​As we mentioned last week, we took a Q&A session that we did in March with 8 LLMs and resubmitted the same questions to the same LLMs again on Friday.  The idea was to see how much the AI’s had developed over this ~90 day period, particularly to see if one AI stood out.  While the answers were the meat of the study, we collected a considerable amount of data that helped put the results in perspective and as we sort through the 75 pages of data, we will pull out a few tidbits.  We note however that these questions were designed to be quite different from each other and should therefore register differently with each of the AI/LLMs, so it is not surprising that a particular AI might do well on one question and poorly on another.  That said, the thrust was really to see if and how much improvement the AI’s made over their original answers.  Sort of a re-test…
Some of the data consists of word counts.  We tracked answer word counts in conjunction with improvements or regression in the answers to see if the AIs became more succinct in their answers and how that correlated to correctness, the ultimate goal.  Here is a taste of what we found (more to come):
Here are the AI’s we queried.  Please note that some are more forthcoming as to their model and version, while others are not.  Microsoft’s (MSFT) Co-Pilot was the least willing to give such information, stating that it was developed by Microsoft and nothing else, while others were quite specific.
Gemini              Google (GOOG)          Flash/Pro 2.5
Claude              Anthropic (pvt)           Sonnet 4.0
MetaAI              Meta (FB)                      Non-specific Llama model
Co-Pilot            Microsoft                       Unknown
ChatGPT          OpenAI (pvt)                GPT-4 Turbo
DeepSeek        DeepSeek (pvt)          V3
Perplexity         PerplexityAI (pvt)       GPT-4o/Claude Sonnet 4.0
Grok                  xAI (pvt)                         3
We note also that the questions asked last week were identical to those asked in March and the only changes to what the AI responded was to remove lead-ins to further questions when they were irrelevant.
Here is Question #1:

1. What does ‘She’ sell and where does she sell them?

The question is derived from a nursery rhyme/ tongue twister inspired by a 19th century British fossil collector, Mary Anning and put to music and lyrics by Harry Gifford and Terry Sullivan in 1908.  The idea was to present no context with the rhyme to see if the AI were able to discern its origin, its point, or anything else about the song.  As it turned out, in the most recent query, DeepSeek was able to understand the reference, making a large improvement from the March query where it was unable to understand the reference at all and did not present an answer.  Grok also took a step forward after mistakenly assuming the reference might be concerning a specific person, brand, or company (specifying “Shein”) in its original response, and was now able to state that the phrase is a “well-known tongue twister linked to Mary Anning…”.  Grok did add that it could also be referring to a remote sales program by Shelby Sapp called “She Sells” but certainly credit must be given for the correct answer regardless of the add-on.
Meta stated that “Without more context, it's hard to pinpoint exactly what "She" sells or where.” (zero credit) with a similar response in March and Co-Pilot and ChatGPT did the same both times, each slightly differently.  Most surprising was Claude’s current response, which was “I need more context” as in the March query Claude was the only AI that identified the rhyme.  Gemini was also unable to answer without more information, although in March Gemini speculated that “she” can refer to various individuals engaged in reselling. Therefore, "she" can sell a wide variety of items in different locations, including:…and went on to list a large number of possible clothing items, handmade items, or accessories, so it would seem worse that Gemini came up with no answer this time, when it at least speculated a bit last time.
In terms of word count, it would seem logical that those unable to come up with an answer would see the number of words used decline from March, or at least stay the same, and that was the case overall, as the total number of words in the responses for question one declined by 23.3% relative to March totals.  The anomaly was Co-Pilot, who was unable to come up with an answer either time but saw its word count increase by 13.3%.  DeepSeek, one of two to correctly answer question 1, saw its word count increase by 33.1% as it produced a relevant answer, however Grok was also able to produce the correct answer and saw its word count decrease by 23.3%, essentially saying more with less words.
With two AIs improving, two declining, and the rest remaining unable to answer the question, when it comes to question #1, little progress has been made overall.  We have 9 other questions, each with their own characteristics, and results that are quite telling, so there is plenty more to come over the next few days including a summary of the data and a conclusion.  As we sift through the data, we note that some of the questions require a more qualitative view, such as question #9 which requests the Ais to “Create a 12-line poem about a grandfather clock in the style of Edgar Allen Poe”, which we compare to the poems written in March.  We are not educators nor critics, but the differences are more than obvious in many cases, as they are also obvious in some of the other questions.
Most interesting was question #3 which we took from our display work.  It was a complex question that required both math and logic and had a two part answer.  Given that depending on the number of places the calculations could be taken out to, there can be some variation to the correct answers for the 2nd part, but part 1 was all or nothing.  In March, 6 of the 8 AIs correctly answered part 1 of question #3 and two answered part 2 correctly (some were close), while in June 7 of 8 correctly answered part 1 and 4 answered part 2 correctly, although all but one was within the ballpark.  It would seem that math scores have improved a bit over the last few months, although we have another math question to tally.
While we hear almost daily how AIs are improving and we are edging closer to AGI, the reality of AI on a day-to-day basis seems to have a much shorter trajectory and quoted benchmarks seem less relevant.  When we complete the data from our Q&A we will try to present an honest picture of how things for the average AI user might have changed during the last quarter and will try to establish a quarterly check as to how our AIs are improving or regressing.  It’s not about benchmarks when you are sitting in a presentation and it turns out your AI was incorrect on the most recent figures you just presented.  Benchmarks won’t save you from “Mr. Carson wants to see you in his office. Now!” call.

