The economic and social trends that have shuttered family-run pubs across Ireland appear remorseless, leaving many communities with nowhere to meet, have a drink and share stories.

But this week Kilteely bucked the trend. The doomed bar reopened after 26 villagers clubbed together to buy it. “We felt we were going to be annexed into other communities if we didn’t have a place to meet and call our own,” said Liam Carroll. “So here we are, we’re publicans.”

The new owners pooled their savings and formed a syndicate to buy the pub, which otherwise faced probable demolition and conversion into accommodation.

The eclectic group – which includes a barrister, a solicitor, a pharmacist, a clinical psychologist, a carpenter, an accountant, a teacher, a signmaker, builders, farmers and electricians – bought the bar and licence for €300,000 (£260,000) and used its varied skills to reestablish and refurbish the business.

Previously Ahern’s, the pub is now called the Street Bar, a reference to the local expression “heading up the street”, a euphemism for heading out for a pint. (Some syndicate members wanted to call it the Ambush, after a famous 1921 attack during the British-Irish war that killed 11 soldiers and police, but that was vetoed.)

It has new wiring, a cool room for beer, Sky Sports and, tucked between wine and whiskey, a sign: “Welcome to the Street Bar. A community working together in Kilteely.” Another sign lists syndicate members.

“We hope other communities will see that it can be done,” said Carroll, the barrister. “All these closures – it doesn’t have to happen.”

Since 2005, Ireland has lost a quarter of its pubs: more than 2,100, averaging 112 a year, according to a study commissioned by the Drinks Industry Group of Ireland. The phenomenon is highest in rural areas. At county level, Limerick, in the south-west, recorded the highest decrease of 37.2%.

Multiple reasons are cited: the cost of living, high taxes on alcohol, drink-driving laws, young people drinking less, preferences to drink at home, the Covid pandemic and shrinking profit margins. A similar trend has closed 15,000 pubs in England, Wales and Scotland since 2000, according to the British Beer and Pub Association.

Having already lost post offices, shops and other pubs, the closure of Ahern’s – which had been kept afloat by Noreen Ahern, who is near retirement age, working an estimated 90 hours a week – would have left Kilteely’s main street empty and abandoned, save for a recycling business.

Rather than mourn, some residents proposed a rescue based on an example from the County Waterford village of Rathgormack: in 2021, 19 local people formed a syndicate to buy and run the last pub, which was due to close permanently.

A group of 20 Kilteely residents each contributed €15,000 to meet the €300,000 price and turn the pub into a private limited company with a social enterprise ethos. “We made clear to all who invested that they should expect not to see a return of their money,” said Carroll.

skip past newsletter promotion

after newsletter promotion

The group divided tasks such as paperwork, accounts, electrics and carpentry according to skillset, said Gerry O’Dea, 54, a farmer and financial adviser. “Everybody brought something to the table.”

Heeding advice from the Rathgormack syndicate, the Kilteely group separated ownership and management. “You couldn’t run a pub where you have 20 people with different opinions about the price of drinks,” said Carroll.

The shareholders appointed a five-member board that in turn hired a manager who, unlike them, had experience running a pub. “You hire the best people and get out of their way,” said Eoin English, 50, an engineer. The pub would remain a gathering point, he said. “This is where people have birthday parties or post-funeral receptions.”

Daniel Kreith, 29, a solicitor and syndicate member originally from Galway, said his home village had lost nine of its 13 pubs. Kilteely showed that decline and oblivion was not inevitable, he said: “Some of them could have been saved.”

The model appears to be spreading – County Kerry’s first community-owned pub is due to open soon.

Bosco Ryan, 56, another Kilteely syndicate member, said stakeholders and their friends and families formed a network that could help sustain the Street. “We all have a responsibility to support it.”

