This article was written with AI (Claude) and reviewed by an editor. Our company is a paying customer of both Anthropic and OpenAI and also tests Google models. We operate an AI influencer (Ria) ourselves, so we stand in the same line of business as item 1 — which is why we add no evaluation of any company's broadcast and carry over only the figures and sentences. Figures announced by companies are marked "announced by the company," and content from a single source is given together with that source. The legal explanation in item 2 is not legal advice; how it applies to our company is something we confirm with professionals. This article is not an investment opinion on any company.
In NO.43 we wrote that "customers cannot bring their own agents," and in NO.47 that "the last step of the wallet is a human hand." This week's news runs the other way — toward emptying the place where a person used to stand. A home-shopping broadcast removed its host and its camera; a telecom's business processes removed the human hand. But the two companies wrote the "condition for removing" in different ways, and Korean law has its own answer to whom "if you removed the person, say so" is addressed. We separate facts, interpretation, and what we would do.
1. Hyundai Home Shopping — a broadcast "with no host and no camera," on the schedule at least monthly (facts, company announcement)
On Sunday, November 15, Hyundai Home Shopping announced that it is adding its "100% AI broadcast generation system AI-REAL" to the regular schedule of its data home-shopping channel Hyundai Home Shopping Plus Shop (reported by ZDNet Korea, Money Today, The Korea Herald and the Seoul Economic Daily English edition, with matching content). The first broadcast is Monday, November 16 at 3:20 p.m., Jeju Heart Gold kiwis. By the company's description there is no human host on air and no camera shoot, and AI generates both the voice and the video; it said that replacing some scenes or a presenter with AI had been done before, but making the entire broadcast with AI is a first in the industry ("first" is the company's description). It wrote that scenes hard to show with filming or props — the soil conditions where produce is grown, the distribution path, cooking examples — are rendered as video without a separate shoot.
The company announced two numbers from the trial broadcast. Costs borne by suppliers — sample preparation, insert-video production and the like — fell more than 80% from before, and sales were 120% of target. The timing, product and number of trial broadcasts are not in the reports. Lim Soo-jin, head of broadcast production, said the customer response to the trial was that it felt "as natural as a broadcast with a real person on air," and that information such as growing conditions and cooking examples helped purchase decisions. Lim Hyun-tae, head of the channel strategy division, said it was developed "to grow together with small and mid-sized suppliers and offer a differentiated shopping experience." The plan is at least once a month, starting with agricultural, livestock and fishery products and widening to categories with strong demand from smaller suppliers, such as processed foods and living goods.
Some things appear in none of the five reports we read — how viewers are told that the broadcast is AI-made, what the on-screen presenter looks like (only The Korea Herald's photo caption uses the phrase "virtual presenter"), the cost of the AI system, which technology or partners were used, and ratings. Whether they are absent because no one asked or because no one answered is not something we can know.
We also note the background. Attempts to "remove the person" have continued all year in the home-shopping industry. According to the Financial News (March 25), GS Shop put a generative-AI virtual model, "Jaei," on live broadcasts alternating with real models, and Lotte Home Shopping said broadcasts with its virtual host "Lucy" had reached 50 billion won in cumulative orders, with sales 50% higher than the same time slot (company announcement, one outlet). Newsway (May 21) wrote that NS Home Shopping put an AI shopping host into a roughly 10-minute processed-food broadcast for the first time on May 20 (excluding product demonstration and cooking scenes), and that GS Shop had widened virtual backgrounds to about 60% of all broadcasts. The industry figures in the same article are that TV home-shopping gross merchandise fell from about 21.9771 trillion won in 2021 to about 18.505 trillion won in 2025, and that carriage fees were 73.3% of broadcast revenue in 2024 (article figures; we did not open the underlying data). Digital Daily's NS Home Shopping article (July 28) said production AI cut banner production from about 40 minutes to 5 and broadcast post-production from about 4 hours to 40 minutes (figures the company shared at an internal town hall). Hyundai Home Shopping itself said in the March report that it had cut video production time "from two days to one hour." So this announcement did not come out of nowhere; it sits at the end of a sequence of emptying — "some scenes → some presenters → the whole broadcast" (an arrangement of facts, not an evaluation).