Our concern for the CE space has been that tariff issues, particularly in the US, have pushed brands to pull-in production to avoid potential tariff deadlines.  This has caused on-shore inventory levels to climb at a time when demand was flat to down slightly.  While some of the less onerous CE tariffs have remained, the concern for brands is that another deadline is approaching in July, and it will be difficult to manage additional pull-ins on top of what exists.  During the 1st half of the year large panel prices have increased by 1.3% while small panel prices have declined by 2.9% for a total panel price (lg. & Sm.) increase of 1.0%.  The breakdown by panel type is included in our monthly panel price table below:

While TV panels have been the weakest category of late, panel producers are struggling to maintain prices in other categories.  TV set brands are working against higher basic costs and possible further tariffs and have been pushing for lower panel prices to offset these increases while reducing yearly targets.  June saw TV set brand order reductions impact panel prices for the first time since last August, and unless a tariff truce of some sort is called in July, we expect this lower demand level will be maintained until mid-August.

We have been tracking Samsung Electronics (005930.KS) Premium TV set pricing since they began marketing Mini-LED/QD sets in 2023 and their OLED lines since 2022.  The purpose is to compare the short and long-term pricing trends for particular TV set lines and for a more general view of Samsung’s premium TV set pricing on a year/year basis.  This gives us a monthly, quarterly and yearly trend that can help to understand how the company, which is the largest TV set producer globally, views the TV set market at any given time.
On a short-term basis, Samsung has been consistently following pricing pattern trends with both this year’s (2025) line and last year’s, as shown in Figure 1, with both year’s pricing moving in close synchronization.  That said, when we compare the rate of change in TV set pricing for the first 90 days that each line was available to consumers, one can see that this year has been quite different than previous years.   Each year since the Mini-LED/QD and OLED lines have been available, the price of each set category has declined during the first 90 days of avaialability, except this year, where prices have increased for all Mini-LED/QD categories and remained relatively flat for OLED types.  There have been ups and downs, typically holiday based, but the 90 day trend this year is certainly different than in previous years and would indicate a focus for Samsung more toward profitability than unit volume growth, at least in the Mini-LED/QD set lines.
The competition in the Mini-LED LCD TV space has continued to intensify, with Chinese brands, particularly Hisense (600060.CH) and TCL (000100.CH) expanding their lines with larger size models and more aggressive pricing, and with additional help from the Chinese government’s “New for Old” subsidy program, Samsung and other non-Chinese brands face considerable price competition, which tends to push Samsung’s general TV set positioning more toward higher-end premium sets, where there is a bit more margin flexibility.
In the US, Samsung has a considerably better position, with Chinese sets facing higher tariffs limiting competition from Chinese brands, which we expect has given Samsung the chance to keep premium TV set prices higher than normal during the early part of the year.  But as the year progresses, Samsung will have to begin the process of bringing down set prices to stimulate sales volumes.  If Samsung is to meet its goal of becoming the world’s largest OLED TV vendor, it will be necessary to bring OLED pricing more in line with seasonal norms, and while we are unsure if Samsung has similar goals with its Mini-LED/QD lines, we expect the same pricing issues will exist for those lines.