The initial search yielded a total of 150 studies across four databases. After removing duplicates, 120 unique records were screened based on titles and abstracts. Of these, 30 full-text articles were assessed for eligibility. Following a full-text review, 14 studies were included in the final synthesis from PubMed, Scopus, Web of Science, and Google Scholar (Fig. 1). The remaining studies were excluded for reasons including lack of educational context, absence of AI-related content, or unavailability of full text. A summary of the study selection process is presented in the PRISMA flow diagram (Fig. 1).

Study characteristics

The narrative review includes 14 studies covering diverse geographic locations including the United States, India, Australia, Germany, Saudi Arabia, and multi-national collaborations. The selected studies employed a mix of methodologies: scoping reviews, narrative reviews, cross-sectional surveys, educational interventions, integrative reviews, and qualitative case studies. The target population across studies ranged from undergraduate medical students and postgraduate trainees to faculty members and in-service professionals.

Artificial Intelligence (AI) was applied across various educational contexts, including admissions, diagnostics, teaching and assessment, clinical decision-making, and curriculum development. Several studies focused on stakeholder perceptions, ethical implications, and the need for standardized curricular frameworks. Notably, interventions such as the Four-Week Modular AI Elective [13] and the Four-Dimensional AI Literacy Framework [12] were evaluated for their impact on learner outcomes.

Table 1 provides a comprehensive summary of each study, outlining country/region, study type, education level targeted, AI application domain, frameworks or interventions used, major outcomes, barriers to implementation, and ethical concerns addressed.

Risk of bias assessment

A comprehensive risk of bias assessment was conducted using appropriate tools tailored to each study design. For systematic and scoping reviews (e.g., Gordon et al. [2], Khalifa & Albadawy [8], Crotty et al. [11]), the AMSTAR 2 tool was applied, revealing a moderate risk of bias, primarily due to the lack of formal appraisal of included studies and incomplete reporting on funding sources. Observational studies such as that by Parsaiyan & Mansouri [9] were assessed using the Newcastle-Ottawa Scale (NOS) and showed a low risk of bias, with clear selection methods and outcome assessment. For cross-sectional survey designs (e.g., Narayanan et al. [10], Ma et al. [12], Wood et al. [14], Salih [20]), the AXIS tool was used. These showed low to moderate risk depending on sampling clarity, non-response bias, and data reporting. Qualitative and mixed-methods studies such as those by Krive et al. [13] and Weidener & Fischer [15] were appraised using a combination of the CASP checklist and NOS, showing overall low to moderate risk, particularly for their methodological rigor and triangulation. One study [19], which employed a quasi-experimental design, was evaluated using ROBINS-I and was found to have a moderate risk of bias, primarily due to concerns about confounding and deviations from intended interventions. Lastly, narrative reviews like Mondal & Mondal [17] were categorized as high risk due to their lack of systematic methodology and critical appraisal Table 2.

Characteristics of included studies

A total of 14 studies were included in this systematic review, published between 2019 and 2024. These comprised a range of study designs: 5 systematic or scoping reviews, 4 cross-sectional survey studies, 2 mixed-methods or qualitative studies, 1 quasi-experimental study, 1 narrative review, and 1 conceptual framework development paper. The majority of the studies were conducted in high-income countries, particularly the United States, United Kingdom, and Canada, while others included contributions from Asia and Europe, highlighting a growing global interest in the integration of artificial intelligence (AI) in medical education.

The key themes addressed across these studies included: the use of AI for enhancing clinical reasoning and decision-making skills, curriculum integration of AI tools, attitudes and readiness of faculty and students, AI-based educational interventions and simulations, and ethical and regulatory considerations in AI-driven learning. Sample sizes in survey-based studies ranged from fewer than 100 to over 1,000 participants, representing diverse medical student populations and teaching faculty.

All included studies explored the potential of AI to transform undergraduate and postgraduate medical education through improved personalization, automation of feedback, and development of clinical competencies. However, variability in methodology, focus, and outcome reporting was observed, reinforcing the importance of structured synthesis and cautious interpretation.

