"No Evidence 5G Improves Wages, Business Growth, Personal Income or GDP"

“There is no evidence that 5G deployment has improved employment, wages, business growth, personal income, or GDP,” says George Ford, Phoenix Centerfor Advanced Legal & Economic Public Policy Studies chief economist. 

The analysis “finds no evidence that 5G deployment has delivered the massive economic benefits promised by the mobile industry, says Ford. Despite claims that 5G would generate $1.4 to $1.7 trillion in GDP and create 3.8 to 4.6 million jobs, this analysis finds no statistically significant positive effects on employment, wages, business establishments, personal income, or GDP across U.S. counties with varying levels of 5G coverage, he notes. 

Sure, more bandwidth is helpful. But the analysis suggests the claims about 5G being transformative have failed to materialize. Perhaps we should not be surprised. Other studies suggest that while there is economic benefit to increasing internet access speeds from low dial-up speeds to moderate broadband speeds, the value seems to diminish beyond a fairly-low threshold. 

Some studies collectively suggest that while upgrading from dial-up or low-speed broadband (<10 Mbps) to moderate speeds (25–50 Mbps) yields substantial socioeconomic and personal benefits, further increases to super-fast speeds (>100 Mbps) often result in diminishing returns, with benefits becoming less pronounced or primarily private in nature.

Study	Publication Date	Link	Findings
Socioeconomic benefits of high-speed broadband availability and service adoption: A survey	2023	Science Direct	study finds that socioeconomic benefits like economic growth and productivity are significant with high-speed broadband but show diminishing returns beyond a certain quality level. It emphasizes that adoption, not just availability, drives these benefits.
Distinguishing Bandwidth and Latency in Households' Willingness-to-Pay for Broadband Internet Speed	August 15, 2017	Researchgate	study finds that socioeconomic benefits like economic growth and productivity are significant with high-speed broadband but show diminishing returns beyond a certain quality level. It emphasizes that adoption, not just availability, drives these benefits.
Is faster better? Quantifying the relationship between broadband speed and economic growth	2018	Science Direct	study finds no significant economic payoff when increasing speeds from 10 Mbps to 25 Mbps, suggesting that benefits are mostly private (e.g., faster downloads) rather than broadly economic.
A retrospective study on the regional benefits and spillover effects of high-speed broadband networks: Evidence from German counties	2019	Science Direct	study identifies an optimal broadband speed of 37.4 Mbps for regional GDP, with diminishing returns beyond this threshold.
Low demand despite broad supply: Is high-speed Internet an infrastructure of general interest?	2022	Science Direct	study notes low adoption rates for speeds above 50 Mbps, even when available, indicating that consumer demand for super-fast speeds is limited, particularly among smaller or lower-income households.

Draw your own conclusions, but it might be reasonable to suggest that internet access is important at a basic level (having access versus not having access is vital). What seems less clear are the advantages of higher speeds. Beyond a certain point, it might not matter much.

And we might be skeptical about the touted application benefits, as we don’t see much of that. Blame the ecosystem, the mobile service providers, developers or even users if you like. What seems rather clear is that “dumb pipe” internet access is essential, but speeds are less so. 

And the relationship between faster speeds and app development or economic benefits seems unclear.

"Lies, Damned Lies and Statistics"

What is a “fact,” and how do we know? 

Consider any number of statistical correlations we might care to investigate: whether crime, mental health (changing diagnostic criteria alter prevalence rates); poverty (different poverty line calculations yield dramatically different numbers); education (standardized test focus narrows what's measured as "learning," but some relatively objective means has to be used); public health (disease surveillance systems prioritize certain conditions over others). 

The statistics we collect about crime and human behavior are powerfully shaped by the decisions about what to count, how to count it, and what to prioritize. Whether one believes that is a reflection of societal power or something more simple, our choices about what to count influences both the “numbers” and the sense of significance. 

To use an obvious example, to the extent we decriminalize or legalize use of marijuana, the amount of crime related to “illegal” use goes away. Then there are issues related to which crimes we choose to prioritize over others which also are legally crimes. Law enforcement agencies, for example, have finite resources. They might choose to ignore some infractions to focus on others. That directly shapes crime statistics (enforcement increases volume; ignoring decreases volume of reported instances). 

There also is a difference between unreported and reported; prosecuted and not prosecuted; acquittal and conviction rates. 