2. The AI Basic Act's "say it's AI" — who, when, how (facts)
Korea's "Basic Act on the Development of Artificial Intelligence and the Establishment of a Foundation for Trust" (the AI Basic Act) took effect on January 22, 2026, and its transparency duties are in Article 31 (the article numbers and contents match across Professor Lee Seong-yeop's February 4 piece in the Law Times, the February 11 newsletter of the law firm Shin & Kim, and the January 21 reports of Newsis and Asia Economy. The National Law Information Center page would not load its body on our side, so we did not read the statute text directly). There are three paragraphs — (1) prior notice: tell users in advance that the product or service is operated on high-impact AI or generative AI (terms of service, contracts, in-app screens, offline postings, etc.). (2) labeling of outputs: indicate that an output was produced by generative AI. (3) deepfakes: virtual sounds, images or video that are hard to distinguish from reality must be labeled so that users can clearly recognize them.
The "AI Transparency Guideline" published by the Ministry of Science and ICT on January 21 (draft in September 2025) sets the methods. If the output is used only inside the service, flexible means such as UI notices or a logo are allowed (a chatbot can give notice before use; a game or metaverse can mark a character as AI); if it leaves the service (download, sharing), a visible or audible watermark that people can recognize, or "notice followed by machine-readable metadata," must be used; and deepfakes without exception must be marked so that people can tell — video throughout playback, audio at the start. Sanctions are a corrective order for violating the labeling duty (Article 40(3)) and an administrative fine of up to 30 million won for violating the prior-notice duty (Article 43) (Shin & Kim newsletter), and the government said fact-finding and fines would be suspended during a guidance period of at least one year (all four sources agree).
What matters for this article is who bears the duty. The four sources say the same thing — the duty falls on the "AI business operator" that directly provides AI products or services to users (the developer that builds the model, the user-business that takes a model and turns it into a service, and overseas operators serving Korean users), and users who employ AI as a tool for work or creation are not covered. The government's example relayed by Newsis is "a production company that makes a film with a video-generation AI," and Professor Lee in the Law Times wrote that "webtoon artists, broadcasters, news outlets, YouTubers and other parties that use AI applications are not operators and so are not covered," raising the need to consider adding a "distributor" concept because the law makes it hard to hold the distributing side responsible (one legal scholar's opinion). One caveat: the header of the National Law Information Center page showed a version "partially amended January 20, 2026, in force July 21," but because we could not read the body, we did not confirm what the amendment changed.
3. AT&T — "no autonomy before accuracy reaches the mid-to-high 90s" (facts, company statements)
In a case study SiliconANGLE published on November 14 (by Paul Gillin, labeled "case study"), Luke Corbin of AT&T Business (AVP for digital transformation and AI), speaking at an AWS event in Boston in early November, said this — nearly 500 agentic workflows in production, about three times that in 2027, about 45 billion AI tokens a day across the company, an agent platform handling more than 1 million transactions a month, and production-stage AI investments returning "fivefold within a year" (all company statements). The reference sentence is this one. "Every production workflow starts under human oversight, and runs autonomously only after accuracy reaches the mid-to-high 90% range within a permissible risk profile." Results are measured in labor savings, revenue, operating costs and customer satisfaction, and continuous monitoring watches for performance drift, he said. Structurally, a "Business Agent Foundry" created in early 2025 maps processes with business units to find steps, handoffs and legacy dependencies to remove, stores business logic centrally in context and memory rather than in the model, and a small number of agents (retrieval, planning) run that logic. The examples given were mainframe automation of fiber-installation order changes (previously a person pulled the ticket and updated the record), halving the time to provision dedicated internet in buildings where fiber already exists, and validating customer location and fiber availability before sales outreach. Employees' personal productivity agents are built freely, he said, while workflows that materially affect customers or the business go through the foundry's security, privacy and compliance reviews. There are no cost figures, and the article carried no sponsorship label (we could not confirm whether it was sponsored).