1. A.

   Applications of AI in Medical Education

AI serves multiple educational functions. Gordon et al. identified its use in admissions, diagnostics, assessments, clinical simulations, and predictive analytics [2]. Khalifa and Albadawy reported improvements in diagnostic imaging accuracy and workflow efficiency [8]. Narrative reviews by Parsaiyan et al. [9] and Narayanan et al. [10] highlighted AI’s impact on virtual simulations, personalized learning, and competency-based education.

1. B.

   Curricular innovations and interventions

Several studies introduced innovative curricular designs. Crotty et al. advocated for a modular curriculum incorporating machine learning, ethics, and governance [11], while Ma et al. proposed a Four-Dimensional Framework to cultivate AI literacy [12]. Krive et al. [13] reported significant learning gains through a four-week elective, emphasizing the value of early, practical exposure.

Studies evaluating AI-focused educational interventions primarily reported improvements in knowledge acquisition, diagnostic reasoning, and ethical awareness. For instance, Krive et al. [13] documented substantial gains in students’ ability to apply AI in clinical settings, with average quiz and assignment scores of 97% and 89%, respectively. Ma et al. highlighted enhanced conceptual understanding through their framework, though outcomes were primarily self-reported [12]. However, few studies included objective or longitudinal assessments of educational impact. None evaluated whether improvements were sustained over time or translated into clinical behavior or patient care. This reveals a critical gap and underscores the need for robust, multi-phase evaluation of AI education interventions.

1. C.

   Stakeholder perceptions

Both students and faculty showed interest and concern about AI integration. Wood et al. [14] and Weidener and Fischer [15] noted a scarcity of formal training opportunities, despite growing awareness of AI’s importance. Ethical dilemmas, fears of job displacement, and insufficient preparation emerged as key concerns.

1. D.

   Ethical and regulatory challenges

Critical ethical issues were raised by Mennella et al. [16] and Mondal and Mondal [17], focusing on data privacy, transparency, and patient autonomy. Multiple studies called for international regulatory standards and the embedding of AI ethics within core curricula.

While several reviewed studies acknowledged the importance of ethical training in AI, the discussion of ethics often remained surface-level. A more critical lens reveals deeper tensions that must be addressed in AI-integrated medical education. One such tension lies between technological innovation and equity AI tools, if not designed and deployed with care, risk widening disparities by favoring data-rich, high-resource settings while neglecting underrepresented populations. Moreover, AI’s potential to entrench existing biases—due to skewed training datasets or uncritical deployment of algorithms—poses a threat to fair and inclusive healthcare delivery.

Another pressing concern is algorithmic opacity. As future physicians are expected to work alongside AI systems in high-stakes clinical decisions, the inability to fully understand or challenge these systems’ inner workings raises accountability dilemmas and undermines trust. Educational interventions must therefore go beyond theoretical awareness and cultivate critical engagement with the socio-technical dimensions of AI, emphasizing ethical reasoning, bias recognition, and equity-oriented decision-making.

1. E.

   Barriers to implementation

Implementation hurdles included limited empirical evidence [18], infrastructural constraints [19], context-specific applicability challenges [20], and an over-reliance on conceptual frameworks [10]. The lack of unified teaching models and outcome-based assessments remains a significant obstacle.

These findings informed the creation of a conceptual framework for integrating artificial intelligence into medical education, depicted in Fig. 1. A cross-theme synthesis revealed that while AI integration strategies were broadly similar across countries, their implementation success varied significantly by geographic and economic context. High-income countries (e.g., USA, Australia, Germany) demonstrated more comprehensive curricular pilots, infrastructure support, and faculty readiness, whereas studies from LMICs (e.g., India, Saudi Arabia) emphasized conceptual interest but lacked institutional capacity and access to AI technologies. Contextual barriers such as resource limitations, cultural sensitivity, and institutional inertia appeared more pronounced in LMIC settings, influencing the feasibility and depth of AI adoption in medical education.

