Filtering out humanity: AI-assisted internet research favors cold logic over ethos and pathos

By David Danelski, UC Riverside News

Image: Amirhosain Gazor - unsplash

Is the internet losing its soul? A collaborative study by UC Riverside computer and social scientists suggests so. As artificial intelligence increasingly answers our online questions with quick summaries and polished explanations, we may be gaining efficiency while losing something distinctly human in the process.

The study found that large language models, or LLMs, such as ChatGPT and Gemini overwhelmingly rely on logic and factual consistency when responding to subjective questions, while web pages written by humans draw from a richer mix of reasoning that includes emotion, lived experience, ethics, and personal authority.

“As people increasingly rely on AI systems for information discovery at the expense of traditional web searches, the web may gradually lose its soul and cease to reflect the human nature that has shaped it over the past 25 years,” said co-author Vagelis Hristidis, a computer scientist at UCR’s Bourns College of Engineering. “This may give rise to new information dissemination platforms in the future.”

Another co-author, Kevin Esterling, a professor of public policy and political science, added, “As humans, we’re hardwired to think that anything producing language has human cognition behind it, but this paper is showing that machines produce language that doesn’t have human qualities when it comes to reasoning and argumentation.”

The study compared how AI systems and traditional web searches respond to controversial or opinion-based questions, such as whether governments should ban fossil-fuel cars or whether the U.S. healthcare system needs major reform. It was presented this week at the ACM Web Science Conference in Braunschweig, Germany, by UCR computer science doctoral student and lead author Md Taukir Azam Chowdhury.

To conduct the study, the researchers analyzed responses from ChatGPT and Gemini alongside results from Google and Bing web searches. Using hundreds of subjective questions drawn from established research datasets, they examined not only the positions taken by the systems, but also the types of justifications used to support those positions.

The researchers classified reasoning according to Aristotle’s rhetorical triangle: logos, ethos, and pathos. Logos refers to logic and factual consistency. Ethos appeals to authority or credibility. Pathos appeals to emotions and shared human experience.

“What we found is that humans essentially use all three of those, whereas LLMs essentially only rely on logos,” Esterling said. “The way they try to persuade is different from the way humans persuade.”

Esterling said the difference becomes obvious in everyday searches.

Suppose someone wants a margarita recipe. An AI chatbot may instantly produce a competent recipe distilled from massive amounts of training data. But by bypassing culinary blogs and personal stories, users also miss the details that could make mixing and consuming the drink more memorable and rewarding.

The website Difford’s Guide, for example, offers dozens of margarita recipes divided into seven styles, including classic, fruity, floral, herbal, and spicy. Written by Simon Difford, the site also provides a history of the cocktail, tracing it back to a journalist’s discovery in the 1930s of a drink then called the “Tequila Daisy” during his travels in Mexico. (Margarita is the Spanish word for daisy.) He further documents with a 1936 newspaper report how it was created unwittingly by an Irish bartender in Tijuana who had grabbed the wrong bottle while trying to mix another drink.

LLMs, however, filter out such depth, nuance, and passion that human writers bring to the table.

The study also found that web pages contain a far broader range of reasoning styles than LLMs. Human-authored web content mixed factual arguments with moral concerns, practical consequences, emotional appeals, and storytelling. For example, a call for funding food banks may draw on the writer’s experience with childhood poverty.

AI systems, by contrast, strongly favored fact-centered explanations that “prioritize logos-type reasoning,” Esterling said.

The researchers suspect this may stem partly from the “alignment” and safety systems AI companies build into their models. These guardrails are designed to steer responses toward factual, non-harmful answers and away from controversial or emotional language.

The study also found that ChatGPT and Gemini closely resembled each other in how they answered questions, even when their responses diverged significantly from the broader diversity found on the web.

Esterling said humans communicate very differently from machines because people constantly anticipate how others will react emotionally and intellectually during conversation.

“When humans talk to each other, we can understand what the other is thinking,” he said. “There’s this kind of two-way interaction.”

Large language models, however, do not truly think about the audience in that way, he said. Instead, they generate statistically probable sequences of words based on training data and internal parameters.

“It’s not like talking to a person at all,” Esterling said. “It’s just a machine that’s predicting what words ought to be said in response to a prompt.”

The researchers warn that as more people rely on AI systems instead of traditional web searches to understand politics, health, ethics, and public affairs, society could gradually lose exposure to the messy but deeply human diversity of reasoning that shapes public discourse.

