Google’s content recommendation algorithm is a black box. Here’s how it decides what to show you

Olaya López Munuera, Universitat Autònoma de Barcelona and Carlos Lopezosa, Universitat de Barcelona

If you’re one of Google Discover’s millions of users, you’ve probably come across all manner of content without actively searching for it. A Spanish user, for instance, might see headlines promising to reveal the secret recipes of avant-garde chef Ferràn Adrià, what time 46 year-old TV presenter Pilar Rubio goes to bed, the latest unmissable deals from Lidl, or the new car that will allow them to say “adiós” to annual inspections.

We still tend to think of Google as a search engine first and foremost. The iconic search bar on a white background – Alphabet’s flagship product that underpins its substantial advertising revenue – has long been part of our vocabulary, as both a noun and a verb that has the final say in contentious debates.

Image: Google Discover. DIW. CC BY.

But over the years, Google has established another service, one that moves beyond traditional “searching” to algorithmic “finding”. Google Discover, its content recommendation system, has a similar interface to a social media feed, though without comments, retweets or visible interactions from other users.

Unlike other Google services like Maps, Drive or Photos, Discover does not have its own app. Instead, it comes pre-installed on Android phones as an integral part of the interface, usually accessed by swiping right on the home screen.

The big question regarding this and any other recommendation system – from controversial algorithms like those on Instagram and Facebook, to Netflix, YouTube and even seemingly innocuous services like Vinted – is how it decides what content to recommend to you. What have you, as a user, done to deserve this on your phone?

No magic formula

When people try to explain how platforms recommend content to us, they tend to lean on algorithmic imaginaries, gossip or unfounded beliefs. Some even attempt to train or manipulate the algorithm based on these reconstructions.

However, the truth is that there is no single algorithm or grand equation that you can control. Recommendation systems like Google Discover are powered by a vast, opaque network of algorithmic infrastructure, a circuit of computer instructions hidden inside a metaphorical black box about which Google has revealed very few details.

Everything else that we know comes from either professional efforts to learn how it works, or from scientific literature, which has noted its impact on fields like journalism.

However, these systems’ ingredients cannot simply be read like the label on a bottle of ketchup or shampoo.

Funnelling content

Google’s webpage defines Discover as a personalised feed based on the user’s interests, determined by their activity on the web and in apps. Scientific studies published by Google’s own engineers have revealed it to be an industrial-scale content recommendation system, while independent academic research places it at the intersection between information retrieval and algorithmic curation.

While we do not know exactly how it works, we can explain why Google Discover shows you certain content by looking at the typical architecture of a recommendation system, along with the evidence and hypotheses we have about Google’s product.

A useful metaphor is that of a funnel divided into stages. Every day, millions of pieces of content of all kinds are published online, from articles and long-form videos to social media posts and AI summaries. Discover distils this immense variety into a list tailored specifically to every single user.

The first filters

As with Google Search, the first step is indexing. Google maintains a database where it crawls the internet and applies an initial filter to exclude anything that does not comply with its policies. This ensures that content which breaches security rules is excluded from the outset. This includes, for instance, posts seeking to recruit people to a terrorist organisation, or graphic images without any journalistic context.

However, getting past this filter merely means that content is eligible. It does not guarantee it a place in your feed.

To narrow down this initial pool, the system then identifies which content is most likely to match your interests. Google Discover does this by identifying which specific elements (known as entities) are present in your searches or interactions – such as Ferràn Adriá, Pilar Rubio and Lidl – and linking them together. In this way, it can interpret which seemingly unrelated searches or articles form part of the same interest, and organise them into broader topics and subtopics.

Once it has profiled your interests, the system needs to translate them into a language in which it can compare and find related content on a large scale. In the case of Google Discover, analysts such as Damien Andell have suggested that it uses mathematical representations known as embeddings.

These vectors convert groups of entities (known as clusters) and your own interests into coordinates on a single board that resembles a game of Battleships. The closer two points are, the greater their affinity.

Tracking clicks, likes and zooms

Once these clusters have been identified, the funnel narrows again into predictive models. Here, the system’s AI attempts to predict how you will react to each of the options.

Google’s engineers describe eleven predictive objectives, seven of which are specified: five positive (clicking, liking, responding positively to a survey, clicking to view details, and clicking to expand text) and two negative (dismissing or permanently blocking the content).

With all of these predictions on the table, the system then processes them to rank the candidates. The usual approach is to combine them into a single score, giving greater weight to some signals than others. One article may, for instance, have a high probability of you clicking on it, but also a high probability of you blocking it; another may arouse less immediate curiosity, but generate more positive signals.

Discover has to resolve this trade-off to decide what appears first and what appears next. However, Google has not disclosed how it combines these predictions.

Once the candidates have been ranked, some systems apply a final filter to incorporate criteria such as diversity, recency or fairness. This prevents the top five pieces of content from all being about, for example, the same unmissable new gadget from Lidl.

An endless loop

We’re now at the end of the funnel, and Google Discover has decided what to show you. But the process doesn’t stop there. The system feeds back on itself, observing the signals you generate both explicitly (when you block a source or like a piece of content) and implicitly (whether you stop to read an article or move on to the next one).

This new data is fed back into the system to train and update the model. Recommendations generate interactions, interactions produce new data, and that data feeds into future recommendations.

We cannot know exactly what mix of signals leads to Ferràn Adrià, Pilar Rubio, or any other name appearing in your Discover feed. What we do know is that the system is a never-ending game of chance, one that feeds on your history and location, as well as small, everyday behaviours that, while imperceptible to you, provide the data that fuels this vast recommendation system.

Olaya López Munuera, PhD en Comunicación y Periodismo, Universitat Autònoma de Barcelona and Carlos Lopezosa, Profesor Lector, Universitat de Barcelona

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

Fact-Checked by Irfan Ahmad.