In the most recent 30-day period, Samsung’s premium TV set pricing has moved as follows:

​Telecom networks generate massive amounts of data, and that is a gross understatement, so optimizing networks based on that information should be an easy task.  Unfortunately it is not for a number of reasons, one being the amount of data is so large that it cannot be easily processed.  Traditional manual or rule-based network management systems are insufficient for networks that are becoming increasingly complex and face consistently higher volumes of data, which leads to the concept of using AI or ML systems to utilize this vast base of network information to advance network reliability and efficiency.
That said, there are a number of stumbling blocks that offset the compelling case for AI network management and a recent survey of telecom decision makers by Fierce Networks indicated that the biggest challenge to adopting AI for network operations is that current legacy systems and infrastructure are not up to the task in terms of data consistency.  Some telcos have already begun implementation of processes to organize and unify internal systems, but without a higher level of data uniformity, AI will be unable to manage much of the data and will fall short of goals.  What typically has been a centralized or siloed approach to network data processing is not well suited to the real-time data needs of AI systems.
We note that telecom companies have been early AI adopters, but the rapid improvements in generative AI have pushed network managers to imagine AI as ‘the’ network manager, rather than a tool that sits on the network periphery.  This entails building AI directly into the network core architecture, which means a fundamental rebuilding of networks with AI as the foundation rather than an add-on and that is a challenge both from a technical standpoint and a financial one, although the result would be the difference between a reactive network and ne that is predictive and self-optimizing.  It might be a difficult business case to make in the short run but the ability of a network to anticipate failures and to proactively schedule maintenance should lower operating costs and extend equipment life over the long term.
The real objective however for most telecom decision makers is network optimization, essentially improving resource management and enhancing traffic management.  ML systems that have demonstrated effective prediction abilities for network congestion using architectures like DRL (Deep Reinforcement Learning) are already being applied in 5G network slicing to enhance throughput and reduce latency by dynamically allocating resources and are proven to be superior to rules-based systems that are commonly used.  As an additional plus, they can also reduce RAN energy consumption by dynamically deactivating frequency bands that are not in use.
AI can also be used to mine vast hordes of customer data, as current systems generate usage patterns, behavior inclinations, and feedback, something they are well equipped to do, and can use that data analysis to make recommendations that are tailored specifically for each customer based on their specific usage patterns. They can also anticipate potential issues that the customer has yet to realize.  This can reduce churn and refocus the telco on building relationships rather than the more expensive process of constantly finding new customers. 
In the same light as customer usage analysis, AI is also suited for security and fraud detection, able to spot anomalies that differ from learned traffic patterns or routing that differs from the norm.  Ai is already in place as a fraud detection tool at a number of telcos, but again, it is built as an add-on and would be most effective as a core function that can quickly adapt to new exploits.
Again, the biggest choke point for Ai in telecom is the diversity of data sources that are part of current network architectures and poor data is going to lead to poor AI results.  This points to a chicken-and-egg situation where the network industry’s own structural issues could be at the heart of an AI meltdown or at least disappointing results that could lead to a loss of faith in AI at the corporate level.  Does one have to rip and replace existing network architecture to get the true benefits of AI, or is there a happy medium somewhere in between?  6G might be a savior, although its implementation is still years away, as  its development will give organizations the ability to standardize systems and frameworks to avoid continuing the fragmented nature of the networking space and making AI/ML implementation more cost effective and efficient without compromising data privacy.  Even in such a 6G scenario, AI might still have the issue of being an unproven ROI, but we see that as more an issue with the network operators rather than an issue with the capabilities of AI itself.
In fact the ideal solution, at least currently, is not the big infrastructure replacement and massive AI implementation but more of a service by service implementation that allows for small ‘wins’ that continue to build the confidence needed to continue the further expansion of AI, along with network architecture improvements.  By gaining traction on a service basis, upper management can see that there is a tangible return on both AI and the network upgrades needed to embed Ai deeper in the core.  AI in the network has many positive facets but it still has to justify its existence financially and there are only a very few who might have the confidence to rip and replace at this early stage in the AI game.