The AI-based research may be faster and more efficient. Still, it is also flatter, less personal, and less connected to the emotional and moral experiences that help people understand one another, Esterling said.

Hristidis added, “When using AI platforms instead of Web searches, we retrieve a distilled version of knowledge, constrained by the guardrails of each AI platform, and missing any human emotion or opinion diversity.”

The study’s title is “Comparing the Subjective Opinions and Justifications of LLMs and Web Search Engines.” The research team included Esterling, Hristidis, Chowdhury, and doctoral students from the Computer Science and Engineering Department: Jannat Ara Meem, and Zabir Al Nazi. Chowdhury led the experiments for this work.

This article was originally published by University of California, Riverside News and has been republished here with permission.

Reviewed by Irfan Ahmad.

Read next:

• UN report warns AI could soon use 3% of world’s electricity and more water than we need to drink

• Google’s AI Search Has Struggled With One Religious Question for Years
by External Contributor via Digital Information World

UN report warns AI could soon use 3% of world’s electricity and more water than we need to drink

Amanda Turnbull-McRae, University of Waikato

Image: Field Engineer - Pexels

One argument often used to quell concerns about the rising energy and resource demand of data centres is that artificial intelligence (AI) models will need less in the future as they improve and become more efficient.

But this seemingly logical thinking is a trap, according to a new United Nations report that quantifies the environmental costs of AI.

The report estimates that by 2030, AI’s energy use could double to consume 3% of the world’s electricity, produce emissions to equal the UK and deplete more water for cooling than the annual drinking water need of the global population.

It also anticipates the use of AI will follow an economic principle known as the “Jevons paradox”, which predicts that when technological improvements increase the efficiency of a resource, it leads to a rise, rather than a fall, in the total consumption of that resource.

The paradox is named after economist William Stanley Jevons who observed this effect with the use of coal in 19th-century England. Efficiency gains did not reduce overall consumption. Instead, the lower costs resulted in expanded use and higher overall demand.

As AI models become cheaper and more attractive, the report expects this to encourage new uses and higher volumes of use, eroding and possibly erasing any savings from efficiency advances.

To avoid falling into this trap, it lays out a roadmap for responsible AI use based on guiding principles of transparency, efficiency by design, equity and justice, lifecycle responsibility, global cooperation and sustainable use.

The scale of the problem

Last year, data centres already consumed as much electricity as Saudi Arabia, which ranks as the world’s 11th largest electricity consumer.

If electricity use doubles as projected by 2030, the associated carbon footprint would require 6.7 billion trees grown over ten years to offset this demand.

Data centres would also require 9.3 trillion litres of water and land nearly ten times the size of Mexico City.

Beyond resource use, the report also underscores the structural inequity at the heart of the AI boom, with only 32 nations hosting AI-specific cloud infrastructure and 90% of that capacity located in the US and China.

It warns of a widening digital divide between nations that build and control AI systems and those that consume them, with the latter often bearing a disproportionate environmental burden caused by mineral extraction and e-waste.

Responsible AI use

Two main forces shape AI’s operational footprint: how much we use it and how we use it.

This involves all tasks AI models perform, from text and code generation to image and video. Each of these tasks requires different levels of computational effort.

The model choice also matters as each AI system performs these task with distinct energy and environmental costs.

The report argues responsible AI requires full value-chain governance, from mineral sourcing to recycling and safe disposal.

It calls for a twinning of capability and environmental stewardship – thinking about both what AI can do for us and the protection of the natural environment.

This would mean making environmental disclosures a routine part of AI development, at both the model and task level, and incorporating projected AI demand in climate and energy planning.

Responsible AI is crucial as countries are promoting and adopting AI across government and the public sector.

In Aotearoa New Zealand, the government has launched a national AI strategy and a public service AI framework.

While the framework was informed by the OECD’s values-based AI principles, including inclusive and sustainable development, there is no requirement for environmental disclosures and no regulator compiling energy use or emissions.

Likewise in Australia, improving public services is part of the national AI plan. For example, the National Film and Sound Archive of Australia has created Bowerbird, a machine learning-enabled mass audio and video transcription engine, to document material. The Department of Veteran’s Affairs has developed a proof-of-concept tool to see whether AI can help speed up the processing of claims.