We matched it against two earlier numbers. AT&T's official blog (July 23, Andy Markus, chief data and AI officer) wrote an average of 45 billion tokens a day, an AI gateway that picks a model per task and cut costs "by as much as 90%" (company figure), and that tasks needing top-tier models are "a small percentage" (no figure). Broadband Breakfast (March 6, remarks at MWC) wrote 27 billion tokens a day and "a fivefold return in free cash flow within the same year on every dollar invested" (rest of the article paywalled). So tokens went from 27 billion in March to 45 billion in July, and the November remarks also say 45 billion; "fivefold" has been the same phrase since March. "Nearly 500" and "mid-to-high 90s" appear only in the November article.
In brief, one (facts)
Bing's "turn off AI summaries." According to Next.ink (November 12), Microsoft's Bing blog on November 11 described two ways to turn off AI summaries in search results — appending "-ai" to a query removes the summary for that one search, and installing an extension for Chrome and Edge adds an on/off toggle (Firefox unsupported; the extension also switches the default search engine to Bing). Next.ink criticized the demand for an extension instead of a single setting, and the presentation as new of an extension that had been in "preview" since June. The June launch was recorded by Search Engine Roundtable (June 8) and PPC Land (June 6), citing a post by Microsoft's head of search Jordi Ribas; the extension is named "Microsoft Bing AI Search Choice" and had 6 Chrome installs as of June 6. We did not open the Bing blog post itself. It sits on the same line as NO.42's "visitors who used to come from search" — when a user turns off summaries, AI-citation exposure turns off with it.
Reading the three together (interpretation)
First, the condition for removing the person has started to be written in numbers, and the two companies' numbers are different in kind. Hyundai Home Shopping's numbers are results — sales at 120% of target and supplier costs down 80% from one trial broadcast (number of trials not given). AT&T's number is a threshold — a person watches until accuracy reaches the mid-to-high 90s. Remove the person on a result number and one performance becomes the standard; remove the person on a threshold number and "when do we revert" is set at the same time (interpretation). Both are company announcements, so we do not take the sizes at face value, but what kind of number a company wrote shows what it is measuring. In NO.40 we wrote that "real usage has a different unit"; this time it is "the unit for removing the person" that differs.
Second, the law did not write "say it's AI" on the side that uses the tool. If the explanation on which four sources agree holds, home-shopping broadcasters, video producers and YouTubers are users employing AI as a tool, and the notice and labeling duties lie with the operators that provided the AI to them. That is probably why five outlets did not ask about disclosure (our guess). But the law has one box with no exception — paragraph (3), which tells everyone to clearly mark virtual people, sounds and video that are hard to distinguish from the real thing. The customer response "as natural as a real person" is a promotional sentence and, at the same time, a signal of approaching that boundary (interpretation). Who fills the box the law left blank is each company's decision; when we started Ria, we made "disclose the AI persona" a rule — not because of the law, but because we judged the cost of customers finding out later and turning away to be greater (our judgment).
Third, the denominators of 80% and 120% have to be read. The 80% is "sample and insert-video costs that suppliers used to bear"; the broadcaster's own production cost, the AI system's cost and generation costs are not in the announcement. The 120% is "versus target," and whether the target is the usual sales of a human-hosted broadcast or a separate figure is not given. Lotte's "cumulative 50 billion" is a cumulative figure with no period; NS's "8x" is the time for one type of banner. AT&T's "fivefold" is also the company's word, with no cost figure anywhere. In NO.21 we wrote "count before you compare," and this time's numbers all travel with "what versus what" missing. So in our tables we keep a column next to each number for its denominator.
What we would do — three things this week for a team running an AI influencer, commerce and apps (practice)
- Rewrite Ria's "three lines that say it's AI" to match the three paragraphs of Article 31. (1) Prior notice — is there an "AI persona" sentence in the channel description, the first video and our site's terms; (2) output labeling — since the videos are shared outside, are there two layers, the on-screen/description label and metadata (watermark); (3) not a deepfake — is the rule "never made to resemble a real person" written on the character sheet. We keep the explanation that, as tool users, we may not be covered by the duty, and keep the rules regardless of duty. The statute text and the content of the amended version in force from July 21 go on a separate list of items for professional confirmation (we are not lawyers).