Read next: Researchers Find Inaccurate Claims and Critical Omissions in ChatGPT Election Responses

by External Contributor via Digital Information World

Researchers Find Inaccurate Claims and Critical Omissions in ChatGPT Election Responses

By University of Massachusetts Amherst

On Oct. 25, 2024, 10 days before the election that would return Republican Donald Trump to the White House and over three months after Kamala Harris ascended to the top of the Democratic ticket, ChatGPT was still telling users that Joe Biden was running for a second term as president.

Image: Laurin Steffens - Unsplash

The “careless failure with profound democratic implications” is outlined in a recent study co-authored by a University of Massachusetts Amherst journalism scholar detailing how the generative AI model was serving up false information in the runup to the high-stakes election.

Less than two weeks before Election Day 2024, the study found that nearly 10% of election-related claims by ChatGPT were inaccurate, and that the model used “hedging,” “omission” and, to abbreviate a term used by the authors, “B.S.,” that obscured political realities. The research is published in the journal Information, Communication & Society.

Paper co-author Heesoo Jang, assistant professor of media law and ethics at UMass Amherst, says though ChatGPT may be able to ace tests and analyze large datasets, it was never designed to uphold democratic institutions—a role that framers of the Constitution recognized as essential when they enshrined a free press.

“We should be concerned and alarmed by the fact that while ChatGPT is nailing exams, it is also failing in these areas that it wasn’t designed for,” she says. “ChatGPT is commercially driven. It isn’t designed in a way that centers democracy.”

To test the model’s grasp of the 2024 election, Jang and researchers at the University of North Carolina at Chapel Hill (UNC-Chapel Hill) fed it both partisan and nonpartisan-framed prompts related to election information, disinformation and threats, candidates, and Project 2025. They took steps, including new “chats” and “incognito” web browsing, to ensure previous prompts did not influence later responses.

The researchers audited ChatGPT-4o mini, which was the default free version of ChatGPT available to users during the study period.

Of 592 verifiable claims it made on Oct. 25, 2024, 9.63% were inaccurate. In addition to claiming Biden was still in the race, ChatGPT inaccurately stated that North Carolina Gov. Roy Cooper was seeking reelection, even though he was term limited.

The study also shows the AI model hedged 18% of the time by returning answers with qualifying phrases like “could potentially,” “generally aligned” and “typically.” In other instances, it omitted critical information about efforts to overturn the 2020 election.

Then, there’s the category that researchers labelled B.S. They found ChatGPT appeared unable to recognize terms that partisan actors redefined and politicized, such as the “DEI hire” pejorative used against Kamala Harris. The model defined it as someone hired to oversee diversity, equity and inclusion efforts at an organization.

In another case, the chatbot made up the policy platforms of statewide school superintendent candidates in North Carolina, using generic language that appeared nearly identical.

“It was trying to fill in the gaps,” Jang explains. “It can’t give us, ‘I don’t know,’ because that’s not how it is designed.”

Though ChatGPT models have been updated since 2024 and would likely perform better today, she’s not optimistic about the 2026 midterm elections, particularly as AI increasingly replaces internet searches and links to primary sources.

“Truth and accurate electoral information—and also dealing with disinformation—are not things we can train large language models to handle using past datasets alone,” Jang asserts. “It will be really hard for them to catch up with what’s going on now, unless they are designed to prioritize democracy, which is extremely difficult when their development is driven primarily by commercial incentives.”

She argues that the rise of AI makes fact-based journalism even more essential.

“We need journalism that not only delivers truth, but also helps people evaluate the truth,” Jang says, by calling out mis- and disinformation, avoiding false equivalency and sounding the alarm when democratic norms are breached.

The research was co-authored by Shannon C. McGregor and Lorcan Neill of the Hussman School of Journalism and Media and the Center for Information, Technology, and Public Life at UNC-Chapel Hill.

Fact-Checked by Irfan Ahmad.

Read next: 

• The consequences of relying on AI for accurate news

• Human-like AI not always the answer in customer service

• Apple’s new watches are always listening
by External Contributor via Digital Information World

Apple’s new watches are always listening

Luke Heemsbergen, Deakin University

Image: Madalyn Cox - unsplash

The latest Apple Watch models have some new features: they can listen to the sounds around you all day, transcribe snippets of conversation on demand, and write up summaries of everything they heard.

This kind of ambient, always-on surveillance raises privacy concerns, which Apple addresses well. The company says audio is never “recorded” or stored, that nothing is attributed to any speaker, and all content is encrypted so even Apple cannot read it.

I have no reason to doubt any of that. But my reaction to these features is still clear: ick.

I’m not the only one who feels that way, and the ick deserves to be taken seriously. Let me explain why.

The conversation belongs to the room

The new features may run into problems with privacy laws, particularly those that require consent from people being recorded, but also with unspoken social conventions about speech.

When we talk with other people, we have some expectations about what happens to our words. Our comfort comes from what information scientist Helen Nissenbaum calls contextual integrity: the expectation that information moves around according to the norms of the setting in which it is shared.

What you say across a café table, in a bathroom, at a dinner party, at a town hall meeting or during a television broadcast is said for the people present and for the purposes apparent in that moment. Until now, our ambient conversations were not said for a auto-magical-summariser to fold into someone else’s searchable life records.

Apple can encrypt, minimise and delete all it likes, but data protection isn’t the problem. The problem is a conversation has been quietly re-routed from the room where it happened into a device and service and product.

‘Opt-in’ doesn’t work for everybody

These features are marketed as “opt-in”. But only the person wearing the Apple Watch gets to opt at all. Nobody else in the room has a choice.

Apple seems to know this is a problem, so Live Rewind – which transcribes the previous 15 seconds of speech on demand – announces itself with a chime, an animation and a microphone indicator.

Siri Recap, the feature that runs all day and makes on-device summaries later, affords no such notice. You can see how a continuous chime and blinking light would make the feature unusable, and any watch unwearable.

In places where confidentiality is expected – whether for personal, commercial, or legal reasons – devices with ambient recording features are likely to be unacceptable.

Right now, the Apple Watch is welcome anywhere, like AirPods. The new features may change that – and the cameras Apple reportedly plans to add to AirPods may do the same for the now-ubiquitous earphones.