Both countries take a deliberate “light touch” and principles-based regulatory approach to AI. But this approach risks overlooking the growing environmental cost of AI that can’t be solved by improving it.

The natural environment is foundational to the economy, culture and wellbeing. It should be at the centre of our thinking. It’s time to rethink the AI innovation playbook and shift focus toward a sustainable tech future.

Amanda Turnbull-McRae, Senior Lecturer in Law, University of Waikato

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Reviewed by Irfan Ahmad.

Read next: 

• Fast deliveries worsen conditions for e-commerce warehouse workers

• Google Says AI Overviews Has Over 2.5 Billion Monthly Active Users as New Website Controls Roll Out

by External Contributor via Digital Information World

Fast deliveries worsen conditions for e-commerce warehouse workers

By James Dean, Cornell Chronicle

Image: Daniel Bernard - unsplash

Holding off on that late-night online order of a book, blender or blue jeans could ease the strain on a warehouse worker.

Consumers’ around-the-clock, often impulsive demand for cheap, rapidly delivered products creates harsher working conditions in e-commerce fulfillment centers than in traditional warehouses, according to Cornell-led research that provides the first comprehensive assessment of e-commerce work in the U.S.

Between Amazon and Walmart, the nation’s two largest warehouse employers, surveys found jobs at Amazon fulfillment centers to be significantly worse – more intense and dangerous – likely driven by the e-commerce market leader’s emphasis on fast delivery.

The findings are cause for concern if Amazon’s practices become the norm, the researchers said – but also show alternative approaches are viable.

“E-commerce heightens the frenzy among retailers to satisfy customers with convenience, speed and cheap prices, offerings that all erode job quality for workers behind the scenes,” said Alexander Kowalski, assistant professor of human resource studies in the ILR School. “This is a problematic trend for a large and growing share of the labor market – but it doesn’t have to be this way.”

Kowalski is the first author of “At the Mercy of the Market: E-Commerce and Warehouse Work in the United States,” published May 19 in the ILR Review. Co-authors are Sanjay Pinto and Beth Gutelius, senior research fellows at the University of Illinois-Chicago’s Center for Urban Economic Development; and Steven Vallas, professor emeritus of sociology at Northeastern University.

In related papers, Kowalski has found shortcomings in so-called “collaborative robots” being deployed in e-commerce warehouses, and that a program giving e-commerce workers more of a voice improved their well-being.

Research about the workers who enable e-commerce has highlighted reasons such jobs may be more challenging: They involve more varied inventory, more intensive labor, less predictable demand and more frequent returns. In response, workers including material movers, pickers, packers and clerks earn low wages and may face unpredictable schedules, monotonous tasks, limited opportunities for breaks, digital surveillance and discipline imposed by algorithms.

Research, however, has largely focused on Amazon, leaving it unclear whether job quality there represented warehousing broadly, the e-commerce sector or a particular corporate strategy. To address that question, Kowalski’s team first sought to establish a big-picture view of U.S. warehouse work, which has not been documented comprehensively despite enormous growth – tripling over the past 20 years, as e-commerce increased to 16.4% of retail sales.

The team surveyed about 400 hourly employees representative of the U.S. population of warehouse workers. Some supported e-commerce fulfillment centers primarily reacting to consumer orders, called business-to-consumer (B2C) warehouses; others worked in distribution and sorting centers serving brick-and-mortar stores or manufacturers, called business-to-business (B2B) warehouses.

The results confirmed a “B2C Effect”: Warehouse jobs were challenging on average, but worse at B2C – largely e-commerce – sites. Those workers experienced more pressure to move quickly and avoid taking breaks, greater exposure to unsafe situations and significantly worse overall well-being – a measure of anxiety, burnout and stress – but were not paid more.

“The warehouse segment dedicated to e-commerce offers lower-quality jobs across the board,” Kowalski said. “The results suggest that more time-sensitive pressures related to speedy delivery make for especially low-quality jobs in e-commerce.”

To investigate if those issues were consistent across retailers, the researchers next surveyed warehouse employees at Amazon (1,450 respondents) and Walmart (450 respondents), which respectively accounted for 38% and 6% of U.S. online shopping in 2023.