Based on the five synthesized themes, we developed a Comprehensive Framework for the Strategic Integration of AI in Medical Education (Fig. 2). This model incorporates components such as foundational AI literacy, ethical preparedness, faculty development, curriculum redesign, and contextual adaptability. It builds on and extends existing models such as the FACETS framework, the Technology Acceptance Model (TAM), and the Diffusion of Innovation theory. Unlike FACETS, which primarily categorizes existing studies, our framework is action-oriented and aligned with Kern’s curriculum development process, making it suitable for practical implementation. Compared to TAM and Diffusion of Innovation, which focus on user behavior and adoption dynamics, our model integrates educational design elements with implementation feasibility across diverse economic and institutional settings.

Table 3 shows a comparative synthesis of included studies evaluating AI integration in medical and health professions education using Kern’s six-step curriculum development framework. The analysis reveals that most studies effectively identify the need for AI literacy (Step 1) and conduct some form of needs assessment (Step 2), often through surveys, literature reviews, or scoping exercises. However, only a subset of studies explicitly define measurable educational goals and objectives (Step 3), and even fewer describe detailed instructional strategies (Step 4) or implement their proposed curricula (Step 5). Evaluation and feedback mechanisms (Step 6) were rarely reported, and when included, they were typically limited to short-term student feedback or pre-post knowledge assessments. Longitudinal evaluations and outcome-based assessments remain largely absent. The findings underscore a critical implementation gap and emphasize the need for structured, theory-informed, and empirically evaluated AI education models tailored to medical and allied health curricula.

This conceptual model is informed by thematic synthesis and integrates principles from existing frameworks (FACETS, TAM, Diffusion of Innovation) while aligning with Kern’s six-step approach for curriculum design.

In a new and dramatic move, Meta CEO Mark Zuckerberg is making the most aggressive move in his AI arm race. Zuckerberg is now reorganising the company’s artificial intelligence operations by restricting the Meta Superintelligence Labs (MSL) into four distinct teams. As reported by Business Insider the company outlined the move in an internal memo from the newly appointed chief of AI Alexandr Wang. The restructuring is taking place after the aggressive hiring process in which Meta poached dozens of top AI researchers from its rivals companies. The internal email, sent by Alexandr Wang, the 28-year-old head of Meta Superintelligence Labs (MSL), outlines a dramatic restructuring aimed at accelerating Meta’s pursuit of “personal superintelligence”—AI that can outperform humans across intellectual domains.

Four pillars of Meta’s new AI strategy

In the internal memo shared with employees, Wang talks about the four specialised teams:

Most of these leaders now report directly to Wang, signaling a centralization of power within MSL.

FAIR and TBD: Meta’s innovation engine

FAIR, led by Rob Fergus and chief scientist Yann LeCun will now play an important role in the development of MSL’s model. Whereas, TBD Lab will look for new directions, including the enigmatic “omni” model—believed to be a multimodal system capable of understanding text, audio, video, and more.The research wing at MSL will be headed by Shengjia Zhao, co-creator of ChatGPT, who notably does not report directly to Wang.

Read Alexandr Wang’s full memo here

Superintelligence is coming, and in order to take it seriously, we need to organize around the key areas that will be critical to reach it — research, product and infra. We are building a world-class organization around these areas, and have brought in some incredible leaders to drive the work forward.

As we previously announced, Shengjia Zhao will direct our research efforts as Chief Scientist for MSL, Nat Friedman will lead our product effort and Rob Fergus will continue to lead FAIR. Today, I’m pleased to announce that Aparna Ramani will be moving over to MSL to lead the infrastructure necessary to support our ambitious research and product bets.

As part of this, we are dissolving the AGI Foundations organization and moving the talent from that team into the right areas. Teams whose work naturally aligns with and serves our products will move to Nat’s team. Some of the researchers will move to FAIR to double down on our long term research while teams working on infra will transition into Aparna’s org. Anyone who is changing teams will get an update from their manager or HRBP today, if you haven’t already.

We’re making three key changes to our organizational design that will help us to accelerate our efforts.

The work will map to four teams:

TBD Lab will be a small team focused on training and scaling large models to achieve superintelligence across pre-training, reasoning, and post-training, and explore new directions such as an omni model.