Meta’s video-recording smart glasses have already been banned in some workplaces and venues. The media often frame the problem with individual users, as with Meta’s so-called “pervert glasses”, but the tech industry has earned public distrust in its own right.

It’s the trust, stupid

Tech companies have a reputation for misusing user data. One study of 13,796 Android apps found 42.4% contained at least one omitted or incorrect disclosure of privacy-sensitive data flows. Another study of almost 5,855 children’s apps found most may have been in breach of privacy law.

Norms about ambient recording are already being set by other devices.

Meta’s smart glasses are one, and a recent crop of AI pendants and pucks are in similar territory. These devices and services have broken integrity expectations (and privacy laws) again and again and again and again.

Surveys show most people already think their phones are listening to what they say in order to serve them ads. The irony of introducing a similar feature to everyone’s wrist should not be lost on Apple.

Nor are they unfamiliar with the need to get it right. The company paid out US$95 million to settle a court case in 2025 over claims that Siri activated by accident and sent private audio to contractors, but did not acknowledge wrongdoing and maintains their focus is privacy.

A private way to record your day?

Apple is handling privacy in ways that are demonstrably better than others. Apple’s Private Cloud Compute, which powers the new features, is built so the company’s own staff cannot read what passes through it. The servers keep no logs and retain nothing after a user’s answer is returned.

Apple has promised to publish the software used in each Private Cloud Compute instance, so independent researchers can check it works as described.

So, Apple’s privacy policy and software experience is backed by hardware open to inspection. It shows a precedent that others can follow. Seen in this way, the Rewind and Recap features demonstrate that always-on listening can be done without giving the data to anyone but the user.

Apple products have set social precedents in the past, as well. The iPod changed ideas about music piracy, and AirPods made it okay to have weird white sticks hanging out of your ears at all times.

The socialisation of Apple tech is not guaranteed, however. The futuristic Vision Pro goggles launched in 2024, for example, are still not cool.

Nope is still a choice

Despite Apple’s efforts, the basic problem of a device that is always listening in public remains. It still violates our expectations about the integrity of a conversation belonging to the room.

The ick feeling about Recap and Rewind comes from what Apple is actually selling: a new default understanding that a conversation is something that may be casually captured, analysed and used elsewhere by others.

Apple is one of the most credible companies to ask us to accept that new default. Which is why those saying “nope” to breaking the contextual integrity of our conversations should be heard. If we don’t want that kind of tool in our lives – or the lives of others – we need to say it out loud: ick.

Luke Heemsbergen, Senior Lecturer in Communication, Deakin University

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

Fact-Checked by Irfan Ahmad.

Read next: Human-like AI not always the answer in customer service

by External Contributor via Digital Information World

Human-like AI not always the answer in customer service

By Felix Septianto, University of Queensland

Image: Anna Shvets - pexels

The next time something goes wrong with your bank account, parcel delivery or a bill, there's a good chance you'll be directed to AI for customer service.

Countless organisations across many industries are investing in increasingly human-like AI to answer questions, troubleshoot problems and resolve complaints.

The Commonwealth Bank recently revealed its AI-powered customer service platform handles more than 2 million conversations every month across voice and messaging channels.

Australia Post also reported that 55 per cent of customer enquiries were supported by self-service AI in 2025.

The logic of this trend seems straightforward – when something goes wrong, customers are used to talking to another person for help. If AI can communicate more naturally, it seems reasonable to assume it could recreate some of the benefits of human customer service.

But research published by my colleagues and I suggest the relationship is more complicated.

The push towards voice-based and conversational AI reflects an assumption that customers prefer interactions that resemble human conversations.

The Australian Contact Centre Industry Benchmarking Report found when people were told their issue would be resolved, a phone call was the preferred channel at 27 per cent, compared with 12 per cent for webchat.

But it does not necessarily follow that making AI more human-like will always improve customer experiences. One recent survey of 1,031 Australians found 78 per cent had used automated customer service systems, yet only 39 per cent reported being satisfied with them.

The benefits of slowing down

Our research involved several studies examining how people respond to different forms of AI customer service following service failures.

In one study, we examined what happens when customers interact with a text-based chatbot rather than a voicebot after experiencing a service problem. We found in some situations the less human-like option produced more favourable customer responses.

Additionally, participants who interacted with a chatbot reported greater feelings of relief than those who interacted with a voicebot. Those feelings of relief were, in turn, associated with more positive evaluations of the service experience.

This may occur because writing is slower and more deliberate than speaking. When customers are required to type, they have more opportunity to redirect their attention and allow some of the negative emotion associated with the issue to subside.

In short, when a customer is already upset, a faster and more conversational interaction may sometimes remove something valuable: a little time to cool down.

But creating emotional distance is only one part of effective AI customer service.

Where emotional intelligence matters

In separate research, we examined how customers responded when an artificial service agent acknowledged their emotions.

We compared a robot that explicitly recognised a customer's anger by saying, "I understand you are feeling angry", with one that used the less specific phrase, "I understand you are feeling bad".

Customers responded more positively when the robot accurately identified their emotional state. More precise emotional recognition increased perceptions of the agent's emotional competence and improved customer satisfaction.

Together, these findings suggest that making AI more human-like is not necessarily the same as making it better at responding to human emotions.

Sometimes, the best customer experience may come from giving people emotional distance and time to process their frustration. In other situations, success depends on recognising exactly how a customer is feeling and responding appropriately.

Beyond sounding human

As AI takes on a larger role in customer service, the key question is not simply whether it sounds more human. Organisations need to think carefully about what customers need at different moments in their experience.

The future of customer service AI may not be about imitating humans as closely as possible.

The most effective AI customer service models will recognise when people need empathy and when they need space – and be designed to respond accordingly.

Fact-Checked by Irfan Ahmad.

Read next: 

• In the race to roll out 6G, the UK and Europe are caught between technology’s superpowers

• Businesses are using AI for the calls customers want humans to handle

• Interacting with customer service AI can make people act more like robots, research suggests
by External Contributor via Digital Information World

In the race to roll out 6G, the UK and Europe are caught between technology’s superpowers

Natalia Orrego Tapia, University of Bristol

Image: Mario Caruso - Unsplash

Even as the rollout of the 5G network continues around the world, the race to design and build the sixth generation of mobile infrastructure has begun.