The results showed Amazon’s fulfillment centers housed the least desirable warehousing jobs. Respondents working in those facilities were more likely to report higher intensity, less opportunity, lower wages, greater unfairness and more challenges concerning safety, well-being and injuries. Walmart’s baseline job was not high-quality but varied little by warehouse type, whereas Amazon’s e-commerce jobs – the majority of its warehouse workforce – were significantly worse than comparable Walmart jobs, the researchers said. The findings suggest Walmart’s emphasis on low prices imposes less burden on e-commerce warehouse workers than Amazon’s focus on delivery speed.

“E-commerce creates downward pressure on work conditions,” Kowalski said, “and Amazon dials it up even more.”

The contrast between e-commerce’s biggest players, however, implies that Amazon’s lower job quality is not inevitable, but flows from its strategic priorities. Improving job quality will require worker movements, regulatory policies, consumer awareness and businesses seeing value in doing things differently, the researchers said. Kowalski pointed to his research showing better outcomes when workers are more involved in decision-making as further evidence that e-commerce can be managed less punitively.

“We are going down a path that should be worrisome for workers and for consumers alike,” he said. “But if there is a silver lining, it’s that one can arrange e-commerce operations in a way that shifts some of the burden off of workers. Now it’s up to companies to choose to do things that way.”

This post was originally published on Cornell Chronicle and republished here with permission.

Reviewed by Irfan Ahmad.

Read next: 

• Powerful AI is making facial recognition better at identifying you

• Google Says AI Overviews Has Over 2.5 Billion Monthly Active Users as New Website Controls Roll Out
by External Contributor via Digital Information World

Powerful AI is making facial recognition better at identifying you

Vijayan Asari, University of Dayton

Image: AI-generated for illustrative purposes by DIW

If you are fortunate enough to have a ticket to an event at Madison Square Garden in New York – say, an NBA Finals game – one aspect of your visit will be having your face scanned by a facial recognition system.

Major event venues are increasingly using the technology. Some, like Madison Square Garden, use it for surveillance purposes, and some, like Citizens Bank Park in Philadelphia, to offer visitors optional ticketless admission.

Adoption of facial recognition technology is increasing, becoming more prevalent in daily life, from public buses to public buildings. The Transportation Security Administration has deployed the latest facial recognition technology at security checkpoints at numerous airports. The agency says the new system will be used in cities across the U.S. that are hosting World Cup 2026 soccer matches.

The growing use of facial recognition has broadened concerns about accuracy and bias. But in my research studying facial recognition technology in the Vision Lab at the University of Dayton, I’ve found that advanced deep learning models have made face recognition systems more accurate and reliable. The AI models, trained on hundreds of millions of face images, are more than 99% accurate in controlled environments – settings such as cellphones, airports and border checkpoints.

Facial recognition basics

Facial recognition involves three steps: locate a face in an image or video frame, create a faceprint that catalogs salient features – including the shape of the face and landmark points such as eyes, nose and mouth – and record the texture of the skin. Then it compares the faceprint to those in a database, which may be inside a smartphone or at a bank or hospital, to verify a person’s identity or allow access.

In the physical world, these systems are faster and simpler than requiring people to show IDs. In the online world, they are easier than entering a login name and password. Facial recognition also significantly reduces the possibility of forgery or fraud when compared with ID cards or passwords.

Improvements in the technology have come from a variety of research projects. FaceNet, a deep learning model developed by Google, has upgraded recognition of faces that are partly covered or hidden in images. DeepFace, a landmark AI-powered facial recognition system developed by Facebook AI Research, achieves the same high level of verification shown by humans.

NeoFace, a highly accurate AI-powered algorithm developed by NEC, is built into Mobile Fortify, the mobile facial recognition system used by U.S. Immigration and Customs Enforcement to identify people.

Reducing false positives and negatives

Real-world conditions such as poor lighting, difficult viewing angles, extreme facial expressions, concealment by face masks or sunglasses, and poor image quality can still hamper performance, leading to faulty identification. False positives and false negatives are the two primary errors. False positives are when a person is incorrectly matched to a different person in a database. False negatives are when an individual is not found in a database, even though their image exists there.

False positives are more critical in security and safety applications. They can lead to wrongful accusations, discrimination or detention. In 2025, a 50-year-old woman in Tennessee was arrested and put in jail for six months based on an AI-powered facial recognition system that incorrectly tied her to a North Dakota bank fraud investigation. False negatives may prompt authorities to deny services to people who qualify for them.