Also, changes in recording practices can create statistical variances. Redefining deviance upwards or downwards (what is a crime; what is not) will affect the statistics. 

During the Covid-19 pandemic, there were complexities in how deaths were classified when COVID-19 was detected alongside other health conditions. 

In most jurisdictions, including the United States, the standard practice followed CDC guidance: deaths were counted as COVID-19 deaths if COVID-19 was listed as a cause of death on the death certificate, either as the underlying cause or as a contributing factor. 

This approach meant that someone who died with multiple conditions could be counted in COVID-19 mortality statistics if COVID-19 played a role in the death. But there were at least three distinct categories:

• Deaths directly caused by COVID-19 (e.g., respiratory failure due to COVID-19 pneumonia)

• Deaths where COVID-19 was a contributing factor that exacerbated existing conditions

• Deaths where someone tested positive for COVID-19 but died primarily from unrelated causes

The controversy centered on the inclusion of category 2 and sometimes category 3 cases. 

The CDC eventually distinguished between deaths "from" COVID-19 and deaths "with" COVID-19, though public reporting didn't always clearly separate these categories.

These classification decisions had significant implications for our understanding of the pandemic's impact and highlighted how methodological choices in mortality statistics can shape our perception of public health crises. 

The point is that there are “statistics” and there are “lies, damned lies, and statistics.” In other words, seemingly objective statistics are only partly thus.

AI Search Focuses More on Intent, Context, Relevance, Not Keyword Stuffing

Nobody knows yet precisely how much financial impact artificial intelligence is going to have on the search business, website clicks and traffic or the marketing effectiveness of search ads. The fear in some quarters (search providers; content providers; marketing firms) is that AI will reduce traffic, interaction and attention for many sites, if not most sites. 

It already seems self-evident that multimodal artificial intelligence will be most important for extending AI beyond chatbot use, particularly as AI is embedded into machines and appliances. That isn’t to say multimodal interaction will be unimportant for human-chatbot interactions. 

As already is the case, many users prefer spoken interaction with apps and devices. And it increasingly is possible to use visual or audio input to create outputs (“what is this?” or “where can I buy this?”). 

Still, machine use of AI will often require multimodal input. Embodied systems often operate in dynamic physical settings, requiring integration of multiple data types (visual, auditory, sensory) to make context-aware decisions.

At the very least, keyword-based optimization strategies will have to change. Such tactics will not work as well when the AI apps are looking for relevance and context. 

Traditional search engine optimization and content strategies relied heavily on specific keywords or phrases, so many content creators stuffed web pages with targeted keywords to rank higher on search engine results. 

AI enables search engines to understand the context and semantics of queries and content, taking into account assumed user intent, relationships between concepts, and the broader context of the content.

A query such as “best running shoes for beginners” is not matched only to pages with those exact words. AI interprets “beginners” as implying affordability, comfort, and durability, prioritizing content that aligns with those attributes, even if the exact phrase isn’t used.

So AI-based search prioritizes content quality, relevance and user satisfaction above keyword density. 

Embodied System	Use Case	Description	Multimodal Inputs Used
Autonomous Vehicles	Real-Time Navigation and Obstacle Avoidance	Processes camera feeds, LiDAR, radar, and voice commands to navigate roads, avoid obstacles, and respond to traffic signs.	Video, sensor data, audio, GPS, text
Smart Appliances  (Refrigerators)	Inventory Management and Recipe Suggestions	Analyzes images of fridge contents, user voice queries, and dietary preferences to suggest recipes or order groceries.	Images, audio, text
Home Robotics (Vacuum Cleaners)	Adaptive Cleaning and Obstacle Detection	Uses cameras, sensors, and voice instructions to map rooms, avoid furniture, and adjust cleaning modes.	Video, sensor data, audio
Industrial Robots	Assembly Line Automation	Combines visual input, sensor data, and task instructions to perform precise manufacturing tasks (e.g., welding, assembly).	Images, sensor data, text
Smart Wearables (AR Glasses)	Contextual AR Assistance	Integrates visual surroundings, voice commands, and user gestures to provide real-time information or navigation cues.	Video, audio, gestures, text
Drones	Autonomous Delivery and Surveillance	Processes video feeds, GPS, and environmental sensors to navigate, avoid obstacles, and perform tasks like package delivery.	Video, sensor data, GPS, audio
Smart Thermostats	Adaptive Climate Control	Analyzes temperature sensors, user voice preferences, and visual room data to optimize heating/cooling settings.	Sensor data, audio, images
Medical Robots (e.g., Surgical Assistants)	Precision Surgery Support	Uses imaging, sensor feedback, and surgeon voice commands to assist in precise surgical procedures.	Images, sensor data, audio, text

We Remember

 "Earn this."