FAIR will be an innovation engine for MSL and we will aim to integrate and scale many of the research ideas and projects from FAIR into the larger model runs conducted by TBD Lab. Rob will continue to lead FAIR and Yann will continue to serve as Chief Scientist for FAIR, with both reporting to me.

Products & Applied Research will bring our product-focused research efforts closer to product development. This will include teams previously working on Assistant, Voice, Media, Trust, Embodiment and Developer pillars in AI Tech. Nat will continue to lead this work reporting to me.

MSL Infra team will unify elements of Infra and MSL’s infrastructure teams into one. This team will focus on accelerating AI research and production by building advanced infrastructure, optimized GPU clusters, comprehensive environments, data infrastructure, and developer tools to support state-of-the-art research, products and AI development across Meta. Aparna will lead this team reporting to me.

Ahmad and Amir will continue reporting to me focusing on strategic MSL initiatives they will share more about later.

I recognize that org changes can be disruptive, but I truly believe that taking the time to get this structure right now will allow us to reach superintelligence with more velocity over the long term. We’re still working through updated rhythms and our collaboration model across teams, including when we’ll come together as a full MSL org.

Thank you all for your flexibility as we adapt to this new structure. Every team in MSL plays a critical role and I’m excited to get to work with all of you.

Obesity is a complex disease that not only affects an individual’s appearance but also has multifaceted negative impacts on physical health, including cardiovascular diseases, type 2 diabetes, hypertension, fatty liver disease, and certain types of cancer¹. Additionally, obesity contributes to psychological distress, increasing the risk of depression and anxiety, and imposes a significant economic burden on public health systems^2,3. Given the rising prevalence of obesity, understanding how consumers estimate food calories and make dietary choices is essential for both public health and marketing strategies. Previous research has strongly linked obesity with the underestimation of food calories, highlighting the need for interventions that help consumers make more informed choices^4,5,6.

Food choices are influenced by various factors, including economic conditions, sociocultural influences, psychological aspects, marketing strategies, and food packaging design^7,8,9. Among these, food packaging color plays a crucial role in shaping consumer perceptions and decision-making. Colors can act as implicit cues that influence caloric estimations, thereby affecting consumers’ food selections and dietary behaviors¹⁰. As a key element of sensory marketing, packaging color has the potential to nudge consumers toward healthier choices, making it a valuable tool for both businesses and policymakers.

Despite extensive research on food packaging and consumer behavior, there remains a significant research gap regarding how color influences calorie estimation and the mediating role of perceived healthiness in this relationship. Furthermore, most existing studies rely on self-reported surveys or controlled laboratory experiments, which may lack ecological validity. This study uses virtual reality technology to investigate how food packaging color (red vs. green) influences consumers’ calorie estimations and whether perceived healthiness mediates this effect. By uncovering these mechanisms, this research provides insights that are relevant not only to public health interventions but also to food industry marketing strategies and regulatory policies. Specifically, these findings can inform front-of-pack labeling regulations and guide food manufacturers in designing health-oriented packaging that encourages better consumer choices^11,12. Given the increasing emphasis on behavioral nudges in public policy, understanding the impact of color-coded cues on food perception can help develop more effective marketing and regulatory frameworks, ultimately contributing to obesity prevention and healthier consumer behavior.

Color, caloric estimation, and perceived healthiness

Color is one of the most immediate and influential visual cues in food packaging, shaping consumer perceptions and decision-making processes. Previous research has highlighted the significant impact of front-of-pack (FOP) visual cues on consumer food choices and eating behaviors, emphasizing the role of packaging color as a critical element in shaping dietary decisions¹³. Consumers frequently rely on packaging color to infer the healthiness and caloric content of food rather than consulting detailed nutritional labels^5,14,15. However, the specific mechanisms by which color affects food health perception and caloric estimation remain unclear, particularly regarding the mediating role of perceived healthiness in this relationship.