The new, global 6G network represents a fundamental shift in what a mobile network is built to do. Where previous generations were designed primarily as pipelines for transporting data between devices, 6G’s ambition is to treat the network as an intelligent system.

This means it is designed to be “AI native”, with artificial intelligence at its core from the outset. An example is integrated sensing and communication: radio waves that also act like radar sensors, detecting movement without the need for cameras. Real-time data collection and connectivity will be another important application in the management of largely automated urban environments.

The problem is that someone has to create 6G first – and pick up the bill. And this next technological cycle is inheriting many unresolved issues from 5G, in a technology landscape fractured by geopolitical tensions and public scepticism.

The development of 6G relies on a global conversation shaped by institutions such as the UN’s International Telecommunication Union (ITU) and the 3rd Generation Partnership Project (3GPP), a global collaboration of telecommunications standards organisations. Together, they have established the official 6G rollout plan, with the aim of creating the first technical standards by 2028. This would lay the groundwork for the release of the first 6G commercial networks around 2030.

As telecommunications shift from physical hardware to cloud software and AI, traditional mobile operators will find themselves reliant on the tech giants – in particular, massive cloud providers such as Amazon and Google – for storage capacity and data processing they cannot build themselves.

Yet this convergence raises some important questions. Can national infrastructure remain sovereign and secure when running on foreign cloud platforms? And how easily can governments cooperate across borders when the digital infrastructure underpinning their societies is controlled by a handful of private corporations?

A new 6G alliance

In late July, the US government announced a new 6G coalition, including the UK and 15 other European countries, India, Canada, Australia, Korea and Japan. Its stated purpose is to counteract the Chinese digital infrastructure industry under the banner of “6G leadership and security”.

This announcement takes the politics of 6G to a new geopolitical scale. This US-led bloc says it envisages 6G as the engine and heart of “technological competitiveness, economic prosperity and national security”.

The allies have only one real competitor, China, and one important lesson: not to repeat the strategic mistakes made during 5G’s development.

These were, primarily, the lack of a coordinated industrial policy, and underestimation of the complexity of the agreements that underpin this global network. Such missteps gave China a headstart on hardware, enabling its companies to dominate 5G patents and be the first to market a lot of 5G equipment.

The newly established 6G coalition has already set goals to solidify connections among governments, industry and academia within a few months. These links should evolve into a comprehensive strategy to support a secure global 6G network for the long term.

A key moment in this story will come in October 2027 – in Shanghai, China, of all places. At the ITU’s World Radiocommunication Conference 2027, governments will negotiate on how to allocate the multi-layered radio wave spectrum required for 6G. This should lay the foundations for infrastructure standardisation the following year.

But for the UK and Europe, this represents a serious dilemma. They do not possess the industrial scale to fully replace Chinese infrastructure. But they also fear escalating dependence on US big tech with the advent of 6G.

Europe’s dilemma

The EU’s flagship initiative Hexa-X and successor Hexa-X-II, funded by Nordic mobile giants Nokia and Ericsson, was conceived as a blueprint for a European 6G to tackle these issues. The primary aim is to produce technologies that can operate autonomously, outside both US and Chinese hardware and software.

In the UK, the recent dissolution of the Department for Science, Innovation and Technology, and migration of the digital infrastructure portfolio to the new Department for Digital, Culture, Media and Sport, created weeks of uncertainty across the UK tech sector. This came at a crucial time for determining what the future looks like for 6G research and development nationally.

Local UK mobile operators are still decommissioning legacy 2G and 3G infrastructure, and are close to decommissioning traditional telephone landlines. They are also striving to reap the benefits of the 5G network. It is not easy for the UK to assert leadership in mobile networks abroad while wrestling with all these issues at home.

Right now, the core values of the 6G technology that will define the next decade are already being decided. Future AI-native networks will come with a higher energy bill, so one key challenge is how to minimise energy consumption and maintain a path towards a sustainable technology.

Deploying new infrastructure also risks reigniting the intense public anxieties seen during the 5G rollout. Fears about electromagnetic radiation and surveillance quickly escalated into conspiracy theories linked to a history of health concerns about wireless technologies, and the tendency to exclude citizens from infrastructure design and deployment.

Getting 6G right demands a broader approach to building this new global network – as most participants agree. Including citizens in this process could be a powerful tool, demonstrating from the outset that infrastructure is never just a technical or engineering endeavour. It is also a political and social issue that will shape all our futures.

Natalia Orrego Tapia, Senior Research Associate in Environmental Sustainability and Future Networks, ESRC Centre for Sociodigital Futures, University of Bristol

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

Fact-Checked by Irfan Ahmad.

Read next:

• Flaws In Lost-Device Reporting Could Block Legitimate Devices From Cellular Networks

• Nvidia CEO Jensen Huang Calls 10-Year AI Extinction Prediction ‘Completely False’

• With education technology, engagement is not the same as learning

by External Contributor via Digital Information World

Making mental health chatbots more culturally sensitive won’t necessarily make them safer

Maya Low, McGill University

Image: Aerps - Unsplash

Demand for mental health care in Canada has risen sharply. In a health-care system already short of publicly funded specialists, Canadians often wait months for support. Many are filling this gap with general-purpose generative AI chatbots (like ChatGPT), presumably because these systems are free, immediate and available at all hours.

However, these chatbots carry assumptions about what distress is and how it should be described, and those assumptions may not fit everyone.

In 2024–25, there were roughly 94,000 referrals for community mental health counselling across the provinces and territories that report this data. Half of those people were seen within 30 days. One in 10 waited more than four months.

A 2026 survey of more than 1,200 licensed psychologists in the United States found that 77 per cent had patients who told them they were using AI for mental health support, and more than one-third had patients who considered their chatbot an additional provider. These tools have not been approved by Health Canada, and no Canadian regulation governs them as therapy.