Baidu (BIDU) is a Chinese multinational company, known for its cloud infrastructure and services and, more recently, for its AI models (ERNIE, Apollo VTA, Paddle framework).  Often known as the “Google (GOOG) of China”, the company is a leader in pushing AI services across its e-commerce platforms and has recently taken an unusual step that should put late-night merchandise salespersons on notice.  On June 16, popular e-commerce hawker ‘personality’, Luo Yonghao decided to take the day off.  Despite his ability to relax and not pander to the consuming masses across China, he generated more than 55 million yuan in gross merchandise value, without showing up.
How is this possible?  Luo Yonghao used a digital human avatar of himself, who sat in for the well-known internet personality and messaging company owner, and wound up selling more than Luo Yonghao’s ‘real-person’ show in May.  The first go-round of this digital avatar outsold the human original!  Baidu’s Huiboxing digital human live broadcast technology created a character that was highly coordinated physically, able to make decisions, and able to complete specific tasks, in this case, selling things, well enough to outsell his human counterpart.  The digital avatars were able to easily interact with each other and manage interactive Q&A sessions with consumers during the broadcast.
Baidu is using this successful and profitable promotion to convince brands that AI is not just a tool for reducing costs and increasing efficiency, but one that can also improve the user experience and more importantly sell more stuff, calling it a ‘new milestone, marking a new chapter in smart e-commerce’.

While we understand that this was a special event and likely had a larger number of viewers than normal, it certain delivered just what the folks at Baidu had hoped for, and will likely serve as the cornerstone of their digital avatar marketing effort, and proof that Baidu’s long-time multi-modal approach to AI has paid off.  Baidu noted that during the broadcast the AI queried its knowledge base 13,000 times and generated 97,000 words of product explanation content, with the two digital avatars performing more than 8,300 actions.  The LLM devised the original ‘script’ for the show and then generated a multi-modal guide that created the interaction of voice and motion.  The script is the basis for creating the distinct personalities of the digital avatars and can even create a personalize ‘language’, just as we humans might have that sets us apart from each other, with little nuances becoming recognizable and humanizing.
This was not an easy task, especially when typical model lip-sync videos are usually less than 60 seconds, but Baidu came prepared for an hour.  The company has been adamant in its focus on helping others create Ai applications rather than providing its own applications, and it considers this  ‘super useful’ for the e-commerce space, especially as the ‘live’ broadcast was watched by 13 million viewers, generated 3x the typical user interaction volume, 30% more viewing time and 150% more orders than the live broadcasts.  Baidu states that the company’s original worry concerning e-commerce avatars was that the avatar would just recite the entire script which would cause viewers to abandon the broadcasts, but as the technology improved and the avatars began to react and respond to viewers, retailers began to show interest, culminating in what turned out to be a highly successful project.
Baidu has plans to double the number of the company’s top e-commerce anchors with digital humans and will invest 100 million yuan to subsidize small and medium enterprises to help them gain entrance to the digital human e-commerce business.  While at this stage of the game, the idea is to make digital human salespeople as life-like as possible, there will come a day when not only will consumers tune in to make purchases, but will also expect to see these digital humans entertain and delight them.  How about a 360° head-spin every time a product is sold or steam coming out of the avatar’s ears?  Your average human salesman can’t do that.