Food packaging color plays a crucial role in shaping consumer perceptions of healthiness. Research has shown that color-coded packaging influences consumer expectations about food attributes, often more strongly than textual nutritional information¹⁶. Green is typically associated with health, natural ingredients, and lower calories, whereas red is often linked to high-calorie, indulgent, and less healthy foods^17,18. These associations arise not only from everyday experiences but also from strategic marketing practices. Many health-oriented products use green packaging to emphasize their nutritional benefits, while high-calorie snacks or fast foods often employ red to attract attention and stimulate impulse purchases^19,20. Moreover, color influences sensory expectations, with red enhancing perceptions of richness and indulgence, whereas green can make food seem healthier but potentially less flavorful²¹. Recent research has expanded from examining singular color effects to investigating the interactions between color and other visual cues, such as shape and labeling²². Recent research by Hallez et al. (2023) further supports the significant role of packaging color in shaping consumer perceptions of healthiness, sustainability, and taste among young consumers. Their study demonstrates that color, in interaction with packaging claims, influences product evaluations, with findings closely aligning with the current research on color-driven health perceptions. This highlights the importance of considering interactive effects between color and other visual or textual cues in food packaging design, an area warranting further exploration in the context of calorie estimation²³. For example, Grunert & Wills (2007) found that consumers prefer simplified front-of-pack labeling formats but differ in their reliance on color-based health cues depending on the product category and context²⁴. However, the precise mechanisms by which color influences food health perception require further empirical validation.

Studies suggest that consumer perceptions of a food product’s healthiness significantly influence their caloric judgments²⁵. Prior research indicates that consumers tend to overestimate the caloric content of foods they perceive as unhealthy while underestimating those they deem healthy²⁵. This phenomenon has been observed in various contexts; for example, when consumers compare two identical yogurt products labeled as “low-fat” and “full-fat,” they consistently judge the full-fat yogurt as having higher calories, despite both products containing the same caloric content²⁶. Similarly, fast food consumers—especially those frequenting health-branded chains such as Subway—tend to underestimate the caloric content of meals, leading to higher caloric intake²⁷. In one study, participants were asked to evaluate the healthiness and caloric content of eight different foods. The results showed that foods perceived as healthy or beneficial for weight loss were typically underestimated in caloric content, whereas foods perceived as unhealthy or detrimental to weight loss were overestimated¹⁷. Furthermore, the health-oriented branding of a restaurant can lead consumers to underestimate the caloric content of its meals, resulting in increased caloric intake¹⁴. These findings indicate that consumers do not estimate calorie content directly based on the food itself but rather indirectly through their judgments of healthiness. Thus, food packaging color may first influence health perception, which in turn affects caloric estimation^19,20. While perceived healthiness has been widely studied in consumer decision-making, its specific mediating role in the color–caloric estimation relationship remains underexplored.

Overall, food packaging color may influence caloric estimation by first shaping consumers’ perceptions of healthiness. However, most existing research has focused on color’s impact on food choice and consumption behavior, with limited direct investigation into how color affects caloric estimation²². Additionally, prior studies have relied primarily on self-reported surveys or controlled laboratory experiments, which may lack ecological validity. With the development of Virtual Reality (VR) technology, this research integrates VR to further enrich and develop its application in consumer behavior studies. Compared to traditional research methodologies, VR technology offers low costs, reusability of experimental scenarios, and immersive experiences that more accurately replicate real-world environments²⁸, thereby holding significant advantages in consumer research. Therefore, this experiment leverages VR technology to perform experimental manipulations, aiming to explore the underlying mechanisms more profoundly.

Theoretical framework: association theory and embodied cognition theory

Sensory marketing research frequently utilizes Association Theory and Embodied Cognition Theory to explain how sensory cues influence consumer perceptions and behaviors. These theories provide a foundation for understanding how color cues (red vs. green) affect consumers’ calorie estimations through their perceptions of healthiness.