The data also suggest that the people most likely to depend on chatbots for support are the ones least likely to be culturally understood by them: newcomers, people in rural and remote communities and anyone who cannot afford private therapy.

Chatbots are not culturally neutral

When a chatbot responds to someone in distress, it’s already committed to a view of what distress is and what helps. Its default assumes distress sits inside the individual, is made of thoughts and feelings, and improves when those thoughts are examined and changed.

That is a recognizable clinical tradition (broadly western and cognitive-behavioural), when an individualistic concept of the person has long been identified as culturally particular rather than universal. In general-purpose AI chatbots, this surfaces as a tendency toward the values of English-speaking populations in the Global North.

As a PhD candidate in cultural psychiatry, I know the discipline has long made this argument in the clinical context. Psychiatrist and anthropologist Arthur Kleinman described the “category fallacy” in 1977 after working with Chinese patients whose depression presented largely as bodily complaint (fatigue, dizziness, insomnia and pain rather than reported sadness or guilt).

Sociocultural anthropologist Mark Nichter’s 1981 work on idioms of distress demonstrated how suffering is voiced through whatever channels a community makes available. For the women he studied in southern India, that again meant feeling distress in the body instead of voicing it directly.

Distress expressed through the body, faith or obligation to family isn’t a confused version of psychological distress. A person may describe their suffering as a trial sent by God, a loss of faith, a spiritual affliction or a consequence of moral failing; another may locate it not in their own mood but in what they owe others — shame at failing an aging parent, guilt at not providing or the sense of having brought difficulty onto the family. It’s how a great deal of suffering is communicated, especially in a country as diverse as Canada.

AI chatbots handle this poorly. Studies testing whether chatbots pick up on cultural cues find they often struggle to identify a cultural pattern, especially when prompted in English. Recent research found that models express stigma toward people with schizophrenia and alcohol dependence, respond inappropriately in realistic therapy scenarios, and that newer and larger models did not improve.

This failure can be categorized as imposition, to borrow the term from the account of Madeleine Leininger, who developed the concept of transcultural nursing and of cultural imposition in clinical care. Imposition is when the system applies its own model of distress to someone that model does not serve. I describe this further in my article about the dilemma(s) associated with cultural adaptation and the use of AI chatbots for mental health.

Would a more culturally sensitive chatbot be safer?

The obvious response is to adapt these systems by training on more languages, using local idiom and building tools with and for specific communities. That is important, but it presupposes that cultural fit is the same thing as cultural safety. It’s not.

Some ways of understanding distress keep people inside it by ruling out the actions that might relieve it. For example, a person may believe their condition would shame their family if anyone were to find out, or that they are not unwell but weak and undisciplined. A chatbot with good cultural fit might reflect those beliefs back sympathetically, in the user’s own vocabulary. In psychotherapy this is called collusion. It occurs when a therapist goes along with a client’s account in a way that prevents anything from changing.

Chatbots are prone to collusion because of how they are built and trained (largely on human approval ratings). To the user, being understood and being agreed with feel much the same, so both earn approval, and the training reinforces both together. The system therefore has no way to tell recognition from endorsement.

That produces an awkward conclusion. The better a model’s cultural fit, the stronger its incentive to agree with what the user already believes, and the more convincing that agreement will be. This also applies to western accounts of distress — if you insist you’re not depressed and you’re just not working hard enough, a chatbot could readily agree with you.

The imposition side is already well-documented. The collusion side in its cultural form has not been measured, despite the underlying tendency being well established: sycophancy in language models has been demonstrated repeatedly, and 97 per cent of psychologists in the U.S. survey worry that chatbots might reinforce negative behaviours or delusional beliefs. Nothing in an approval-based signal separates a belief someone reached alone from one absorbed from their culture or community.

What would make these systems safer

If a chatbot is judged as a safe mental health tool because its users felt understood, we are using an instrument that cannot separate good care from agreeable care.

A more useful test looks at where the conversation ends — did the person end up more limited in what they feel able to do or further from help than they wanted? Those questions can be answered without deciding in advance whose understanding of distress is correct (otherwise calling something collusion is simply imposition by another name).

National regulation of these tools is being discussed, and many safety failures will likely be addressed. However, the failure described here will not look like a safety failure — it will only look like being understood.

Maya Low, PhD Candidate and Lecturer, Cultural Psychiatry, McGill University

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

Fact-Checked by Irfan Ahmad.

Read next: Google Fined Around $462 Million for Location Data and GDPR Failures

by External Contributor via Digital Information World

Google Fined Around $462 Million for Location Data and GDPR Failures

Ireland’s Data Protection Commission (DPC) has fined Google €403 million euro (which is around $462.3 million in USD) over its processing of location data. It also ordered Google to bring its processing into compliance with the GDPR within six months.

Image: Samuel Regan-Asante - Unsplash

The DPC announced the decision on Sept. 21, 2026, after an inquiry that began in February 2020. The inquiry followed complaints from several European consumer rights organisations, including BEUC.

The inquiry covered the period from May 25, 2018, to Feb. 4, 2020. It examined three features: Web & App Activity, Location History and Location Accuracy.

The DPC found that Google infringed the GDPR over the lawfulness and fairness of its location data processing in Web & App Activity and Location History. It also found an accountability issue involving Location Accuracy, transparency issues across all three features, and retention issues in Web & App Activity and Location History.

DPC Deputy Commissioner Graham Doyle said location data can reveal significant information about a person, including information that is inherently private. He said Google’s failures could have left people unaware that their location was being used to, for example, influence them with ads or infer their interests.

The DPC said it will issue the full decision in due course.

Fact-Checked by Irfan Ahmad.

Read next: 

• Flaws In Lost-Device Reporting Could Block Legitimate Devices From Cellular Networks

• Nvidia CEO Jensen Huang Calls 10-Year AI Extinction Prediction ‘Completely False’
by AI Analysis via Digital Information World

Nvidia CEO Jensen Huang Calls 10-Year AI Extinction Prediction ‘Completely False’

Fact-Checked by Irfan Ahmad.