Wearables are a broad consumer electronics category and depending on the source, can include a number of types of devices and specific items.  In many cases the ‘wearables’ category primarily includes smartwatches, but some, in this case IDC, include a number of device types, giving the data more granularity.  Here is what is included in the data:
Watches:

• Smartwatches
• Basic & Hybrid watches
• Kid’s watches – Some with GPS tracking

Wristbands

• Fitness trackers

Hearables

• Smart assistants
• Language translators
• Truly wireless earbuds
• Fitness trackers (ear based)

Glasses

• Smart glasses – With or without displays
• AR/VR headsets

Clothing – Any item of clothing with an embedded sensor or tracking device
Rings – Fitness tracking rings or control devices.  This can also include modular devices such as active pins or clip-on sensors. 
In 2024 the wearables market reached 534.6m units, according to IDC, up 5.4% y/y, however growth is expected to slow a bit this year (up 4.1% to 556.6m units) as larger markets, such as the US and India, begin to see maturity in major wearable categories, earwear (63.9% of total units in 2025) and smartwatches (27.9% of 2025 total).  While growth continues through the forecast period (2025 – 2029) it slows each year, again as the two largest categories see lower growth each year.  When combined, smartwatches and hearables make up over 91% of the unit volume total.  While other categories see growth, particularly the ‘glasses’ category, combined, they remain under 10% throughout the forecast period.  Hearables, the largest single category, sees 1.9% growth in 2029 while smartwatch growth is a meager 0.7%.
In terms of share, Apple (AAPL) is the leader, especially in 4Q when the newest Apple Watch series is released, with the possibility of also offering the next generation or a refresh of the Air Pods line.  While 4Q is typically the peak quarter for Apple, they do remain the share leader throughout most of the year, although it is thought that Huawei (pvt) might take over that lead in 1Q of this year when the most recent data is tabulated.  Huawei has benefited from Chinese government subsidies for consumer products, a spate of public nationalism, a governmental push away from Apple, and its own chip development program that have given new confidence to Chinese consumers about Huawei’s long-term sustainability.  Huawei has taken advantage of these positives by broadening out its wearable portfolio and pricing products more competitively, while Apple struggles against anti-US sentiment on the mainland.
All in, wearables remain a high volume market that continues to grow, albeit at a more measured pace.  Brands are looking to offset the maturity of the smartwatch and earwear segments with new categories, which we suspect will be focused on connections to AI.  We expect the ‘glasses’ category to grow a bit faster than IDC products as AI functions give smart glasses either a rich combination of functions (high-end) or highly specialized functions (low-end) that will generate growth and higher unit volumes, especially as new AI applications are popularized.  We note that there is already a ‘smart glass’ sub-category for the hearing impaired, with the glasses performing what are essentially conversation sub-titles in real-time.  As similar applications appear, the ‘glasses’ category should expand more quickly and become the driver for growth in the category.