Association Theory posits that repeated co-occurrences of stimuli lead to learned associations, allowing consumers to anticipate one event based on the presence of another²⁹. In food packaging, red and green are commonly used in health-related messaging, shaping consumer expectations. Red is often associated with cautionary signals, energy, and intensity, while green is frequently linked to health, natural ingredients, and lower-calorie options. These learned associations influence consumer judgments, leading to expectations that green-packaged foods are healthier and lower in calories, whereas red-packaged foods may be perceived as less healthy and higher in calories. This heuristic processing may guide consumer decision-making, particularly in contexts where detailed nutritional information is not immediately considered.

Embodied Cognition Theory, in contrast, suggests that cognitive processes are deeply rooted in bodily interactions with the environment, meaning that sensory experiences—such as visual exposure to color—can directly shape cognitive evaluations and perceptions. Research has demonstrated that bodily perceptions significantly impact consumer experiences; for instance, tactile sensations influence service perception, with soft textures enhancing tolerance for service failures³⁰, while rough textures evoke empathy and generosity³¹. Similarly, product shape affects size perception, as consumers tend to perceive round pizzas as smaller than square pizzas of the same surface area³². In the context of food perception, color can elicit immediate cognitive and affective responses, influencing how consumers assess healthiness and caloric content. Red may enhance perceptions of higher energy content, whereas green may reinforce associations with health and lower caloric density.

By integrating Association Theory and Embodied Cognition Theory, this study examines how packaging color influences consumers’ calorie estimations through their perceptions of healthiness, providing insight into the cognitive mechanisms underlying food-related judgments.

Research objectives

This study aims to investigate how food packaging color (red vs. green) influences consumers’ calorie estimations and whether perceived healthiness mediates this effect. While previous research has demonstrated that color cues shape consumer perceptions, the underlying cognitive mechanisms—particularly the mediating role of perceived healthiness—remain underexplored. Drawing on Association Theory and Embodied Cognition Theory, this research seeks to clarify how learned associations (e.g., red with unhealthiness and green with health) and direct sensory experiences influence food-related judgments.

Specifically, this study examines whether red packaging leads to higher calorie estimations and green packaging leads to lower ones, particularly in the context of unhealthy foods. Since consumers often rely on heuristic cues, color may serve as a visual shortcut for evaluating a product’s healthiness, subsequently influencing calorie estimation. However, if the association between color and healthiness is disrupted—such as through an experimental manipulation where color is framed as unrelated to health—the effect of color on calorie estimation should diminish. Additionally, this research explores whether altering the color-health association affects consumers’ food choices, potentially leading to an increase in food selection when color no longer serves as a heuristic signal for health.

To address these research objectives, the following hypotheses are proposed:

H1: For unhealthy foods, red packaging will increase consumer calorie estimations compared to green packaging.

H2: Perceived healthiness mediates the effect of packaging color on calorie estimation of unhealthy foods.

H3: For unhealthy foods, the association of red with unhealthiness will lead to higher calorie estimations of foods in red packaging. After manipulating “color unrelated to health,” the impact of color on calorie estimation will disappear in the manipulated group.

H4: Compared to the control group, the number of food choices will increase in the manipulated group where “color is unrelated to health.”

By testing these hypotheses, this study seeks to provide both theoretical contributions to sensory marketing and practical implications for food packaging design and public health communication. Understanding how color influences calorie estimation through perceived healthiness can inform strategies for promoting healthier eating behaviors and improving consumer awareness of nutritional content.

Study 1: the impact of food packaging color on calorie Estimation

Study 1 employs a single-factor between-subjects design, with the independent variable being packaging color, divided into two levels: red and green. The dependent variable is the numerical estimation of calories, and the mediating variable is the perceived level of healthiness.

Experimental objectives and hypotheses

The purpose of Study 1 is to explore the impact of food packaging color on calorie estimation and to examine the mediating mechanism of perceived healthiness. The following hypotheses are proposed for this experiment:

H1: For unhealthy foods, red packaging will increase calorie estimations compared to green packaging.

H2: Perceived healthiness mediates the effect of packaging color on calorie estimation for unhealthy foods.