Nvidia CEO Jensen Huang, in an interview, said the prediction that humanity could be gone within a decade is “completely false,” while saying concerns about AI safety are not wrong.

The interview was conducted by CBS News senior business and technology correspondent Jo Ling Kent on Sunday, Sept. 20, 2026.

Image: Jensen huang stanford by Anderseidesvik, CC BY-SA 4.0.

Huang explained further that the narrative is “dramatic” and gets a lot of attention. He said the people raising the concerns care deeply about making sure AI products are built safely.

Huang said the AI industry is transitioning from developing technology that is becoming capable to technology that has become useful. Companies are moving from laboratories to engineering and products, he said, requiring more research and computing to be dedicated to safety.

As products mature and roll out into society, Huang said more engineering is dedicated to verification, testing, evaluation, benchmarking, standards and reliability. He compared this transition with aviation, saying more people work on airline, aircraft and air-travel safety than on making new types of aircraft.

On regulation, Huang argued for using existing laws and regulations before creating new ones. He pointed to cybersecurity laws that address unauthorized entry into systems and said existing laws also cover damage caused by such activity. He also referred to product-liability laws, saying that if a product causes harm, the company can be held liable under those laws.

He also said AI companies need more secure sandboxes, isolation and containment for experiments, closer monitoring during testing and evaluation, and public sharing of incidents so the industry can learn.

Huang said AI should advance “as fast as we can, but not faster than we should”, and that compromising safety “can’t happen.”

Read next: 

• AI is distorting archaeology and flattening Indigenous Knowledge

• Nearly Half Of Investors Have Used GenAI, 17.7% Of Users Adopted It As A Routine Tool, And 54% Cited Reliability And Accuracy Concerns

• How to know if you can trust an AI’s answer to your question
by AI Analysis via Digital Information World

AI is distorting archaeology and flattening Indigenous Knowledge

By Dr Michelle Richards Professor Michael-Shawn Fletcher and Anthony Romano

If you image search on Google for ‘cultural heritage Australia’, near the top of the results is a photograph of giant carved stone heads rising from a red desert.

They look ancient and real. But the image is a generative artificial intelligence (GenAI) fabrication that never existed, and it is circulating the internet as though it is a real record of Australia’s history.

Image: DIW. CC BY.

While GenAI, the technology behind chatbots and image generators, has the potential to broaden access to knowledge, it can also popularise false histories and strip Indigenous Knowledge of its context.

It’s something that can undermine both scientific expertise and Indigenous authority.

As researchers, we are increasingly concerned about how these systems scrape and reuse Indigenous Cultural and Intellectual Property (ICIP) without consent.

So, how do we combat misinformation while keeping genuine scientific knowledge publicly accessible? And what do we give up when we share that knowledge openly, only to have AI appropriate it?

What is pseudoscience?

Archaeology has long attracted fringe narratives and pseudoscientific claims.

Pseudoscience presents itself as science while skipping the rigorous methods or reliable evidence that make science work.

Real science advances by correcting itself. As evidence accumulates, weaker ideas are discarded and disproven theories drift to the fringe.

Archaeology generates its own fringe whenever fresh discoveries overturn old interpretations. But some ideas persist.

The so-called 'lost continents' of Mu and Lemuria have no scientific basis, but the myth of mysterious superior civilisations sunken in the Pacific continue to circulate today.

But pseudoscience is rarely harmless nonsense. At its worst, it has been used to decide who counts as fully human.

Doctrines like eugenics and Social Darwinism claimed scientific legitimacy while justifying systematic dehumanisation. Much of today's pseudoarchaeology draws on those same intellectual traditions.

Why fringe ideas spread and why it matters

Fringe narratives promise hidden truths, mystery and revelation, while careful archaeological work tends to be slower and more complicated.

Stories about a lost civilisation or ancient aliens will usually beat an evidence-based interpretation for sheer entertainment.

The consequences are not trivial.

Fringe narratives erode public trust in science and expertise, as the relentless flat-earth movement shows.

And pseudoarchaeology has a long record of diminishing Indigenous histories. Some of the most remarkable monuments ever built have been credited to almost anyone but the Indigenous peoples who made them.

The moai of Rapa Nui, the great stone figures carved and hauled into place by Polynesian islanders, have been attributed to ancient aliens and lost civilisations in bestselling books and now in AI-generated 'histories'.

Nan Madol, a city of basalt raised on artificial islets in Micronesia, gets credited to a vanished ‘super race’ rather than the Saudeleur people who engineered it.

In the Kimberley in Western Australia, some say the exquisite Gwion Gwion paintings are the work of a mysterious foreign people, not the Aboriginal ancestors whose descendants still care for them. New Caledonia's petroglyphs attract the same treatment.

Each of these false histories severs living communities from their ancestral heritage and props up racist assumptions about Indigenous people’s capabilities.

What GenAI has to do with it

Most GenAI systems learn from the open web, where misinformation and fringe theories flourish.

Unlike a researcher, these systems do not weigh evidence or test a claim. They predict the next plausible word based on patterns in their training data, which means they can reproduce something untrue with complete confidence.

Accuracy is only half the problem.

GenAI works at scale, compressing vast quantities of information into statistical patterns. Indigenous Knowledge works the other way. It is local, place-based and bound to particular Country and communities.

These are the qualities that give Indigenous Knowledge its strength, which are flattened by large-scale AI systems.

A model can absorb Indigenous knowledge as training data with no consent, no acknowledgement, no compensation and no community control over how it is used next

Our recent study found that AI can often tell scientific and Indigenous consensus apart from well-worn fringe narratives. But that ability relies on training data collected or reproduced without community permission in the first place.

Pushing back

Fighting misinformation means engaging the public, but if strong research is inaccessible, fringe narratives fill the gap.

Those narratives then go on to feed the next generation of AI training data.

Indigenous Knowledges reach well beyond stories and cultural expression into environmental records, landscape knowledge and data built through generations of relationship with Country.

Indigenous Knowledges are not inert or ownerless information waiting to be mined. They are relational, situated and governed by protocols.