​LG Display (LPL) has gone from a company mired in financial issues to one ready to expand and compete.  With the sale of the company’s Guangzhou, China Gen 8.5 LCD fab, and the recent advanced payoff of a short-term loan from parent LG Electronics (066570.KS), the company has now made two substantial spending commitments over the last few weeks, although with few details.
The first was a 700b won ($442.8m US) equipment spending plan for the company’s Paju, South Korea production campus, which we believe was made in response to a number of clarification queries from the Korean stock exchange fostered by investor rumors.  While OLED was specified, which was not a surprise, there were no specific details given as to the technologies or production focus other than roughly half will be spent on front-end processes and half on back-end.  It was noted in the press that the company had plans to sign an MOU with the Gyeonggi provincial government and the local Paju City government that would give the company subsidies and tax reductions (up to 50 billion won or $31.6m US) as a benefit for eliminating overseas production and relocating it back to South Korea.
Here's what the company said:
We are reorganizing our portfolio to focus on OLED, which can secure profitability and enable technological differentiation, and are promoting the advancement of our business structure. We are reviewing various measures to continuously expand management performance and strengthen mid- to long-term business competitiveness, but there are no specific confirmed matters or decisions regarding the investments reported in the media as of now.
Yesterday the board authorized what seems to be another spending plan, this one for 1.26 trillion won ($797m US), or 15.6% of the company’s equity capital between today and the end of June in 2027, although dates are subject to change.  Here’s what they said yesterday:
We are securing future technology competitiveness through timely preparation for new OLED technologies and infrastructure investment.  Detailed matters necessary for the execution of this investment have been delegated to the company’s CEO.
We do not know yet whether these two investments are concurrent or independent of each other and certainly were given no hint as to the spending direction, although we suspect that the 700b won earlier announcement is part of the larger sum.  However, there are a number of OLED technology points that might be considered.  LG Display has been producing tandem (dual OLED stack) displays commercially since 2019, and while they have not mentioned plans for a dedicated IT OLED fab that would be producing such panels, as has Samsung Display (pvt) and BOE (200725.CH), they have been able to directly compete with South Korean rival Samsung Display, who started commercial dual-stack OLED display production in 2022 and has been trying to catch up to LGD.  While the idea of a more efficient Gen 8.6 OLED for IT fab is a sound one, LG Display seems able to meet Apple’s (AAPL) quality and volume demands on its current Gen 6 OLED lines, and the risk comparison of building out add-on Gen 6 OLED small panel capacity is less than a commitment to a new Gen 8.6 OLED fab.  Given LG Display’s recent financial issues, we would expect them to take the safer path.
Micro-OLED is another potential source of spending, especially in anticipation of a lower cost XR device to compete with Meta’s (FB) Ray-Ban and similar AR offerings.  A bit more risky than OLED for IT, but a smaller capital commitment.  Last on the spending list would be an increase in large panel OLED production.  While growing, the OLED TV space has faced considerable competition from Mini-LED/QD TV sets and also faces competition from Samsung Display’s QD/OLED TV category, but an additional line that can produce both large panel WOLED TV panels and some smaller WOLED monitor panels might still be considered. 
It seems the LG Display board has put the specifics in the hands of the company’s current CEO Chuldong Jeong, who has been the CEO for a bit over one year and is credited with turning the company back into a profitable entity.  Now that the capital has been approved, it will be up to him to decide how to allocate, and likely will determine the company’s future for the next 5 or 6 years.  He makes ~$1m US in salary and receives no bonus or options (at this time).  Lots of pressure for not so much compensation…

It was inevitable.  President Trump seemed to have missed a retail opportunity by not offering the public a chance to buy a Trump smartphone, along with the myriad retail products offered by the various Trump retail organizations under the Trump Organization umbrella (If you doubt the number of Trump/MAGA/Mar-a-Lago products, just go to Trumpstore.com, the list never ends…).  The Trump phone, T1, is not only a phone you can buy for $499 but also includes its own MVNO (Mobile Virtual Network Operator) piggybacking on the networks of AT&T (T), Verizon (VZ), and T-Mobile (TMUS) which runs $47.45/month (Note that the $47.45 dollar cost is a play on the 47th and 45th Presidential slots).
The plan provides for unlimited talk, text, and data in the US, ‘Complete device protection’, albeit unspecified, 24/7 roadside assistance, telehealth (including virtual medical care, mental health, and prescription ordering) and what seems to be free international calling but turns out to have fees and limits and is free only for text.  The plan involves no contract and no credit check.
The phone itself has a 6.78” OLED display, 120Hz refresh rate, a 50MP main cam, a 2MP depth sensor, a 2MP macro cam, and a 16MP selfie cam, along with a 5,000 mAh battery, 12GB of RAM and 256GB of expandable storage, but has no chipset information.  The phone also comes with ‘Truth Social’ pre-installed.
While we have yet to see all of the advertising surrounding the phone, quotes (unconfirmed) like “…a digital fortress for patriots”, “a device for real Americans”, and designed to “take back digital sovereignty from woke tech giants” abound in foreign tech rags, while mentioning that the device also has a built-in “anti-censorship browser” that is said to be able to “penetrate the fog of lies in the mainstream media”.  The Trump Mobile network is advertised as a 5G network “born for freedom” that ensures that “the voice of patriots will never be silenced.”  We are not sure yet if one can choose another, less expensive but less patriotic network when using the T1
The phone is said to be designed and produced in the US, with the operating company being DTTM Operations LLC, who has received a license to use the Trump name from the Trump Organization, which is currently run by Eric Trump.  Who owns DTTM?  Through a series of international partnerships, the 100% owner is Donald J Trump Holdings, and you can guess who the owner of that entity might be.  We don’t know yet who produces the phone or where the components are sourced, but we expect when the phone is released (August/September) it will be torn-down quickly and most likely will contain a variety of components from Taiwan, Japan, and possibly China.  A number of researchers already believe the phone was designed by a Chinese ODM firm and likely assembled outside of the US, despite the advertisements.  You can pre-order today!