Participants

Participants were recruited through the distribution of a survey link on social media platforms using Questionnaire Star. A total of 159 respondents completed the survey, including 67 males and 92 females, with an age range of 18–32 years (M = 23.55, SD = 2.46). The studies involving human participants were reviewed and approved by Fudan University Ethics Committee (FDU-SSDPP-IRB-2024-2-103). They were performed by the ethical standards laid down in the 1964 Declaration of Helsinki and its later amendments. All participants in the study provided informed consent, which involves the consent to publish their data.

Experimental materials

Food packaging images

Following previous studies, the products were categorized into two broad categories: healthy and unhealthy, with a total of six products selected (see Appendix 1).

This design allows for a controlled and focused examination of the influence of color on perceived calorie content, leveraging both visual stimuli and participant self-report measures to gather data on perception dynamics influenced by packaging color (see Fig. 1).

Healthy categories include yogurt, nuts, and fruit cereal, while unhealthy categories include potato chips, chocolate, and milk tea instant beverages. These six products were processed through Photoshop, keeping all aspects identical except for the packaging color. To maintain consistency, the RGB color values for red are: Red: 213; Green: 32; Blue: 53. The RGB color values for green are: Red: 123; Green: 225; Blue: 47.

To exclude other possible explanations such as taste appeal and attractiveness of packaging, participants were required to evaluate the taste appeal of the food, the attractiveness of the packaging, and their perception of the healthiness associated with the colors. The taste appeal was rated on a scale from 1 (not tasty at all) to 7 (very tasty). The attractiveness of the packaging was rated from 1 (not attractive at all) to 7 (very attractive).

Results (see Table 1) indicate that participants did not perceive a significant difference in taste between red and green packaging (M_red = 4.81, SD_red = 1.45; M_green = 4.80, SD_green = 1.47; t(475) = 0.05, p = 0.96), thus excluding taste appeal as another possible explanation.

Additionally, as shown in Table 2, there is no significant difference in the perceived attractiveness of the packaging between the red and green options (M_red = 4.16, SD_red = 1.50; M_green = 4.13, SD_green = 1.48; t(475) = 0.23, p = 0.82). Therefore, the potential explanation based on the attractiveness of the packaging is also ruled out.

Caloric Estimation measurement

In this experiment, participants were asked to estimate the caloric content of the displayed foods, using the calorie content of walnuts as a reference. Specifically, participants were informed: “The calorie content of walnuts we commonly consume is 574 cal per 100 g. Please estimate the calorie content per 100 g for the food items shown in the picture above.”

Measurement of perceived healthiness

Perceived healthiness refers to consumers’ immediate judgment on the healthiness of a food item. This was measured using two items: perceived healthiness of the food (1 = very unhealthy, 7 = very healthy) and perceived increase in body fat after consuming the food (1 = very little, 7 = very much). The scores for perceived increase in body fat were reverse-scored and then averaged with the perceived healthiness score to construct a healthiness variable for the food (α = 0.787), where higher scores indicate greater perceived healthiness.

Experimental procedure

Study 1 employed a questionnaire method. Participants were instructed: “Hello: Thank you very much for your participation. We are currently conducting a study on food packaging. Please fill out the questionnaire on your phone or computer based on your actual situation, which will take about 5 minutes of your time. We solemnly promise that all data is anonymous and will only be used for this research.” Participants were first required to fill out personal information, then view product images presented in the questionnaire, estimate the calorie content of the products based on the reference calorie information provided, and answer questions about their perceived healthiness of the food. Subsequently, participants evaluated the taste of the food (1 = not tasty at all, 7 = very tasty), the attractiveness of the packaging (1 = not attractive at all, 7 = very attractive), and the healthiness perception related to the color (measured by the item “Color is related to people’s physical and mental health” (1 = strongly disagree, 7 = strongly agree)). To avoid practice and fatigue effects, the order of image presentation was balanced as “unhealthy-healthy-unhealthy.”

The study distributed 159 questionnaires, retrieved 159 effective responses, and then processed the data and performed statistical analysis.