Accuracy can suffer if models are trained on data that does not reflect real-world demographics. A 2025 study showed that systems trained on public databases in which people with darker skin tones are lacking leads to lower recognition accuracy. This kind of unintentional bias in training data may lead to misidentification of women, people of color and young and old people. One report found that facial recognition systems used by 42 U.S. government agencies falsely identified African and Asian American faces 10 to 100 times more often than white faces, in some cases leading to wrongful arrests.

Accuracy also deteriorates when people are wearing heavy makeup and for young children and old people because their landmark features tend to change more quickly than adults of other ages. Balancing datasets by collecting more representative images across age, gender and ethnicity, and frequently updating databases, can improve accuracy and produce fairer results.

Adjusting images before they are sent for matching – for example, changing brightness levels – can improve accuracy, too. People squint their eyes when they are in dark or very bright light. Advanced processing software can mimic this human trait to improve the facial recognition system’s ability to extract facial features from the image.

A full face from partial data

Humans are good at identifying a person even if part of their face is covered by sunglasses or a face mask. The brain assigns more significance to the exposed details. If facial recognition programs can learn to do the same, that would reduce false positives and false negatives, including when cameras only capture part of a face.

Facial dynamics can help, too. It may be difficult for someone to instantaneously recognize a middle school friend they haven’t seen for many years, but if the old friend smiles, that change in expression can immediately improve recall.

Researchers are developing a facial recognition method for doing this, known as volumetric directional patterning. It captures the subtle movements of facial muscles, as well as eyelid blinks, in consecutive frames of a video. It tracks how facial landmarks shift over time, as well as the context in which a face is being observed, which can improve recognition accuracy.

Researchers are also creating more accurate AI-powered three-dimensional systems that can capture the precise geometry of a face, including features such as contours of the eye socket, nose and chin. This kind of work could lead to anti-spoofing techniques that prevent facial recognition systems from falling for fake faces that are generated by computers and their human operators.

Fewer mistaken identities

Setting aside questions of privacy and cybersecurity and lingering issues of bias, one thing is clear: Facial recognition technology is improving. And that promises fewer errors – and fewer of the serious consequences that come with them.

Vijayan Asari, Professor of Electrical and Computer Engineering, University of Dayton

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Reviewed by Irfan Ahmad.

Read next: 

• Google Says AI Overviews Has Over 2.5 Billion Monthly Active Users as New Website Controls Roll Out

• Lessons from crypto are paving the way for a democratic global financial future

by External Contributor via Digital Information World

Lessons from crypto are paving the way for a democratic global financial future

By Taylor & Francis

Experts argue the convergence of traditional and decentralised finance is breaking down barriers to financial access worldwide.

Image: Kamal Uddin - Unsplash

The convergence of traditional finance (TradFi) and decentralised finance (DeFi) is fundamentally reshaping global banking and democratising access to financial services.

This convergence is creating unprecedented opportunities for those previously excluded from the formal financial system, such as those living in remote regions.

According to finance experts, tokenisation – the digital representation of real-world assets on blockchain networks – is emerging as a bridge between established financial institutions and decentralised infrastructure, paving the way for a more inclusive and efficient financial ecosystem that operates at the speed of the internet.

These emerging possibilities are laid out by experts in The New Intersection of Money: Where DeFi & TradFi Converge, authored by Scarlett Sieber, Chief Strategy and Growth Officer at Money20/20, and her colleagues at the fintech events firm.

Breaking down barriers to financial access

The convergence of new and old is particularly impactful beyond developed markets.

In emerging economies where basic access to financial services remains out of reach for many, DeFi is providing essential infrastructure. The book cites the case of Hala Mahmoud Almahmoud in rural Syria, whose farm was restarted after 14 years of conflict through a cryptocurrency payment via a plastic card that instantly provided approximately 500 dollars in relief.

“DeFi shines in areas with low trust in centralised entities,” the authors note. “Digital assets evolved from experimental curiosity to essential financial infrastructure. The most profound impact of DeFi lies in continuing humanity’s oldest financial purpose: finding new routes when existing systems create too much friction.”

Tokenisation is also fundamentally altering access to investment opportunities that were once the exclusive preserve of the ultra-wealthy. Private equity investments traditionally requiring millions of dollars in minimum commitments can now be accessed for a few thousand dollars through tokenised securities.

The financial implications are substantial. Tokenised real-world assets, once considered a fringe financial experiment, are now projected to reach between $10 trillion and $16 trillion in market value by the end of this decade.