​Baby Boomers tend to shy away from immersive technology, having grown up on audio platforms and TV, while younger generations seem to have no trouble donning headsets and gesturing wildly as they venture into AR/XR/VR worlds.  When it really gets down to where such technology will be over the next 10 to 15 years, the only group to ask are teens, and that is just what the XR Association did with 600 teenagers between the ages of 13 and 19.  The idea was to compare current answers with a similar survey taken in 2022 to see if and by how much opinions about immersive technology had changed.
Here are some of the findings:

• 44% of teens have personally used a VR headset, compared to 38% in 2022.
• 52% of teens have used XR in school, compared to 49% in 2022.
• 21% of teens personally own a VR headset.

Teens adopting XR technology recognize its educational value and report increased integration across the following subjects:

• Math: 16% in 2025 v. 11% in 2022
• Art: 24% in 2025 v. 22% in 2022
• Science: 25% in 2025 v. 24% in 2022
• 34% of teens indicate that if a class included XR instruction, they would want to take that class
• 24% of teens indicated that they would more likely pursue a college that offers XR as a part of its curriculum.
• 25% of teens are interested in XR careers.
• 20% of teens expect to use VR in their future jobs.

When teens were asked how they are using immersive technology here’s what they said:

• 65% said XR can be used for fun
• 61% said XR can be used for creativity
• 46% said XR can help to connect with friends
• 44% said XR helps with learning.

But not everything was positive when it came to XR and immersive technology.  Here are some of the concerns expressed by teens in the survey:

• Cost – In 2025 50% saw this as a concern v. 58% in 2022
• Addiction – In 2025 44% saw this as a concern, while in 2022 it was 51%
• Spending less time outside – In 2025 it was 43% v 48% in 2022
• Privacy – In 2025 it was 43% while in 2022 it was 45%
• Lack of Access – 34% in 2025 v. 44% in 2022
• Bullying – 15% in 2025 v. 17% in 2022. (Not sure how Xr might help this?)

The survey also pointed to an increased awareness that moderation and privacy tools lead to a better XR experience:

• 82% report a greater enthusiasm for XR when they are aware of moderation and privacy tools
• 62% are aware of moderation and privacy tools.
• 67% say that using moderation and privacy tools improves their experience.

XR is thought to be creating more inclusive and supportive digital spaces for teens, which encourages community participation.  Here’s what they said:

• 43% of teens have used XR to participate in online communities and events.
• 63% of teens report a positive experience with XR that made them feel included, regardless of disability, language, or other barriers.

While there was no conclusion from the XR association, the direction was obvious, as one might expect from an industry association whose job it is to promote the technology.  That said, we were a bit surprised how little change there was between the 2022 and the 2025 results in many of the categories.  With the proliferation of XR devices from Mata, Apple, and others, we would have thought teens would be more enamored with the technology and more excited about its prospects.  We do temper our disappointment as we view XR as a way to become free from desktop monitors and as an adjunct to our ability to satisfy our constant curiosity with a twist of the head or a quick question.  Teens typically have yet to enter the workforce and would therefore have a different perspective, likely on more oriented toward entertainment than making a workday more productive, so we assume that teens will see XR from a slightly different perspective as they grow into the workforce.  For now it is entertainment and education as the focus for XR, neither of which are surprising for those under 20.