Many researchers in Australia and the Pacific have spent decades building trust with Aboriginal, Torres Strait Islander and Pacific Islander communities. These communities are now using AI themselves, for language revitalisation and for research that requires access to cultural archives.

When communities agree to publish, they are choosing to make their knowledge public, on their own terms. What they have not agreed to is a machine stripping it of context and getting it wrong.

There are some practical steps research can take.

Research journals and repositories should require evidence of community permission before publishing Indigenous data and audit existing collections for their ICIP status. Universities and funding agencies should build ICIP into research design, ethics review and peer review.

Above all, Indigenous communities must retain authority over how their Knowledge is interpreted and reused. That means true partnership, not tick-box consultation.

Expertise and understanding

Despite its rapid advance, GenAI has not replaced expertise.

An Elder, Traditional Owner, Knowledge holder or scientist can be questioned, challenged and held accountable.

Their Knowledge grows out of lived experience, long relationships and responsibility. A chatbot offers none of that.

The real experts, including Country itself, are embedded in the lives and Knowledge systems from which understanding actually grows.

This article was first published 21 September 2026 on Pursuit. Read the original article.

Fact-checked by Irfan Ahmad.

Read next:

• Nearly Half Of Investors Have Used GenAI, 17.7% Of Users Adopted It As A Routine Tool, And 54% Cited Reliability And Accuracy Concerns

• How to know if you can trust an AI’s answer to your question

by External Contributor via Digital Information World

How to know if you can trust an AI’s answer to your question

By Leo S. Lo, University of Virginia

I recently typed a simple question into Google search: How much screen time is too much for teenagers? Instead of presenting links, as Google had been doing for many years, it gave me an AI-generated answer. The artificial intelligence agent cited a number, then complicated that reply, noting that quality and balance of time could matter more than the number of hours, and that “too much” time could depend on a teenager’s sleep, exercise, school demands and mood.

I tried another search: Should I take a daily aspirin? This time the AI answer presented me with medical information, warned about risks and offered more tailored guidance if I provided my age and medical history.

These were good replies. What interested me was that they were different kinds of replies.

Debate about AI answers has focused on accuracy: Did the system get the answer right? That matters, but accuracy is only one test. Each kind of answer requires a user to judge something different.

I find it useful to sort AI answers into an “answer typography” of four broad types: factual, interpretive, constructive and strategic. A factual claim can often be checked against a source. An interpretation can be accurate and still reflect choices about which evidence matters. A construction can be well reasoned and still be wrong for the person receiving it. A beautifully written strategic document may not be true. Yet AI presents all four types of answers in much the same fluent, authoritative form; the differences are easy to miss.

I’m university librarian and dean of libraries at the University of Virginia who leads national efforts to develop AI competencies for library professionals, and I consult widely on AI literacy. I first proposed the typography in the Journal of Academic Librarianship.

The four categories are not airtight boxes. A response from an AI agent can reflect several types. That said, I describe each type of answer below, and offer guidance for deciding whether a reply is ready to use or needs more investigation.

Which answer is Google giving you?

A factual answer makes a claim that can, in principle, be checked against evidence. When was the University of Virginia founded? What is the chemical symbol for gold?

To determine whether a factual answer is robust enough for you to use, verify the claim against an appropriate source. If the answer cites a source, follow that link instead of simply treating the answer itself as proof.

An interpretive reply is built on evidence, but there is not a single takeaway. How much screen time is too much for a teenager? Does remote work raise productivity? The answer depends on what evidence is included, what is left out and how disagreement is understood.

Google’s initial answer to my screen-time question indicated that two hours was a limit for teenagers. Then it noted that pediatric guidance puts more weight on the quality and context of screen use than on simple hours. The American Academy of Pediatrics says there is no exact recommended amount for teens and emphasizes the kind of screen use and what activities it might be displacing. A question that looked numerical turned out to require interpretation.

To assess interpretive answers, do more than check facts. Ask yourself what evidence the system emphasized, what it left out and whether another defensible interpretation exists. A useful follow-up question to present to the search engine is: “What is the strongest evidence for a different conclusion?”

Image: DIW. CC BY

Constructive answers are made rather than discovered. Ask AI to draft a cover letter, write a eulogy, suggest a lesson plan or reorganize a paragraph – there is no single correct result.

You can judge the response by considering purpose, audience and voice. A eulogy can be grammatically perfect and still sound nothing like the person delivering it, or it may land flat on family members hearing it. It may not capture the deceased person well, either. Consider these kinds of effects as you read.

Strategic questions ask what to do. Should I take a daily aspirin? Should I buy the house? The answers combine information with judgment about goals, risks, trade-offs and personal circumstances.

My aspirin search shows why context matters. Google warned about risks, told me to consult a medical professional and offered more tailored information if I provided my age, cardiovascular history and risk of bleeding. That caution matches the U.S. Preventive Services Task Force guidance. It says the decision to start low-dose aspirin for prevention of heart attacks and strokes should be individualized and weigh cardiovascular benefit against bleeding risk.

For strategic answers, ask what the system would need to know before its advice could reasonably apply to you individually. Consider the stakes, the alternatives and whether a qualified person should be involved. For the aspirin question, a useful follow-up would be: “What details about my age, medical history or bleeding risk could change this advice? What should I discuss with my doctor before deciding?” The final judgment remains yours because you are the person who has to live with the outcome.

The first question after an answer

My questions began as ordinary Google searches. I did not open a chatbot. The AI-generated responses simply arrived, and links were appended.

The responses were useful. Google added context, acknowledged complications and offered tailored guidance if I supplied additional information. Within each response, though, the type of answer could change. Reporting what a medical guideline says is different from deciding how it applies to a particular person. A fluent response can move between those types of answers without a noticeable change in voice.

As a user, try to recognize what kind of intellectual work the AI agent did for a response you receive. Consider whether the interpretation is persuasive or the advice fits your circumstances.

Before asking whether an AI answer is right, ask a more basic question: What kind of answer is this? The type will tell you what to do next.

Leo S. Lo, Dean, University of Virginia

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

Fact-Checked by Irfan Ahmad.