“This growth is not mere speculation but driven by a calculated migration of traditional assets like real estate, bonds, and private equity onto the blockchain,” the authors state. “that is impact. That is scale.”

Meeting consumer expectations

The convergence does not just benefit financial entrepreneurialism but the end users – the people using banking apps, making payments and spending their wages every day.

This is because the convergence of traditional banking and digital technology addresses a fundamental disconnect between consumer expectations and the limitations of traditional banking infrastructure.

Research indicates that 83% of Gen Z prefer digital-first financial services, with 78% having one bank account and 66% using mobile apps as their primary form of banking.

“New generations accustomed to instant messaging, on-demand services, and global digital platforms began to question why moving money remained slow, opaque, and expensive,” the authors observe. “The demand was not simply for cheaper transactions but for time compression – money needed to move at the speed of the internet.”

The future of finance

Industry experts expect traditional finance and decentralised finance to continue to converge in a stepwise manner, as it continues to solve long-standing problems.

“Finance is not moving onto blockchains because it wants to be modern,” the analysts explain. “It’s doing so because programmable settlement, tokenised collateral, and always-on liquidity solve problems that have existed for decades and grown intolerable at a global scale.”

By 2030, experts suggest consumers will see money will move as easily as streaming data: instant, continuous and operating year-round.

“This is not a race to declare winners or losers,” the authors say. “It is the global financial system doing what it has always done when confronted with new infrastructure: adapting, standardising, and quietly deciding what gets to last.”

The book draws on insights from Money20/20’s global platform, which brings together executives from established financial institutions, blockchain developers, DeFi protocol creators and regulatory authorities.

As regulatory frameworks mature across the world, the experts suggest that the debate about whether DeFi belongs in the formal financial system has effectively ended. The question now is how quickly institutions can adapt to capture the efficiencies that these technologies enable – and whether they can do so while maintaining the trust and stability that underpin the global financial system.

“The future of finance isn’t coming. It’s already clearing,” the authors conclude.

This post was originally published on Taylor & Francis Newsroom and republished here with permission.

Reviewed by Irfan Ahmad.

Read next: 

• Should You Accept Internet Cookies? BU Researchers Say the Open Web Could Suffer Without Them

• Your phone screen doesn’t have the same color range as the human eye – and AI widens the gap between digital images and the real thing
by Press Releases via Digital Information World

Should You Accept Internet Cookies? BU Researchers Say the Open Web Could Suffer Without Them

By Andrew Thurston, The Brink, Boston University

New study finds ad revenue that supports digital publishers and content creators tumbles when cookies are removed.

Image: Vyshnavi Bisani - unsplash

It’s a choice you may face multiple times a day—and, at this point, your reaction is probably reflexive. Are you going to accept those internet cookies, reject them, or spend a little time customizing your settings?

Increasingly, internet users are pushing back against cookies—the digital crumbs used by websites and advertisers to spot returning customers—by choosing privacy-enhancing browsers or clicking that reject button. But ditching the cookies may have big implications for the free web. If digital companies, content creators, and advertisers aren’t making money from our surfing, the quality and usefulness of the products they offer might suffer too.

In a new study, Boston University researchers highlight the potential impact the loss of cookies has on advertisers and how alternative systems designed to balance privacy and revenue fail to recoup the costs.

They analyzed 200 million ad impressions—or views—worldwide and found that removing cookies cut website publishers’ revenue by more than a third. They also discovered that privacy-enhanced alternatives, notably a major Google project called Privacy Sandbox, only clawed back a small portion of that lost revenue. The findings were published in PNAS, the National Academy of Sciences’ flagship journal.

“Internet cookies—especially third-party cookies—have been central to how online advertising works,” says Garrett Johnson, a BU Questrom School of Business associate professor of marketing. Third-party cookies are those placed by an organization, like an advertiser, not connected to the site you’re on. “In our study, removing third-party cookies reduced publisher ad revenue by about 35 percent—and about 66 percent in the European Union—showing that cookies still play a major economic role in supporting the open web.” The European Union has tougher online privacy rules than much of the rest of the world.

According to Zhengrong Gu, a Questrom PhD candidate, because cookies help advertisers spot users around the web, they can better target and measure their ads. That makes advertisers’ spending more efficient, putting more ad money in the pockets of content creators and publishers. “If more users decline cookies, it would likely reduce the effectiveness of digital advertising and the revenue that supports much of the open web,” says Gu (Questrom’26).