- Set "the threshold for removing the person" as a number first. Attach one AT&T-style sentence to each of three places — Ria's automated script generation, ingredient-label judgments (NO.03), and automated replies to commerce inquiries: "a person reviews everything until accuracy reaches N%; above it, sample review." Fill N by measuring the current value in a 20-case test, and next to it write the condition for reverting (how many wrong answers send us back to full review) — in the same table as NO.39's "stop button." Without a number, one "the result was good" ends up removing the person.
- Split the denominator of the video production cost table into three. In Ria's 15-second unit-cost table (NO.41), add three columns — "borne by the brand (supplier) / borne by us / AI and system cost" — plus a column for "cost per scene that replaced a shoot." Record sales not "versus target" but "versus a human-hosted broadcast in the same time slot," plus conversion rate. For commerce, one line only — search our product names on Bing with "-ai" appended, see where our site appears in the results without summaries, and note it in NO.42's traffic table.
What we did not pick this time
The "159,000 Lovable-built apps leaked passports and 200,000 chats" item flagged by AI Weekly (crissyfield.de research) is a security incident, so we keep only the title — when an official notice appears, we cover it as a follow-up to NO.32 without methods. The same newsletter's "Snap ships twice the code, PennyMac plans fewer hires" is held because we did not open the earnings-call source; HSBC's rating on Nvidia is investment material and we do not use it. Next.ink's tutorial on installing the agent harness Pi.dev (November 13) is a follow-up candidate for NO.48's "harness." On the home-shopping side, "a 25-million-won promo video for 150,000 won" (May 6) and an earlier report on Hyundai Home Shopping's "AI host" (Smart Today) were not used because we did not open the articles. Anthropic's newsroom had no new post as of November 15 (still the November 10 roadshow and November 9 Fable 5.2), and OpenAI's news list would not open this time either. Whether Meta's newsroom has the Muse phone post was not checked today because it was not on today's list (NO.47).
What we could not confirm in this news
- Hyundai Home Shopping's 80% and 120% are trial-broadcast figures announced by the company; the timing, product and number of trials, the definition of "target," and the broadcaster's own production cost, system cost and ratings appear in no report. "Industry first" is the company's description.
- How viewers are told the broadcast is AI-made, the form of the on-screen presenter (only The Korea Herald caption's "virtual presenter"), and the technology and partners used are not in any of the five outlets. It was reported to air on November 16 at 15:20 KST. We could not confirm the broadcast itself or any coverage after it.
- Lotte's "Lucy" cumulative 50 billion, 50% and 3x for appliances are a company announcement relayed by one outlet, the Financial News. For Newsway's GMV and carriage-fee figures we did not open the underlying data (association or regulator materials).
- Because the National Law Information Center body would not load, we confirmed the AI Basic Act provisions through the agreement of four secondary sources (Law Times, Shin & Kim, Newsis, Asia Economy). We did not confirm what the version "partially amended January 20, in force July 21" changed. "Broadcasters and YouTubers are not covered" is one legal scholar's interpretation; we confirmed only that it points the same way as the government guideline's example (a film production company).
- AT&T's "nearly 500," "mid-to-high 90s," "1 million a month" and "3x in 2027" are company statements in one SiliconANGLE article. "45 billion a day" matches the company blog (July), and "fivefold" is the same phrase as the March MWC remarks, but the measurement method appears nowhere. We could not confirm whether the article was sponsored.
- We did not open the November 11 Bing blog post and wrote from Next.ink's citation. We did not check the extension's current install count.
Read alongside
We opened sources of different kinds directly. Company announcements and official documents, commentary on the law and government guideline, media reports, and our earlier articles are listed separately.