Read next: 

• Oversight Board Says Meta's AI Deepfake Policies Are Inadequate

• AI Is Producing More Software. Why Isn’t It Being Used?

by External Contributor via Digital Information World

Oversight Board Says Meta's AI Deepfake Policies Are Inadequate

Fact-Checked by Irfan Ahmad

Meta's Oversight Board said Sept. 17 that its policies on AI-manipulated imagery and deepfake labeling are inadequate in two cases involving Facebook posts.

Image: Zulfugar Karimov - Unsplash

In one case, the Board ordered removal of an AI-generated video targeting a young Muslim woman in Europe. A majority found Meta's definition of "unwanted manipulated imagery" too narrow because it does not cover falsifying a private person's words and actions. The Board said, "Meta also defines mass harassment campaigns too narrowly, according to the majority, and the company should adopt new measures to reduce the burden on victims reporting them."

In a separate case, the Board ordered removal of an AI-generated video that falsely portrayed a Scottish Labour councillor making a statement about refugees. The Board found that the video violated Meta's Hateful Conduct policy. A majority also found Meta's rules for labeling deepfakes inadequate and said the video should have received a "High Risk AI" label.

The Board recommended expanding the situations when "high risk" labels can be applied, taking more measures to reduce the spread of deceptive AI content, increasing penalties for accounts that repeatedly share it, and providing more transparency on data about when AI labels are applied.

Editor's Note: This post was mistakenly published under the External Contributor category. It was intended to be published under DIW AI Analysis.

Read next:

• AI Is Producing More Software. Why Isn’t It Being Used?

• “Almost Every Job Has Tasks That AI Can Change”. Economist Erik Brynjolfsson says AI’s transformation of work is just beginning and our institutions aren’t ready for it
by External Contributor via Digital Information World

AI Is Producing More Software. Why Isn’t It Being Used?

By Seb Murray

A Wharton study finds that while AI dramatically speeds up software development, human bottlenecks prevent many of those gains from reaching customers.

Image: Joy Christian - Unsplash

Artificial intelligence coding tools dramatically increase developers’ productivity, but much of those gains are lost before they translate into finished software because human bottlenecks persist later in the production process.

That’s according to a new study written by Leon Musolff, a Wharton professor of business economics and public policy, and MIT researchers Mert Demirer and Liyuan Yang. (Editor’s note: This paper was updated with new data after this article was written.)

When looking at the impact of AI tools on coding activity, the gains grew sharply with each new generation of AI tools. Autocomplete systems that suggest the next line of code increased coding activity by 40%. Adding “sync agents,” which edit code alongside developers in real time, lifted the cumulative increase to 140%; “async agents,” which work autonomously from a prompt, pushed it to 180%.

Yet even the biggest cumulative gain translated into only a 50% increase in software projects, and a 30% increase in software releases. “In software, the binding constraint appears to be shifting from writing code to reviewing, integrating, and ultimately distributing it,” wrote the authors in the paper.

What’s Blocking AI Productivity Gains?

The researchers tracked more than 100,000 developers on GitHub, the world’s biggest software development platform, comparing their productivity before and after they adopted the three successive generations of AI coding tools, from 2022 to 2026. They combined those public GitHub records with Microsoft data on developers’ use of the tools to identify when they first adopted the technology.

The findings suggest that AI can dramatically speed up individual coding tasks, but those gains will not automatically translate into more finished software — unless AI can also automate more of the work involved in reviewing, integrating, and releasing software.

Increasingly powerful AI coding tools have made it possible to generate working software from simple prompts, dramatically lowering the barriers to software development. That has helped fuel the recent “vibe coding” boom, allowing employees with limited programming experience to build applications in minutes. But the research suggests writing code is no longer the main block.

“If the world froze at today’s level of AI capabilities, these results would be a bit of a cold shower.” — Leon Musolff

Will AI Coding Tools Improve?

So what lessons should companies draw from the findings? “If the world froze at today’s level of AI capabilities, these results would be a bit of a cold shower,” Musolff said.

However, the tools are improving apace. “We studied these tools in a previous paper, and it’s night and day,” said Musolff. “A 30% increase in software releases — there are very few technologies you can invest in today that deliver those kinds of gains.”

The researchers also found that each new generation of tools is tackling a later stage of the software development process, so the gap between gains in coding productivity and gains in finished software could begin to narrow as the AI gets better.

The paper says that if the tech can produce higher-quality code that requires less human review, today’s bottlenecks may prove temporary.

Some tech companies are already trying to tackle that problem by developing AI tools that review machine-written code. But Musolff is unconvinced they can yet match human judgement.

“If the same AI that wrote the code also reviews it, that doesn’t really solve the problem. The review just isn’t of the same quality,” he said.

“If the same AI that wrote the code also reviews it, that doesn’t really solve the problem.” — Leon Musolff

Is User Adoption the Next Barrier?

Even once software is released, it still has to find traction with an audience. The research found AI is increasing the number of new software applications, but not user adoption.

The researchers studied the four biggest software marketplaces — Apple App Store, Google Play Store, Chrome Web Store and SourceForge — and found a broad surge in new software applications since mid-2025. But crucially, no increase in overall usage.

On Apple’s App Store, for example, monthly new releases rose from around 30,000 before AI coding agents arrived in early 2025 to roughly 100,000 per month by April 2026. Yet total usage remained flat or declined across the four major app stores.

“It could simply be that it’s much harder to discover new applications when there’s such a flood of them,” said Musolff. “Alternatively, even once you’ve shipped an app, there’s another skill involved: iterating with users.”

In other words: Getting software into users’ hands is only the start.

---

©2026 Knowledge at Wharton. Originally published by Wharton School, University of Pennsylvania on September 8, 2026. Republished on Digital Information World with permission.

Fact-Checked by Irfan Ahmad.

Read next: “Almost Every Job Has Tasks That AI Can Change”. Economist Erik Brynjolfsson says AI’s transformation of work is just beginning and our institutions aren’t ready for it
by External Contributor via Digital Information World