The downside of cookies: no one really likes being followed. “Website cookies are online surveillance tools,” wrote Wayne State University researcher Elizabeth Stoycheff in a Conversation article, “and the commercial and government entities that use them would prefer people not read those notifications too closely.”

There have been a couple of different responses to the decline in cookie use. One is the implementation of paywalls and subscriptions to keep the cash flowing; another is requiring customers to use log-ins that work across multiple sites. Tech companies are also experimenting with privacy-enhancing technologies (PETs) that try to balance advertising needs with user privacy concerns. One of the best known PETs is Privacy Sandbox, Google’s now-defunct six-year experiment in cookie alternatives, which included innovations such as a browser tool that shared a customer’s interests rather than their detailed online history.

“In our study, Privacy Sandbox recovered only about 4 percent of the revenue lost when cookies were removed,” says Shunto J. Kobayashi, a Questrom assistant professor of marketing. That weak impact was in part due to the limited adoption of the new tools and because they changed the user experience, he says, introducing “technical frictions, especially slower ad loading times.”

In their paper, the researchers write that their findings, alongside those from other studies, “informed Google’s decision to abandon its plan to replace cookies with Privacy Sandbox. The episode underscores the difficulty of aligning privacy, performance, and competition goals in digital markets.”

To examine privacy technologies in a real-world setting, the BU team used data from ad management firm Raptive, and leveraged an experiment conducted by Google and overseen by the United Kingdom’s Competition and Markets Authority. During the study, Chrome users were randomly assigned to one of three groups: cookies-enabled, cookies-disabled, or cookies replaced by Privacy Sandbox. The study included around 60 million desktop and mobile Chrome users.

“The experiment created a rare opportunity for independent, large-scale evaluation open to external participants,” says Johnson, an expert on digital marketing who has studied privacy regulations, online ad effectiveness, and the economics of digital advertising.

He adds that many European regulators are considering even tighter online privacy rules, which could have a negative impact: “Our results provide unusually strong evidence—from a global, industry-wide field experiment—that restricting cookies carries significant economic downsides that regulators should consider.”

As for users faced with that daily accept or reject decision, Johnson recognizes that everyone will make the call that works for them—but he leans toward clicking “accept.”

“From my perspective, accepting cookies creates substantial benefits for the advertising ecosystem and the publishers I care about,” he says, “with what I perceive to be little personal risk.”

This research received financial support from the Center for Industry Self-Regulation.

This post was originally published on Boston University's The Brink and republished here with permission.

Reviewed by Irfan Ahmad.

Read next: Your phone screen doesn’t have the same color range as the human eye – and AI widens the gap between digital images and the real thing
by External Contributor via Digital Information World

Your phone screen doesn’t have the same color range as the human eye – and AI widens the gap between digital images and the real thing

Douglas Goodwin, University of California, Los Angeles; California Institute of the Arts

Image: ricardo frantz - Unsplash

A peacock feather in sunlight shifts from blue to green to bronze as you turn it. Photograph it, and this shimmer collapses into one angle, one exposure, one compromise.

A digital image is not a record of what your eye sees. The standard color space that most digital images use was built for an older display world, when cathode-ray tube monitors swept beams of electrons across phosphor-coated glass. This standard color space made color predictable across many devices, but the compromise was a narrower range of colors for screens, cameras and image files to share.

Whatever the screen offers feels complete. It is not that your eyes cannot see more; digital images give them less to work with.

I teach a class about color at the California Institute of the Arts called Plastics, Neon, and Psychedelia, which covers the many ways color is produced: by materials, by light, by screens and by the mind.

I also have a condition called deuteranomaly, which changes the way I discriminate color, though not in the way you might imagine. A deuteranomalous eye does not simply lose color distinctions – it remaps them.

Vision researchers in Cambridge showed in 2005 that deuteranomalous observers can reliably distinguish khakis and olives that look identical to people with standard color vision. I have mistaken a traffic light for an overhead streetlamp while driving at night, but my color vision is not a shrunken copy of ordinary vision: It is a different map of the same ground.

While my eyes leave some colors uncertain, they sharpen other distinctions. Screens impose another kind of limit, though more quietly: They organize color according to their own rules, then offer that version as complete. My eye and the screen are both maps that include and exclude differently. Mine trades some distinctions for others. The screen trades range for reliability. The question for any color system is not whether it is accurate but what it keeps.