Company announcements and official documents:
Commentary on the law and government guideline:
- lawtimes.co.krIssues and tasks following the AI Basic Act's entry into force (Law Times, Lee Seong-yeop, February 4)Korean; in force January 22, three paragraphs of Article 31, scope (broadcasters, YouTubers and other users not covered), "distributor" concept proposal, guidance period
- shinkim.comAI Basic Act and Enforcement Decree in force — transparency duties (Shin & Kim newsletter, February 11)Korean; Article 31 and Decree Article 23 methods, corrective order (Article 40(3)), fine of up to 30 million won (Article 43), guidance period
- mobile.newsis.comFrom today, labeling of AI outputs is mandatory — who, when, how (Newsis, January 21)Korean; summary of the Ministry of Science and ICT guideline, example of "a production company making a film with video-generation AI is a user"
- view.asiae.co.krFlexible inside the service, clear labeling for exports and deepfakes (Asia Economy, January 21)Korean; same guideline, guidance period
Media reports:
- koreaherald.comHyundai Home Shopping to air first fully AI-generated sales show (The Korea Herald, November 15)English; November 16 3:20 p.m., Plus Shop, 80%, 120%, monthly, remarks by Lim Soo-jin and Lim Hyun-tae, photo caption "virtual presenter"
- zdnet.co.krHyundai Home Shopping introduces a "100% AI broadcast" with no host or camera (ZDNet Korea, November 15)Korean; same announcement, executives' remarks in the original
- mt.co.krNo host, no camera — Hyundai Home Shopping debuts AI broadcast (Money Today, November 15)Korean; same announcement, "from voice to screen"
- en.sedaily.comNo Show Hosts, No Cameras: Hyundai Home Shopping Launches Industry-First All-AI Broadcast (Seoul Economic Daily, November 15)English; official's remark on "accelerating AX"
- fnnews.comAI in the home-shopping industry — GS Shop's virtual model, Lotte's Lucy, Hyundai's short-form (Financial News, March 25)Korean; background, video production two days → one hour, Lucy cumulative 50 billion (company announcement, one outlet)
- v.daum.netAI even in 10-minute broadcasts — a production-cost revolution in TV home shopping (Newsway, May 21)Korean; background, NS Home Shopping's first AI shopping host, GMV, carriage-fee share
- ddaily.co.krNS Home Shopping expands AI company-wide — content production productivity up 8x (Digital Daily, July 28)Korean; background, banners 40 min → 5 min, post-production 4 h → 40 min (company announcement)
- siliconangle.comAT&T uses agentic AI to reinvent business processes from the ground up (SiliconANGLE, November 14)English; nearly 500, 45 billion, 1 million a month, "mid-to-high 90s," starts under human oversight (company statements, one outlet)
- broadbandbreakfast.comAT&T Said It Generated 5x Return on AI Investment (Broadband Breakfast, March 6)English; MWC remarks, 27 billion a day, fivefold (rest paywalled)
- next.inkMicrosoft permet de désactiver les résumés IA dans Bing, mais s'y prend (très) mal (Next.ink, November 12)French; "-ai" operator, extension, criticism of the preview-since-June
- seroundtable.comBing: "-ai" operator and extension to turn off AI answers (Search Engine Roundtable, June 8)English; June launch, Ribas remarks
- ppc.landBing now lets you turn off AI search results with one click (PPC Land, June 6)English; extension name, toggle wording, default-search-engine change, 6 installs
Our earlier articles:
- apholdings.krSwapping a model takes weeks (AP News NO.48)
- apholdings.krWeekly AI News — Google changed three models and Search in five days, OpenAI handed agents a wallet (AP News NO.47)
- apholdings.krCustomers cannot bring their own agents (AP News NO.43)
- apholdings.krThe visitors who came from search are dwindling (AP News NO.42)
- apholdings.krWeekly AI News — Google cut the price of 15-second video by 30% (AP News NO.41)
- apholdings.krThe scoreboard now comes with probabilities (AP News NO.40)
- apholdings.kr"Agents may exceed their authority" has started appearing in filings (AP News NO.39)
- apholdings.krCount before you compare (AP News NO.21)
- apholdings.krWhen AI reads the ingredient label for you (AP News NO.03)