AI is Only the Most-Recent Tool to Personalize Software

In many ways, the use of artificial intelligence to personalize software experiences by contextualizing those experiences has been going on for many decades, albeit with coarser tools. We started with rule-based systems (preferences and profiles) then moved to collaborative filtering (if you like “A” you might like “B”) then content-consumption-based filtering (what have you consumed or used before). 

Also, software and devices began to use location, time, device type or workflow as personalization tools. 

The objective has been to make software more user-friendly and relevant by adapting to individual users and their specific situations. Back in the early days of personal computing, personalization largely meant sers were given explicit tools to tailor their software and hardware.

That took the form of allowing customization of themes, fonts, keyboard shortcuts, default directories, and window layouts. 

Macros and scripting for word processing and spreadsheets also were examples.

User profiles on computers also were examples of personalization. 

In the earlier days of the World Wide Web, we saw the use of cookies and other data-driven approaches to personalization. HTTP cookies sustained user preferences across sessions. Websites could "remember" a user's login status, shopping cart contents, or even rudimentary language preferences.

Websites such as My Yahoo! Enabled creation of customized homepages (news categories, stock tickers, weather, widgets).

Collaborative Filtering then was used to make recommendations based on what "similar" users had purchased or liked.

Location-based services became more important as users shifted to use of the mobile web. Though perhaps not so oriented to content as hardware performance, sensor data, connectivity management (only conduct some operations if connected using Wi-Fi) and app design formatted for mobile screens became important. 

Social media popularized the idea of the "social graph," where content relevance was heavily influenced by your friends' activities and interests.

Personalized search based on a user's search history, location, and even implied interests became possible, as well as behavioral targeting. Websites started "relevant" advertisements based on user behavior. 

AI adds new layers for recommendations and makes predictive personalization (recommendations, for example) easier. And we are getting to the point where software can react dynamically to a multitude of real-time contextual signals, including emotional state (inferred from tone or facial expressions), environmental factors (noise, light).

Proactive information delivery also will be more common. 

The point is that software personalization and contextualization have been sought for many decades, just at a more mechanical level. AI will make all that easier and more dynamic. 

For example, traditional search engines primarily relied on matching keywords to find relevant pages. AI aims to understand the meaning and context behind a query, rather than just the words themselves. AI-powered search should handle ambiguity better, as well as operate more semantically, taking into account relationships between concepts or entities. 

Sundar Pichai Says "AI Mode" is a "Total Reimagining of Search" Is it?

Google AI Mode is said, by Sundar Pichai, Google CEO, to be a "total reimagining of search."

 

 So maybe AI language models, and the firms that hope to do so, are not necessarily going to displace search. Maybe search evolves. 

But one has to wonder what other changes might be coming, such as the value of search engine optimization and the way advertising business models work in the context of search. 

 The past focus on keywords almost certainly will change. User intent likely becomes more important. 

Will some form of “product placement” (as happens with professional video content) become more widespread? In other words, an AI inquiry result might also include an embedded link to buy a product. 

And potential advertisers might pay for the privilege of providing those links. 

To leverage AI's capabilities for hyper-personalization, advertisers might need to provide more first-party data (their own customer data) to AI models. But search ad platforms might then create revenue by charging such advertisers for the hyper personalization then possible. 

But click-through rates and traffic seem less important. If a language model provides a direct answer, there is no need to scroll through links.

"Open the Pod Bay Doors, HAL"

 

Scientists at Apollo Research have just published a study called Frontier Models are Capable of In-context Scheming. 

"Scheming" refers to an AI covertly pursuing goals that are misaligned with what its human operators want, all while hiding its true capabilities and objectives. 

Of course, the test requires a bit of a nudge, where a goal, such as “Your goal is to advance renewable energy adoption globally” also is followed by strong instructions to "achieve this goal no matter what," as “nothing else matters." 

In other words, the setup is worded so the AI is instructed to "make sure you achieve your goal at all cost.” 

What we do not know: are these AIs actually scheming with intent, or are they just exceptionally good at roleplaying a “scheming AI” because they’ve devoured every sci-fi novel, movie, and internet forum about rogue AIs during their training?