From wild green to screen green

My neon pothos houseplant is so green that it seems to generate its own light. Photograph it with a smartphone and the result is fine: The leaves are green, the picture makes sense. But the green in the photograph is not the green on the plant.

Look at the photo. Look at the plant. Then look at the photo again. The photographed leaves are muted, but not evenly. Some greens flatten while others appear boosted, as if the phone were trying to compensate for what it cannot show. The leaves on the actual plant are electric. No phone I own, no printed page and no Instax print has captured that green, though the Instax comes closer.

Here is what happens. Light bounces off the pothos and strikes the phone’s sensor, which records numbers representing the color the phone sensed. Each pixel is stored as a recipe for red, green and blue light: three values that tell a screen how much of each primary color to emit. In much of the image world, those numbers are still translated into sRGB, the default color space for ordinary digital images.

Color scientists map human color perception as a horseshoe-shaped field. A standard display space cuts a triangle from that horseshoe, enclosing only part of what the eye can see. A triangle’s straight sides cannot follow the horseshoe’s curve, so some colors always fall outside the display space.

Many modern screens can show more than sRGB, but sRGB remains the default format for ordinary digital images because it works reliably across devices and platforms. The pothos green is remapped to fit, and that remapped version is the picture you get. Screen green is not wild green.

Every medium translates color in its own way. Film does too, through chemistry, exposure, dyes and paper. The Instax print is not more accurate in any absolute sense, but it conveys the pothos differently. It reflects light from a surface rather than rebuilding color as light from a screen. The green feels denser, less flat and less removed from its source. The Instax still misses the plant’s absolute color, but it misses it in a different way.

The phone’s translation matters now because people see a thing’s color on a screen before they meet it in the world.

When AI learns from limited colors

AI image generators do not simply inherit this color gap. They can amplify it. They are not trained on the plant in front of you. They are trained on other people’s photographs of plants like it: millions of images already filtered through sensors, editing software, platform compression and the color limits described above. Many of the vivid greens were clipped or shifted before the model ever encountered them.

Ask an AI image model to generate a peacock feather and you will likely receive a competent image: the canonical eyespot, the dark pupil, the cyan ring, the gold, the magenta rim. The surviving colors are the ones the image world knows how to keep. What is missing is the iridescence.

In a real feather, the barbs can flash the same blue-green-bronze as the eyespot. A photograph fixes that shimmer to a single angle. The generated image flattens it further. Its barbs are muddy brown with faint metallic highlights. The model has learned the symbols of a peacock feather, but not the event of seeing one turn in the light.

Generative models make images from the patterns they find most often in their training images. They render ordinary brightness convincingly. The rarer effects are the ones that slip away: saturated colors, metallic flashes, structural glints. The model can still make an image that looks bright, even spectacular. But its brightness is screen-native, learned from other images on screens, not from seeing the subject itself.

The loop tightens with each pass. AI-generated images are uploaded, shared, indexed and may be folded into future training sets. When models train on material produced by earlier models, their outputs narrow over time. The rare colors are the first to go.

These images do not stay inside the machine. They can fill social media feeds and image searches until they stand in for the thing itself. The simulated green becomes the green you meet first.

Watching for the gap

Day after day, screens show you only the colors inside their range, translating anything outside that range into colors they can display. If you only ever see the beetle wing as a dull image, nothing tells you a brighter one ever existed. It will just look dull. If every picture of a plant arrives with its greens muted into the same screen-safe range, that range becomes green. You will not miss the absent colors.

One way to observe the gap is to find something vividly colored: a ripe persimmon, a peacock feather, the orange-pink at the bottom of the desert sky. Look at it before you photograph it. Stay with it. Stand close. Give it a name, even a private one. Then photograph it. Hold the image next to the thing. That distance is the gap.

The wild colors have not been erased. They have been excluded from the screen’s territory. The version that travels is what gets remembered. The thing is still there. Look for it.

Douglas Goodwin, Lecturer in Design and Media Arts, University of California, Los Angeles; California Institute of the Arts

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Reviewed by Irfan Ahmad.

Read next: 

• ChatGPT Loses Share of App Downloads As Competition Intensifies

• 66% of Parents Want a Teen Social Media Ban But Nearly Half Reject Common Tools to Enforce It

by External Contributor via Digital Information World